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:$rA, i16ImmSExt10:$val))]>;
588 // v4i32, i32 add instruction:
590 class AInst<dag OOL, dag IOL, list<dag> pattern>:
591 RRForm<0b00000011000, OOL, IOL,
592 "a\t$rT, $rA, $rB", IntegerOp,
595 class AVecInst<ValueType vectype>:
596 AInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
597 [(set (vectype VECREG:$rT), (add (vectype VECREG:$rA),
598 (vectype VECREG:$rB)))]>;
600 class ARegInst<RegisterClass rclass>:
601 AInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
602 [(set rclass:$rT, (add rclass:$rA, rclass:$rB))]>;
604 multiclass AddInstruction {
605 def v4i32: AVecInst<v4i32>;
606 def v16i8: AVecInst<v16i8>;
608 def r32: ARegInst<R32C>;
611 defm A : AddInstruction;
613 class AIInst<dag OOL, dag IOL, list<dag> pattern>:
614 RI10Form<0b00111000, OOL, IOL,
615 "ai\t$rT, $rA, $val", IntegerOp,
618 class AIVecInst<ValueType vectype, PatLeaf immpred>:
619 AIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
620 [(set (vectype VECREG:$rT), (add (vectype VECREG:$rA), immpred:$val))]>;
622 class AIFPVecInst<ValueType vectype, PatLeaf immpred>:
623 AIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
626 class AIRegInst<RegisterClass rclass, PatLeaf immpred>:
627 AIInst<(outs rclass:$rT), (ins rclass:$rA, s10imm_i32:$val),
628 [(set rclass:$rT, (add rclass:$rA, immpred:$val))]>;
630 // This is used to add epsilons to floating point numbers in the f32 fdiv code:
631 class AIFPInst<RegisterClass rclass, PatLeaf immpred>:
632 AIInst<(outs rclass:$rT), (ins rclass:$rA, s10imm_i32:$val),
635 multiclass AddImmediate {
636 def v4i32: AIVecInst<v4i32, v4i32SExt10Imm>;
638 def r32: AIRegInst<R32C, i32ImmSExt10>;
640 def v4f32: AIFPVecInst<v4f32, v4i32SExt10Imm>;
641 def f32: AIFPInst<R32FP, i32ImmSExt10>;
644 defm AI : AddImmediate;
647 RRForm<0b00010010000, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
648 "sfh\t$rT, $rA, $rB", IntegerOp,
649 [(set (v8i16 VECREG:$rT), (sub (v8i16 VECREG:$rA),
650 (v8i16 VECREG:$rB)))]>;
653 RRForm<0b00010010000, (outs R16C:$rT), (ins R16C:$rA, R16C:$rB),
654 "sfh\t$rT, $rA, $rB", IntegerOp,
655 [(set R16C:$rT, (sub R16C:$rA, R16C:$rB))]>;
658 RI10Form<0b10110000, (outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
659 "sfhi\t$rT, $rA, $val", IntegerOp,
660 [(set (v8i16 VECREG:$rT), (sub v8i16SExt10Imm:$val,
661 (v8i16 VECREG:$rA)))]>;
663 def SFHIr16 : RI10Form<0b10110000, (outs R16C:$rT), (ins R16C:$rA, s10imm:$val),
664 "sfhi\t$rT, $rA, $val", IntegerOp,
665 [(set R16C:$rT, (sub i16ImmSExt10:$val, R16C:$rA))]>;
667 def SFvec : RRForm<0b00000010000, (outs VECREG:$rT),
668 (ins VECREG:$rA, VECREG:$rB),
669 "sf\t$rT, $rA, $rB", IntegerOp,
670 [(set (v4i32 VECREG:$rT), (sub (v4i32 VECREG:$rA), (v4i32 VECREG:$rB)))]>;
672 def SFr32 : RRForm<0b00000010000, (outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
673 "sf\t$rT, $rA, $rB", IntegerOp,
674 [(set R32C:$rT, (sub R32C:$rA, R32C:$rB))]>;
677 RI10Form<0b00110000, (outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
678 "sfi\t$rT, $rA, $val", IntegerOp,
679 [(set (v4i32 VECREG:$rT), (sub v4i32SExt10Imm:$val,
680 (v4i32 VECREG:$rA)))]>;
682 def SFIr32 : RI10Form<0b00110000, (outs R32C:$rT),
683 (ins R32C:$rA, s10imm_i32:$val),
684 "sfi\t$rT, $rA, $val", IntegerOp,
685 [(set R32C:$rT, (sub i32ImmSExt10:$val, R32C:$rA))]>;
687 // ADDX: only available in vector form, doesn't match a pattern.
688 class ADDXInst<dag OOL, dag IOL, list<dag> pattern>:
689 RRForm<0b00000010110, OOL, IOL,
690 "addx\t$rT, $rA, $rB",
693 class ADDXVecInst<ValueType vectype>:
694 ADDXInst<(outs VECREG:$rT),
695 (ins VECREG:$rA, VECREG:$rB, VECREG:$rCarry),
696 [(set (vectype VECREG:$rT),
697 (SPUaddx (vectype VECREG:$rA), (vectype VECREG:$rB),
698 (vectype VECREG:$rCarry)))]>,
699 RegConstraint<"$rCarry = $rT">,
702 class ADDXRegInst<RegisterClass rclass>:
703 ADDXInst<(outs rclass:$rT),
704 (ins rclass:$rA, rclass:$rB, rclass:$rCarry),
706 (SPUaddx rclass:$rA, rclass:$rB, rclass:$rCarry))]>,
707 RegConstraint<"$rCarry = $rT">,
710 multiclass AddExtended {
711 def v2i64 : ADDXVecInst<v2i64>;
712 def v4i32 : ADDXVecInst<v4i32>;
713 def r64 : ADDXRegInst<R64C>;
714 def r32 : ADDXRegInst<R32C>;
717 defm ADDX : AddExtended;
719 // CG: Generate carry for add
720 class CGInst<dag OOL, dag IOL, list<dag> pattern>:
721 RRForm<0b01000011000, OOL, IOL,
725 class CGVecInst<ValueType vectype>:
726 CGInst<(outs VECREG:$rT),
727 (ins VECREG:$rA, VECREG:$rB),
728 [(set (vectype VECREG:$rT),
729 (SPUcarry_gen (vectype VECREG:$rA), (vectype VECREG:$rB)))]>;
731 class CGRegInst<RegisterClass rclass>:
732 CGInst<(outs rclass:$rT),
733 (ins rclass:$rA, rclass:$rB),
735 (SPUcarry_gen rclass:$rA, rclass:$rB))]>;
737 multiclass CarryGenerate {
738 def v2i64 : CGVecInst<v2i64>;
739 def v4i32 : CGVecInst<v4i32>;
740 def r64 : CGRegInst<R64C>;
741 def r32 : CGRegInst<R32C>;
744 defm CG : CarryGenerate;
746 // SFX: Subract from, extended. This is used in conjunction with BG to subtract
747 // with carry (borrow, in this case)
748 class SFXInst<dag OOL, dag IOL, list<dag> pattern>:
749 RRForm<0b10000010110, OOL, IOL,
750 "sfx\t$rT, $rA, $rB",
753 class SFXVecInst<ValueType vectype>:
754 SFXInst<(outs VECREG:$rT),
755 (ins VECREG:$rA, VECREG:$rB, VECREG:$rCarry),
756 [(set (vectype VECREG:$rT),
757 (SPUsubx (vectype VECREG:$rA), (vectype VECREG:$rB),
758 (vectype VECREG:$rCarry)))]>,
759 RegConstraint<"$rCarry = $rT">,
762 class SFXRegInst<RegisterClass rclass>:
763 SFXInst<(outs rclass:$rT),
764 (ins rclass:$rA, rclass:$rB, rclass:$rCarry),
766 (SPUsubx rclass:$rA, rclass:$rB, rclass:$rCarry))]>,
767 RegConstraint<"$rCarry = $rT">,
770 multiclass SubtractExtended {
771 def v2i64 : SFXVecInst<v2i64>;
772 def v4i32 : SFXVecInst<v4i32>;
773 def r64 : SFXRegInst<R64C>;
774 def r32 : SFXRegInst<R32C>;
777 defm SFX : SubtractExtended;
779 // BG: only available in vector form, doesn't match a pattern.
780 class BGInst<dag OOL, dag IOL, list<dag> pattern>:
781 RRForm<0b01000010000, OOL, IOL,
785 class BGVecInst<ValueType vectype>:
786 BGInst<(outs VECREG:$rT),
787 (ins VECREG:$rA, VECREG:$rB),
788 [(set (vectype VECREG:$rT),
789 (SPUborrow_gen (vectype VECREG:$rA), (vectype VECREG:$rB)))]>;
791 class BGRegInst<RegisterClass rclass>:
792 BGInst<(outs rclass:$rT),
793 (ins rclass:$rA, rclass:$rB),
795 (SPUborrow_gen rclass:$rA, rclass:$rB))]>;
797 multiclass BorrowGenerate {
798 def v4i32 : BGVecInst<v4i32>;
799 def v2i64 : BGVecInst<v2i64>;
800 def r64 : BGRegInst<R64C>;
801 def r32 : BGRegInst<R32C>;
804 defm BG : BorrowGenerate;
806 // BGX: Borrow generate, extended.
808 RRForm<0b11000010110, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB,
810 "bgx\t$rT, $rA, $rB", IntegerOp,
812 RegConstraint<"$rCarry = $rT">,
815 // Halfword multiply variants:
816 // N.B: These can be used to build up larger quantities (16x16 -> 32)
819 RRForm<0b00100011110, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
820 "mpy\t$rT, $rA, $rB", IntegerMulDiv,
824 RRForm<0b00100011110, (outs R16C:$rT), (ins R16C:$rA, R16C:$rB),
825 "mpy\t$rT, $rA, $rB", IntegerMulDiv,
826 [(set R16C:$rT, (mul R16C:$rA, R16C:$rB))]>;
828 // Unsigned 16-bit multiply:
830 class MPYUInst<dag OOL, dag IOL, list<dag> pattern>:
831 RRForm<0b00110011110, OOL, IOL,
832 "mpyu\t$rT, $rA, $rB", IntegerMulDiv,
836 MPYUInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
840 MPYUInst<(outs R32C:$rT), (ins R16C:$rA, R16C:$rB),
841 [(set R32C:$rT, (mul (zext R16C:$rA), (zext R16C:$rB)))]>;
844 MPYUInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
847 // mpyi: multiply 16 x s10imm -> 32 result.
849 class MPYIInst<dag OOL, dag IOL, list<dag> pattern>:
850 RI10Form<0b00101110, OOL, IOL,
851 "mpyi\t$rT, $rA, $val", IntegerMulDiv,
855 MPYIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
856 [(set (v8i16 VECREG:$rT),
857 (mul (v8i16 VECREG:$rA), v8i16SExt10Imm:$val))]>;
860 MPYIInst<(outs R16C:$rT), (ins R16C:$rA, s10imm:$val),
861 [(set R16C:$rT, (mul R16C:$rA, i16ImmSExt10:$val))]>;
863 // mpyui: same issues as other multiplies, plus, this doesn't match a
864 // pattern... but may be used during target DAG selection or lowering
866 class MPYUIInst<dag OOL, dag IOL, list<dag> pattern>:
867 RI10Form<0b10101110, OOL, IOL,
868 "mpyui\t$rT, $rA, $val", IntegerMulDiv,
872 MPYUIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
876 MPYUIInst<(outs R16C:$rT), (ins R16C:$rA, s10imm:$val),
879 // mpya: 16 x 16 + 16 -> 32 bit result
880 class MPYAInst<dag OOL, dag IOL, list<dag> pattern>:
881 RRRForm<0b0011, OOL, IOL,
882 "mpya\t$rT, $rA, $rB, $rC", IntegerMulDiv,
886 MPYAInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
887 [(set (v4i32 VECREG:$rT),
888 (add (v4i32 (bitconvert (mul (v8i16 VECREG:$rA),
889 (v8i16 VECREG:$rB)))),
890 (v4i32 VECREG:$rC)))]>;
893 MPYAInst<(outs R32C:$rT), (ins R16C:$rA, R16C:$rB, R32C:$rC),
894 [(set R32C:$rT, (add (sext (mul R16C:$rA, R16C:$rB)),
898 MPYAInst<(outs R32C:$rT), (ins R16C:$rA, R16C:$rB, R32C:$rC),
899 [(set R32C:$rT, (add (mul (sext R16C:$rA), (sext R16C:$rB)),
902 def MPYAr32_sextinreg:
903 MPYAInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB, R32C:$rC),
904 [(set R32C:$rT, (add (mul (sext_inreg R32C:$rA, i16),
905 (sext_inreg R32C:$rB, i16)),
908 // mpyh: multiply high, used to synthesize 32-bit multiplies
909 class MPYHInst<dag OOL, dag IOL, list<dag> pattern>:
910 RRForm<0b10100011110, OOL, IOL,
911 "mpyh\t$rT, $rA, $rB", IntegerMulDiv,
915 MPYHInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
919 MPYHInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
922 // mpys: multiply high and shift right (returns the top half of
923 // a 16-bit multiply, sign extended to 32 bits.)
925 class MPYSInst<dag OOL, dag IOL>:
926 RRForm<0b11100011110, OOL, IOL,
927 "mpys\t$rT, $rA, $rB", IntegerMulDiv,
931 MPYSInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB)>;
934 MPYSInst<(outs R32C:$rT), (ins R16C:$rA, R16C:$rB)>;
936 // mpyhh: multiply high-high (returns the 32-bit result from multiplying
937 // the top 16 bits of the $rA, $rB)
939 class MPYHHInst<dag OOL, dag IOL>:
940 RRForm<0b01100011110, OOL, IOL,
941 "mpyhh\t$rT, $rA, $rB", IntegerMulDiv,
945 MPYHHInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB)>;
948 MPYHHInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB)>;
950 // mpyhha: Multiply high-high, add to $rT:
952 class MPYHHAInst<dag OOL, dag IOL>:
953 RRForm<0b01100010110, OOL, IOL,
954 "mpyhha\t$rT, $rA, $rB", IntegerMulDiv,
958 MPYHHAInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB)>;
961 MPYHHAInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB)>;
963 // mpyhhu: Multiply high-high, unsigned
965 class MPYHHUInst<dag OOL, dag IOL>:
966 RRForm<0b01110011110, OOL, IOL,
967 "mpyhhu\t$rT, $rA, $rB", IntegerMulDiv,
971 MPYHHUInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB)>;
974 MPYHHUInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB)>;
976 // mpyhhau: Multiply high-high, unsigned
978 class MPYHHAUInst<dag OOL, dag IOL>:
979 RRForm<0b01110010110, OOL, IOL,
980 "mpyhhau\t$rT, $rA, $rB", IntegerMulDiv,
984 MPYHHAUInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB)>;
987 MPYHHAUInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB)>;
989 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
990 // clz: Count leading zeroes
991 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
992 class CLZInst<dag OOL, dag IOL, list<dag> pattern>:
993 RRForm_1<0b10100101010, OOL, IOL, "clz\t$rT, $rA",
996 class CLZRegInst<RegisterClass rclass>:
997 CLZInst<(outs rclass:$rT), (ins rclass:$rA),
998 [(set rclass:$rT, (ctlz rclass:$rA))]>;
1000 class CLZVecInst<ValueType vectype>:
1001 CLZInst<(outs VECREG:$rT), (ins VECREG:$rA),
1002 [(set (vectype VECREG:$rT), (ctlz (vectype VECREG:$rA)))]>;
1004 multiclass CountLeadingZeroes {
1005 def v4i32 : CLZVecInst<v4i32>;
1006 def r32 : CLZRegInst<R32C>;
1009 defm CLZ : CountLeadingZeroes;
1011 // cntb: Count ones in bytes (aka "population count")
1013 // NOTE: This instruction is really a vector instruction, but the custom
1014 // lowering code uses it in unorthodox ways to support CTPOP for other
1018 RRForm_1<0b00101101010, (outs VECREG:$rT), (ins VECREG:$rA),
1019 "cntb\t$rT, $rA", IntegerOp,
1020 [(set (v16i8 VECREG:$rT), (SPUcntb (v16i8 VECREG:$rA)))]>;
1023 RRForm_1<0b00101101010, (outs VECREG:$rT), (ins VECREG:$rA),
1024 "cntb\t$rT, $rA", IntegerOp,
1025 [(set (v8i16 VECREG:$rT), (SPUcntb (v8i16 VECREG:$rA)))]>;
1028 RRForm_1<0b00101101010, (outs VECREG:$rT), (ins VECREG:$rA),
1029 "cntb\t$rT, $rA", IntegerOp,
1030 [(set (v4i32 VECREG:$rT), (SPUcntb (v4i32 VECREG:$rA)))]>;
1032 // gbb: Gather the low order bits from each byte in $rA into a single 16-bit
1033 // quantity stored into $rT's slot 0, upper 16 bits are zeroed, as are
1036 // Note: This instruction "pairs" with the fsmb instruction for all of the
1037 // various types defined here.
1039 // Note 2: The "VecInst" and "RegInst" forms refer to the result being either
1040 // a vector or register.
1042 class GBBInst<dag OOL, dag IOL, list<dag> pattern>:
1043 RRForm_1<0b01001101100, OOL, IOL, "gbb\t$rT, $rA", GatherOp, pattern>;
1045 class GBBRegInst<RegisterClass rclass, ValueType vectype>:
1046 GBBInst<(outs rclass:$rT), (ins VECREG:$rA),
1047 [(set rclass:$rT, (SPUgatherbits (vectype VECREG:$rA)))]>;
1049 class GBBVecInst<ValueType vectype>:
1050 GBBInst<(outs VECREG:$rT), (ins VECREG:$rA),
1051 [(set (vectype VECREG:$rT), (SPUgatherbits (vectype VECREG:$rA)))]>;
1053 multiclass GatherBitsFromBytes {
1054 def v16i8_r32: GBBRegInst<R32C, v16i8>;
1055 def v16i8_r16: GBBRegInst<R16C, v16i8>;
1056 def v16i8: GBBVecInst<v16i8>;
1059 defm GBB: GatherBitsFromBytes;
1061 // gbh: Gather all low order bits from each halfword in $rA into a single
1062 // 8-bit quantity stored in $rT's slot 0, with the upper bits of $rT set to 0
1063 // and slots 1-3 also set to 0.
1065 // See notes for GBBInst, above.
1067 class GBHInst<dag OOL, dag IOL, list<dag> pattern>:
1068 RRForm_1<0b10001101100, OOL, IOL, "gbh\t$rT, $rA", GatherOp,
1071 class GBHRegInst<RegisterClass rclass, ValueType vectype>:
1072 GBHInst<(outs rclass:$rT), (ins VECREG:$rA),
1073 [(set rclass:$rT, (SPUgatherbits (vectype VECREG:$rA)))]>;
1075 class GBHVecInst<ValueType vectype>:
1076 GBHInst<(outs VECREG:$rT), (ins VECREG:$rA),
1077 [(set (vectype VECREG:$rT),
1078 (SPUgatherbits (vectype VECREG:$rA)))]>;
1080 multiclass GatherBitsHalfword {
1081 def v8i16_r32: GBHRegInst<R32C, v8i16>;
1082 def v8i16_r16: GBHRegInst<R16C, v8i16>;
1083 def v8i16: GBHVecInst<v8i16>;
1086 defm GBH: GatherBitsHalfword;
1088 // gb: Gather all low order bits from each word in $rA into a single
1089 // 4-bit quantity stored in $rT's slot 0, upper bits in $rT set to 0,
1090 // as well as slots 1-3.
1092 // See notes for gbb, above.
1094 class GBInst<dag OOL, dag IOL, list<dag> pattern>:
1095 RRForm_1<0b00001101100, OOL, IOL, "gb\t$rT, $rA", GatherOp,
1098 class GBRegInst<RegisterClass rclass, ValueType vectype>:
1099 GBInst<(outs rclass:$rT), (ins VECREG:$rA),
1100 [(set rclass:$rT, (SPUgatherbits (vectype VECREG:$rA)))]>;
1102 class GBVecInst<ValueType vectype>:
1103 GBInst<(outs VECREG:$rT), (ins VECREG:$rA),
1104 [(set (vectype VECREG:$rT),
1105 (SPUgatherbits (vectype VECREG:$rA)))]>;
1107 multiclass GatherBitsWord {
1108 def v4i32_r32: GBRegInst<R32C, v4i32>;
1109 def v4i32_r16: GBRegInst<R16C, v4i32>;
1110 def v4i32: GBVecInst<v4i32>;
1113 defm GB: GatherBitsWord;
1115 // avgb: average bytes
1117 RRForm<0b11001011000, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1118 "avgb\t$rT, $rA, $rB", ByteOp,
1121 // absdb: absolute difference of bytes
1123 RRForm<0b11001010000, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1124 "absdb\t$rT, $rA, $rB", ByteOp,
1127 // sumb: sum bytes into halfwords
1129 RRForm<0b11001010010, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1130 "sumb\t$rT, $rA, $rB", ByteOp,
1133 // Sign extension operations:
1134 class XSBHInst<dag OOL, dag IOL, list<dag> pattern>:
1135 RRForm_1<0b01101101010, OOL, IOL,
1136 "xsbh\t$rDst, $rSrc",
1137 IntegerOp, pattern>;
1139 class XSBHVecInst<ValueType vectype>:
1140 XSBHInst<(outs VECREG:$rDst), (ins VECREG:$rSrc),
1141 [(set (v8i16 VECREG:$rDst), (sext (vectype VECREG:$rSrc)))]>;
1143 class XSBHInRegInst<RegisterClass rclass>:
1144 XSBHInst<(outs rclass:$rDst), (ins rclass:$rSrc),
1145 [(set rclass:$rDst, (sext_inreg rclass:$rSrc, i8))]>;
1147 multiclass ExtendByteHalfword {
1148 def v16i8: XSBHVecInst<v8i16>;
1149 def r16: XSBHInRegInst<R16C>;
1150 def r8: XSBHInst<(outs R16C:$rDst), (ins R8C:$rSrc),
1151 [(set R16C:$rDst, (sext R8C:$rSrc))]>;
1153 // 32-bit form for XSBH: used to sign extend 8-bit quantities to 16-bit
1154 // quantities to 32-bit quantities via a 32-bit register (see the sext 8->32
1155 // pattern below). Intentionally doesn't match a pattern because we want the
1156 // sext 8->32 pattern to do the work for us, namely because we need the extra
1158 def r32: XSBHInRegInst<R32C>;
1161 defm XSBH : ExtendByteHalfword;
1163 // Sign extend halfwords to words:
1165 RRForm_1<0b01101101010, (outs VECREG:$rDest), (ins VECREG:$rSrc),
1166 "xshw\t$rDest, $rSrc", IntegerOp,
1167 [(set (v4i32 VECREG:$rDest), (sext (v8i16 VECREG:$rSrc)))]>;
1170 RRForm_1<0b01101101010, (outs R32C:$rDst), (ins R32C:$rSrc),
1171 "xshw\t$rDst, $rSrc", IntegerOp,
1172 [(set R32C:$rDst, (sext_inreg R32C:$rSrc, i16))]>;
1175 RRForm_1<0b01101101010, (outs R32C:$rDst), (ins R16C:$rSrc),
1176 "xshw\t$rDst, $rSrc", IntegerOp,
1177 [(set R32C:$rDst, (sext R16C:$rSrc))]>;
1180 RRForm_1<0b01100101010, (outs VECREG:$rDst), (ins VECREG:$rSrc),
1181 "xswd\t$rDst, $rSrc", IntegerOp,
1182 [(set (v2i64 VECREG:$rDst), (sext (v4i32 VECREG:$rSrc)))]>;
1185 RRForm_1<0b01100101010, (outs R64C:$rDst), (ins R64C:$rSrc),
1186 "xswd\t$rDst, $rSrc", IntegerOp,
1187 [(set R64C:$rDst, (sext_inreg R64C:$rSrc, i32))]>;
1190 RRForm_1<0b01100101010, (outs R64C:$rDst), (ins R32C:$rSrc),
1191 "xswd\t$rDst, $rSrc", IntegerOp,
1192 [(set R64C:$rDst, (SPUsext32_to_64 R32C:$rSrc))]>;
1194 def : Pat<(sext R32C:$inp),
1195 (XSWDr32 R32C:$inp)>;
1199 class ANDInst<dag OOL, dag IOL, list<dag> pattern> :
1200 RRForm<0b10000011000, OOL, IOL, "and\t$rT, $rA, $rB",
1201 IntegerOp, pattern>;
1203 class ANDVecInst<ValueType vectype>:
1204 ANDInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1205 [(set (vectype VECREG:$rT), (and (vectype VECREG:$rA),
1206 (vectype VECREG:$rB)))]>;
1208 class ANDRegInst<RegisterClass rclass>:
1209 ANDInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
1210 [(set rclass:$rT, (and rclass:$rA, rclass:$rB))]>;
1212 multiclass BitwiseAnd
1214 def v16i8: ANDVecInst<v16i8>;
1215 def v8i16: ANDVecInst<v8i16>;
1216 def v4i32: ANDVecInst<v4i32>;
1217 def v2i64: ANDVecInst<v2i64>;
1219 def r128: ANDRegInst<GPRC>;
1220 def r64: ANDRegInst<R64C>;
1221 def r32: ANDRegInst<R32C>;
1222 def r16: ANDRegInst<R16C>;
1223 def r8: ANDRegInst<R8C>;
1225 //===---------------------------------------------
1226 // Special instructions to perform the fabs instruction
1227 def fabs32: ANDInst<(outs R32FP:$rT), (ins R32FP:$rA, R32C:$rB),
1228 [/* Intentionally does not match a pattern */]>;
1230 def fabs64: ANDInst<(outs R64FP:$rT), (ins R64FP:$rA, VECREG:$rB),
1231 [/* Intentionally does not match a pattern */]>;
1233 // Could use v4i32, but won't for clarity
1234 def fabsvec: ANDInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1235 [/* Intentionally does not match a pattern */]>;
1237 //===---------------------------------------------
1239 // Hacked form of AND to zero-extend 16-bit quantities to 32-bit
1240 // quantities -- see 16->32 zext pattern.
1242 // This pattern is somewhat artificial, since it might match some
1243 // compiler generated pattern but it is unlikely to do so.
1245 def i16i32: ANDInst<(outs R32C:$rT), (ins R16C:$rA, R32C:$rB),
1246 [(set R32C:$rT, (and (zext R16C:$rA), R32C:$rB))]>;
1249 defm AND : BitwiseAnd;
1251 // N.B.: vnot_conv is one of those special target selection pattern fragments,
1252 // in which we expect there to be a bit_convert on the constant. Bear in mind
1253 // that llvm translates "not <reg>" to "xor <reg>, -1" (or in this case, a
1254 // constant -1 vector.)
1256 class ANDCInst<dag OOL, dag IOL, list<dag> pattern>:
1257 RRForm<0b10000011010, OOL, IOL, "andc\t$rT, $rA, $rB",
1258 IntegerOp, pattern>;
1260 class ANDCVecInst<ValueType vectype>:
1261 ANDCInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1262 [(set (vectype VECREG:$rT), (and (vectype VECREG:$rA),
1263 (vnot (vectype VECREG:$rB))))]>;
1265 class ANDCRegInst<RegisterClass rclass>:
1266 ANDCInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
1267 [(set rclass:$rT, (and rclass:$rA, (not rclass:$rB)))]>;
1269 multiclass AndComplement
1271 def v16i8: ANDCVecInst<v16i8>;
1272 def v8i16: ANDCVecInst<v8i16>;
1273 def v4i32: ANDCVecInst<v4i32>;
1274 def v2i64: ANDCVecInst<v2i64>;
1276 def r128: ANDCRegInst<GPRC>;
1277 def r64: ANDCRegInst<R64C>;
1278 def r32: ANDCRegInst<R32C>;
1279 def r16: ANDCRegInst<R16C>;
1280 def r8: ANDCRegInst<R8C>;
1283 defm ANDC : AndComplement;
1285 class ANDBIInst<dag OOL, dag IOL, list<dag> pattern>:
1286 RI10Form<0b01101000, OOL, IOL, "andbi\t$rT, $rA, $val",
1289 multiclass AndByteImm
1291 def v16i8: ANDBIInst<(outs VECREG:$rT), (ins VECREG:$rA, u10imm:$val),
1292 [(set (v16i8 VECREG:$rT),
1293 (and (v16i8 VECREG:$rA),
1294 (v16i8 v16i8U8Imm:$val)))]>;
1296 def r8: ANDBIInst<(outs R8C:$rT), (ins R8C:$rA, u10imm_i8:$val),
1297 [(set R8C:$rT, (and R8C:$rA, immU8:$val))]>;
1300 defm ANDBI : AndByteImm;
1302 class ANDHIInst<dag OOL, dag IOL, list<dag> pattern> :
1303 RI10Form<0b10101000, OOL, IOL, "andhi\t$rT, $rA, $val",
1306 multiclass AndHalfwordImm
1308 def v8i16: ANDHIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
1309 [(set (v8i16 VECREG:$rT),
1310 (and (v8i16 VECREG:$rA), v8i16SExt10Imm:$val))]>;
1312 def r16: ANDHIInst<(outs R16C:$rT), (ins R16C:$rA, u10imm:$val),
1313 [(set R16C:$rT, (and R16C:$rA, i16ImmUns10:$val))]>;
1315 // Zero-extend i8 to i16:
1316 def i8i16: ANDHIInst<(outs R16C:$rT), (ins R8C:$rA, u10imm:$val),
1317 [(set R16C:$rT, (and (zext R8C:$rA), i16ImmUns10:$val))]>;
1320 defm ANDHI : AndHalfwordImm;
1322 class ANDIInst<dag OOL, dag IOL, list<dag> pattern> :
1323 RI10Form<0b00101000, OOL, IOL, "andi\t$rT, $rA, $val",
1324 IntegerOp, pattern>;
1326 multiclass AndWordImm
1328 def v4i32: ANDIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
1329 [(set (v4i32 VECREG:$rT),
1330 (and (v4i32 VECREG:$rA), v4i32SExt10Imm:$val))]>;
1332 def r32: ANDIInst<(outs R32C:$rT), (ins R32C:$rA, s10imm_i32:$val),
1333 [(set R32C:$rT, (and R32C:$rA, i32ImmSExt10:$val))]>;
1335 // Hacked form of ANDI to zero-extend i8 quantities to i32. See the zext 8->32
1337 def i8i32: ANDIInst<(outs R32C:$rT), (ins R8C:$rA, s10imm_i32:$val),
1339 (and (zext R8C:$rA), i32ImmSExt10:$val))]>;
1341 // Hacked form of ANDI to zero-extend i16 quantities to i32. See the
1342 // zext 16->32 pattern below.
1344 // Note that this pattern is somewhat artificial, since it might match
1345 // something the compiler generates but is unlikely to occur in practice.
1346 def i16i32: ANDIInst<(outs R32C:$rT), (ins R16C:$rA, s10imm_i32:$val),
1348 (and (zext R16C:$rA), i32ImmSExt10:$val))]>;
1351 defm ANDI : AndWordImm;
1353 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
1354 // Bitwise OR group:
1355 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
1357 // Bitwise "or" (N.B.: These are also register-register copy instructions...)
1358 class ORInst<dag OOL, dag IOL, list<dag> pattern>:
1359 RRForm<0b10000010000, OOL, IOL, "or\t$rT, $rA, $rB",
1360 IntegerOp, pattern>;
1362 class ORVecInst<ValueType vectype>:
1363 ORInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1364 [(set (vectype VECREG:$rT), (or (vectype VECREG:$rA),
1365 (vectype VECREG:$rB)))]>;
1367 class ORRegInst<RegisterClass rclass>:
1368 ORInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
1369 [(set rclass:$rT, (or rclass:$rA, rclass:$rB))]>;
1371 // ORCvtForm: OR conversion form
1373 // This is used to "convert" the preferred slot to its vector equivalent, as
1374 // well as convert a vector back to its preferred slot.
1376 // These are effectively no-ops, but need to exist for proper type conversion
1377 // and type coercion.
1379 class ORCvtForm<dag OOL, dag IOL>
1380 : SPUInstr<OOL, IOL, "or\t$rT, $rA, $rA", IntegerOp> {
1384 let Pattern = [/* no pattern */];
1386 let Inst{0-10} = 0b10000010000;
1387 let Inst{11-17} = RA;
1388 let Inst{18-24} = RA;
1389 let Inst{25-31} = RT;
1392 class ORPromoteScalar<RegisterClass rclass>:
1393 ORCvtForm<(outs VECREG:$rT), (ins rclass:$rA)>;
1395 class ORExtractElt<RegisterClass rclass>:
1396 ORCvtForm<(outs rclass:$rT), (ins VECREG:$rA)>;
1398 class ORCvtRegGPRC<RegisterClass rclass>:
1399 ORCvtForm<(outs GPRC:$rT), (ins rclass:$rA)>;
1402 ORCvtForm<(outs GPRC:$rT), (ins VECREG:$rA)>;
1404 class ORCvtGPRCReg<RegisterClass rclass>:
1405 ORCvtForm<(outs rclass:$rT), (ins GPRC:$rA)>;
1408 ORCvtForm<(outs VECREG:$rT), (ins GPRC:$rA)>;
1410 multiclass BitwiseOr
1412 def v16i8: ORVecInst<v16i8>;
1413 def v8i16: ORVecInst<v8i16>;
1414 def v4i32: ORVecInst<v4i32>;
1415 def v2i64: ORVecInst<v2i64>;
1417 def v4f32: ORInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1418 [(set (v4f32 VECREG:$rT),
1419 (v4f32 (bitconvert (or (v4i32 VECREG:$rA),
1420 (v4i32 VECREG:$rB)))))]>;
1422 def v2f64: ORInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1423 [(set (v2f64 VECREG:$rT),
1424 (v2f64 (bitconvert (or (v2i64 VECREG:$rA),
1425 (v2i64 VECREG:$rB)))))]>;
1427 def r64: ORRegInst<R64C>;
1428 def r32: ORRegInst<R32C>;
1429 def r16: ORRegInst<R16C>;
1430 def r8: ORRegInst<R8C>;
1432 // OR instructions used to copy f32 and f64 registers.
1433 def f32: ORInst<(outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB),
1434 [/* no pattern */]>;
1436 def f64: ORInst<(outs R64FP:$rT), (ins R64FP:$rA, R64FP:$rB),
1437 [/* no pattern */]>;
1439 // scalar->vector promotion, prefslot2vec:
1440 def v16i8_i8: ORPromoteScalar<R8C>;
1441 def v8i16_i16: ORPromoteScalar<R16C>;
1442 def v4i32_i32: ORPromoteScalar<R32C>;
1443 def v2i64_i64: ORPromoteScalar<R64C>;
1444 def v4f32_f32: ORPromoteScalar<R32FP>;
1445 def v2f64_f64: ORPromoteScalar<R64FP>;
1447 // vector->scalar demotion, vec2prefslot:
1448 def i8_v16i8: ORExtractElt<R8C>;
1449 def i16_v8i16: ORExtractElt<R16C>;
1450 def i32_v4i32: ORExtractElt<R32C>;
1451 def i64_v2i64: ORExtractElt<R64C>;
1452 def f32_v4f32: ORExtractElt<R32FP>;
1453 def f64_v2f64: ORExtractElt<R64FP>;
1455 // Conversion from GPRC to register
1456 def i128_r64: ORCvtRegGPRC<R64C>;
1457 def i128_f64: ORCvtRegGPRC<R64FP>;
1458 def i128_r32: ORCvtRegGPRC<R32C>;
1459 def i128_f32: ORCvtRegGPRC<R32FP>;
1460 def i128_r16: ORCvtRegGPRC<R16C>;
1461 def i128_r8: ORCvtRegGPRC<R8C>;
1463 // Conversion from GPRC to vector
1464 def i128_vec: ORCvtVecGPRC;
1466 // Conversion from register to GPRC
1467 def r64_i128: ORCvtGPRCReg<R64C>;
1468 def f64_i128: ORCvtGPRCReg<R64FP>;
1469 def r32_i128: ORCvtGPRCReg<R32C>;
1470 def f32_i128: ORCvtGPRCReg<R32FP>;
1471 def r16_i128: ORCvtGPRCReg<R16C>;
1472 def r8_i128: ORCvtGPRCReg<R8C>;
1474 // Conversion from vector to GPRC
1475 def vec_i128: ORCvtGPRCVec;
1478 defm OR : BitwiseOr;
1480 // scalar->vector promotion patterns (preferred slot to vector):
1481 def : Pat<(v16i8 (SPUprefslot2vec R8C:$rA)),
1482 (ORv16i8_i8 R8C:$rA)>;
1484 def : Pat<(v8i16 (SPUprefslot2vec R16C:$rA)),
1485 (ORv8i16_i16 R16C:$rA)>;
1487 def : Pat<(v4i32 (SPUprefslot2vec R32C:$rA)),
1488 (ORv4i32_i32 R32C:$rA)>;
1490 def : Pat<(v2i64 (SPUprefslot2vec R64C:$rA)),
1491 (ORv2i64_i64 R64C:$rA)>;
1493 def : Pat<(v4f32 (SPUprefslot2vec R32FP:$rA)),
1494 (ORv4f32_f32 R32FP:$rA)>;
1496 def : Pat<(v2f64 (SPUprefslot2vec R64FP:$rA)),
1497 (ORv2f64_f64 R64FP:$rA)>;
1499 // ORi*_v*: Used to extract vector element 0 (the preferred slot), otherwise
1500 // known as converting the vector back to its preferred slot
1502 def : Pat<(SPUvec2prefslot (v16i8 VECREG:$rA)),
1503 (ORi8_v16i8 VECREG:$rA)>;
1505 def : Pat<(SPUvec2prefslot (v8i16 VECREG:$rA)),
1506 (ORi16_v8i16 VECREG:$rA)>;
1508 def : Pat<(SPUvec2prefslot (v4i32 VECREG:$rA)),
1509 (ORi32_v4i32 VECREG:$rA)>;
1511 def : Pat<(SPUvec2prefslot (v2i64 VECREG:$rA)),
1512 (ORi64_v2i64 VECREG:$rA)>;
1514 def : Pat<(SPUvec2prefslot (v4f32 VECREG:$rA)),
1515 (ORf32_v4f32 VECREG:$rA)>;
1517 def : Pat<(SPUvec2prefslot (v2f64 VECREG:$rA)),
1518 (ORf64_v2f64 VECREG:$rA)>;
1520 // Load Register: This is an assembler alias for a bitwise OR of a register
1521 // against itself. It's here because it brings some clarity to assembly
1524 let hasCtrlDep = 1 in {
1525 class LRInst<dag OOL, dag IOL>
1526 : SPUInstr<OOL, IOL, "lr\t$rT, $rA", IntegerOp> {
1530 let Pattern = [/*no pattern*/];
1532 let Inst{0-10} = 0b10000010000; /* It's an OR operation */
1533 let Inst{11-17} = RA;
1534 let Inst{18-24} = RA;
1535 let Inst{25-31} = RT;
1538 class LRVecInst<ValueType vectype>:
1539 LRInst<(outs VECREG:$rT), (ins VECREG:$rA)>;
1541 class LRRegInst<RegisterClass rclass>:
1542 LRInst<(outs rclass:$rT), (ins rclass:$rA)>;
1544 multiclass LoadRegister {
1545 def v2i64: LRVecInst<v2i64>;
1546 def v2f64: LRVecInst<v2f64>;
1547 def v4i32: LRVecInst<v4i32>;
1548 def v4f32: LRVecInst<v4f32>;
1549 def v8i16: LRVecInst<v8i16>;
1550 def v16i8: LRVecInst<v16i8>;
1552 def r128: LRRegInst<GPRC>;
1553 def r64: LRRegInst<R64C>;
1554 def f64: LRRegInst<R64FP>;
1555 def r32: LRRegInst<R32C>;
1556 def f32: LRRegInst<R32FP>;
1557 def r16: LRRegInst<R16C>;
1558 def r8: LRRegInst<R8C>;
1561 defm LR: LoadRegister;
1564 // ORC: Bitwise "or" with complement (c = a | ~b)
1566 class ORCInst<dag OOL, dag IOL, list<dag> pattern>:
1567 RRForm<0b10010010000, OOL, IOL, "orc\t$rT, $rA, $rB",
1568 IntegerOp, pattern>;
1570 class ORCVecInst<ValueType vectype>:
1571 ORCInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1572 [(set (vectype VECREG:$rT), (or (vectype VECREG:$rA),
1573 (vnot (vectype VECREG:$rB))))]>;
1575 class ORCRegInst<RegisterClass rclass>:
1576 ORCInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
1577 [(set rclass:$rT, (or rclass:$rA, (not rclass:$rB)))]>;
1579 multiclass BitwiseOrComplement
1581 def v16i8: ORCVecInst<v16i8>;
1582 def v8i16: ORCVecInst<v8i16>;
1583 def v4i32: ORCVecInst<v4i32>;
1584 def v2i64: ORCVecInst<v2i64>;
1586 def r64: ORCRegInst<R64C>;
1587 def r32: ORCRegInst<R32C>;
1588 def r16: ORCRegInst<R16C>;
1589 def r8: ORCRegInst<R8C>;
1592 defm ORC : BitwiseOrComplement;
1594 // OR byte immediate
1595 class ORBIInst<dag OOL, dag IOL, list<dag> pattern>:
1596 RI10Form<0b01100000, OOL, IOL, "orbi\t$rT, $rA, $val",
1597 IntegerOp, pattern>;
1599 class ORBIVecInst<ValueType vectype, PatLeaf immpred>:
1600 ORBIInst<(outs VECREG:$rT), (ins VECREG:$rA, u10imm:$val),
1601 [(set (v16i8 VECREG:$rT), (or (vectype VECREG:$rA),
1602 (vectype immpred:$val)))]>;
1604 multiclass BitwiseOrByteImm
1606 def v16i8: ORBIVecInst<v16i8, v16i8U8Imm>;
1608 def r8: ORBIInst<(outs R8C:$rT), (ins R8C:$rA, u10imm_i8:$val),
1609 [(set R8C:$rT, (or R8C:$rA, immU8:$val))]>;
1612 defm ORBI : BitwiseOrByteImm;
1614 // OR halfword immediate
1615 class ORHIInst<dag OOL, dag IOL, list<dag> pattern>:
1616 RI10Form<0b10100000, OOL, IOL, "orhi\t$rT, $rA, $val",
1617 IntegerOp, pattern>;
1619 class ORHIVecInst<ValueType vectype, PatLeaf immpred>:
1620 ORHIInst<(outs VECREG:$rT), (ins VECREG:$rA, u10imm:$val),
1621 [(set (vectype VECREG:$rT), (or (vectype VECREG:$rA),
1624 multiclass BitwiseOrHalfwordImm
1626 def v8i16: ORHIVecInst<v8i16, v8i16Uns10Imm>;
1628 def r16: ORHIInst<(outs R16C:$rT), (ins R16C:$rA, u10imm:$val),
1629 [(set R16C:$rT, (or R16C:$rA, i16ImmUns10:$val))]>;
1631 // Specialized ORHI form used to promote 8-bit registers to 16-bit
1632 def i8i16: ORHIInst<(outs R16C:$rT), (ins R8C:$rA, s10imm:$val),
1633 [(set R16C:$rT, (or (anyext R8C:$rA),
1634 i16ImmSExt10:$val))]>;
1637 defm ORHI : BitwiseOrHalfwordImm;
1639 class ORIInst<dag OOL, dag IOL, list<dag> pattern>:
1640 RI10Form<0b00100000, OOL, IOL, "ori\t$rT, $rA, $val",
1641 IntegerOp, pattern>;
1643 class ORIVecInst<ValueType vectype, PatLeaf immpred>:
1644 ORIInst<(outs VECREG:$rT), (ins VECREG:$rA, u10imm:$val),
1645 [(set (vectype VECREG:$rT), (or (vectype VECREG:$rA),
1648 // Bitwise "or" with immediate
1649 multiclass BitwiseOrImm
1651 def v4i32: ORIVecInst<v4i32, v4i32Uns10Imm>;
1653 def r32: ORIInst<(outs R32C:$rT), (ins R32C:$rA, u10imm_i32:$val),
1654 [(set R32C:$rT, (or R32C:$rA, i32ImmUns10:$val))]>;
1656 // i16i32: hacked version of the ori instruction to extend 16-bit quantities
1657 // to 32-bit quantities. used exclusively to match "anyext" conversions (vide
1658 // infra "anyext 16->32" pattern.)
1659 def i16i32: ORIInst<(outs R32C:$rT), (ins R16C:$rA, s10imm_i32:$val),
1660 [(set R32C:$rT, (or (anyext R16C:$rA),
1661 i32ImmSExt10:$val))]>;
1663 // i8i32: Hacked version of the ORI instruction to extend 16-bit quantities
1664 // to 32-bit quantities. Used exclusively to match "anyext" conversions (vide
1665 // infra "anyext 16->32" pattern.)
1666 def i8i32: ORIInst<(outs R32C:$rT), (ins R8C:$rA, s10imm_i32:$val),
1667 [(set R32C:$rT, (or (anyext R8C:$rA),
1668 i32ImmSExt10:$val))]>;
1671 defm ORI : BitwiseOrImm;
1673 // ORX: "or" across the vector: or's $rA's word slots leaving the result in
1674 // $rT[0], slots 1-3 are zeroed.
1676 // FIXME: Needs to match an intrinsic pattern.
1678 RRForm<0b10010010000, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1679 "orx\t$rT, $rA, $rB", IntegerOp,
1684 class XORInst<dag OOL, dag IOL, list<dag> pattern> :
1685 RRForm<0b10010010000, OOL, IOL, "xor\t$rT, $rA, $rB",
1686 IntegerOp, pattern>;
1688 class XORVecInst<ValueType vectype>:
1689 XORInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1690 [(set (vectype VECREG:$rT), (xor (vectype VECREG:$rA),
1691 (vectype VECREG:$rB)))]>;
1693 class XORRegInst<RegisterClass rclass>:
1694 XORInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
1695 [(set rclass:$rT, (xor rclass:$rA, rclass:$rB))]>;
1697 multiclass BitwiseExclusiveOr
1699 def v16i8: XORVecInst<v16i8>;
1700 def v8i16: XORVecInst<v8i16>;
1701 def v4i32: XORVecInst<v4i32>;
1702 def v2i64: XORVecInst<v2i64>;
1704 def r128: XORRegInst<GPRC>;
1705 def r64: XORRegInst<R64C>;
1706 def r32: XORRegInst<R32C>;
1707 def r16: XORRegInst<R16C>;
1708 def r8: XORRegInst<R8C>;
1710 // Special forms for floating point instructions.
1711 // fneg and fabs require bitwise logical ops to manipulate the sign bit.
1713 def fneg32: XORInst<(outs R32FP:$rT), (ins R32FP:$rA, R32C:$rB),
1714 [/* no pattern */]>;
1716 def fneg64: XORInst<(outs R64FP:$rT), (ins R64FP:$rA, VECREG:$rB),
1717 [/* no pattern */]>;
1719 def fnegvec: XORInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1720 [/* no pattern, see fneg{32,64} */]>;
1723 defm XOR : BitwiseExclusiveOr;
1725 //==----------------------------------------------------------
1727 class XORBIInst<dag OOL, dag IOL, list<dag> pattern>:
1728 RI10Form<0b01100000, OOL, IOL, "xorbi\t$rT, $rA, $val",
1729 IntegerOp, pattern>;
1731 multiclass XorByteImm
1734 XORBIInst<(outs VECREG:$rT), (ins VECREG:$rA, u10imm:$val),
1735 [(set (v16i8 VECREG:$rT), (xor (v16i8 VECREG:$rA), v16i8U8Imm:$val))]>;
1738 XORBIInst<(outs R8C:$rT), (ins R8C:$rA, u10imm_i8:$val),
1739 [(set R8C:$rT, (xor R8C:$rA, immU8:$val))]>;
1742 defm XORBI : XorByteImm;
1745 RI10Form<0b10100000, (outs VECREG:$rT), (ins VECREG:$rA, u10imm:$val),
1746 "xorhi\t$rT, $rA, $val", IntegerOp,
1747 [(set (v8i16 VECREG:$rT), (xor (v8i16 VECREG:$rA),
1748 v8i16SExt10Imm:$val))]>;
1751 RI10Form<0b10100000, (outs R16C:$rT), (ins R16C:$rA, s10imm:$val),
1752 "xorhi\t$rT, $rA, $val", IntegerOp,
1753 [(set R16C:$rT, (xor R16C:$rA, i16ImmSExt10:$val))]>;
1756 RI10Form<0b00100000, (outs VECREG:$rT), (ins VECREG:$rA, s10imm_i32:$val),
1757 "xori\t$rT, $rA, $val", IntegerOp,
1758 [(set (v4i32 VECREG:$rT), (xor (v4i32 VECREG:$rA),
1759 v4i32SExt10Imm:$val))]>;
1762 RI10Form<0b00100000, (outs R32C:$rT), (ins R32C:$rA, s10imm_i32:$val),
1763 "xori\t$rT, $rA, $val", IntegerOp,
1764 [(set R32C:$rT, (xor R32C:$rA, i32ImmSExt10:$val))]>;
1768 RRForm<0b10010010000, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1769 "nand\t$rT, $rA, $rB", IntegerOp,
1770 [(set (v16i8 VECREG:$rT), (vnot (and (v16i8 VECREG:$rA),
1771 (v16i8 VECREG:$rB))))]>;
1774 RRForm<0b10010010000, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1775 "nand\t$rT, $rA, $rB", IntegerOp,
1776 [(set (v8i16 VECREG:$rT), (vnot (and (v8i16 VECREG:$rA),
1777 (v8i16 VECREG:$rB))))]>;
1780 RRForm<0b10010010000, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1781 "nand\t$rT, $rA, $rB", IntegerOp,
1782 [(set (v4i32 VECREG:$rT), (vnot (and (v4i32 VECREG:$rA),
1783 (v4i32 VECREG:$rB))))]>;
1786 RRForm<0b10010010000, (outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
1787 "nand\t$rT, $rA, $rB", IntegerOp,
1788 [(set R32C:$rT, (not (and R32C:$rA, R32C:$rB)))]>;
1791 RRForm<0b10010010000, (outs R16C:$rT), (ins R16C:$rA, R16C:$rB),
1792 "nand\t$rT, $rA, $rB", IntegerOp,
1793 [(set R16C:$rT, (not (and R16C:$rA, R16C:$rB)))]>;
1796 RRForm<0b10010010000, (outs R8C:$rT), (ins R8C:$rA, R8C:$rB),
1797 "nand\t$rT, $rA, $rB", IntegerOp,
1798 [(set R8C:$rT, (not (and R8C:$rA, R8C:$rB)))]>;
1802 RRForm<0b10010010000, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1803 "nor\t$rT, $rA, $rB", IntegerOp,
1804 [(set (v16i8 VECREG:$rT), (vnot (or (v16i8 VECREG:$rA),
1805 (v16i8 VECREG:$rB))))]>;
1808 RRForm<0b10010010000, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1809 "nor\t$rT, $rA, $rB", IntegerOp,
1810 [(set (v8i16 VECREG:$rT), (vnot (or (v8i16 VECREG:$rA),
1811 (v8i16 VECREG:$rB))))]>;
1814 RRForm<0b10010010000, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1815 "nor\t$rT, $rA, $rB", IntegerOp,
1816 [(set (v4i32 VECREG:$rT), (vnot (or (v4i32 VECREG:$rA),
1817 (v4i32 VECREG:$rB))))]>;
1820 RRForm<0b10010010000, (outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
1821 "nor\t$rT, $rA, $rB", IntegerOp,
1822 [(set R32C:$rT, (not (or R32C:$rA, R32C:$rB)))]>;
1825 RRForm<0b10010010000, (outs R16C:$rT), (ins R16C:$rA, R16C:$rB),
1826 "nor\t$rT, $rA, $rB", IntegerOp,
1827 [(set R16C:$rT, (not (or R16C:$rA, R16C:$rB)))]>;
1830 RRForm<0b10010010000, (outs R8C:$rT), (ins R8C:$rA, R8C:$rB),
1831 "nor\t$rT, $rA, $rB", IntegerOp,
1832 [(set R8C:$rT, (not (or R8C:$rA, R8C:$rB)))]>;
1835 class SELBInst<dag OOL, dag IOL, list<dag> pattern>:
1836 RRRForm<0b1000, OOL, IOL, "selb\t$rT, $rA, $rB, $rC",
1837 IntegerOp, pattern>;
1839 class SELBVecInst<ValueType vectype>:
1840 SELBInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
1841 [(set (vectype VECREG:$rT),
1842 (or (and (vectype VECREG:$rC), (vectype VECREG:$rB)),
1843 (and (vnot (vectype VECREG:$rC)),
1844 (vectype VECREG:$rA))))]>;
1846 class SELBVecVCondInst<ValueType vectype>:
1847 SELBInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
1848 [(set (vectype VECREG:$rT),
1849 (select (vectype VECREG:$rC),
1850 (vectype VECREG:$rB),
1851 (vectype VECREG:$rA)))]>;
1853 class SELBVecCondInst<ValueType vectype>:
1854 SELBInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, R32C:$rC),
1855 [(set (vectype VECREG:$rT),
1857 (vectype VECREG:$rB),
1858 (vectype VECREG:$rA)))]>;
1860 class SELBRegInst<RegisterClass rclass>:
1861 SELBInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB, rclass:$rC),
1863 (or (and rclass:$rB, rclass:$rC),
1864 (and rclass:$rA, (not rclass:$rC))))]>;
1866 class SELBRegCondInst<RegisterClass rcond, RegisterClass rclass>:
1867 SELBInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB, rcond:$rC),
1869 (select rcond:$rC, rclass:$rB, rclass:$rA))]>;
1871 multiclass SelectBits
1873 def v16i8: SELBVecInst<v16i8>;
1874 def v8i16: SELBVecInst<v8i16>;
1875 def v4i32: SELBVecInst<v4i32>;
1876 def v2i64: SELBVecInst<v2i64>;
1878 def r128: SELBRegInst<GPRC>;
1879 def r64: SELBRegInst<R64C>;
1880 def r32: SELBRegInst<R32C>;
1881 def r16: SELBRegInst<R16C>;
1882 def r8: SELBRegInst<R8C>;
1884 def v16i8_cond: SELBVecCondInst<v16i8>;
1885 def v8i16_cond: SELBVecCondInst<v8i16>;
1886 def v4i32_cond: SELBVecCondInst<v4i32>;
1887 def v2i64_cond: SELBVecCondInst<v2i64>;
1889 def v16i8_vcond: SELBVecCondInst<v16i8>;
1890 def v8i16_vcond: SELBVecCondInst<v8i16>;
1891 def v4i32_vcond: SELBVecCondInst<v4i32>;
1892 def v2i64_vcond: SELBVecCondInst<v2i64>;
1895 SELBInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
1896 [(set (v4f32 VECREG:$rT),
1897 (select (v4i32 VECREG:$rC),
1899 (v4f32 VECREG:$rA)))]>;
1901 // SELBr64_cond is defined further down, look for i64 comparisons
1902 def r32_cond: SELBRegCondInst<R32C, R32C>;
1903 def f32_cond: SELBRegCondInst<R32C, R32FP>;
1904 def r16_cond: SELBRegCondInst<R16C, R16C>;
1905 def r8_cond: SELBRegCondInst<R8C, R8C>;
1908 defm SELB : SelectBits;
1910 class SPUselbPatVec<ValueType vectype, SPUInstr inst>:
1911 Pat<(SPUselb (vectype VECREG:$rA), (vectype VECREG:$rB), (vectype VECREG:$rC)),
1912 (inst VECREG:$rA, VECREG:$rB, VECREG:$rC)>;
1914 def : SPUselbPatVec<v16i8, SELBv16i8>;
1915 def : SPUselbPatVec<v8i16, SELBv8i16>;
1916 def : SPUselbPatVec<v4i32, SELBv4i32>;
1917 def : SPUselbPatVec<v2i64, SELBv2i64>;
1919 class SPUselbPatReg<RegisterClass rclass, SPUInstr inst>:
1920 Pat<(SPUselb rclass:$rA, rclass:$rB, rclass:$rC),
1921 (inst rclass:$rA, rclass:$rB, rclass:$rC)>;
1923 def : SPUselbPatReg<R8C, SELBr8>;
1924 def : SPUselbPatReg<R16C, SELBr16>;
1925 def : SPUselbPatReg<R32C, SELBr32>;
1926 def : SPUselbPatReg<R64C, SELBr64>;
1928 // EQV: Equivalence (1 for each same bit, otherwise 0)
1930 // Note: There are a lot of ways to match this bit operator and these patterns
1931 // attempt to be as exhaustive as possible.
1933 class EQVInst<dag OOL, dag IOL, list<dag> pattern>:
1934 RRForm<0b10010010000, OOL, IOL, "eqv\t$rT, $rA, $rB",
1935 IntegerOp, pattern>;
1937 class EQVVecInst<ValueType vectype>:
1938 EQVInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1939 [(set (vectype VECREG:$rT),
1940 (or (and (vectype VECREG:$rA), (vectype VECREG:$rB)),
1941 (and (vnot (vectype VECREG:$rA)),
1942 (vnot (vectype VECREG:$rB)))))]>;
1944 class EQVRegInst<RegisterClass rclass>:
1945 EQVInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
1946 [(set rclass:$rT, (or (and rclass:$rA, rclass:$rB),
1947 (and (not rclass:$rA), (not rclass:$rB))))]>;
1949 class EQVVecPattern1<ValueType vectype>:
1950 EQVInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1951 [(set (vectype VECREG:$rT),
1952 (xor (vectype VECREG:$rA), (vnot (vectype VECREG:$rB))))]>;
1954 class EQVRegPattern1<RegisterClass rclass>:
1955 EQVInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
1956 [(set rclass:$rT, (xor rclass:$rA, (not rclass:$rB)))]>;
1958 class EQVVecPattern2<ValueType vectype>:
1959 EQVInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1960 [(set (vectype VECREG:$rT),
1961 (or (and (vectype VECREG:$rA), (vectype VECREG:$rB)),
1962 (vnot (or (vectype VECREG:$rA), (vectype VECREG:$rB)))))]>;
1964 class EQVRegPattern2<RegisterClass rclass>:
1965 EQVInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
1967 (or (and rclass:$rA, rclass:$rB),
1968 (not (or rclass:$rA, rclass:$rB))))]>;
1970 class EQVVecPattern3<ValueType vectype>:
1971 EQVInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1972 [(set (vectype VECREG:$rT),
1973 (not (xor (vectype VECREG:$rA), (vectype VECREG:$rB))))]>;
1975 class EQVRegPattern3<RegisterClass rclass>:
1976 EQVInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
1977 [(set rclass:$rT, (not (xor rclass:$rA, rclass:$rB)))]>;
1979 multiclass BitEquivalence
1981 def v16i8: EQVVecInst<v16i8>;
1982 def v8i16: EQVVecInst<v8i16>;
1983 def v4i32: EQVVecInst<v4i32>;
1984 def v2i64: EQVVecInst<v2i64>;
1986 def v16i8_1: EQVVecPattern1<v16i8>;
1987 def v8i16_1: EQVVecPattern1<v8i16>;
1988 def v4i32_1: EQVVecPattern1<v4i32>;
1989 def v2i64_1: EQVVecPattern1<v2i64>;
1991 def v16i8_2: EQVVecPattern2<v16i8>;
1992 def v8i16_2: EQVVecPattern2<v8i16>;
1993 def v4i32_2: EQVVecPattern2<v4i32>;
1994 def v2i64_2: EQVVecPattern2<v2i64>;
1996 def v16i8_3: EQVVecPattern3<v16i8>;
1997 def v8i16_3: EQVVecPattern3<v8i16>;
1998 def v4i32_3: EQVVecPattern3<v4i32>;
1999 def v2i64_3: EQVVecPattern3<v2i64>;
2001 def r128: EQVRegInst<GPRC>;
2002 def r64: EQVRegInst<R64C>;
2003 def r32: EQVRegInst<R32C>;
2004 def r16: EQVRegInst<R16C>;
2005 def r8: EQVRegInst<R8C>;
2007 def r128_1: EQVRegPattern1<GPRC>;
2008 def r64_1: EQVRegPattern1<R64C>;
2009 def r32_1: EQVRegPattern1<R32C>;
2010 def r16_1: EQVRegPattern1<R16C>;
2011 def r8_1: EQVRegPattern1<R8C>;
2013 def r128_2: EQVRegPattern2<GPRC>;
2014 def r64_2: EQVRegPattern2<R64C>;
2015 def r32_2: EQVRegPattern2<R32C>;
2016 def r16_2: EQVRegPattern2<R16C>;
2017 def r8_2: EQVRegPattern2<R8C>;
2019 def r128_3: EQVRegPattern3<GPRC>;
2020 def r64_3: EQVRegPattern3<R64C>;
2021 def r32_3: EQVRegPattern3<R32C>;
2022 def r16_3: EQVRegPattern3<R16C>;
2023 def r8_3: EQVRegPattern3<R8C>;
2026 defm EQV: BitEquivalence;
2028 //===----------------------------------------------------------------------===//
2029 // Vector shuffle...
2030 //===----------------------------------------------------------------------===//
2031 // SPUshuffle is generated in LowerVECTOR_SHUFFLE and gets replaced with SHUFB.
2032 // See the SPUshuffle SDNode operand above, which sets up the DAG pattern
2033 // matcher to emit something when the LowerVECTOR_SHUFFLE generates a node with
2034 // the SPUISD::SHUFB opcode.
2035 //===----------------------------------------------------------------------===//
2037 class SHUFBInst<dag OOL, dag IOL, list<dag> pattern>:
2038 RRRForm<0b1000, OOL, IOL, "shufb\t$rT, $rA, $rB, $rC",
2039 IntegerOp, pattern>;
2041 class SHUFBVecInst<ValueType resultvec, ValueType maskvec>:
2042 SHUFBInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
2043 [(set (resultvec VECREG:$rT),
2044 (SPUshuffle (resultvec VECREG:$rA),
2045 (resultvec VECREG:$rB),
2046 (maskvec VECREG:$rC)))]>;
2048 class SHUFBGPRCInst:
2049 SHUFBInst<(outs VECREG:$rT), (ins GPRC:$rA, GPRC:$rB, VECREG:$rC),
2050 [/* no pattern */]>;
2052 multiclass ShuffleBytes
2054 def v16i8 : SHUFBVecInst<v16i8, v16i8>;
2055 def v16i8_m32 : SHUFBVecInst<v16i8, v4i32>;
2056 def v8i16 : SHUFBVecInst<v8i16, v16i8>;
2057 def v8i16_m32 : SHUFBVecInst<v8i16, v4i32>;
2058 def v4i32 : SHUFBVecInst<v4i32, v16i8>;
2059 def v4i32_m32 : SHUFBVecInst<v4i32, v4i32>;
2060 def v2i64 : SHUFBVecInst<v2i64, v16i8>;
2061 def v2i64_m32 : SHUFBVecInst<v2i64, v4i32>;
2063 def v4f32 : SHUFBVecInst<v4f32, v16i8>;
2064 def v4f32_m32 : SHUFBVecInst<v4f32, v4i32>;
2066 def v2f64 : SHUFBVecInst<v2f64, v16i8>;
2067 def v2f64_m32 : SHUFBVecInst<v2f64, v4i32>;
2069 def gprc : SHUFBGPRCInst;
2072 defm SHUFB : ShuffleBytes;
2074 //===----------------------------------------------------------------------===//
2075 // Shift and rotate group:
2076 //===----------------------------------------------------------------------===//
2078 class SHLHInst<dag OOL, dag IOL, list<dag> pattern>:
2079 RRForm<0b11111010000, OOL, IOL, "shlh\t$rT, $rA, $rB",
2080 RotateShift, pattern>;
2082 class SHLHVecInst<ValueType vectype>:
2083 SHLHInst<(outs VECREG:$rT), (ins VECREG:$rA, R16C:$rB),
2084 [(set (vectype VECREG:$rT),
2085 (SPUvec_shl (vectype VECREG:$rA), R16C:$rB))]>;
2087 // $rB gets promoted to 32-bit register type when confronted with
2088 // this llvm assembly code:
2090 // define i16 @shlh_i16_1(i16 %arg1, i16 %arg2) {
2091 // %A = shl i16 %arg1, %arg2
2095 multiclass ShiftLeftHalfword
2097 def v8i16: SHLHVecInst<v8i16>;
2098 def r16: SHLHInst<(outs R16C:$rT), (ins R16C:$rA, R16C:$rB),
2099 [(set R16C:$rT, (shl R16C:$rA, R16C:$rB))]>;
2100 def r16_r32: SHLHInst<(outs R16C:$rT), (ins R16C:$rA, R32C:$rB),
2101 [(set R16C:$rT, (shl R16C:$rA, R32C:$rB))]>;
2104 defm SHLH : ShiftLeftHalfword;
2106 //===----------------------------------------------------------------------===//
2108 class SHLHIInst<dag OOL, dag IOL, list<dag> pattern>:
2109 RI7Form<0b11111010000, OOL, IOL, "shlhi\t$rT, $rA, $val",
2110 RotateShift, pattern>;
2112 class SHLHIVecInst<ValueType vectype>:
2113 SHLHIInst<(outs VECREG:$rT), (ins VECREG:$rA, u7imm:$val),
2114 [(set (vectype VECREG:$rT),
2115 (SPUvec_shl (vectype VECREG:$rA), (i16 uimm7:$val)))]>;
2117 multiclass ShiftLeftHalfwordImm
2119 def v8i16: SHLHIVecInst<v8i16>;
2120 def r16: SHLHIInst<(outs R16C:$rT), (ins R16C:$rA, u7imm:$val),
2121 [(set R16C:$rT, (shl R16C:$rA, (i16 uimm7:$val)))]>;
2124 defm SHLHI : ShiftLeftHalfwordImm;
2126 def : Pat<(SPUvec_shl (v8i16 VECREG:$rA), (i32 uimm7:$val)),
2127 (SHLHIv8i16 VECREG:$rA, uimm7:$val)>;
2129 def : Pat<(shl R16C:$rA, (i32 uimm7:$val)),
2130 (SHLHIr16 R16C:$rA, uimm7:$val)>;
2132 //===----------------------------------------------------------------------===//
2134 class SHLInst<dag OOL, dag IOL, list<dag> pattern>:
2135 RRForm<0b11111010000, OOL, IOL, "shl\t$rT, $rA, $rB",
2136 RotateShift, pattern>;
2138 multiclass ShiftLeftWord
2141 SHLInst<(outs VECREG:$rT), (ins VECREG:$rA, R16C:$rB),
2142 [(set (v4i32 VECREG:$rT),
2143 (SPUvec_shl (v4i32 VECREG:$rA), R16C:$rB))]>;
2145 SHLInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
2146 [(set R32C:$rT, (shl R32C:$rA, R32C:$rB))]>;
2149 defm SHL: ShiftLeftWord;
2151 //===----------------------------------------------------------------------===//
2153 class SHLIInst<dag OOL, dag IOL, list<dag> pattern>:
2154 RI7Form<0b11111010000, OOL, IOL, "shli\t$rT, $rA, $val",
2155 RotateShift, pattern>;
2157 multiclass ShiftLeftWordImm
2160 SHLIInst<(outs VECREG:$rT), (ins VECREG:$rA, u7imm_i32:$val),
2161 [(set (v4i32 VECREG:$rT),
2162 (SPUvec_shl (v4i32 VECREG:$rA), (i32 uimm7:$val)))]>;
2165 SHLIInst<(outs R32C:$rT), (ins R32C:$rA, u7imm_i32:$val),
2166 [(set R32C:$rT, (shl R32C:$rA, (i32 uimm7:$val)))]>;
2169 defm SHLI : ShiftLeftWordImm;
2171 //===----------------------------------------------------------------------===//
2172 // SHLQBI vec form: Note that this will shift the entire vector (the 128-bit
2173 // register) to the left. Vector form is here to ensure type correctness.
2175 // The shift count is in the lowest 3 bits (29-31) of $rB, so only a bit shift
2176 // of 7 bits is actually possible.
2178 // Note also that SHLQBI/SHLQBII are used in conjunction with SHLQBY/SHLQBYI
2179 // to shift i64 and i128. SHLQBI is the residual left over after shifting by
2180 // bytes with SHLQBY.
2182 class SHLQBIInst<dag OOL, dag IOL, list<dag> pattern>:
2183 RRForm<0b11011011100, OOL, IOL, "shlqbi\t$rT, $rA, $rB",
2184 RotateShift, pattern>;
2186 class SHLQBIVecInst<ValueType vectype>:
2187 SHLQBIInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2188 [(set (vectype VECREG:$rT),
2189 (SPUshlquad_l_bits (vectype VECREG:$rA), R32C:$rB))]>;
2191 multiclass ShiftLeftQuadByBits
2193 def v16i8: SHLQBIVecInst<v16i8>;
2194 def v8i16: SHLQBIVecInst<v8i16>;
2195 def v4i32: SHLQBIVecInst<v4i32>;
2196 def v4f32: SHLQBIVecInst<v4f32>;
2197 def v2i64: SHLQBIVecInst<v2i64>;
2198 def v2f64: SHLQBIVecInst<v2f64>;
2201 defm SHLQBI : ShiftLeftQuadByBits;
2203 // See note above on SHLQBI. In this case, the predicate actually does then
2204 // enforcement, whereas with SHLQBI, we have to "take it on faith."
2205 class SHLQBIIInst<dag OOL, dag IOL, list<dag> pattern>:
2206 RI7Form<0b11011111100, OOL, IOL, "shlqbii\t$rT, $rA, $val",
2207 RotateShift, pattern>;
2209 class SHLQBIIVecInst<ValueType vectype>:
2210 SHLQBIIInst<(outs VECREG:$rT), (ins VECREG:$rA, u7imm_i32:$val),
2211 [(set (vectype VECREG:$rT),
2212 (SPUshlquad_l_bits (vectype VECREG:$rA), (i32 bitshift:$val)))]>;
2214 multiclass ShiftLeftQuadByBitsImm
2216 def v16i8 : SHLQBIIVecInst<v16i8>;
2217 def v8i16 : SHLQBIIVecInst<v8i16>;
2218 def v4i32 : SHLQBIIVecInst<v4i32>;
2219 def v4f32 : SHLQBIIVecInst<v4f32>;
2220 def v2i64 : SHLQBIIVecInst<v2i64>;
2221 def v2f64 : SHLQBIIVecInst<v2f64>;
2224 defm SHLQBII : ShiftLeftQuadByBitsImm;
2226 // SHLQBY, SHLQBYI vector forms: Shift the entire vector to the left by bytes,
2227 // not by bits. See notes above on SHLQBI.
2229 class SHLQBYInst<dag OOL, dag IOL, list<dag> pattern>:
2230 RI7Form<0b11111011100, OOL, IOL, "shlqby\t$rT, $rA, $rB",
2231 RotateShift, pattern>;
2233 class SHLQBYVecInst<ValueType vectype>:
2234 SHLQBYInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2235 [(set (vectype VECREG:$rT),
2236 (SPUshlquad_l_bytes (vectype VECREG:$rA), R32C:$rB))]>;
2238 multiclass ShiftLeftQuadBytes
2240 def v16i8: SHLQBYVecInst<v16i8>;
2241 def v8i16: SHLQBYVecInst<v8i16>;
2242 def v4i32: SHLQBYVecInst<v4i32>;
2243 def v4f32: SHLQBYVecInst<v4f32>;
2244 def v2i64: SHLQBYVecInst<v2i64>;
2245 def v2f64: SHLQBYVecInst<v2f64>;
2246 def r128: SHLQBYInst<(outs GPRC:$rT), (ins GPRC:$rA, R32C:$rB),
2247 [(set GPRC:$rT, (SPUshlquad_l_bytes GPRC:$rA, R32C:$rB))]>;
2250 defm SHLQBY: ShiftLeftQuadBytes;
2252 class SHLQBYIInst<dag OOL, dag IOL, list<dag> pattern>:
2253 RI7Form<0b11111111100, OOL, IOL, "shlqbyi\t$rT, $rA, $val",
2254 RotateShift, pattern>;
2256 class SHLQBYIVecInst<ValueType vectype>:
2257 SHLQBYIInst<(outs VECREG:$rT), (ins VECREG:$rA, u7imm_i32:$val),
2258 [(set (vectype VECREG:$rT),
2259 (SPUshlquad_l_bytes (vectype VECREG:$rA), (i32 uimm7:$val)))]>;
2261 multiclass ShiftLeftQuadBytesImm
2263 def v16i8: SHLQBYIVecInst<v16i8>;
2264 def v8i16: SHLQBYIVecInst<v8i16>;
2265 def v4i32: SHLQBYIVecInst<v4i32>;
2266 def v4f32: SHLQBYIVecInst<v4f32>;
2267 def v2i64: SHLQBYIVecInst<v2i64>;
2268 def v2f64: SHLQBYIVecInst<v2f64>;
2269 def r128: SHLQBYIInst<(outs GPRC:$rT), (ins GPRC:$rA, u7imm_i32:$val),
2271 (SPUshlquad_l_bytes GPRC:$rA, (i32 uimm7:$val)))]>;
2274 defm SHLQBYI : ShiftLeftQuadBytesImm;
2276 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2278 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2279 class ROTHInst<dag OOL, dag IOL, list<dag> pattern>:
2280 RRForm<0b00111010000, OOL, IOL, "roth\t$rT, $rA, $rB",
2281 RotateShift, pattern>;
2283 class ROTHVecInst<ValueType vectype>:
2284 ROTHInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
2285 [(set (vectype VECREG:$rT),
2286 (SPUvec_rotl VECREG:$rA, VECREG:$rB))]>;
2288 class ROTHRegInst<RegisterClass rclass>:
2289 ROTHInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
2290 [(set rclass:$rT, (rotl rclass:$rA, rclass:$rB))]>;
2292 multiclass RotateLeftHalfword
2294 def v8i16: ROTHVecInst<v8i16>;
2295 def r16: ROTHRegInst<R16C>;
2298 defm ROTH: RotateLeftHalfword;
2300 def ROTHr16_r32: ROTHInst<(outs R16C:$rT), (ins R16C:$rA, R32C:$rB),
2301 [(set R16C:$rT, (rotl R16C:$rA, R32C:$rB))]>;
2303 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2304 // Rotate halfword, immediate:
2305 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2306 class ROTHIInst<dag OOL, dag IOL, list<dag> pattern>:
2307 RI7Form<0b00111110000, OOL, IOL, "rothi\t$rT, $rA, $val",
2308 RotateShift, pattern>;
2310 class ROTHIVecInst<ValueType vectype>:
2311 ROTHIInst<(outs VECREG:$rT), (ins VECREG:$rA, u7imm:$val),
2312 [(set (vectype VECREG:$rT),
2313 (SPUvec_rotl VECREG:$rA, (i16 uimm7:$val)))]>;
2315 multiclass RotateLeftHalfwordImm
2317 def v8i16: ROTHIVecInst<v8i16>;
2318 def r16: ROTHIInst<(outs R16C:$rT), (ins R16C:$rA, u7imm:$val),
2319 [(set R16C:$rT, (rotl R16C:$rA, (i16 uimm7:$val)))]>;
2320 def r16_r32: ROTHIInst<(outs R16C:$rT), (ins R16C:$rA, u7imm_i32:$val),
2321 [(set R16C:$rT, (rotl R16C:$rA, (i32 uimm7:$val)))]>;
2324 defm ROTHI: RotateLeftHalfwordImm;
2326 def : Pat<(SPUvec_rotl VECREG:$rA, (i32 uimm7:$val)),
2327 (ROTHIv8i16 VECREG:$rA, imm:$val)>;
2329 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2331 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2333 class ROTInst<dag OOL, dag IOL, list<dag> pattern>:
2334 RRForm<0b00011010000, OOL, IOL, "rot\t$rT, $rA, $rB",
2335 RotateShift, pattern>;
2337 class ROTVecInst<ValueType vectype>:
2338 ROTInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2339 [(set (vectype VECREG:$rT),
2340 (SPUvec_rotl (vectype VECREG:$rA), R32C:$rB))]>;
2342 class ROTRegInst<RegisterClass rclass>:
2343 ROTInst<(outs rclass:$rT), (ins rclass:$rA, R32C:$rB),
2345 (rotl rclass:$rA, R32C:$rB))]>;
2347 multiclass RotateLeftWord
2349 def v4i32: ROTVecInst<v4i32>;
2350 def r32: ROTRegInst<R32C>;
2353 defm ROT: RotateLeftWord;
2355 // The rotate amount is in the same bits whether we've got an 8-bit, 16-bit or
2357 def ROTr32_r16_anyext:
2358 ROTInst<(outs R32C:$rT), (ins R32C:$rA, R16C:$rB),
2359 [(set R32C:$rT, (rotl R32C:$rA, (i32 (anyext R16C:$rB))))]>;
2361 def : Pat<(rotl R32C:$rA, (i32 (zext R16C:$rB))),
2362 (ROTr32_r16_anyext R32C:$rA, R16C:$rB)>;
2364 def : Pat<(rotl R32C:$rA, (i32 (sext R16C:$rB))),
2365 (ROTr32_r16_anyext R32C:$rA, R16C:$rB)>;
2367 def ROTr32_r8_anyext:
2368 ROTInst<(outs R32C:$rT), (ins R32C:$rA, R8C:$rB),
2369 [(set R32C:$rT, (rotl R32C:$rA, (i32 (anyext R8C:$rB))))]>;
2371 def : Pat<(rotl R32C:$rA, (i32 (zext R8C:$rB))),
2372 (ROTr32_r8_anyext R32C:$rA, R8C:$rB)>;
2374 def : Pat<(rotl R32C:$rA, (i32 (sext R8C:$rB))),
2375 (ROTr32_r8_anyext R32C:$rA, R8C:$rB)>;
2377 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2378 // Rotate word, immediate
2379 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2381 class ROTIInst<dag OOL, dag IOL, list<dag> pattern>:
2382 RI7Form<0b00011110000, OOL, IOL, "roti\t$rT, $rA, $val",
2383 RotateShift, pattern>;
2385 class ROTIVecInst<ValueType vectype, Operand optype, ValueType inttype, PatLeaf pred>:
2386 ROTIInst<(outs VECREG:$rT), (ins VECREG:$rA, optype:$val),
2387 [(set (vectype VECREG:$rT),
2388 (SPUvec_rotl (vectype VECREG:$rA), (inttype pred:$val)))]>;
2390 class ROTIRegInst<RegisterClass rclass, Operand optype, ValueType inttype, PatLeaf pred>:
2391 ROTIInst<(outs rclass:$rT), (ins rclass:$rA, optype:$val),
2392 [(set rclass:$rT, (rotl rclass:$rA, (inttype pred:$val)))]>;
2394 multiclass RotateLeftWordImm
2396 def v4i32: ROTIVecInst<v4i32, u7imm_i32, i32, uimm7>;
2397 def v4i32_i16: ROTIVecInst<v4i32, u7imm, i16, uimm7>;
2398 def v4i32_i8: ROTIVecInst<v4i32, u7imm_i8, i8, uimm7>;
2400 def r32: ROTIRegInst<R32C, u7imm_i32, i32, uimm7>;
2401 def r32_i16: ROTIRegInst<R32C, u7imm, i16, uimm7>;
2402 def r32_i8: ROTIRegInst<R32C, u7imm_i8, i8, uimm7>;
2405 defm ROTI : RotateLeftWordImm;
2407 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2408 // Rotate quad by byte (count)
2409 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2411 class ROTQBYInst<dag OOL, dag IOL, list<dag> pattern>:
2412 RRForm<0b00111011100, OOL, IOL, "rotqby\t$rT, $rA, $rB",
2413 RotateShift, pattern>;
2415 class ROTQBYVecInst<ValueType vectype>:
2416 ROTQBYInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2417 [(set (vectype VECREG:$rT),
2418 (SPUrotbytes_left (vectype VECREG:$rA), R32C:$rB))]>;
2420 multiclass RotateQuadLeftByBytes
2422 def v16i8: ROTQBYVecInst<v16i8>;
2423 def v8i16: ROTQBYVecInst<v8i16>;
2424 def v4i32: ROTQBYVecInst<v4i32>;
2425 def v4f32: ROTQBYVecInst<v4f32>;
2426 def v2i64: ROTQBYVecInst<v2i64>;
2427 def v2f64: ROTQBYVecInst<v2f64>;
2430 defm ROTQBY: RotateQuadLeftByBytes;
2432 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2433 // Rotate quad by byte (count), immediate
2434 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2436 class ROTQBYIInst<dag OOL, dag IOL, list<dag> pattern>:
2437 RI7Form<0b00111111100, OOL, IOL, "rotqbyi\t$rT, $rA, $val",
2438 RotateShift, pattern>;
2440 class ROTQBYIVecInst<ValueType vectype>:
2441 ROTQBYIInst<(outs VECREG:$rT), (ins VECREG:$rA, u7imm:$val),
2442 [(set (vectype VECREG:$rT),
2443 (SPUrotbytes_left (vectype VECREG:$rA), (i16 uimm7:$val)))]>;
2445 multiclass RotateQuadByBytesImm
2447 def v16i8: ROTQBYIVecInst<v16i8>;
2448 def v8i16: ROTQBYIVecInst<v8i16>;
2449 def v4i32: ROTQBYIVecInst<v4i32>;
2450 def v4f32: ROTQBYIVecInst<v4f32>;
2451 def v2i64: ROTQBYIVecInst<v2i64>;
2452 def vfi64: ROTQBYIVecInst<v2f64>;
2455 defm ROTQBYI: RotateQuadByBytesImm;
2457 // See ROTQBY note above.
2458 class ROTQBYBIInst<dag OOL, dag IOL, list<dag> pattern>:
2459 RI7Form<0b00110011100, OOL, IOL,
2460 "rotqbybi\t$rT, $rA, $shift",
2461 RotateShift, pattern>;
2463 class ROTQBYBIVecInst<ValueType vectype, RegisterClass rclass>:
2464 ROTQBYBIInst<(outs VECREG:$rT), (ins VECREG:$rA, rclass:$shift),
2465 [(set (vectype VECREG:$rT),
2466 (SPUrotbytes_left_bits (vectype VECREG:$rA), rclass:$shift))]>;
2468 multiclass RotateQuadByBytesByBitshift {
2469 def v16i8_r32: ROTQBYBIVecInst<v16i8, R32C>;
2470 def v8i16_r32: ROTQBYBIVecInst<v8i16, R32C>;
2471 def v4i32_r32: ROTQBYBIVecInst<v4i32, R32C>;
2472 def v2i64_r32: ROTQBYBIVecInst<v2i64, R32C>;
2475 defm ROTQBYBI : RotateQuadByBytesByBitshift;
2477 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2478 // See ROTQBY note above.
2480 // Assume that the user of this instruction knows to shift the rotate count
2482 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2484 class ROTQBIInst<dag OOL, dag IOL, list<dag> pattern>:
2485 RRForm<0b00011011100, OOL, IOL, "rotqbi\t$rT, $rA, $rB",
2486 RotateShift, pattern>;
2488 class ROTQBIVecInst<ValueType vectype>:
2489 ROTQBIInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2490 [/* no pattern yet */]>;
2492 class ROTQBIRegInst<RegisterClass rclass>:
2493 ROTQBIInst<(outs rclass:$rT), (ins rclass:$rA, R32C:$rB),
2494 [/* no pattern yet */]>;
2496 multiclass RotateQuadByBitCount
2498 def v16i8: ROTQBIVecInst<v16i8>;
2499 def v8i16: ROTQBIVecInst<v8i16>;
2500 def v4i32: ROTQBIVecInst<v4i32>;
2501 def v2i64: ROTQBIVecInst<v2i64>;
2503 def r128: ROTQBIRegInst<GPRC>;
2504 def r64: ROTQBIRegInst<R64C>;
2507 defm ROTQBI: RotateQuadByBitCount;
2509 class ROTQBIIInst<dag OOL, dag IOL, list<dag> pattern>:
2510 RI7Form<0b00011111100, OOL, IOL, "rotqbii\t$rT, $rA, $val",
2511 RotateShift, pattern>;
2513 class ROTQBIIVecInst<ValueType vectype, Operand optype, ValueType inttype,
2515 ROTQBIIInst<(outs VECREG:$rT), (ins VECREG:$rA, optype:$val),
2516 [/* no pattern yet */]>;
2518 class ROTQBIIRegInst<RegisterClass rclass, Operand optype, ValueType inttype,
2520 ROTQBIIInst<(outs rclass:$rT), (ins rclass:$rA, optype:$val),
2521 [/* no pattern yet */]>;
2523 multiclass RotateQuadByBitCountImm
2525 def v16i8: ROTQBIIVecInst<v16i8, u7imm_i32, i32, uimm7>;
2526 def v8i16: ROTQBIIVecInst<v8i16, u7imm_i32, i32, uimm7>;
2527 def v4i32: ROTQBIIVecInst<v4i32, u7imm_i32, i32, uimm7>;
2528 def v2i64: ROTQBIIVecInst<v2i64, u7imm_i32, i32, uimm7>;
2530 def r128: ROTQBIIRegInst<GPRC, u7imm_i32, i32, uimm7>;
2531 def r64: ROTQBIIRegInst<R64C, u7imm_i32, i32, uimm7>;
2534 defm ROTQBII : RotateQuadByBitCountImm;
2536 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2537 // ROTHM v8i16 form:
2538 // NOTE(1): No vector rotate is generated by the C/C++ frontend (today),
2539 // so this only matches a synthetically generated/lowered code
2541 // NOTE(2): $rB must be negated before the right rotate!
2542 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2544 class ROTHMInst<dag OOL, dag IOL, list<dag> pattern>:
2545 RRForm<0b10111010000, OOL, IOL, "rothm\t$rT, $rA, $rB",
2546 RotateShift, pattern>;
2549 ROTHMInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2550 [/* see patterns below - $rB must be negated */]>;
2552 def : Pat<(SPUvec_srl (v8i16 VECREG:$rA), R32C:$rB),
2553 (ROTHMv8i16 VECREG:$rA, (SFIr32 R32C:$rB, 0))>;
2555 def : Pat<(SPUvec_srl (v8i16 VECREG:$rA), R16C:$rB),
2556 (ROTHMv8i16 VECREG:$rA,
2557 (SFIr32 (XSHWr16 R16C:$rB), 0))>;
2559 def : Pat<(SPUvec_srl (v8i16 VECREG:$rA), R8C:$rB),
2560 (ROTHMv8i16 VECREG:$rA,
2561 (SFIr32 (XSHWr16 (XSBHr8 R8C:$rB) ), 0))>;
2563 // ROTHM r16 form: Rotate 16-bit quantity to right, zero fill at the left
2564 // Note: This instruction doesn't match a pattern because rB must be negated
2565 // for the instruction to work. Thus, the pattern below the instruction!
2568 ROTHMInst<(outs R16C:$rT), (ins R16C:$rA, R32C:$rB),
2569 [/* see patterns below - $rB must be negated! */]>;
2571 def : Pat<(srl R16C:$rA, R32C:$rB),
2572 (ROTHMr16 R16C:$rA, (SFIr32 R32C:$rB, 0))>;
2574 def : Pat<(srl R16C:$rA, R16C:$rB),
2576 (SFIr32 (XSHWr16 R16C:$rB), 0))>;
2578 def : Pat<(srl R16C:$rA, R8C:$rB),
2580 (SFIr32 (XSHWr16 (XSBHr8 R8C:$rB) ), 0))>;
2582 // ROTHMI v8i16 form: See the comment for ROTHM v8i16. The difference here is
2583 // that the immediate can be complemented, so that the user doesn't have to
2586 class ROTHMIInst<dag OOL, dag IOL, list<dag> pattern>:
2587 RI7Form<0b10111110000, OOL, IOL, "rothmi\t$rT, $rA, $val",
2588 RotateShift, pattern>;
2591 ROTHMIInst<(outs VECREG:$rT), (ins VECREG:$rA, rothNeg7imm:$val),
2592 [/* no pattern */]>;
2594 def : Pat<(SPUvec_srl (v8i16 VECREG:$rA), (i32 imm:$val)),
2595 (ROTHMIv8i16 VECREG:$rA, imm:$val)>;
2597 def: Pat<(SPUvec_srl (v8i16 VECREG:$rA), (i16 imm:$val)),
2598 (ROTHMIv8i16 VECREG:$rA, imm:$val)>;
2600 def: Pat<(SPUvec_srl (v8i16 VECREG:$rA), (i8 imm:$val)),
2601 (ROTHMIv8i16 VECREG:$rA, imm:$val)>;
2604 ROTHMIInst<(outs R16C:$rT), (ins R16C:$rA, rothNeg7imm:$val),
2605 [/* no pattern */]>;
2607 def: Pat<(srl R16C:$rA, (i32 uimm7:$val)),
2608 (ROTHMIr16 R16C:$rA, uimm7:$val)>;
2610 def: Pat<(srl R16C:$rA, (i16 uimm7:$val)),
2611 (ROTHMIr16 R16C:$rA, uimm7:$val)>;
2613 def: Pat<(srl R16C:$rA, (i8 uimm7:$val)),
2614 (ROTHMIr16 R16C:$rA, uimm7:$val)>;
2616 // ROTM v4i32 form: See the ROTHM v8i16 comments.
2617 class ROTMInst<dag OOL, dag IOL, list<dag> pattern>:
2618 RRForm<0b10011010000, OOL, IOL, "rotm\t$rT, $rA, $rB",
2619 RotateShift, pattern>;
2622 ROTMInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2623 [/* see patterns below - $rB must be negated */]>;
2625 def : Pat<(SPUvec_srl VECREG:$rA, R32C:$rB),
2626 (ROTMv4i32 VECREG:$rA, (SFIr32 R32C:$rB, 0))>;
2628 def : Pat<(SPUvec_srl VECREG:$rA, R16C:$rB),
2629 (ROTMv4i32 VECREG:$rA,
2630 (SFIr32 (XSHWr16 R16C:$rB), 0))>;
2632 def : Pat<(SPUvec_srl VECREG:$rA, R8C:$rB),
2633 (ROTMv4i32 VECREG:$rA,
2634 (SFIr32 (XSHWr16 (XSBHr8 R8C:$rB)), 0))>;
2637 ROTMInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
2638 [/* see patterns below - $rB must be negated */]>;
2640 def : Pat<(srl R32C:$rA, R32C:$rB),
2641 (ROTMr32 R32C:$rA, (SFIr32 R32C:$rB, 0))>;
2643 def : Pat<(srl R32C:$rA, R16C:$rB),
2645 (SFIr32 (XSHWr16 R16C:$rB), 0))>;
2647 def : Pat<(srl R32C:$rA, R8C:$rB),
2649 (SFIr32 (XSHWr16 (XSBHr8 R8C:$rB)), 0))>;
2651 // ROTMI v4i32 form: See the comment for ROTHM v8i16.
2653 RI7Form<0b10011110000, (outs VECREG:$rT), (ins VECREG:$rA, rotNeg7imm:$val),
2654 "rotmi\t$rT, $rA, $val", RotateShift,
2655 [(set (v4i32 VECREG:$rT),
2656 (SPUvec_srl VECREG:$rA, (i32 uimm7:$val)))]>;
2658 def : Pat<(SPUvec_srl VECREG:$rA, (i16 uimm7:$val)),
2659 (ROTMIv4i32 VECREG:$rA, uimm7:$val)>;
2661 def : Pat<(SPUvec_srl VECREG:$rA, (i8 uimm7:$val)),
2662 (ROTMIv4i32 VECREG:$rA, uimm7:$val)>;
2664 // ROTMI r32 form: know how to complement the immediate value.
2666 RI7Form<0b10011110000, (outs R32C:$rT), (ins R32C:$rA, rotNeg7imm:$val),
2667 "rotmi\t$rT, $rA, $val", RotateShift,
2668 [(set R32C:$rT, (srl R32C:$rA, (i32 uimm7:$val)))]>;
2670 def : Pat<(srl R32C:$rA, (i16 imm:$val)),
2671 (ROTMIr32 R32C:$rA, uimm7:$val)>;
2673 def : Pat<(srl R32C:$rA, (i8 imm:$val)),
2674 (ROTMIr32 R32C:$rA, uimm7:$val)>;
2676 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2677 // ROTQMBYvec: This is a vector form merely so that when used in an
2678 // instruction pattern, type checking will succeed. This instruction assumes
2679 // that the user knew to negate $rB.
2680 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2682 class ROTQMBYInst<dag OOL, dag IOL, list<dag> pattern>:
2683 RRForm<0b10111011100, OOL, IOL, "rotqmby\t$rT, $rA, $rB",
2684 RotateShift, pattern>;
2686 class ROTQMBYVecInst<ValueType vectype>:
2687 ROTQMBYInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2688 [/* no pattern, $rB must be negated */]>;
2690 class ROTQMBYRegInst<RegisterClass rclass>:
2691 ROTQMBYInst<(outs rclass:$rT), (ins rclass:$rA, R32C:$rB),
2692 [/* no pattern */]>;
2694 multiclass RotateQuadBytes
2696 def v16i8: ROTQMBYVecInst<v16i8>;
2697 def v8i16: ROTQMBYVecInst<v8i16>;
2698 def v4i32: ROTQMBYVecInst<v4i32>;
2699 def v2i64: ROTQMBYVecInst<v2i64>;
2701 def r128: ROTQMBYRegInst<GPRC>;
2702 def r64: ROTQMBYRegInst<R64C>;
2705 defm ROTQMBY : RotateQuadBytes;
2707 class ROTQMBYIInst<dag OOL, dag IOL, list<dag> pattern>:
2708 RI7Form<0b10111111100, OOL, IOL, "rotqmbyi\t$rT, $rA, $val",
2709 RotateShift, pattern>;
2711 class ROTQMBYIVecInst<ValueType vectype>:
2712 ROTQMBYIInst<(outs VECREG:$rT), (ins VECREG:$rA, rotNeg7imm:$val),
2713 [/* no pattern */]>;
2715 class ROTQMBYIRegInst<RegisterClass rclass, Operand optype, ValueType inttype, PatLeaf pred>:
2716 ROTQMBYIInst<(outs rclass:$rT), (ins rclass:$rA, optype:$val),
2717 [/* no pattern */]>;
2719 multiclass RotateQuadBytesImm
2721 def v16i8: ROTQMBYIVecInst<v16i8>;
2722 def v8i16: ROTQMBYIVecInst<v8i16>;
2723 def v4i32: ROTQMBYIVecInst<v4i32>;
2724 def v2i64: ROTQMBYIVecInst<v2i64>;
2726 def r128: ROTQMBYIRegInst<GPRC, rotNeg7imm, i32, uimm7>;
2727 def r64: ROTQMBYIRegInst<R64C, rotNeg7imm, i32, uimm7>;
2730 defm ROTQMBYI : RotateQuadBytesImm;
2732 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2733 // Rotate right and mask by bit count
2734 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2736 class ROTQMBYBIInst<dag OOL, dag IOL, list<dag> pattern>:
2737 RRForm<0b10110011100, OOL, IOL, "rotqmbybi\t$rT, $rA, $rB",
2738 RotateShift, pattern>;
2740 class ROTQMBYBIVecInst<ValueType vectype>:
2741 ROTQMBYBIInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2742 [/* no pattern, */]>;
2744 multiclass RotateMaskQuadByBitCount
2746 def v16i8: ROTQMBYBIVecInst<v16i8>;
2747 def v8i16: ROTQMBYBIVecInst<v8i16>;
2748 def v4i32: ROTQMBYBIVecInst<v4i32>;
2749 def v2i64: ROTQMBYBIVecInst<v2i64>;
2752 defm ROTQMBYBI: RotateMaskQuadByBitCount;
2754 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2755 // Rotate quad and mask by bits
2756 // Note that the rotate amount has to be negated
2757 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2759 class ROTQMBIInst<dag OOL, dag IOL, list<dag> pattern>:
2760 RRForm<0b10011011100, OOL, IOL, "rotqmbi\t$rT, $rA, $rB",
2761 RotateShift, pattern>;
2763 class ROTQMBIVecInst<ValueType vectype>:
2764 ROTQMBIInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2765 [/* no pattern */]>;
2767 class ROTQMBIRegInst<RegisterClass rclass>:
2768 ROTQMBIInst<(outs rclass:$rT), (ins rclass:$rA, R32C:$rB),
2769 [/* no pattern */]>;
2771 multiclass RotateMaskQuadByBits
2773 def v16i8: ROTQMBIVecInst<v16i8>;
2774 def v8i16: ROTQMBIVecInst<v8i16>;
2775 def v4i32: ROTQMBIVecInst<v4i32>;
2776 def v2i64: ROTQMBIVecInst<v2i64>;
2778 def r128: ROTQMBIRegInst<GPRC>;
2779 def r64: ROTQMBIRegInst<R64C>;
2782 defm ROTQMBI: RotateMaskQuadByBits;
2784 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2785 // Rotate quad and mask by bits, immediate
2786 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2788 class ROTQMBIIInst<dag OOL, dag IOL, list<dag> pattern>:
2789 RI7Form<0b10011111100, OOL, IOL, "rotqmbii\t$rT, $rA, $val",
2790 RotateShift, pattern>;
2792 class ROTQMBIIVecInst<ValueType vectype>:
2793 ROTQMBIIInst<(outs VECREG:$rT), (ins VECREG:$rA, rotNeg7imm:$val),
2794 [/* no pattern */]>;
2796 class ROTQMBIIRegInst<RegisterClass rclass>:
2797 ROTQMBIIInst<(outs rclass:$rT), (ins rclass:$rA, rotNeg7imm:$val),
2798 [/* no pattern */]>;
2800 multiclass RotateMaskQuadByBitsImm
2802 def v16i8: ROTQMBIIVecInst<v16i8>;
2803 def v8i16: ROTQMBIIVecInst<v8i16>;
2804 def v4i32: ROTQMBIIVecInst<v4i32>;
2805 def v2i64: ROTQMBIIVecInst<v2i64>;
2807 def r128: ROTQMBIIRegInst<GPRC>;
2808 def r64: ROTQMBIIRegInst<R64C>;
2811 defm ROTQMBII: RotateMaskQuadByBitsImm;
2813 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2814 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2817 RRForm<0b01111010000, (outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2818 "rotmah\t$rT, $rA, $rB", RotateShift,
2819 [/* see patterns below - $rB must be negated */]>;
2821 def : Pat<(SPUvec_sra VECREG:$rA, R32C:$rB),
2822 (ROTMAHv8i16 VECREG:$rA, (SFIr32 R32C:$rB, 0))>;
2824 def : Pat<(SPUvec_sra VECREG:$rA, R16C:$rB),
2825 (ROTMAHv8i16 VECREG:$rA,
2826 (SFIr32 (XSHWr16 R16C:$rB), 0))>;
2828 def : Pat<(SPUvec_sra VECREG:$rA, R8C:$rB),
2829 (ROTMAHv8i16 VECREG:$rA,
2830 (SFIr32 (XSHWr16 (XSBHr8 R8C:$rB)), 0))>;
2833 RRForm<0b01111010000, (outs R16C:$rT), (ins R16C:$rA, R32C:$rB),
2834 "rotmah\t$rT, $rA, $rB", RotateShift,
2835 [/* see patterns below - $rB must be negated */]>;
2837 def : Pat<(sra R16C:$rA, R32C:$rB),
2838 (ROTMAHr16 R16C:$rA, (SFIr32 R32C:$rB, 0))>;
2840 def : Pat<(sra R16C:$rA, R16C:$rB),
2841 (ROTMAHr16 R16C:$rA,
2842 (SFIr32 (XSHWr16 R16C:$rB), 0))>;
2844 def : Pat<(sra R16C:$rA, R8C:$rB),
2845 (ROTMAHr16 R16C:$rA,
2846 (SFIr32 (XSHWr16 (XSBHr8 R8C:$rB)), 0))>;
2849 RRForm<0b01111110000, (outs VECREG:$rT), (ins VECREG:$rA, rothNeg7imm:$val),
2850 "rotmahi\t$rT, $rA, $val", RotateShift,
2851 [(set (v8i16 VECREG:$rT),
2852 (SPUvec_sra (v8i16 VECREG:$rA), (i32 uimm7:$val)))]>;
2854 def : Pat<(SPUvec_sra (v8i16 VECREG:$rA), (i16 uimm7:$val)),
2855 (ROTMAHIv8i16 (v8i16 VECREG:$rA), (i32 uimm7:$val))>;
2857 def : Pat<(SPUvec_sra (v8i16 VECREG:$rA), (i8 uimm7:$val)),
2858 (ROTMAHIv8i16 (v8i16 VECREG:$rA), (i32 uimm7:$val))>;
2861 RRForm<0b01111110000, (outs R16C:$rT), (ins R16C:$rA, rothNeg7imm_i16:$val),
2862 "rotmahi\t$rT, $rA, $val", RotateShift,
2863 [(set R16C:$rT, (sra R16C:$rA, (i16 uimm7:$val)))]>;
2865 def : Pat<(sra R16C:$rA, (i32 imm:$val)),
2866 (ROTMAHIr16 R16C:$rA, uimm7:$val)>;
2868 def : Pat<(sra R16C:$rA, (i8 imm:$val)),
2869 (ROTMAHIr16 R16C:$rA, uimm7:$val)>;
2872 RRForm<0b01011010000, (outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2873 "rotma\t$rT, $rA, $rB", RotateShift,
2874 [/* see patterns below - $rB must be negated */]>;
2876 def : Pat<(SPUvec_sra VECREG:$rA, R32C:$rB),
2877 (ROTMAv4i32 (v4i32 VECREG:$rA), (SFIr32 R32C:$rB, 0))>;
2879 def : Pat<(SPUvec_sra VECREG:$rA, R16C:$rB),
2880 (ROTMAv4i32 (v4i32 VECREG:$rA),
2881 (SFIr32 (XSHWr16 R16C:$rB), 0))>;
2883 def : Pat<(SPUvec_sra VECREG:$rA, R8C:$rB),
2884 (ROTMAv4i32 (v4i32 VECREG:$rA),
2885 (SFIr32 (XSHWr16 (XSBHr8 R8C:$rB)), 0))>;
2888 RRForm<0b01011010000, (outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
2889 "rotma\t$rT, $rA, $rB", RotateShift,
2890 [/* see patterns below - $rB must be negated */]>;
2892 def : Pat<(sra R32C:$rA, R32C:$rB),
2893 (ROTMAr32 R32C:$rA, (SFIr32 R32C:$rB, 0))>;
2895 def : Pat<(sra R32C:$rA, R16C:$rB),
2897 (SFIr32 (XSHWr16 R16C:$rB), 0))>;
2899 def : Pat<(sra R32C:$rA, R8C:$rB),
2901 (SFIr32 (XSHWr16 (XSBHr8 R8C:$rB)), 0))>;
2903 class ROTMAIInst<dag OOL, dag IOL, list<dag> pattern>:
2904 RRForm<0b01011110000, OOL, IOL,
2905 "rotmai\t$rT, $rA, $val",
2906 RotateShift, pattern>;
2908 class ROTMAIVecInst<ValueType vectype, Operand intop, ValueType inttype>:
2909 ROTMAIInst<(outs VECREG:$rT), (ins VECREG:$rA, intop:$val),
2910 [(set (vectype VECREG:$rT),
2911 (SPUvec_sra VECREG:$rA, (inttype uimm7:$val)))]>;
2913 class ROTMAIRegInst<RegisterClass rclass, Operand intop, ValueType inttype>:
2914 ROTMAIInst<(outs rclass:$rT), (ins rclass:$rA, intop:$val),
2915 [(set rclass:$rT, (sra rclass:$rA, (inttype uimm7:$val)))]>;
2917 multiclass RotateMaskAlgebraicImm {
2918 def v2i64_i32 : ROTMAIVecInst<v2i64, rotNeg7imm, i32>;
2919 def v4i32_i32 : ROTMAIVecInst<v4i32, rotNeg7imm, i32>;
2920 def r64_i32 : ROTMAIRegInst<R64C, rotNeg7imm, i32>;
2921 def r32_i32 : ROTMAIRegInst<R32C, rotNeg7imm, i32>;
2924 defm ROTMAI : RotateMaskAlgebraicImm;
2926 //===----------------------------------------------------------------------===//
2927 // Branch and conditionals:
2928 //===----------------------------------------------------------------------===//
2930 let isTerminator = 1, isBarrier = 1 in {
2931 // Halt If Equal (r32 preferred slot only, no vector form)
2933 RRForm_3<0b00011011110, (outs), (ins R32C:$rA, R32C:$rB),
2934 "heq\t$rA, $rB", BranchResolv,
2935 [/* no pattern to match */]>;
2938 RI10Form_2<0b11111110, (outs), (ins R32C:$rA, s10imm:$val),
2939 "heqi\t$rA, $val", BranchResolv,
2940 [/* no pattern to match */]>;
2942 // HGT/HGTI: These instructions use signed arithmetic for the comparison,
2943 // contrasting with HLGT/HLGTI, which use unsigned comparison:
2945 RRForm_3<0b00011010010, (outs), (ins R32C:$rA, R32C:$rB),
2946 "hgt\t$rA, $rB", BranchResolv,
2947 [/* no pattern to match */]>;
2950 RI10Form_2<0b11110010, (outs), (ins R32C:$rA, s10imm:$val),
2951 "hgti\t$rA, $val", BranchResolv,
2952 [/* no pattern to match */]>;
2955 RRForm_3<0b00011011010, (outs), (ins R32C:$rA, R32C:$rB),
2956 "hlgt\t$rA, $rB", BranchResolv,
2957 [/* no pattern to match */]>;
2960 RI10Form_2<0b11111010, (outs), (ins R32C:$rA, s10imm:$val),
2961 "hlgti\t$rA, $val", BranchResolv,
2962 [/* no pattern to match */]>;
2965 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2966 // Comparison operators for i8, i16 and i32:
2967 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2969 class CEQBInst<dag OOL, dag IOL, list<dag> pattern> :
2970 RRForm<0b00001011110, OOL, IOL, "ceqb\t$rT, $rA, $rB",
2973 multiclass CmpEqualByte
2976 CEQBInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
2977 [(set (v16i8 VECREG:$rT), (seteq (v8i16 VECREG:$rA),
2978 (v8i16 VECREG:$rB)))]>;
2981 CEQBInst<(outs R8C:$rT), (ins R8C:$rA, R8C:$rB),
2982 [(set R8C:$rT, (seteq R8C:$rA, R8C:$rB))]>;
2985 class CEQBIInst<dag OOL, dag IOL, list<dag> pattern> :
2986 RI10Form<0b01111110, OOL, IOL, "ceqbi\t$rT, $rA, $val",
2989 multiclass CmpEqualByteImm
2992 CEQBIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm_i8:$val),
2993 [(set (v16i8 VECREG:$rT), (seteq (v16i8 VECREG:$rA),
2994 v16i8SExt8Imm:$val))]>;
2996 CEQBIInst<(outs R8C:$rT), (ins R8C:$rA, s10imm_i8:$val),
2997 [(set R8C:$rT, (seteq R8C:$rA, immSExt8:$val))]>;
3000 class CEQHInst<dag OOL, dag IOL, list<dag> pattern> :
3001 RRForm<0b00010011110, OOL, IOL, "ceqh\t$rT, $rA, $rB",
3004 multiclass CmpEqualHalfword
3006 def v8i16 : CEQHInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3007 [(set (v8i16 VECREG:$rT), (seteq (v8i16 VECREG:$rA),
3008 (v8i16 VECREG:$rB)))]>;
3010 def r16 : CEQHInst<(outs R16C:$rT), (ins R16C:$rA, R16C:$rB),
3011 [(set R16C:$rT, (seteq R16C:$rA, R16C:$rB))]>;
3014 class CEQHIInst<dag OOL, dag IOL, list<dag> pattern> :
3015 RI10Form<0b10111110, OOL, IOL, "ceqhi\t$rT, $rA, $val",
3018 multiclass CmpEqualHalfwordImm
3020 def v8i16 : CEQHIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
3021 [(set (v8i16 VECREG:$rT),
3022 (seteq (v8i16 VECREG:$rA),
3023 (v8i16 v8i16SExt10Imm:$val)))]>;
3024 def r16 : CEQHIInst<(outs R16C:$rT), (ins R16C:$rA, s10imm:$val),
3025 [(set R16C:$rT, (seteq R16C:$rA, i16ImmSExt10:$val))]>;
3028 class CEQInst<dag OOL, dag IOL, list<dag> pattern> :
3029 RRForm<0b00000011110, OOL, IOL, "ceq\t$rT, $rA, $rB",
3032 multiclass CmpEqualWord
3034 def v4i32 : CEQInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3035 [(set (v4i32 VECREG:$rT),
3036 (seteq (v4i32 VECREG:$rA), (v4i32 VECREG:$rB)))]>;
3038 def r32 : CEQInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
3039 [(set R32C:$rT, (seteq R32C:$rA, R32C:$rB))]>;
3042 class CEQIInst<dag OOL, dag IOL, list<dag> pattern> :
3043 RI10Form<0b00111110, OOL, IOL, "ceqi\t$rT, $rA, $val",
3046 multiclass CmpEqualWordImm
3048 def v4i32 : CEQIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
3049 [(set (v4i32 VECREG:$rT),
3050 (seteq (v4i32 VECREG:$rA),
3051 (v4i32 v4i32SExt16Imm:$val)))]>;
3053 def r32: CEQIInst<(outs R32C:$rT), (ins R32C:$rA, s10imm_i32:$val),
3054 [(set R32C:$rT, (seteq R32C:$rA, i32ImmSExt10:$val))]>;
3057 class CGTBInst<dag OOL, dag IOL, list<dag> pattern> :
3058 RRForm<0b00001010010, OOL, IOL, "cgtb\t$rT, $rA, $rB",
3061 multiclass CmpGtrByte
3064 CGTBInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3065 [(set (v16i8 VECREG:$rT), (setgt (v8i16 VECREG:$rA),
3066 (v8i16 VECREG:$rB)))]>;
3069 CGTBInst<(outs R8C:$rT), (ins R8C:$rA, R8C:$rB),
3070 [(set R8C:$rT, (setgt R8C:$rA, R8C:$rB))]>;
3073 class CGTBIInst<dag OOL, dag IOL, list<dag> pattern> :
3074 RI10Form<0b01110010, OOL, IOL, "cgtbi\t$rT, $rA, $val",
3077 multiclass CmpGtrByteImm
3080 CGTBIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm_i8:$val),
3081 [(set (v16i8 VECREG:$rT), (setgt (v16i8 VECREG:$rA),
3082 v16i8SExt8Imm:$val))]>;
3084 CGTBIInst<(outs R8C:$rT), (ins R8C:$rA, s10imm_i8:$val),
3085 [(set R8C:$rT, (setgt R8C:$rA, immSExt8:$val))]>;
3088 class CGTHInst<dag OOL, dag IOL, list<dag> pattern> :
3089 RRForm<0b00010010010, OOL, IOL, "cgth\t$rT, $rA, $rB",
3092 multiclass CmpGtrHalfword
3094 def v8i16 : CGTHInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3095 [(set (v8i16 VECREG:$rT), (setgt (v8i16 VECREG:$rA),
3096 (v8i16 VECREG:$rB)))]>;
3098 def r16 : CGTHInst<(outs R16C:$rT), (ins R16C:$rA, R16C:$rB),
3099 [(set R16C:$rT, (setgt R16C:$rA, R16C:$rB))]>;
3102 class CGTHIInst<dag OOL, dag IOL, list<dag> pattern> :
3103 RI10Form<0b10110010, OOL, IOL, "cgthi\t$rT, $rA, $val",
3106 multiclass CmpGtrHalfwordImm
3108 def v8i16 : CGTHIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
3109 [(set (v8i16 VECREG:$rT),
3110 (setgt (v8i16 VECREG:$rA),
3111 (v8i16 v8i16SExt10Imm:$val)))]>;
3112 def r16 : CGTHIInst<(outs R16C:$rT), (ins R16C:$rA, s10imm:$val),
3113 [(set R16C:$rT, (setgt R16C:$rA, i16ImmSExt10:$val))]>;
3116 class CGTInst<dag OOL, dag IOL, list<dag> pattern> :
3117 RRForm<0b00000010010, OOL, IOL, "cgt\t$rT, $rA, $rB",
3120 multiclass CmpGtrWord
3122 def v4i32 : CGTInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3123 [(set (v4i32 VECREG:$rT),
3124 (setgt (v4i32 VECREG:$rA), (v4i32 VECREG:$rB)))]>;
3126 def r32 : CGTInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
3127 [(set R32C:$rT, (setgt R32C:$rA, R32C:$rB))]>;
3130 class CGTIInst<dag OOL, dag IOL, list<dag> pattern> :
3131 RI10Form<0b00110010, OOL, IOL, "cgti\t$rT, $rA, $val",
3134 multiclass CmpGtrWordImm
3136 def v4i32 : CGTIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
3137 [(set (v4i32 VECREG:$rT),
3138 (setgt (v4i32 VECREG:$rA),
3139 (v4i32 v4i32SExt16Imm:$val)))]>;
3141 def r32: CGTIInst<(outs R32C:$rT), (ins R32C:$rA, s10imm_i32:$val),
3142 [(set R32C:$rT, (setgt R32C:$rA, i32ImmSExt10:$val))]>;
3144 // CGTIv4f32, CGTIf32: These are used in the f32 fdiv instruction sequence:
3145 def v4f32: CGTIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
3146 [(set (v4i32 VECREG:$rT),
3147 (setgt (v4i32 (bitconvert (v4f32 VECREG:$rA))),
3148 (v4i32 v4i32SExt16Imm:$val)))]>;
3150 def f32: CGTIInst<(outs R32C:$rT), (ins R32FP:$rA, s10imm_i32:$val),
3151 [/* no pattern */]>;
3154 class CLGTBInst<dag OOL, dag IOL, list<dag> pattern> :
3155 RRForm<0b00001011010, OOL, IOL, "clgtb\t$rT, $rA, $rB",
3158 multiclass CmpLGtrByte
3161 CLGTBInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3162 [(set (v16i8 VECREG:$rT), (setugt (v8i16 VECREG:$rA),
3163 (v8i16 VECREG:$rB)))]>;
3166 CLGTBInst<(outs R8C:$rT), (ins R8C:$rA, R8C:$rB),
3167 [(set R8C:$rT, (setugt R8C:$rA, R8C:$rB))]>;
3170 class CLGTBIInst<dag OOL, dag IOL, list<dag> pattern> :
3171 RI10Form<0b01111010, OOL, IOL, "clgtbi\t$rT, $rA, $val",
3174 multiclass CmpLGtrByteImm
3177 CLGTBIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm_i8:$val),
3178 [(set (v16i8 VECREG:$rT), (setugt (v16i8 VECREG:$rA),
3179 v16i8SExt8Imm:$val))]>;
3181 CLGTBIInst<(outs R8C:$rT), (ins R8C:$rA, s10imm_i8:$val),
3182 [(set R8C:$rT, (setugt R8C:$rA, immSExt8:$val))]>;
3185 class CLGTHInst<dag OOL, dag IOL, list<dag> pattern> :
3186 RRForm<0b00010011010, OOL, IOL, "clgth\t$rT, $rA, $rB",
3189 multiclass CmpLGtrHalfword
3191 def v8i16 : CLGTHInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3192 [(set (v8i16 VECREG:$rT), (setugt (v8i16 VECREG:$rA),
3193 (v8i16 VECREG:$rB)))]>;
3195 def r16 : CLGTHInst<(outs R16C:$rT), (ins R16C:$rA, R16C:$rB),
3196 [(set R16C:$rT, (setugt R16C:$rA, R16C:$rB))]>;
3199 class CLGTHIInst<dag OOL, dag IOL, list<dag> pattern> :
3200 RI10Form<0b10111010, OOL, IOL, "clgthi\t$rT, $rA, $val",
3203 multiclass CmpLGtrHalfwordImm
3205 def v8i16 : CLGTHIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
3206 [(set (v8i16 VECREG:$rT),
3207 (setugt (v8i16 VECREG:$rA),
3208 (v8i16 v8i16SExt10Imm:$val)))]>;
3209 def r16 : CLGTHIInst<(outs R16C:$rT), (ins R16C:$rA, s10imm:$val),
3210 [(set R16C:$rT, (setugt R16C:$rA, i16ImmSExt10:$val))]>;
3213 class CLGTInst<dag OOL, dag IOL, list<dag> pattern> :
3214 RRForm<0b00000011010, OOL, IOL, "clgt\t$rT, $rA, $rB",
3217 multiclass CmpLGtrWord
3219 def v4i32 : CLGTInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3220 [(set (v4i32 VECREG:$rT),
3221 (setugt (v4i32 VECREG:$rA), (v4i32 VECREG:$rB)))]>;
3223 def r32 : CLGTInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
3224 [(set R32C:$rT, (setugt R32C:$rA, R32C:$rB))]>;
3227 class CLGTIInst<dag OOL, dag IOL, list<dag> pattern> :
3228 RI10Form<0b00111010, OOL, IOL, "clgti\t$rT, $rA, $val",
3231 multiclass CmpLGtrWordImm
3233 def v4i32 : CLGTIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
3234 [(set (v4i32 VECREG:$rT),
3235 (setugt (v4i32 VECREG:$rA),
3236 (v4i32 v4i32SExt16Imm:$val)))]>;
3238 def r32: CLGTIInst<(outs R32C:$rT), (ins R32C:$rA, s10imm_i32:$val),
3239 [(set R32C:$rT, (setugt R32C:$rA, i32ImmSExt10:$val))]>;
3242 defm CEQB : CmpEqualByte;
3243 defm CEQBI : CmpEqualByteImm;
3244 defm CEQH : CmpEqualHalfword;
3245 defm CEQHI : CmpEqualHalfwordImm;
3246 defm CEQ : CmpEqualWord;
3247 defm CEQI : CmpEqualWordImm;
3248 defm CGTB : CmpGtrByte;
3249 defm CGTBI : CmpGtrByteImm;
3250 defm CGTH : CmpGtrHalfword;
3251 defm CGTHI : CmpGtrHalfwordImm;
3252 defm CGT : CmpGtrWord;
3253 defm CGTI : CmpGtrWordImm;
3254 defm CLGTB : CmpLGtrByte;
3255 defm CLGTBI : CmpLGtrByteImm;
3256 defm CLGTH : CmpLGtrHalfword;
3257 defm CLGTHI : CmpLGtrHalfwordImm;
3258 defm CLGT : CmpLGtrWord;
3259 defm CLGTI : CmpLGtrWordImm;
3261 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
3262 // For SETCC primitives not supported above (setlt, setle, setge, etc.)
3263 // define a pattern to generate the right code, as a binary operator
3264 // (in a manner of speaking.)
3267 // 1. This only matches the setcc set of conditionals. Special pattern
3268 // matching is used for select conditionals.
3270 // 2. The "DAG" versions of these classes is almost exclusively used for
3271 // i64 comparisons. See the tblgen fundamentals documentation for what
3272 // ".ResultInstrs[0]" means; see TargetSelectionDAG.td and the Pattern
3273 // class for where ResultInstrs originates.
3274 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
3276 class SETCCNegCondReg<PatFrag cond, RegisterClass rclass, ValueType inttype,
3277 SPUInstr xorinst, SPUInstr cmpare>:
3278 Pat<(cond rclass:$rA, rclass:$rB),
3279 (xorinst (cmpare rclass:$rA, rclass:$rB), (inttype -1))>;
3281 class SETCCNegCondImm<PatFrag cond, RegisterClass rclass, ValueType inttype,
3282 PatLeaf immpred, SPUInstr xorinst, SPUInstr cmpare>:
3283 Pat<(cond rclass:$rA, (inttype immpred:$imm)),
3284 (xorinst (cmpare rclass:$rA, (inttype immpred:$imm)), (inttype -1))>;
3286 def : SETCCNegCondReg<setne, R8C, i8, XORBIr8, CEQBr8>;
3287 def : SETCCNegCondImm<setne, R8C, i8, immSExt8, XORBIr8, CEQBIr8>;
3289 def : SETCCNegCondReg<setne, R16C, i16, XORHIr16, CEQHr16>;
3290 def : SETCCNegCondImm<setne, R16C, i16, i16ImmSExt10, XORHIr16, CEQHIr16>;
3292 def : SETCCNegCondReg<setne, R32C, i32, XORIr32, CEQr32>;
3293 def : SETCCNegCondImm<setne, R32C, i32, i32ImmSExt10, XORIr32, CEQIr32>;
3295 class SETCCBinOpReg<PatFrag cond, RegisterClass rclass,
3296 SPUInstr binop, SPUInstr cmpOp1, SPUInstr cmpOp2>:
3297 Pat<(cond rclass:$rA, rclass:$rB),
3298 (binop (cmpOp1 rclass:$rA, rclass:$rB),
3299 (cmpOp2 rclass:$rA, rclass:$rB))>;
3301 class SETCCBinOpImm<PatFrag cond, RegisterClass rclass, PatLeaf immpred,
3303 SPUInstr binop, SPUInstr cmpOp1, SPUInstr cmpOp2>:
3304 Pat<(cond rclass:$rA, (immtype immpred:$imm)),
3305 (binop (cmpOp1 rclass:$rA, (immtype immpred:$imm)),
3306 (cmpOp2 rclass:$rA, (immtype immpred:$imm)))>;
3308 def : SETCCBinOpReg<setge, R8C, ORr8, CGTBr8, CEQBr8>;
3309 def : SETCCBinOpImm<setge, R8C, immSExt8, i8, ORr8, CGTBIr8, CEQBIr8>;
3310 def : SETCCBinOpReg<setlt, R8C, NORr8, CGTBr8, CEQBr8>;
3311 def : SETCCBinOpImm<setlt, R8C, immSExt8, i8, NORr8, CGTBIr8, CEQBIr8>;
3312 def : Pat<(setle R8C:$rA, R8C:$rB),
3313 (XORBIr8 (CGTBr8 R8C:$rA, R8C:$rB), 0xff)>;
3314 def : Pat<(setle R8C:$rA, immU8:$imm),
3315 (XORBIr8 (CGTBIr8 R8C:$rA, immU8:$imm), 0xff)>;
3317 def : SETCCBinOpReg<setge, R16C, ORr16, CGTHr16, CEQHr16>;
3318 def : SETCCBinOpImm<setge, R16C, i16ImmSExt10, i16,
3319 ORr16, CGTHIr16, CEQHIr16>;
3320 def : SETCCBinOpReg<setlt, R16C, NORr16, CGTHr16, CEQHr16>;
3321 def : SETCCBinOpImm<setlt, R16C, i16ImmSExt10, i16, NORr16, CGTHIr16, CEQHIr16>;
3322 def : Pat<(setle R16C:$rA, R16C:$rB),
3323 (XORHIr16 (CGTHr16 R16C:$rA, R16C:$rB), 0xffff)>;
3324 def : Pat<(setle R16C:$rA, i16ImmSExt10:$imm),
3325 (XORHIr16 (CGTHIr16 R16C:$rA, i16ImmSExt10:$imm), 0xffff)>;
3327 def : SETCCBinOpReg<setge, R32C, ORr32, CGTr32, CEQr32>;
3328 def : SETCCBinOpImm<setge, R32C, i32ImmSExt10, i32,
3329 ORr32, CGTIr32, CEQIr32>;
3330 def : SETCCBinOpReg<setlt, R32C, NORr32, CGTr32, CEQr32>;
3331 def : SETCCBinOpImm<setlt, R32C, i32ImmSExt10, i32, NORr32, CGTIr32, CEQIr32>;
3332 def : Pat<(setle R32C:$rA, R32C:$rB),
3333 (XORIr32 (CGTr32 R32C:$rA, R32C:$rB), 0xffffffff)>;
3334 def : Pat<(setle R32C:$rA, i32ImmSExt10:$imm),
3335 (XORIr32 (CGTIr32 R32C:$rA, i32ImmSExt10:$imm), 0xffffffff)>;
3337 def : SETCCBinOpReg<setuge, R8C, ORr8, CLGTBr8, CEQBr8>;
3338 def : SETCCBinOpImm<setuge, R8C, immSExt8, i8, ORr8, CLGTBIr8, CEQBIr8>;
3339 def : SETCCBinOpReg<setult, R8C, NORr8, CLGTBr8, CEQBr8>;
3340 def : SETCCBinOpImm<setult, R8C, immSExt8, i8, NORr8, CLGTBIr8, CEQBIr8>;
3341 def : Pat<(setule R8C:$rA, R8C:$rB),
3342 (XORBIr8 (CLGTBr8 R8C:$rA, R8C:$rB), 0xff)>;
3343 def : Pat<(setule R8C:$rA, immU8:$imm),
3344 (XORBIr8 (CLGTBIr8 R8C:$rA, immU8:$imm), 0xff)>;
3346 def : SETCCBinOpReg<setuge, R16C, ORr16, CLGTHr16, CEQHr16>;
3347 def : SETCCBinOpImm<setuge, R16C, i16ImmSExt10, i16,
3348 ORr16, CLGTHIr16, CEQHIr16>;
3349 def : SETCCBinOpReg<setult, R16C, NORr16, CLGTHr16, CEQHr16>;
3350 def : SETCCBinOpImm<setult, R16C, i16ImmSExt10, i16, NORr16,
3351 CLGTHIr16, CEQHIr16>;
3352 def : Pat<(setule R16C:$rA, R16C:$rB),
3353 (XORHIr16 (CLGTHr16 R16C:$rA, R16C:$rB), 0xffff)>;
3354 def : Pat<(setule R16C:$rA, i16ImmSExt10:$imm),
3355 (XORHIr16 (CLGTHIr16 R16C:$rA, i16ImmSExt10:$imm), 0xffff)>;
3357 def : SETCCBinOpReg<setuge, R32C, ORr32, CLGTr32, CEQr32>;
3358 def : SETCCBinOpImm<setuge, R32C, i32ImmSExt10, i32,
3359 ORr32, CLGTIr32, CEQIr32>;
3360 def : SETCCBinOpReg<setult, R32C, NORr32, CLGTr32, CEQr32>;
3361 def : SETCCBinOpImm<setult, R32C, i32ImmSExt10, i32, NORr32, CLGTIr32, CEQIr32>;
3362 def : Pat<(setule R32C:$rA, R32C:$rB),
3363 (XORIr32 (CLGTr32 R32C:$rA, R32C:$rB), 0xffffffff)>;
3364 def : Pat<(setule R32C:$rA, i32ImmSExt10:$imm),
3365 (XORIr32 (CLGTIr32 R32C:$rA, i32ImmSExt10:$imm), 0xffffffff)>;
3367 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
3368 // select conditional patterns:
3369 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
3371 class SELECTNegCondReg<PatFrag cond, RegisterClass rclass, ValueType inttype,
3372 SPUInstr selinstr, SPUInstr cmpare>:
3373 Pat<(select (inttype (cond rclass:$rA, rclass:$rB)),
3374 rclass:$rTrue, rclass:$rFalse),
3375 (selinstr rclass:$rTrue, rclass:$rFalse,
3376 (cmpare rclass:$rA, rclass:$rB))>;
3378 class SELECTNegCondImm<PatFrag cond, RegisterClass rclass, ValueType inttype,
3379 PatLeaf immpred, SPUInstr selinstr, SPUInstr cmpare>:
3380 Pat<(select (inttype (cond rclass:$rA, immpred:$imm)),
3381 rclass:$rTrue, rclass:$rFalse),
3382 (selinstr rclass:$rTrue, rclass:$rFalse,
3383 (cmpare rclass:$rA, immpred:$imm))>;
3385 def : SELECTNegCondReg<setne, R8C, i8, SELBr8, CEQBr8>;
3386 def : SELECTNegCondImm<setne, R8C, i8, immSExt8, SELBr8, CEQBIr8>;
3387 def : SELECTNegCondReg<setle, R8C, i8, SELBr8, CGTBr8>;
3388 def : SELECTNegCondImm<setle, R8C, i8, immSExt8, SELBr8, CGTBr8>;
3389 def : SELECTNegCondReg<setule, R8C, i8, SELBr8, CLGTBr8>;
3390 def : SELECTNegCondImm<setule, R8C, i8, immU8, SELBr8, CLGTBIr8>;
3392 def : SELECTNegCondReg<setne, R16C, i16, SELBr16, CEQHr16>;
3393 def : SELECTNegCondImm<setne, R16C, i16, i16ImmSExt10, SELBr16, CEQHIr16>;
3394 def : SELECTNegCondReg<setle, R16C, i16, SELBr16, CGTHr16>;
3395 def : SELECTNegCondImm<setle, R16C, i16, i16ImmSExt10, SELBr16, CGTHIr16>;
3396 def : SELECTNegCondReg<setule, R16C, i16, SELBr16, CLGTHr16>;
3397 def : SELECTNegCondImm<setule, R16C, i16, i16ImmSExt10, SELBr16, CLGTHIr16>;
3399 def : SELECTNegCondReg<setne, R32C, i32, SELBr32, CEQr32>;
3400 def : SELECTNegCondImm<setne, R32C, i32, i32ImmSExt10, SELBr32, CEQIr32>;
3401 def : SELECTNegCondReg<setle, R32C, i32, SELBr32, CGTr32>;
3402 def : SELECTNegCondImm<setle, R32C, i32, i32ImmSExt10, SELBr32, CGTIr32>;
3403 def : SELECTNegCondReg<setule, R32C, i32, SELBr32, CLGTr32>;
3404 def : SELECTNegCondImm<setule, R32C, i32, i32ImmSExt10, SELBr32, CLGTIr32>;
3406 class SELECTBinOpReg<PatFrag cond, RegisterClass rclass, ValueType inttype,
3407 SPUInstr selinstr, SPUInstr binop, SPUInstr cmpOp1,
3409 Pat<(select (inttype (cond rclass:$rA, rclass:$rB)),
3410 rclass:$rTrue, rclass:$rFalse),
3411 (selinstr rclass:$rFalse, rclass:$rTrue,
3412 (binop (cmpOp1 rclass:$rA, rclass:$rB),
3413 (cmpOp2 rclass:$rA, rclass:$rB)))>;
3415 class SELECTBinOpImm<PatFrag cond, RegisterClass rclass, PatLeaf immpred,
3417 SPUInstr selinstr, SPUInstr binop, SPUInstr cmpOp1,
3419 Pat<(select (inttype (cond rclass:$rA, (inttype immpred:$imm))),
3420 rclass:$rTrue, rclass:$rFalse),
3421 (selinstr rclass:$rFalse, rclass:$rTrue,
3422 (binop (cmpOp1 rclass:$rA, (inttype immpred:$imm)),
3423 (cmpOp2 rclass:$rA, (inttype immpred:$imm))))>;
3425 def : SELECTBinOpReg<setge, R8C, i8, SELBr8, ORr8, CGTBr8, CEQBr8>;
3426 def : SELECTBinOpImm<setge, R8C, immSExt8, i8,
3427 SELBr8, ORr8, CGTBIr8, CEQBIr8>;
3429 def : SELECTBinOpReg<setge, R16C, i16, SELBr16, ORr16, CGTHr16, CEQHr16>;
3430 def : SELECTBinOpImm<setge, R16C, i16ImmSExt10, i16,
3431 SELBr16, ORr16, CGTHIr16, CEQHIr16>;
3433 def : SELECTBinOpReg<setge, R32C, i32, SELBr32, ORr32, CGTr32, CEQr32>;
3434 def : SELECTBinOpImm<setge, R32C, i32ImmSExt10, i32,
3435 SELBr32, ORr32, CGTIr32, CEQIr32>;
3437 def : SELECTBinOpReg<setuge, R8C, i8, SELBr8, ORr8, CLGTBr8, CEQBr8>;
3438 def : SELECTBinOpImm<setuge, R8C, immSExt8, i8,
3439 SELBr8, ORr8, CLGTBIr8, CEQBIr8>;
3441 def : SELECTBinOpReg<setuge, R16C, i16, SELBr16, ORr16, CLGTHr16, CEQHr16>;
3442 def : SELECTBinOpImm<setuge, R16C, i16ImmUns10, i16,
3443 SELBr16, ORr16, CLGTHIr16, CEQHIr16>;
3445 def : SELECTBinOpReg<setuge, R32C, i32, SELBr32, ORr32, CLGTr32, CEQr32>;
3446 def : SELECTBinOpImm<setuge, R32C, i32ImmUns10, i32,
3447 SELBr32, ORr32, CLGTIr32, CEQIr32>;
3449 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
3452 // All calls clobber the non-callee-saved registers:
3453 Defs = [R0, R1, R2, R3, R4, R5, R6, R7, R8, R9,
3454 R10,R11,R12,R13,R14,R15,R16,R17,R18,R19,
3455 R20,R21,R22,R23,R24,R25,R26,R27,R28,R29,
3456 R30,R31,R32,R33,R34,R35,R36,R37,R38,R39,
3457 R40,R41,R42,R43,R44,R45,R46,R47,R48,R49,
3458 R50,R51,R52,R53,R54,R55,R56,R57,R58,R59,
3459 R60,R61,R62,R63,R64,R65,R66,R67,R68,R69,
3460 R70,R71,R72,R73,R74,R75,R76,R77,R78,R79],
3461 // All of these instructions use $lr (aka $0)
3463 // Branch relative and set link: Used if we actually know that the target
3464 // is within [-32768, 32767] bytes of the target
3466 BranchSetLink<0b011001100, (outs), (ins relcalltarget:$func, variable_ops),
3467 "brsl\t$$lr, $func",
3468 [(SPUcall (SPUpcrel tglobaladdr:$func, 0))]>;
3470 // Branch absolute and set link: Used if we actually know that the target
3471 // is an absolute address
3473 BranchSetLink<0b011001100, (outs), (ins calltarget:$func, variable_ops),
3474 "brasl\t$$lr, $func",
3475 [(SPUcall (SPUaform tglobaladdr:$func, 0))]>;
3477 // Branch indirect and set link if external data. These instructions are not
3478 // actually generated, matched by an intrinsic:
3479 def BISLED_00: BISLEDForm<0b11, "bisled\t$$lr, $func", [/* empty pattern */]>;
3480 def BISLED_E0: BISLEDForm<0b10, "bisled\t$$lr, $func", [/* empty pattern */]>;
3481 def BISLED_0D: BISLEDForm<0b01, "bisled\t$$lr, $func", [/* empty pattern */]>;
3482 def BISLED_ED: BISLEDForm<0b00, "bisled\t$$lr, $func", [/* empty pattern */]>;
3484 // Branch indirect and set link. This is the "X-form" address version of a
3487 BIForm<0b10010101100, "bisl\t$$lr, $func", [(SPUcall R32C:$func)]>;
3490 // Support calls to external symbols:
3491 def : Pat<(SPUcall (SPUpcrel texternalsym:$func, 0)),
3492 (BRSL texternalsym:$func)>;
3494 def : Pat<(SPUcall (SPUaform texternalsym:$func, 0)),
3495 (BRASL texternalsym:$func)>;
3497 // Unconditional branches:
3498 let isBranch = 1, isTerminator = 1, hasCtrlDep = 1, isBarrier = 1 in {
3500 UncondBranch<0b001001100, (outs), (ins brtarget:$dest),
3504 // Unconditional, absolute address branch
3506 UncondBranch<0b001100000, (outs), (ins brtarget:$dest),
3508 [/* no pattern */]>;
3512 BIForm<0b00010101100, "bi\t$func", [(brind R32C:$func)]>;
3514 // Various branches:
3515 class BRNZInst<dag IOL, list<dag> pattern>:
3516 RI16Form<0b010000100, (outs), IOL, "brnz\t$rCond,$dest",
3517 BranchResolv, pattern>;
3519 class BRNZRegInst<RegisterClass rclass>:
3520 BRNZInst<(ins rclass:$rCond, brtarget:$dest),
3521 [(brcond rclass:$rCond, bb:$dest)]>;
3523 class BRNZVecInst<ValueType vectype>:
3524 BRNZInst<(ins VECREG:$rCond, brtarget:$dest),
3525 [(brcond (vectype VECREG:$rCond), bb:$dest)]>;
3527 multiclass BranchNotZero {
3528 def v4i32 : BRNZVecInst<v4i32>;
3529 def r32 : BRNZRegInst<R32C>;
3532 defm BRNZ : BranchNotZero;
3534 class BRZInst<dag IOL, list<dag> pattern>:
3535 RI16Form<0b000000100, (outs), IOL, "brz\t$rT,$dest",
3536 BranchResolv, pattern>;
3538 class BRZRegInst<RegisterClass rclass>:
3539 BRZInst<(ins rclass:$rT, brtarget:$dest), [/* no pattern */]>;
3541 class BRZVecInst<ValueType vectype>:
3542 BRZInst<(ins VECREG:$rT, brtarget:$dest), [/* no pattern */]>;
3544 multiclass BranchZero {
3545 def v4i32: BRZVecInst<v4i32>;
3546 def r32: BRZRegInst<R32C>;
3549 defm BRZ: BranchZero;
3551 // Note: LLVM doesn't do branch conditional, indirect. Otherwise these would
3554 class BINZInst<dag IOL, list<dag> pattern>:
3555 BICondForm<0b10010100100, (outs), IOL, "binz\t$rA, $dest", pattern>;
3557 class BINZRegInst<RegisterClass rclass>:
3558 BINZInst<(ins rclass:$rA, brtarget:$dest),
3559 [(brcond rclass:$rA, R32C:$dest)]>;
3561 class BINZVecInst<ValueType vectype>:
3562 BINZInst<(ins VECREG:$rA, R32C:$dest),
3563 [(brcond (vectype VECREG:$rA), R32C:$dest)]>;
3565 multiclass BranchNotZeroIndirect {
3566 def v4i32: BINZVecInst<v4i32>;
3567 def r32: BINZRegInst<R32C>;
3570 defm BINZ: BranchNotZeroIndirect;
3572 class BIZInst<dag IOL, list<dag> pattern>:
3573 BICondForm<0b00010100100, (outs), IOL, "biz\t$rA, $func", pattern>;
3575 class BIZRegInst<RegisterClass rclass>:
3576 BIZInst<(ins rclass:$rA, R32C:$func), [/* no pattern */]>;
3578 class BIZVecInst<ValueType vectype>:
3579 BIZInst<(ins VECREG:$rA, R32C:$func), [/* no pattern */]>;
3581 multiclass BranchZeroIndirect {
3582 def v4i32: BIZVecInst<v4i32>;
3583 def r32: BIZRegInst<R32C>;
3586 defm BIZ: BranchZeroIndirect;
3589 class BRHNZInst<dag IOL, list<dag> pattern>:
3590 RI16Form<0b011000100, (outs), IOL, "brhnz\t$rCond,$dest", BranchResolv,
3593 class BRHNZRegInst<RegisterClass rclass>:
3594 BRHNZInst<(ins rclass:$rCond, brtarget:$dest),
3595 [(brcond rclass:$rCond, bb:$dest)]>;
3597 class BRHNZVecInst<ValueType vectype>:
3598 BRHNZInst<(ins VECREG:$rCond, brtarget:$dest), [/* no pattern */]>;
3600 multiclass BranchNotZeroHalfword {
3601 def v8i16: BRHNZVecInst<v8i16>;
3602 def r16: BRHNZRegInst<R16C>;
3605 defm BRHNZ: BranchNotZeroHalfword;
3607 class BRHZInst<dag IOL, list<dag> pattern>:
3608 RI16Form<0b001000100, (outs), IOL, "brhz\t$rT,$dest", BranchResolv,
3611 class BRHZRegInst<RegisterClass rclass>:
3612 BRHZInst<(ins rclass:$rT, brtarget:$dest), [/* no pattern */]>;
3614 class BRHZVecInst<ValueType vectype>:
3615 BRHZInst<(ins VECREG:$rT, brtarget:$dest), [/* no pattern */]>;
3617 multiclass BranchZeroHalfword {
3618 def v8i16: BRHZVecInst<v8i16>;
3619 def r16: BRHZRegInst<R16C>;
3622 defm BRHZ: BranchZeroHalfword;
3625 //===----------------------------------------------------------------------===//
3626 // setcc and brcond patterns:
3627 //===----------------------------------------------------------------------===//
3629 def : Pat<(brcond (i16 (seteq R16C:$rA, 0)), bb:$dest),
3630 (BRHZr16 R16C:$rA, bb:$dest)>;
3631 def : Pat<(brcond (i16 (setne R16C:$rA, 0)), bb:$dest),
3632 (BRHNZr16 R16C:$rA, bb:$dest)>;
3634 def : Pat<(brcond (i32 (seteq R32C:$rA, 0)), bb:$dest),
3635 (BRZr32 R32C:$rA, bb:$dest)>;
3636 def : Pat<(brcond (i32 (setne R32C:$rA, 0)), bb:$dest),
3637 (BRNZr32 R32C:$rA, bb:$dest)>;
3639 multiclass BranchCondEQ<PatFrag cond, SPUInstr brinst16, SPUInstr brinst32>
3641 def r16imm: Pat<(brcond (i16 (cond R16C:$rA, i16ImmSExt10:$val)), bb:$dest),
3642 (brinst16 (CEQHIr16 R16C:$rA, i16ImmSExt10:$val), bb:$dest)>;
3644 def r16 : Pat<(brcond (i16 (cond R16C:$rA, R16C:$rB)), bb:$dest),
3645 (brinst16 (CEQHr16 R16C:$rA, R16:$rB), bb:$dest)>;
3647 def r32imm : Pat<(brcond (i32 (cond R32C:$rA, i32ImmSExt10:$val)), bb:$dest),
3648 (brinst32 (CEQIr32 R32C:$rA, i32ImmSExt10:$val), bb:$dest)>;
3650 def r32 : Pat<(brcond (i32 (cond R32C:$rA, R32C:$rB)), bb:$dest),
3651 (brinst32 (CEQr32 R32C:$rA, R32C:$rB), bb:$dest)>;
3654 defm BRCONDeq : BranchCondEQ<seteq, BRHZr16, BRZr32>;
3655 defm BRCONDne : BranchCondEQ<setne, BRHNZr16, BRNZr32>;
3657 multiclass BranchCondLGT<PatFrag cond, SPUInstr brinst16, SPUInstr brinst32>
3659 def r16imm : Pat<(brcond (i16 (cond R16C:$rA, i16ImmSExt10:$val)), bb:$dest),
3660 (brinst16 (CLGTHIr16 R16C:$rA, i16ImmSExt10:$val), bb:$dest)>;
3662 def r16 : Pat<(brcond (i16 (cond R16C:$rA, R16C:$rB)), bb:$dest),
3663 (brinst16 (CLGTHr16 R16C:$rA, R16:$rB), bb:$dest)>;
3665 def r32imm : Pat<(brcond (i32 (cond R32C:$rA, i32ImmSExt10:$val)), bb:$dest),
3666 (brinst32 (CLGTIr32 R32C:$rA, i32ImmSExt10:$val), bb:$dest)>;
3668 def r32 : Pat<(brcond (i32 (cond R32C:$rA, R32C:$rB)), bb:$dest),
3669 (brinst32 (CLGTr32 R32C:$rA, R32C:$rB), bb:$dest)>;
3672 defm BRCONDugt : BranchCondLGT<setugt, BRHNZr16, BRNZr32>;
3673 defm BRCONDule : BranchCondLGT<setule, BRHZr16, BRZr32>;
3675 multiclass BranchCondLGTEQ<PatFrag cond, SPUInstr orinst16, SPUInstr brinst16,
3676 SPUInstr orinst32, SPUInstr brinst32>
3678 def r16imm: Pat<(brcond (i16 (cond R16C:$rA, i16ImmSExt10:$val)), bb:$dest),
3679 (brinst16 (orinst16 (CLGTHIr16 R16C:$rA, i16ImmSExt10:$val),
3680 (CEQHIr16 R16C:$rA, i16ImmSExt10:$val)),
3683 def r16: Pat<(brcond (i16 (cond R16C:$rA, R16C:$rB)), bb:$dest),
3684 (brinst16 (orinst16 (CLGTHr16 R16C:$rA, R16:$rB),
3685 (CEQHr16 R16C:$rA, R16:$rB)),
3688 def r32imm : Pat<(brcond (i32 (cond R32C:$rA, i32ImmSExt10:$val)), bb:$dest),
3689 (brinst32 (orinst32 (CLGTIr32 R32C:$rA, i32ImmSExt10:$val),
3690 (CEQIr32 R32C:$rA, i32ImmSExt10:$val)),
3693 def r32 : Pat<(brcond (i32 (cond R32C:$rA, R32C:$rB)), bb:$dest),
3694 (brinst32 (orinst32 (CLGTr32 R32C:$rA, R32C:$rB),
3695 (CEQr32 R32C:$rA, R32C:$rB)),
3699 defm BRCONDuge : BranchCondLGTEQ<setuge, ORr16, BRHNZr16, ORr32, BRNZr32>;
3700 defm BRCONDult : BranchCondLGTEQ<setult, ORr16, BRHZr16, ORr32, BRZr32>;
3702 multiclass BranchCondGT<PatFrag cond, SPUInstr brinst16, SPUInstr brinst32>
3704 def r16imm : Pat<(brcond (i16 (cond R16C:$rA, i16ImmSExt10:$val)), bb:$dest),
3705 (brinst16 (CGTHIr16 R16C:$rA, i16ImmSExt10:$val), bb:$dest)>;
3707 def r16 : Pat<(brcond (i16 (cond R16C:$rA, R16C:$rB)), bb:$dest),
3708 (brinst16 (CGTHr16 R16C:$rA, R16:$rB), bb:$dest)>;
3710 def r32imm : Pat<(brcond (i32 (cond R32C:$rA, i32ImmSExt10:$val)), bb:$dest),
3711 (brinst32 (CGTIr32 R32C:$rA, i32ImmSExt10:$val), bb:$dest)>;
3713 def r32 : Pat<(brcond (i32 (cond R32C:$rA, R32C:$rB)), bb:$dest),
3714 (brinst32 (CGTr32 R32C:$rA, R32C:$rB), bb:$dest)>;
3717 defm BRCONDgt : BranchCondGT<setgt, BRHNZr16, BRNZr32>;
3718 defm BRCONDle : BranchCondGT<setle, BRHZr16, BRZr32>;
3720 multiclass BranchCondGTEQ<PatFrag cond, SPUInstr orinst16, SPUInstr brinst16,
3721 SPUInstr orinst32, SPUInstr brinst32>
3723 def r16imm: Pat<(brcond (i16 (cond R16C:$rA, i16ImmSExt10:$val)), bb:$dest),
3724 (brinst16 (orinst16 (CGTHIr16 R16C:$rA, i16ImmSExt10:$val),
3725 (CEQHIr16 R16C:$rA, i16ImmSExt10:$val)),
3728 def r16: Pat<(brcond (i16 (cond R16C:$rA, R16C:$rB)), bb:$dest),
3729 (brinst16 (orinst16 (CGTHr16 R16C:$rA, R16:$rB),
3730 (CEQHr16 R16C:$rA, R16:$rB)),
3733 def r32imm : Pat<(brcond (i32 (cond R32C:$rA, i32ImmSExt10:$val)), bb:$dest),
3734 (brinst32 (orinst32 (CGTIr32 R32C:$rA, i32ImmSExt10:$val),
3735 (CEQIr32 R32C:$rA, i32ImmSExt10:$val)),
3738 def r32 : Pat<(brcond (i32 (cond R32C:$rA, R32C:$rB)), bb:$dest),
3739 (brinst32 (orinst32 (CGTr32 R32C:$rA, R32C:$rB),
3740 (CEQr32 R32C:$rA, R32C:$rB)),
3744 defm BRCONDge : BranchCondGTEQ<setge, ORr16, BRHNZr16, ORr32, BRNZr32>;
3745 defm BRCONDlt : BranchCondGTEQ<setlt, ORr16, BRHZr16, ORr32, BRZr32>;
3747 let isTerminator = 1, isBarrier = 1 in {
3748 let isReturn = 1 in {
3750 RETForm<"bi\t$$lr", [(retflag)]>;
3754 //===----------------------------------------------------------------------===//
3755 // Single precision floating point instructions
3756 //===----------------------------------------------------------------------===//
3758 class FAInst<dag OOL, dag IOL, list<dag> pattern>:
3759 RRForm<0b01011000100, OOL, IOL, "fa\t$rT, $rA, $rB",
3762 class FAVecInst<ValueType vectype>:
3763 FAInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3764 [(set (vectype VECREG:$rT),
3765 (fadd (vectype VECREG:$rA), (vectype VECREG:$rB)))]>;
3769 def v4f32: FAVecInst<v4f32>;
3770 def f32: FAInst<(outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB),
3771 [(set R32FP:$rT, (fadd R32FP:$rA, R32FP:$rB))]>;
3776 class FSInst<dag OOL, dag IOL, list<dag> pattern>:
3777 RRForm<0b01011000100, OOL, IOL, "fs\t$rT, $rA, $rB",
3780 class FSVecInst<ValueType vectype>:
3781 FSInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3782 [(set (vectype VECREG:$rT),
3783 (fsub (vectype VECREG:$rA), (vectype VECREG:$rB)))]>;
3787 def v4f32: FSVecInst<v4f32>;
3788 def f32: FSInst<(outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB),
3789 [(set R32FP:$rT, (fsub R32FP:$rA, R32FP:$rB))]>;
3794 // Floating point reciprocal estimate
3796 class FRESTInst<dag OOL, dag IOL>:
3797 RRForm_1<0b00110111000, OOL, IOL,
3798 "frest\t$rT, $rA", SPrecFP,
3799 [/* no pattern */]>;
3802 FRESTInst<(outs VECREG:$rT), (ins VECREG:$rA)>;
3805 FRESTInst<(outs R32FP:$rT), (ins R32FP:$rA)>;
3807 // Floating point interpolate (used in conjunction with reciprocal estimate)
3809 RRForm<0b00101011110, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3810 "fi\t$rT, $rA, $rB", SPrecFP,
3811 [/* no pattern */]>;
3814 RRForm<0b00101011110, (outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB),
3815 "fi\t$rT, $rA, $rB", SPrecFP,
3816 [/* no pattern */]>;
3818 //--------------------------------------------------------------------------
3819 // Basic single precision floating point comparisons:
3821 // Note: There is no support on SPU for single precision NaN. Consequently,
3822 // ordered and unordered comparisons are the same.
3823 //--------------------------------------------------------------------------
3826 RRForm<0b01000011110, (outs R32C:$rT), (ins R32FP:$rA, R32FP:$rB),
3827 "fceq\t$rT, $rA, $rB", SPrecFP,
3828 [(set R32C:$rT, (setueq R32FP:$rA, R32FP:$rB))]>;
3830 def : Pat<(setoeq R32FP:$rA, R32FP:$rB),
3831 (FCEQf32 R32FP:$rA, R32FP:$rB)>;
3834 RRForm<0b01010011110, (outs R32C:$rT), (ins R32FP:$rA, R32FP:$rB),
3835 "fcmeq\t$rT, $rA, $rB", SPrecFP,
3836 [(set R32C:$rT, (setueq (fabs R32FP:$rA), (fabs R32FP:$rB)))]>;
3838 def : Pat<(setoeq (fabs R32FP:$rA), (fabs R32FP:$rB)),
3839 (FCMEQf32 R32FP:$rA, R32FP:$rB)>;
3842 RRForm<0b01000011010, (outs R32C:$rT), (ins R32FP:$rA, R32FP:$rB),
3843 "fcgt\t$rT, $rA, $rB", SPrecFP,
3844 [(set R32C:$rT, (setugt R32FP:$rA, R32FP:$rB))]>;
3846 def : Pat<(setugt R32FP:$rA, R32FP:$rB),
3847 (FCGTf32 R32FP:$rA, R32FP:$rB)>;
3850 RRForm<0b01010011010, (outs R32C:$rT), (ins R32FP:$rA, R32FP:$rB),
3851 "fcmgt\t$rT, $rA, $rB", SPrecFP,
3852 [(set R32C:$rT, (setugt (fabs R32FP:$rA), (fabs R32FP:$rB)))]>;
3854 def : Pat<(setugt (fabs R32FP:$rA), (fabs R32FP:$rB)),
3855 (FCMGTf32 R32FP:$rA, R32FP:$rB)>;
3857 //--------------------------------------------------------------------------
3858 // Single precision floating point comparisons and SETCC equivalents:
3859 //--------------------------------------------------------------------------
3861 def : SETCCNegCondReg<setune, R32FP, i32, XORIr32, FCEQf32>;
3862 def : SETCCNegCondReg<setone, R32FP, i32, XORIr32, FCEQf32>;
3864 def : SETCCBinOpReg<setuge, R32FP, ORr32, FCGTf32, FCEQf32>;
3865 def : SETCCBinOpReg<setoge, R32FP, ORr32, FCGTf32, FCEQf32>;
3867 def : SETCCBinOpReg<setult, R32FP, NORr32, FCGTf32, FCEQf32>;
3868 def : SETCCBinOpReg<setolt, R32FP, NORr32, FCGTf32, FCEQf32>;
3870 def : Pat<(setule R32FP:$rA, R32FP:$rB),
3871 (XORIr32 (FCGTf32 R32FP:$rA, R32FP:$rB), 0xffffffff)>;
3872 def : Pat<(setole R32FP:$rA, R32FP:$rB),
3873 (XORIr32 (FCGTf32 R32FP:$rA, R32FP:$rB), 0xffffffff)>;
3875 // FP Status and Control Register Write
3876 // Why isn't rT a don't care in the ISA?
3877 // Should we create a special RRForm_3 for this guy and zero out the rT?
3879 RRForm_1<0b01011101110, (outs R32FP:$rT), (ins R32FP:$rA),
3880 "fscrwr\t$rA", SPrecFP,
3881 [/* This instruction requires an intrinsic. Note: rT is unused. */]>;
3883 // FP Status and Control Register Read
3885 RRForm_2<0b01011101110, (outs R32FP:$rT), (ins),
3886 "fscrrd\t$rT", SPrecFP,
3887 [/* This instruction requires an intrinsic */]>;
3889 // llvm instruction space
3890 // How do these map onto cell instructions?
3892 // frest rC rB # c = 1/b (both lines)
3894 // fm rD rA rC # d = a * 1/b
3895 // fnms rB rD rB rA # b = - (d * b - a) --should == 0 in a perfect world
3896 // fma rB rB rC rD # b = b * c + d
3897 // = -(d *b -a) * c + d
3898 // = a * c - c ( a *b *c - a)
3903 // These llvm instructions will actually map to library calls.
3904 // All that's needed, then, is to check that the appropriate library is
3905 // imported and do a brsl to the proper function name.
3906 // frem # fmod(x, y): x - (x/y) * y
3907 // (Note: fmod(double, double), fmodf(float,float)
3911 // Unimplemented SPU instruction space
3912 // floating reciprocal absolute square root estimate (frsqest)
3914 // The following are probably just intrinsics
3915 // status and control register write
3916 // status and control register read
3918 //--------------------------------------
3919 // Floating point multiply instructions
3920 //--------------------------------------
3923 RRForm<0b00100011010, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3924 "fm\t$rT, $rA, $rB", SPrecFP,
3925 [(set (v4f32 VECREG:$rT), (fmul (v4f32 VECREG:$rA),
3926 (v4f32 VECREG:$rB)))]>;
3929 RRForm<0b01100011010, (outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB),
3930 "fm\t$rT, $rA, $rB", SPrecFP,
3931 [(set R32FP:$rT, (fmul R32FP:$rA, R32FP:$rB))]>;
3933 // Floating point multiply and add
3934 // e.g. d = c + (a * b)
3936 RRRForm<0b0111, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
3937 "fma\t$rT, $rA, $rB, $rC", SPrecFP,
3938 [(set (v4f32 VECREG:$rT),
3939 (fadd (v4f32 VECREG:$rC),
3940 (fmul (v4f32 VECREG:$rA), (v4f32 VECREG:$rB))))]>;
3943 RRRForm<0b0111, (outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB, R32FP:$rC),
3944 "fma\t$rT, $rA, $rB, $rC", SPrecFP,
3945 [(set R32FP:$rT, (fadd R32FP:$rC, (fmul R32FP:$rA, R32FP:$rB)))]>;
3947 // FP multiply and subtract
3948 // Subtracts value in rC from product
3951 RRRForm<0b0111, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
3952 "fms\t$rT, $rA, $rB, $rC", SPrecFP,
3953 [(set (v4f32 VECREG:$rT),
3954 (fsub (fmul (v4f32 VECREG:$rA), (v4f32 VECREG:$rB)),
3955 (v4f32 VECREG:$rC)))]>;
3958 RRRForm<0b0111, (outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB, R32FP:$rC),
3959 "fms\t$rT, $rA, $rB, $rC", SPrecFP,
3961 (fsub (fmul R32FP:$rA, R32FP:$rB), R32FP:$rC))]>;
3963 // Floating Negative Mulitply and Subtract
3964 // Subtracts product from value in rC
3965 // res = fneg(fms a b c)
3968 // NOTE: subtraction order
3972 RRRForm<0b1101, (outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB, R32FP:$rC),
3973 "fnms\t$rT, $rA, $rB, $rC", SPrecFP,
3974 [(set R32FP:$rT, (fsub R32FP:$rC, (fmul R32FP:$rA, R32FP:$rB)))]>;
3977 RRRForm<0b1101, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
3978 "fnms\t$rT, $rA, $rB, $rC", SPrecFP,
3979 [(set (v4f32 VECREG:$rT),
3980 (fsub (v4f32 VECREG:$rC),
3981 (fmul (v4f32 VECREG:$rA),
3982 (v4f32 VECREG:$rB))))]>;
3984 //--------------------------------------
3985 // Floating Point Conversions
3986 // Signed conversions:
3988 CVTIntFPForm<0b0101101110, (outs VECREG:$rT), (ins VECREG:$rA),
3989 "csflt\t$rT, $rA, 0", SPrecFP,
3990 [(set (v4f32 VECREG:$rT), (sint_to_fp (v4i32 VECREG:$rA)))]>;
3992 // Convert signed integer to floating point
3994 CVTIntFPForm<0b0101101110, (outs R32FP:$rT), (ins R32C:$rA),
3995 "csflt\t$rT, $rA, 0", SPrecFP,
3996 [(set R32FP:$rT, (sint_to_fp R32C:$rA))]>;
3998 // Convert unsigned into to float
4000 CVTIntFPForm<0b1101101110, (outs VECREG:$rT), (ins VECREG:$rA),
4001 "cuflt\t$rT, $rA, 0", SPrecFP,
4002 [(set (v4f32 VECREG:$rT), (uint_to_fp (v4i32 VECREG:$rA)))]>;
4005 CVTIntFPForm<0b1101101110, (outs R32FP:$rT), (ins R32C:$rA),
4006 "cuflt\t$rT, $rA, 0", SPrecFP,
4007 [(set R32FP:$rT, (uint_to_fp R32C:$rA))]>;
4009 // Convert float to unsigned int
4010 // Assume that scale = 0
4013 CVTIntFPForm<0b1101101110, (outs VECREG:$rT), (ins VECREG:$rA),
4014 "cfltu\t$rT, $rA, 0", SPrecFP,
4015 [(set (v4i32 VECREG:$rT), (fp_to_uint (v4f32 VECREG:$rA)))]>;
4018 CVTIntFPForm<0b1101101110, (outs R32C:$rT), (ins R32FP:$rA),
4019 "cfltu\t$rT, $rA, 0", SPrecFP,
4020 [(set R32C:$rT, (fp_to_uint R32FP:$rA))]>;
4022 // Convert float to signed int
4023 // Assume that scale = 0
4026 CVTIntFPForm<0b1101101110, (outs VECREG:$rT), (ins VECREG:$rA),
4027 "cflts\t$rT, $rA, 0", SPrecFP,
4028 [(set (v4i32 VECREG:$rT), (fp_to_sint (v4f32 VECREG:$rA)))]>;
4031 CVTIntFPForm<0b1101101110, (outs R32C:$rT), (ins R32FP:$rA),
4032 "cflts\t$rT, $rA, 0", SPrecFP,
4033 [(set R32C:$rT, (fp_to_sint R32FP:$rA))]>;
4035 //===----------------------------------------------------------------------==//
4036 // Single<->Double precision conversions
4037 //===----------------------------------------------------------------------==//
4039 // NOTE: We use "vec" name suffix here to avoid confusion (e.g. input is a
4040 // v4f32, output is v2f64--which goes in the name?)
4042 // Floating point extend single to double
4043 // NOTE: Not sure if passing in v4f32 to FESDvec is correct since it
4044 // operates on two double-word slots (i.e. 1st and 3rd fp numbers
4047 RRForm_1<0b00011101110, (outs VECREG:$rT), (ins VECREG:$rA),
4048 "fesd\t$rT, $rA", SPrecFP,
4049 [(set (v2f64 VECREG:$rT), (fextend (v4f32 VECREG:$rA)))]>;
4052 RRForm_1<0b00011101110, (outs R64FP:$rT), (ins R32FP:$rA),
4053 "fesd\t$rT, $rA", SPrecFP,
4054 [(set R64FP:$rT, (fextend R32FP:$rA))]>;
4056 // Floating point round double to single
4058 // RRForm_1<0b10011101110, (outs VECREG:$rT), (ins VECREG:$rA),
4059 // "frds\t$rT, $rA,", SPrecFP,
4060 // [(set (v4f32 R32FP:$rT), (fround (v2f64 R64FP:$rA)))]>;
4063 RRForm_1<0b10011101110, (outs R32FP:$rT), (ins R64FP:$rA),
4064 "frds\t$rT, $rA", SPrecFP,
4065 [(set R32FP:$rT, (fround R64FP:$rA))]>;
4067 //ToDo include anyextend?
4069 //===----------------------------------------------------------------------==//
4070 // Double precision floating point instructions
4071 //===----------------------------------------------------------------------==//
4073 RRForm<0b00110011010, (outs R64FP:$rT), (ins R64FP:$rA, R64FP:$rB),
4074 "dfa\t$rT, $rA, $rB", DPrecFP,
4075 [(set R64FP:$rT, (fadd R64FP:$rA, R64FP:$rB))]>;
4078 RRForm<0b00110011010, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
4079 "dfa\t$rT, $rA, $rB", DPrecFP,
4080 [(set (v2f64 VECREG:$rT), (fadd (v2f64 VECREG:$rA), (v2f64 VECREG:$rB)))]>;
4083 RRForm<0b10100011010, (outs R64FP:$rT), (ins R64FP:$rA, R64FP:$rB),
4084 "dfs\t$rT, $rA, $rB", DPrecFP,
4085 [(set R64FP:$rT, (fsub R64FP:$rA, R64FP:$rB))]>;
4088 RRForm<0b10100011010, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
4089 "dfs\t$rT, $rA, $rB", DPrecFP,
4090 [(set (v2f64 VECREG:$rT),
4091 (fsub (v2f64 VECREG:$rA), (v2f64 VECREG:$rB)))]>;
4094 RRForm<0b01100011010, (outs R64FP:$rT), (ins R64FP:$rA, R64FP:$rB),
4095 "dfm\t$rT, $rA, $rB", DPrecFP,
4096 [(set R64FP:$rT, (fmul R64FP:$rA, R64FP:$rB))]>;
4099 RRForm<0b00100011010, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
4100 "dfm\t$rT, $rA, $rB", DPrecFP,
4101 [(set (v2f64 VECREG:$rT),
4102 (fmul (v2f64 VECREG:$rA), (v2f64 VECREG:$rB)))]>;
4105 RRForm<0b00111010110, (outs R64FP:$rT),
4106 (ins R64FP:$rA, R64FP:$rB, R64FP:$rC),
4107 "dfma\t$rT, $rA, $rB", DPrecFP,
4108 [(set R64FP:$rT, (fadd R64FP:$rC, (fmul R64FP:$rA, R64FP:$rB)))]>,
4109 RegConstraint<"$rC = $rT">,
4113 RRForm<0b00111010110, (outs VECREG:$rT),
4114 (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
4115 "dfma\t$rT, $rA, $rB", DPrecFP,
4116 [(set (v2f64 VECREG:$rT),
4117 (fadd (v2f64 VECREG:$rC),
4118 (fmul (v2f64 VECREG:$rA), (v2f64 VECREG:$rB))))]>,
4119 RegConstraint<"$rC = $rT">,
4123 RRForm<0b10111010110, (outs R64FP:$rT),
4124 (ins R64FP:$rA, R64FP:$rB, R64FP:$rC),
4125 "dfms\t$rT, $rA, $rB", DPrecFP,
4126 [(set R64FP:$rT, (fsub (fmul R64FP:$rA, R64FP:$rB), R64FP:$rC))]>,
4127 RegConstraint<"$rC = $rT">,
4131 RRForm<0b10111010110, (outs VECREG:$rT),
4132 (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
4133 "dfms\t$rT, $rA, $rB", DPrecFP,
4134 [(set (v2f64 VECREG:$rT),
4135 (fsub (fmul (v2f64 VECREG:$rA), (v2f64 VECREG:$rB)),
4136 (v2f64 VECREG:$rC)))]>;
4138 // FNMS: - (a * b - c)
4139 // - (a * b) + c => c - (a * b)
4141 RRForm<0b01111010110, (outs R64FP:$rT),
4142 (ins R64FP:$rA, R64FP:$rB, R64FP:$rC),
4143 "dfnms\t$rT, $rA, $rB", DPrecFP,
4144 [(set R64FP:$rT, (fsub R64FP:$rC, (fmul R64FP:$rA, R64FP:$rB)))]>,
4145 RegConstraint<"$rC = $rT">,
4148 def : Pat<(fneg (fsub (fmul R64FP:$rA, R64FP:$rB), R64FP:$rC)),
4149 (FNMSf64 R64FP:$rA, R64FP:$rB, R64FP:$rC)>;
4152 RRForm<0b01111010110, (outs VECREG:$rT),
4153 (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
4154 "dfnms\t$rT, $rA, $rB", DPrecFP,
4155 [(set (v2f64 VECREG:$rT),
4156 (fsub (v2f64 VECREG:$rC),
4157 (fmul (v2f64 VECREG:$rA),
4158 (v2f64 VECREG:$rB))))]>,
4159 RegConstraint<"$rC = $rT">,
4162 def : Pat<(fneg (fsub (fmul (v2f64 VECREG:$rA), (v2f64 VECREG:$rB)),
4163 (v2f64 VECREG:$rC))),
4164 (FNMSv2f64 VECREG:$rA, VECREG:$rB, VECREG:$rC)>;
4169 RRForm<0b11111010110, (outs R64FP:$rT),
4170 (ins R64FP:$rA, R64FP:$rB, R64FP:$rC),
4171 "dfnma\t$rT, $rA, $rB", DPrecFP,
4172 [(set R64FP:$rT, (fneg (fadd R64FP:$rC, (fmul R64FP:$rA, R64FP:$rB))))]>,
4173 RegConstraint<"$rC = $rT">,
4177 RRForm<0b11111010110, (outs VECREG:$rT),
4178 (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
4179 "dfnma\t$rT, $rA, $rB", DPrecFP,
4180 [(set (v2f64 VECREG:$rT),
4181 (fneg (fadd (v2f64 VECREG:$rC),
4182 (fmul (v2f64 VECREG:$rA),
4183 (v2f64 VECREG:$rB)))))]>,
4184 RegConstraint<"$rC = $rT">,
4187 //===----------------------------------------------------------------------==//
4188 // Floating point negation and absolute value
4189 //===----------------------------------------------------------------------==//
4191 def : Pat<(fneg (v4f32 VECREG:$rA)),
4192 (XORfnegvec (v4f32 VECREG:$rA),
4193 (v4f32 (ILHUv4i32 0x8000)))>;
4195 def : Pat<(fneg R32FP:$rA),
4196 (XORfneg32 R32FP:$rA, (ILHUr32 0x8000))>;
4198 def : Pat<(fneg (v2f64 VECREG:$rA)),
4199 (XORfnegvec (v2f64 VECREG:$rA),
4200 (v2f64 (ANDBIv16i8 (FSMBIv16i8 0x8080), 0x80)))>;
4202 def : Pat<(fneg R64FP:$rA),
4203 (XORfneg64 R64FP:$rA,
4204 (ANDBIv16i8 (FSMBIv16i8 0x8080), 0x80))>;
4206 // Floating point absolute value
4208 def : Pat<(fabs R32FP:$rA),
4209 (ANDfabs32 R32FP:$rA, (IOHLr32 (ILHUr32 0x7fff), 0xffff))>;
4211 def : Pat<(fabs (v4f32 VECREG:$rA)),
4212 (ANDfabsvec (v4f32 VECREG:$rA),
4213 (v4f32 (ANDBIv16i8 (FSMBIv16i8 0xffff), 0x7f)))>;
4215 def : Pat<(fabs R64FP:$rA),
4216 (ANDfabs64 R64FP:$rA, (ANDBIv16i8 (FSMBIv16i8 0xffff), 0x7f))>;
4218 def : Pat<(fabs (v2f64 VECREG:$rA)),
4219 (ANDfabsvec (v2f64 VECREG:$rA),
4220 (v2f64 (ANDBIv16i8 (FSMBIv16i8 0xffff), 0x7f)))>;
4222 //===----------------------------------------------------------------------===//
4223 // Hint for branch instructions:
4224 //===----------------------------------------------------------------------===//
4226 /* def HBR : SPUInstr<(outs), (ins), "hbr\t" */
4228 //===----------------------------------------------------------------------===//
4229 // Execution, Load NOP (execute NOPs belong in even pipeline, load NOPs belong
4230 // in the odd pipeline)
4231 //===----------------------------------------------------------------------===//
4233 def ENOP : SPUInstr<(outs), (ins), "enop", ExecNOP> {
4236 let Inst{0-10} = 0b10000000010;
4237 let Inst{11-17} = 0;
4238 let Inst{18-24} = 0;
4239 let Inst{25-31} = 0;
4242 def LNOP : SPUInstr<(outs), (ins), "lnop", LoadNOP> {
4245 let Inst{0-10} = 0b10000000000;
4246 let Inst{11-17} = 0;
4247 let Inst{18-24} = 0;
4248 let Inst{25-31} = 0;
4251 //===----------------------------------------------------------------------===//
4252 // Bit conversions (type conversions between vector/packed types)
4253 // NOTE: Promotions are handled using the XS* instructions. Truncation
4255 //===----------------------------------------------------------------------===//
4256 def : Pat<(v16i8 (bitconvert (v8i16 VECREG:$src))), (v16i8 VECREG:$src)>;
4257 def : Pat<(v16i8 (bitconvert (v4i32 VECREG:$src))), (v16i8 VECREG:$src)>;
4258 def : Pat<(v16i8 (bitconvert (v2i64 VECREG:$src))), (v16i8 VECREG:$src)>;
4259 def : Pat<(v16i8 (bitconvert (v4f32 VECREG:$src))), (v16i8 VECREG:$src)>;
4260 def : Pat<(v16i8 (bitconvert (v2f64 VECREG:$src))), (v16i8 VECREG:$src)>;
4262 def : Pat<(v8i16 (bitconvert (v16i8 VECREG:$src))), (v8i16 VECREG:$src)>;
4263 def : Pat<(v8i16 (bitconvert (v4i32 VECREG:$src))), (v8i16 VECREG:$src)>;
4264 def : Pat<(v8i16 (bitconvert (v2i64 VECREG:$src))), (v8i16 VECREG:$src)>;
4265 def : Pat<(v8i16 (bitconvert (v4f32 VECREG:$src))), (v8i16 VECREG:$src)>;
4266 def : Pat<(v8i16 (bitconvert (v2f64 VECREG:$src))), (v8i16 VECREG:$src)>;
4268 def : Pat<(v4i32 (bitconvert (v16i8 VECREG:$src))), (v4i32 VECREG:$src)>;
4269 def : Pat<(v4i32 (bitconvert (v8i16 VECREG:$src))), (v4i32 VECREG:$src)>;
4270 def : Pat<(v4i32 (bitconvert (v2i64 VECREG:$src))), (v4i32 VECREG:$src)>;
4271 def : Pat<(v4i32 (bitconvert (v4f32 VECREG:$src))), (v4i32 VECREG:$src)>;
4272 def : Pat<(v4i32 (bitconvert (v2f64 VECREG:$src))), (v4i32 VECREG:$src)>;
4274 def : Pat<(v2i64 (bitconvert (v16i8 VECREG:$src))), (v2i64 VECREG:$src)>;
4275 def : Pat<(v2i64 (bitconvert (v8i16 VECREG:$src))), (v2i64 VECREG:$src)>;
4276 def : Pat<(v2i64 (bitconvert (v4i32 VECREG:$src))), (v2i64 VECREG:$src)>;
4277 def : Pat<(v2i64 (bitconvert (v4f32 VECREG:$src))), (v2i64 VECREG:$src)>;
4278 def : Pat<(v2i64 (bitconvert (v2f64 VECREG:$src))), (v2i64 VECREG:$src)>;
4280 def : Pat<(v4f32 (bitconvert (v16i8 VECREG:$src))), (v4f32 VECREG:$src)>;
4281 def : Pat<(v4f32 (bitconvert (v8i16 VECREG:$src))), (v4f32 VECREG:$src)>;
4282 def : Pat<(v4f32 (bitconvert (v2i64 VECREG:$src))), (v4f32 VECREG:$src)>;
4283 def : Pat<(v4f32 (bitconvert (v4i32 VECREG:$src))), (v4f32 VECREG:$src)>;
4284 def : Pat<(v4f32 (bitconvert (v2f64 VECREG:$src))), (v4f32 VECREG:$src)>;
4286 def : Pat<(v2f64 (bitconvert (v16i8 VECREG:$src))), (v2f64 VECREG:$src)>;
4287 def : Pat<(v2f64 (bitconvert (v8i16 VECREG:$src))), (v2f64 VECREG:$src)>;
4288 def : Pat<(v2f64 (bitconvert (v4i32 VECREG:$src))), (v2f64 VECREG:$src)>;
4289 def : Pat<(v2f64 (bitconvert (v2i64 VECREG:$src))), (v2f64 VECREG:$src)>;
4290 def : Pat<(v2f64 (bitconvert (v2f64 VECREG:$src))), (v2f64 VECREG:$src)>;
4292 def : Pat<(f32 (bitconvert (i32 R32C:$src))), (f32 R32FP:$src)>;
4293 def : Pat<(f64 (bitconvert (i64 R64C:$src))), (f64 R64FP:$src)>;
4295 //===----------------------------------------------------------------------===//
4296 // Instruction patterns:
4297 //===----------------------------------------------------------------------===//
4299 // General 32-bit constants:
4300 def : Pat<(i32 imm:$imm),
4301 (IOHLr32 (ILHUr32 (HI16 imm:$imm)), (LO16 imm:$imm))>;
4303 // Single precision float constants:
4304 def : Pat<(f32 fpimm:$imm),
4305 (IOHLf32 (ILHUf32 (HI16_f32 fpimm:$imm)), (LO16_f32 fpimm:$imm))>;
4307 // General constant 32-bit vectors
4308 def : Pat<(v4i32 v4i32Imm:$imm),
4309 (IOHLv4i32 (v4i32 (ILHUv4i32 (HI16_vec v4i32Imm:$imm))),
4310 (LO16_vec v4i32Imm:$imm))>;
4313 def : Pat<(i8 imm:$imm),
4316 //===----------------------------------------------------------------------===//
4317 // Call instruction patterns:
4318 //===----------------------------------------------------------------------===//
4323 //===----------------------------------------------------------------------===//
4324 // Zero/Any/Sign extensions
4325 //===----------------------------------------------------------------------===//
4327 // sext 8->32: Sign extend bytes to words
4328 def : Pat<(sext_inreg R32C:$rSrc, i8),
4329 (XSHWr32 (XSBHr32 R32C:$rSrc))>;
4331 def : Pat<(i32 (sext R8C:$rSrc)),
4332 (XSHWr16 (XSBHr8 R8C:$rSrc))>;
4334 // zext 8->16: Zero extend bytes to halfwords
4335 def : Pat<(i16 (zext R8C:$rSrc)),
4336 (ANDHIi8i16 R8C:$rSrc, 0xff)>;
4338 // zext 8->32: Zero extend bytes to words
4339 def : Pat<(i32 (zext R8C:$rSrc)),
4340 (ANDIi8i32 R8C:$rSrc, 0xff)>;
4342 // anyext 8->16: Extend 8->16 bits, irrespective of sign
4343 def : Pat<(i16 (anyext R8C:$rSrc)),
4344 (ORHIi8i16 R8C:$rSrc, 0)>;
4346 // anyext 8->32: Extend 8->32 bits, irrespective of sign
4347 def : Pat<(i32 (anyext R8C:$rSrc)),
4348 (ORIi8i32 R8C:$rSrc, 0)>;
4350 // zext 16->32: Zero extend halfwords to words
4351 def : Pat<(i32 (zext R16C:$rSrc)),
4352 (ANDi16i32 R16C:$rSrc, (ILAr32 0xffff))>;
4354 def : Pat<(i32 (zext (and R16C:$rSrc, 0xf))),
4355 (ANDIi16i32 R16C:$rSrc, 0xf)>;
4357 def : Pat<(i32 (zext (and R16C:$rSrc, 0xff))),
4358 (ANDIi16i32 R16C:$rSrc, 0xff)>;
4360 def : Pat<(i32 (zext (and R16C:$rSrc, 0xfff))),
4361 (ANDIi16i32 R16C:$rSrc, 0xfff)>;
4363 // anyext 16->32: Extend 16->32 bits, irrespective of sign
4364 def : Pat<(i32 (anyext R16C:$rSrc)),
4365 (ORIi16i32 R16C:$rSrc, 0)>;
4367 //===----------------------------------------------------------------------===//
4369 // These truncates are for the SPU's supported types (i8, i16, i32). i64 and
4370 // above are custom lowered.
4371 //===----------------------------------------------------------------------===//
4373 def : Pat<(i8 (trunc GPRC:$src)),
4375 (SHUFBgprc GPRC:$src, GPRC:$src,
4376 (IOHLv4i32 (ILHUv4i32 0x0f0f), 0x0f0f)))>;
4378 def : Pat<(i8 (trunc R64C:$src)),
4381 (ORv2i64_i64 R64C:$src),
4382 (ORv2i64_i64 R64C:$src),
4383 (IOHLv4i32 (ILHUv4i32 0x0707), 0x0707)))>;
4385 def : Pat<(i8 (trunc R32C:$src)),
4388 (ORv4i32_i32 R32C:$src),
4389 (ORv4i32_i32 R32C:$src),
4390 (IOHLv4i32 (ILHUv4i32 0x0303), 0x0303)))>;
4392 def : Pat<(i8 (trunc R16C:$src)),
4395 (ORv8i16_i16 R16C:$src),
4396 (ORv8i16_i16 R16C:$src),
4397 (IOHLv4i32 (ILHUv4i32 0x0303), 0x0303)))>;
4399 def : Pat<(i16 (trunc GPRC:$src)),
4401 (SHUFBgprc GPRC:$src, GPRC:$src,
4402 (IOHLv4i32 (ILHUv4i32 0x0e0f), 0x0e0f)))>;
4404 def : Pat<(i16 (trunc R64C:$src)),
4407 (ORv2i64_i64 R64C:$src),
4408 (ORv2i64_i64 R64C:$src),
4409 (IOHLv4i32 (ILHUv4i32 0x0607), 0x0607)))>;
4411 def : Pat<(i16 (trunc R32C:$src)),
4414 (ORv4i32_i32 R32C:$src),
4415 (ORv4i32_i32 R32C:$src),
4416 (IOHLv4i32 (ILHUv4i32 0x0203), 0x0203)))>;
4418 def : Pat<(i32 (trunc GPRC:$src)),
4420 (SHUFBgprc GPRC:$src, GPRC:$src,
4421 (IOHLv4i32 (ILHUv4i32 0x0c0d), 0x0e0f)))>;
4423 def : Pat<(i32 (trunc R64C:$src)),
4426 (ORv2i64_i64 R64C:$src),
4427 (ORv2i64_i64 R64C:$src),
4428 (IOHLv4i32 (ILHUv4i32 0x0405), 0x0607)))>;
4430 //===----------------------------------------------------------------------===//
4431 // Address generation: SPU, like PPC, has to split addresses into high and
4432 // low parts in order to load them into a register.
4433 //===----------------------------------------------------------------------===//
4435 def : Pat<(SPUaform tglobaladdr:$in, 0), (ILAlsa tglobaladdr:$in)>;
4436 def : Pat<(SPUaform texternalsym:$in, 0), (ILAlsa texternalsym:$in)>;
4437 def : Pat<(SPUaform tjumptable:$in, 0), (ILAlsa tjumptable:$in)>;
4438 def : Pat<(SPUaform tconstpool:$in, 0), (ILAlsa tconstpool:$in)>;
4440 def : Pat<(SPUindirect (SPUhi tglobaladdr:$in, 0),
4441 (SPUlo tglobaladdr:$in, 0)),
4442 (IOHLlo (ILHUhi tglobaladdr:$in), tglobaladdr:$in)>;
4444 def : Pat<(SPUindirect (SPUhi texternalsym:$in, 0),
4445 (SPUlo texternalsym:$in, 0)),
4446 (IOHLlo (ILHUhi texternalsym:$in), texternalsym:$in)>;
4448 def : Pat<(SPUindirect (SPUhi tjumptable:$in, 0),
4449 (SPUlo tjumptable:$in, 0)),
4450 (IOHLlo (ILHUhi tjumptable:$in), tjumptable:$in)>;
4452 def : Pat<(SPUindirect (SPUhi tconstpool:$in, 0),
4453 (SPUlo tconstpool:$in, 0)),
4454 (IOHLlo (ILHUhi tconstpool:$in), tconstpool:$in)>;
4457 def : Pat<(SPUindirect R32C:$sp, i32ImmSExt10:$imm),
4458 (AIr32 R32C:$sp, i32ImmSExt10:$imm)>;
4460 def : Pat<(SPUindirect R32C:$sp, imm:$imm),
4462 (IOHLr32 (ILHUr32 (HI16 imm:$imm)), (LO16 imm:$imm)))>;
4465 def : Pat<(add (SPUhi tglobaladdr:$in, 0), (SPUlo tglobaladdr:$in, 0)),
4466 (IOHLlo (ILHUhi tglobaladdr:$in), tglobaladdr:$in)>;
4468 def : Pat<(add (SPUhi texternalsym:$in, 0), (SPUlo texternalsym:$in, 0)),
4469 (IOHLlo (ILHUhi texternalsym:$in), texternalsym:$in)>;
4471 def : Pat<(add (SPUhi tjumptable:$in, 0), (SPUlo tjumptable:$in, 0)),
4472 (IOHLlo (ILHUhi tjumptable:$in), tjumptable:$in)>;
4474 def : Pat<(add (SPUhi tconstpool:$in, 0), (SPUlo tconstpool:$in, 0)),
4475 (IOHLlo (ILHUhi tconstpool:$in), tconstpool:$in)>;
4478 include "CellSDKIntrinsics.td"
4479 // Various math operator instruction sequences
4480 include "SPUMathInstr.td"
4481 // 64-bit "instructions"/support
4482 include "SPU64InstrInfo.td"