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:$rSrc))]>;
1179 // Sign-extend words to doublewords (32->64 bits)
1181 class XSWDInst<dag OOL, dag IOL, list<dag> pattern>:
1182 RRForm_1<0b01100101010, OOL, IOL,
1183 "xswd\t$rDst, $rSrc", IntegerOp,
1186 class XSWDVecInst<ValueType in_vectype, ValueType out_vectype>:
1187 XSWDInst<(outs VECREG:$rDst), (ins VECREG:$rSrc),
1188 [(set (out_vectype VECREG:$rDst),
1189 (sext (out_vectype VECREG:$rSrc)))]>;
1191 class XSWDRegInst<RegisterClass in_rclass, RegisterClass out_rclass>:
1192 XSWDInst<(outs out_rclass:$rDst), (ins in_rclass:$rSrc),
1193 [(set out_rclass:$rDst, (sext in_rclass:$rSrc))]>;
1195 multiclass ExtendWordToDoubleWord {
1196 def v2i64: XSWDVecInst<v4i32, v2i64>;
1197 def r64: XSWDRegInst<R32C, R64C>;
1199 def r64_inreg: XSWDInst<(outs R64C:$rDst), (ins R64C:$rSrc),
1200 [(set R64C:$rDst, (sext_inreg R64C:$rSrc, i32))]>;
1203 defm XSWD : ExtendWordToDoubleWord;
1207 class ANDInst<dag OOL, dag IOL, list<dag> pattern> :
1208 RRForm<0b10000011000, OOL, IOL, "and\t$rT, $rA, $rB",
1209 IntegerOp, pattern>;
1211 class ANDVecInst<ValueType vectype>:
1212 ANDInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1213 [(set (vectype VECREG:$rT), (and (vectype VECREG:$rA),
1214 (vectype VECREG:$rB)))]>;
1216 class ANDRegInst<RegisterClass rclass>:
1217 ANDInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
1218 [(set rclass:$rT, (and rclass:$rA, rclass:$rB))]>;
1220 multiclass BitwiseAnd
1222 def v16i8: ANDVecInst<v16i8>;
1223 def v8i16: ANDVecInst<v8i16>;
1224 def v4i32: ANDVecInst<v4i32>;
1225 def v2i64: ANDVecInst<v2i64>;
1227 def r128: ANDRegInst<GPRC>;
1228 def r64: ANDRegInst<R64C>;
1229 def r32: ANDRegInst<R32C>;
1230 def r16: ANDRegInst<R16C>;
1231 def r8: ANDRegInst<R8C>;
1233 //===---------------------------------------------
1234 // Special instructions to perform the fabs instruction
1235 def fabs32: ANDInst<(outs R32FP:$rT), (ins R32FP:$rA, R32C:$rB),
1236 [/* Intentionally does not match a pattern */]>;
1238 def fabs64: ANDInst<(outs R64FP:$rT), (ins R64FP:$rA, VECREG:$rB),
1239 [/* Intentionally does not match a pattern */]>;
1241 // Could use v4i32, but won't for clarity
1242 def fabsvec: ANDInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1243 [/* Intentionally does not match a pattern */]>;
1245 //===---------------------------------------------
1247 // Hacked form of AND to zero-extend 16-bit quantities to 32-bit
1248 // quantities -- see 16->32 zext pattern.
1250 // This pattern is somewhat artificial, since it might match some
1251 // compiler generated pattern but it is unlikely to do so.
1253 def i16i32: ANDInst<(outs R32C:$rT), (ins R16C:$rA, R32C:$rB),
1254 [(set R32C:$rT, (and (zext R16C:$rA), R32C:$rB))]>;
1257 defm AND : BitwiseAnd;
1259 // N.B.: vnot_conv is one of those special target selection pattern fragments,
1260 // in which we expect there to be a bit_convert on the constant. Bear in mind
1261 // that llvm translates "not <reg>" to "xor <reg>, -1" (or in this case, a
1262 // constant -1 vector.)
1264 class ANDCInst<dag OOL, dag IOL, list<dag> pattern>:
1265 RRForm<0b10000011010, OOL, IOL, "andc\t$rT, $rA, $rB",
1266 IntegerOp, pattern>;
1268 class ANDCVecInst<ValueType vectype>:
1269 ANDCInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1270 [(set (vectype VECREG:$rT), (and (vectype VECREG:$rA),
1271 (vnot (vectype VECREG:$rB))))]>;
1273 class ANDCRegInst<RegisterClass rclass>:
1274 ANDCInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
1275 [(set rclass:$rT, (and rclass:$rA, (not rclass:$rB)))]>;
1277 multiclass AndComplement
1279 def v16i8: ANDCVecInst<v16i8>;
1280 def v8i16: ANDCVecInst<v8i16>;
1281 def v4i32: ANDCVecInst<v4i32>;
1282 def v2i64: ANDCVecInst<v2i64>;
1284 def r128: ANDCRegInst<GPRC>;
1285 def r64: ANDCRegInst<R64C>;
1286 def r32: ANDCRegInst<R32C>;
1287 def r16: ANDCRegInst<R16C>;
1288 def r8: ANDCRegInst<R8C>;
1291 defm ANDC : AndComplement;
1293 class ANDBIInst<dag OOL, dag IOL, list<dag> pattern>:
1294 RI10Form<0b01101000, OOL, IOL, "andbi\t$rT, $rA, $val",
1297 multiclass AndByteImm
1299 def v16i8: ANDBIInst<(outs VECREG:$rT), (ins VECREG:$rA, u10imm:$val),
1300 [(set (v16i8 VECREG:$rT),
1301 (and (v16i8 VECREG:$rA),
1302 (v16i8 v16i8U8Imm:$val)))]>;
1304 def r8: ANDBIInst<(outs R8C:$rT), (ins R8C:$rA, u10imm_i8:$val),
1305 [(set R8C:$rT, (and R8C:$rA, immU8:$val))]>;
1308 defm ANDBI : AndByteImm;
1310 class ANDHIInst<dag OOL, dag IOL, list<dag> pattern> :
1311 RI10Form<0b10101000, OOL, IOL, "andhi\t$rT, $rA, $val",
1314 multiclass AndHalfwordImm
1316 def v8i16: ANDHIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
1317 [(set (v8i16 VECREG:$rT),
1318 (and (v8i16 VECREG:$rA), v8i16SExt10Imm:$val))]>;
1320 def r16: ANDHIInst<(outs R16C:$rT), (ins R16C:$rA, u10imm:$val),
1321 [(set R16C:$rT, (and R16C:$rA, i16ImmUns10:$val))]>;
1323 // Zero-extend i8 to i16:
1324 def i8i16: ANDHIInst<(outs R16C:$rT), (ins R8C:$rA, u10imm:$val),
1325 [(set R16C:$rT, (and (zext R8C:$rA), i16ImmUns10:$val))]>;
1328 defm ANDHI : AndHalfwordImm;
1330 class ANDIInst<dag OOL, dag IOL, list<dag> pattern> :
1331 RI10Form<0b00101000, OOL, IOL, "andi\t$rT, $rA, $val",
1332 IntegerOp, pattern>;
1334 multiclass AndWordImm
1336 def v4i32: ANDIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
1337 [(set (v4i32 VECREG:$rT),
1338 (and (v4i32 VECREG:$rA), v4i32SExt10Imm:$val))]>;
1340 def r32: ANDIInst<(outs R32C:$rT), (ins R32C:$rA, s10imm_i32:$val),
1341 [(set R32C:$rT, (and R32C:$rA, i32ImmSExt10:$val))]>;
1343 // Hacked form of ANDI to zero-extend i8 quantities to i32. See the zext 8->32
1345 def i8i32: ANDIInst<(outs R32C:$rT), (ins R8C:$rA, s10imm_i32:$val),
1347 (and (zext R8C:$rA), i32ImmSExt10:$val))]>;
1349 // Hacked form of ANDI to zero-extend i16 quantities to i32. See the
1350 // zext 16->32 pattern below.
1352 // Note that this pattern is somewhat artificial, since it might match
1353 // something the compiler generates but is unlikely to occur in practice.
1354 def i16i32: ANDIInst<(outs R32C:$rT), (ins R16C:$rA, s10imm_i32:$val),
1356 (and (zext R16C:$rA), i32ImmSExt10:$val))]>;
1359 defm ANDI : AndWordImm;
1361 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
1362 // Bitwise OR group:
1363 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
1365 // Bitwise "or" (N.B.: These are also register-register copy instructions...)
1366 class ORInst<dag OOL, dag IOL, list<dag> pattern>:
1367 RRForm<0b10000010000, OOL, IOL, "or\t$rT, $rA, $rB",
1368 IntegerOp, pattern>;
1370 class ORVecInst<ValueType vectype>:
1371 ORInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1372 [(set (vectype VECREG:$rT), (or (vectype VECREG:$rA),
1373 (vectype VECREG:$rB)))]>;
1375 class ORRegInst<RegisterClass rclass>:
1376 ORInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
1377 [(set rclass:$rT, (or rclass:$rA, rclass:$rB))]>;
1379 // ORCvtForm: OR conversion form
1381 // This is used to "convert" the preferred slot to its vector equivalent, as
1382 // well as convert a vector back to its preferred slot.
1384 // These are effectively no-ops, but need to exist for proper type conversion
1385 // and type coercion.
1387 class ORCvtForm<dag OOL, dag IOL>
1388 : SPUInstr<OOL, IOL, "or\t$rT, $rA, $rA", IntegerOp> {
1392 let Pattern = [/* no pattern */];
1394 let Inst{0-10} = 0b10000010000;
1395 let Inst{11-17} = RA;
1396 let Inst{18-24} = RA;
1397 let Inst{25-31} = RT;
1400 class ORPromoteScalar<RegisterClass rclass>:
1401 ORCvtForm<(outs VECREG:$rT), (ins rclass:$rA)>;
1403 class ORExtractElt<RegisterClass rclass>:
1404 ORCvtForm<(outs rclass:$rT), (ins VECREG:$rA)>;
1406 class ORCvtRegGPRC<RegisterClass rclass>:
1407 ORCvtForm<(outs GPRC:$rT), (ins rclass:$rA)>;
1410 ORCvtForm<(outs GPRC:$rT), (ins VECREG:$rA)>;
1412 class ORCvtGPRCReg<RegisterClass rclass>:
1413 ORCvtForm<(outs rclass:$rT), (ins GPRC:$rA)>;
1416 ORCvtForm<(outs VECREG:$rT), (ins GPRC:$rA)>;
1418 multiclass BitwiseOr
1420 def v16i8: ORVecInst<v16i8>;
1421 def v8i16: ORVecInst<v8i16>;
1422 def v4i32: ORVecInst<v4i32>;
1423 def v2i64: ORVecInst<v2i64>;
1425 def v4f32: ORInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1426 [(set (v4f32 VECREG:$rT),
1427 (v4f32 (bitconvert (or (v4i32 VECREG:$rA),
1428 (v4i32 VECREG:$rB)))))]>;
1430 def v2f64: ORInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1431 [(set (v2f64 VECREG:$rT),
1432 (v2f64 (bitconvert (or (v2i64 VECREG:$rA),
1433 (v2i64 VECREG:$rB)))))]>;
1435 def r64: ORRegInst<R64C>;
1436 def r32: ORRegInst<R32C>;
1437 def r16: ORRegInst<R16C>;
1438 def r8: ORRegInst<R8C>;
1440 // OR instructions used to copy f32 and f64 registers.
1441 def f32: ORInst<(outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB),
1442 [/* no pattern */]>;
1444 def f64: ORInst<(outs R64FP:$rT), (ins R64FP:$rA, R64FP:$rB),
1445 [/* no pattern */]>;
1447 // scalar->vector promotion, prefslot2vec:
1448 def v16i8_i8: ORPromoteScalar<R8C>;
1449 def v8i16_i16: ORPromoteScalar<R16C>;
1450 def v4i32_i32: ORPromoteScalar<R32C>;
1451 def v2i64_i64: ORPromoteScalar<R64C>;
1452 def v4f32_f32: ORPromoteScalar<R32FP>;
1453 def v2f64_f64: ORPromoteScalar<R64FP>;
1455 // vector->scalar demotion, vec2prefslot:
1456 def i8_v16i8: ORExtractElt<R8C>;
1457 def i16_v8i16: ORExtractElt<R16C>;
1458 def i32_v4i32: ORExtractElt<R32C>;
1459 def i64_v2i64: ORExtractElt<R64C>;
1460 def f32_v4f32: ORExtractElt<R32FP>;
1461 def f64_v2f64: ORExtractElt<R64FP>;
1463 // Conversion from GPRC to register
1464 def i128_r64: ORCvtRegGPRC<R64C>;
1465 def i128_f64: ORCvtRegGPRC<R64FP>;
1466 def i128_r32: ORCvtRegGPRC<R32C>;
1467 def i128_f32: ORCvtRegGPRC<R32FP>;
1468 def i128_r16: ORCvtRegGPRC<R16C>;
1469 def i128_r8: ORCvtRegGPRC<R8C>;
1471 // Conversion from GPRC to vector
1472 def i128_vec: ORCvtVecGPRC;
1474 // Conversion from register to GPRC
1475 def r64_i128: ORCvtGPRCReg<R64C>;
1476 def f64_i128: ORCvtGPRCReg<R64FP>;
1477 def r32_i128: ORCvtGPRCReg<R32C>;
1478 def f32_i128: ORCvtGPRCReg<R32FP>;
1479 def r16_i128: ORCvtGPRCReg<R16C>;
1480 def r8_i128: ORCvtGPRCReg<R8C>;
1482 // Conversion from vector to GPRC
1483 def vec_i128: ORCvtGPRCVec;
1486 defm OR : BitwiseOr;
1488 // scalar->vector promotion patterns (preferred slot to vector):
1489 def : Pat<(v16i8 (SPUprefslot2vec R8C:$rA)),
1490 (ORv16i8_i8 R8C:$rA)>;
1492 def : Pat<(v8i16 (SPUprefslot2vec R16C:$rA)),
1493 (ORv8i16_i16 R16C:$rA)>;
1495 def : Pat<(v4i32 (SPUprefslot2vec R32C:$rA)),
1496 (ORv4i32_i32 R32C:$rA)>;
1498 def : Pat<(v2i64 (SPUprefslot2vec R64C:$rA)),
1499 (ORv2i64_i64 R64C:$rA)>;
1501 def : Pat<(v4f32 (SPUprefslot2vec R32FP:$rA)),
1502 (ORv4f32_f32 R32FP:$rA)>;
1504 def : Pat<(v2f64 (SPUprefslot2vec R64FP:$rA)),
1505 (ORv2f64_f64 R64FP:$rA)>;
1507 // ORi*_v*: Used to extract vector element 0 (the preferred slot), otherwise
1508 // known as converting the vector back to its preferred slot
1510 def : Pat<(SPUvec2prefslot (v16i8 VECREG:$rA)),
1511 (ORi8_v16i8 VECREG:$rA)>;
1513 def : Pat<(SPUvec2prefslot (v8i16 VECREG:$rA)),
1514 (ORi16_v8i16 VECREG:$rA)>;
1516 def : Pat<(SPUvec2prefslot (v4i32 VECREG:$rA)),
1517 (ORi32_v4i32 VECREG:$rA)>;
1519 def : Pat<(SPUvec2prefslot (v2i64 VECREG:$rA)),
1520 (ORi64_v2i64 VECREG:$rA)>;
1522 def : Pat<(SPUvec2prefslot (v4f32 VECREG:$rA)),
1523 (ORf32_v4f32 VECREG:$rA)>;
1525 def : Pat<(SPUvec2prefslot (v2f64 VECREG:$rA)),
1526 (ORf64_v2f64 VECREG:$rA)>;
1528 // Load Register: This is an assembler alias for a bitwise OR of a register
1529 // against itself. It's here because it brings some clarity to assembly
1532 let hasCtrlDep = 1 in {
1533 class LRInst<dag OOL, dag IOL>
1534 : SPUInstr<OOL, IOL, "lr\t$rT, $rA", IntegerOp> {
1538 let Pattern = [/*no pattern*/];
1540 let Inst{0-10} = 0b10000010000; /* It's an OR operation */
1541 let Inst{11-17} = RA;
1542 let Inst{18-24} = RA;
1543 let Inst{25-31} = RT;
1546 class LRVecInst<ValueType vectype>:
1547 LRInst<(outs VECREG:$rT), (ins VECREG:$rA)>;
1549 class LRRegInst<RegisterClass rclass>:
1550 LRInst<(outs rclass:$rT), (ins rclass:$rA)>;
1552 multiclass LoadRegister {
1553 def v2i64: LRVecInst<v2i64>;
1554 def v2f64: LRVecInst<v2f64>;
1555 def v4i32: LRVecInst<v4i32>;
1556 def v4f32: LRVecInst<v4f32>;
1557 def v8i16: LRVecInst<v8i16>;
1558 def v16i8: LRVecInst<v16i8>;
1560 def r128: LRRegInst<GPRC>;
1561 def r64: LRRegInst<R64C>;
1562 def f64: LRRegInst<R64FP>;
1563 def r32: LRRegInst<R32C>;
1564 def f32: LRRegInst<R32FP>;
1565 def r16: LRRegInst<R16C>;
1566 def r8: LRRegInst<R8C>;
1569 defm LR: LoadRegister;
1572 // ORC: Bitwise "or" with complement (c = a | ~b)
1574 class ORCInst<dag OOL, dag IOL, list<dag> pattern>:
1575 RRForm<0b10010010000, OOL, IOL, "orc\t$rT, $rA, $rB",
1576 IntegerOp, pattern>;
1578 class ORCVecInst<ValueType vectype>:
1579 ORCInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1580 [(set (vectype VECREG:$rT), (or (vectype VECREG:$rA),
1581 (vnot (vectype VECREG:$rB))))]>;
1583 class ORCRegInst<RegisterClass rclass>:
1584 ORCInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
1585 [(set rclass:$rT, (or rclass:$rA, (not rclass:$rB)))]>;
1587 multiclass BitwiseOrComplement
1589 def v16i8: ORCVecInst<v16i8>;
1590 def v8i16: ORCVecInst<v8i16>;
1591 def v4i32: ORCVecInst<v4i32>;
1592 def v2i64: ORCVecInst<v2i64>;
1594 def r64: ORCRegInst<R64C>;
1595 def r32: ORCRegInst<R32C>;
1596 def r16: ORCRegInst<R16C>;
1597 def r8: ORCRegInst<R8C>;
1600 defm ORC : BitwiseOrComplement;
1602 // OR byte immediate
1603 class ORBIInst<dag OOL, dag IOL, list<dag> pattern>:
1604 RI10Form<0b01100000, OOL, IOL, "orbi\t$rT, $rA, $val",
1605 IntegerOp, pattern>;
1607 class ORBIVecInst<ValueType vectype, PatLeaf immpred>:
1608 ORBIInst<(outs VECREG:$rT), (ins VECREG:$rA, u10imm:$val),
1609 [(set (v16i8 VECREG:$rT), (or (vectype VECREG:$rA),
1610 (vectype immpred:$val)))]>;
1612 multiclass BitwiseOrByteImm
1614 def v16i8: ORBIVecInst<v16i8, v16i8U8Imm>;
1616 def r8: ORBIInst<(outs R8C:$rT), (ins R8C:$rA, u10imm_i8:$val),
1617 [(set R8C:$rT, (or R8C:$rA, immU8:$val))]>;
1620 defm ORBI : BitwiseOrByteImm;
1622 // OR halfword immediate
1623 class ORHIInst<dag OOL, dag IOL, list<dag> pattern>:
1624 RI10Form<0b10100000, OOL, IOL, "orhi\t$rT, $rA, $val",
1625 IntegerOp, pattern>;
1627 class ORHIVecInst<ValueType vectype, PatLeaf immpred>:
1628 ORHIInst<(outs VECREG:$rT), (ins VECREG:$rA, u10imm:$val),
1629 [(set (vectype VECREG:$rT), (or (vectype VECREG:$rA),
1632 multiclass BitwiseOrHalfwordImm
1634 def v8i16: ORHIVecInst<v8i16, v8i16Uns10Imm>;
1636 def r16: ORHIInst<(outs R16C:$rT), (ins R16C:$rA, u10imm:$val),
1637 [(set R16C:$rT, (or R16C:$rA, i16ImmUns10:$val))]>;
1639 // Specialized ORHI form used to promote 8-bit registers to 16-bit
1640 def i8i16: ORHIInst<(outs R16C:$rT), (ins R8C:$rA, s10imm:$val),
1641 [(set R16C:$rT, (or (anyext R8C:$rA),
1642 i16ImmSExt10:$val))]>;
1645 defm ORHI : BitwiseOrHalfwordImm;
1647 class ORIInst<dag OOL, dag IOL, list<dag> pattern>:
1648 RI10Form<0b00100000, OOL, IOL, "ori\t$rT, $rA, $val",
1649 IntegerOp, pattern>;
1651 class ORIVecInst<ValueType vectype, PatLeaf immpred>:
1652 ORIInst<(outs VECREG:$rT), (ins VECREG:$rA, u10imm:$val),
1653 [(set (vectype VECREG:$rT), (or (vectype VECREG:$rA),
1656 // Bitwise "or" with immediate
1657 multiclass BitwiseOrImm
1659 def v4i32: ORIVecInst<v4i32, v4i32Uns10Imm>;
1661 def r32: ORIInst<(outs R32C:$rT), (ins R32C:$rA, u10imm_i32:$val),
1662 [(set R32C:$rT, (or R32C:$rA, i32ImmUns10:$val))]>;
1664 // i16i32: hacked version of the ori instruction to extend 16-bit quantities
1665 // to 32-bit quantities. used exclusively to match "anyext" conversions (vide
1666 // infra "anyext 16->32" pattern.)
1667 def i16i32: ORIInst<(outs R32C:$rT), (ins R16C:$rA, s10imm_i32:$val),
1668 [(set R32C:$rT, (or (anyext R16C:$rA),
1669 i32ImmSExt10:$val))]>;
1671 // i8i32: Hacked version of the ORI instruction to extend 16-bit quantities
1672 // to 32-bit quantities. Used exclusively to match "anyext" conversions (vide
1673 // infra "anyext 16->32" pattern.)
1674 def i8i32: ORIInst<(outs R32C:$rT), (ins R8C:$rA, s10imm_i32:$val),
1675 [(set R32C:$rT, (or (anyext R8C:$rA),
1676 i32ImmSExt10:$val))]>;
1679 defm ORI : BitwiseOrImm;
1681 // ORX: "or" across the vector: or's $rA's word slots leaving the result in
1682 // $rT[0], slots 1-3 are zeroed.
1684 // FIXME: Needs to match an intrinsic pattern.
1686 RRForm<0b10010010000, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1687 "orx\t$rT, $rA, $rB", IntegerOp,
1692 class XORInst<dag OOL, dag IOL, list<dag> pattern> :
1693 RRForm<0b10010010000, OOL, IOL, "xor\t$rT, $rA, $rB",
1694 IntegerOp, pattern>;
1696 class XORVecInst<ValueType vectype>:
1697 XORInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1698 [(set (vectype VECREG:$rT), (xor (vectype VECREG:$rA),
1699 (vectype VECREG:$rB)))]>;
1701 class XORRegInst<RegisterClass rclass>:
1702 XORInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
1703 [(set rclass:$rT, (xor rclass:$rA, rclass:$rB))]>;
1705 multiclass BitwiseExclusiveOr
1707 def v16i8: XORVecInst<v16i8>;
1708 def v8i16: XORVecInst<v8i16>;
1709 def v4i32: XORVecInst<v4i32>;
1710 def v2i64: XORVecInst<v2i64>;
1712 def r128: XORRegInst<GPRC>;
1713 def r64: XORRegInst<R64C>;
1714 def r32: XORRegInst<R32C>;
1715 def r16: XORRegInst<R16C>;
1716 def r8: XORRegInst<R8C>;
1718 // Special forms for floating point instructions.
1719 // fneg and fabs require bitwise logical ops to manipulate the sign bit.
1721 def fneg32: XORInst<(outs R32FP:$rT), (ins R32FP:$rA, R32C:$rB),
1722 [/* no pattern */]>;
1724 def fneg64: XORInst<(outs R64FP:$rT), (ins R64FP:$rA, VECREG:$rB),
1725 [/* no pattern */]>;
1727 def fnegvec: XORInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1728 [/* no pattern, see fneg{32,64} */]>;
1731 defm XOR : BitwiseExclusiveOr;
1733 //==----------------------------------------------------------
1735 class XORBIInst<dag OOL, dag IOL, list<dag> pattern>:
1736 RI10Form<0b01100000, OOL, IOL, "xorbi\t$rT, $rA, $val",
1737 IntegerOp, pattern>;
1739 multiclass XorByteImm
1742 XORBIInst<(outs VECREG:$rT), (ins VECREG:$rA, u10imm:$val),
1743 [(set (v16i8 VECREG:$rT), (xor (v16i8 VECREG:$rA), v16i8U8Imm:$val))]>;
1746 XORBIInst<(outs R8C:$rT), (ins R8C:$rA, u10imm_i8:$val),
1747 [(set R8C:$rT, (xor R8C:$rA, immU8:$val))]>;
1750 defm XORBI : XorByteImm;
1753 RI10Form<0b10100000, (outs VECREG:$rT), (ins VECREG:$rA, u10imm:$val),
1754 "xorhi\t$rT, $rA, $val", IntegerOp,
1755 [(set (v8i16 VECREG:$rT), (xor (v8i16 VECREG:$rA),
1756 v8i16SExt10Imm:$val))]>;
1759 RI10Form<0b10100000, (outs R16C:$rT), (ins R16C:$rA, s10imm:$val),
1760 "xorhi\t$rT, $rA, $val", IntegerOp,
1761 [(set R16C:$rT, (xor R16C:$rA, i16ImmSExt10:$val))]>;
1764 RI10Form<0b00100000, (outs VECREG:$rT), (ins VECREG:$rA, s10imm_i32:$val),
1765 "xori\t$rT, $rA, $val", IntegerOp,
1766 [(set (v4i32 VECREG:$rT), (xor (v4i32 VECREG:$rA),
1767 v4i32SExt10Imm:$val))]>;
1770 RI10Form<0b00100000, (outs R32C:$rT), (ins R32C:$rA, s10imm_i32:$val),
1771 "xori\t$rT, $rA, $val", IntegerOp,
1772 [(set R32C:$rT, (xor R32C:$rA, i32ImmSExt10:$val))]>;
1776 RRForm<0b10010010000, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1777 "nand\t$rT, $rA, $rB", IntegerOp,
1778 [(set (v16i8 VECREG:$rT), (vnot (and (v16i8 VECREG:$rA),
1779 (v16i8 VECREG:$rB))))]>;
1782 RRForm<0b10010010000, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1783 "nand\t$rT, $rA, $rB", IntegerOp,
1784 [(set (v8i16 VECREG:$rT), (vnot (and (v8i16 VECREG:$rA),
1785 (v8i16 VECREG:$rB))))]>;
1788 RRForm<0b10010010000, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1789 "nand\t$rT, $rA, $rB", IntegerOp,
1790 [(set (v4i32 VECREG:$rT), (vnot (and (v4i32 VECREG:$rA),
1791 (v4i32 VECREG:$rB))))]>;
1794 RRForm<0b10010010000, (outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
1795 "nand\t$rT, $rA, $rB", IntegerOp,
1796 [(set R32C:$rT, (not (and R32C:$rA, R32C:$rB)))]>;
1799 RRForm<0b10010010000, (outs R16C:$rT), (ins R16C:$rA, R16C:$rB),
1800 "nand\t$rT, $rA, $rB", IntegerOp,
1801 [(set R16C:$rT, (not (and R16C:$rA, R16C:$rB)))]>;
1804 RRForm<0b10010010000, (outs R8C:$rT), (ins R8C:$rA, R8C:$rB),
1805 "nand\t$rT, $rA, $rB", IntegerOp,
1806 [(set R8C:$rT, (not (and R8C:$rA, R8C:$rB)))]>;
1810 RRForm<0b10010010000, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1811 "nor\t$rT, $rA, $rB", IntegerOp,
1812 [(set (v16i8 VECREG:$rT), (vnot (or (v16i8 VECREG:$rA),
1813 (v16i8 VECREG:$rB))))]>;
1816 RRForm<0b10010010000, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1817 "nor\t$rT, $rA, $rB", IntegerOp,
1818 [(set (v8i16 VECREG:$rT), (vnot (or (v8i16 VECREG:$rA),
1819 (v8i16 VECREG:$rB))))]>;
1822 RRForm<0b10010010000, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1823 "nor\t$rT, $rA, $rB", IntegerOp,
1824 [(set (v4i32 VECREG:$rT), (vnot (or (v4i32 VECREG:$rA),
1825 (v4i32 VECREG:$rB))))]>;
1828 RRForm<0b10010010000, (outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
1829 "nor\t$rT, $rA, $rB", IntegerOp,
1830 [(set R32C:$rT, (not (or R32C:$rA, R32C:$rB)))]>;
1833 RRForm<0b10010010000, (outs R16C:$rT), (ins R16C:$rA, R16C:$rB),
1834 "nor\t$rT, $rA, $rB", IntegerOp,
1835 [(set R16C:$rT, (not (or R16C:$rA, R16C:$rB)))]>;
1838 RRForm<0b10010010000, (outs R8C:$rT), (ins R8C:$rA, R8C:$rB),
1839 "nor\t$rT, $rA, $rB", IntegerOp,
1840 [(set R8C:$rT, (not (or R8C:$rA, R8C:$rB)))]>;
1843 class SELBInst<dag OOL, dag IOL, list<dag> pattern>:
1844 RRRForm<0b1000, OOL, IOL, "selb\t$rT, $rA, $rB, $rC",
1845 IntegerOp, pattern>;
1847 class SELBVecInst<ValueType vectype>:
1848 SELBInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
1849 [(set (vectype VECREG:$rT),
1850 (or (and (vectype VECREG:$rC), (vectype VECREG:$rB)),
1851 (and (vnot (vectype VECREG:$rC)),
1852 (vectype VECREG:$rA))))]>;
1854 class SELBVecVCondInst<ValueType vectype>:
1855 SELBInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
1856 [(set (vectype VECREG:$rT),
1857 (select (vectype VECREG:$rC),
1858 (vectype VECREG:$rB),
1859 (vectype VECREG:$rA)))]>;
1861 class SELBVecCondInst<ValueType vectype>:
1862 SELBInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, R32C:$rC),
1863 [(set (vectype VECREG:$rT),
1865 (vectype VECREG:$rB),
1866 (vectype VECREG:$rA)))]>;
1868 class SELBRegInst<RegisterClass rclass>:
1869 SELBInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB, rclass:$rC),
1871 (or (and rclass:$rB, rclass:$rC),
1872 (and rclass:$rA, (not rclass:$rC))))]>;
1874 class SELBRegCondInst<RegisterClass rcond, RegisterClass rclass>:
1875 SELBInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB, rcond:$rC),
1877 (select rcond:$rC, rclass:$rB, rclass:$rA))]>;
1879 multiclass SelectBits
1881 def v16i8: SELBVecInst<v16i8>;
1882 def v8i16: SELBVecInst<v8i16>;
1883 def v4i32: SELBVecInst<v4i32>;
1884 def v2i64: SELBVecInst<v2i64>;
1886 def r128: SELBRegInst<GPRC>;
1887 def r64: SELBRegInst<R64C>;
1888 def r32: SELBRegInst<R32C>;
1889 def r16: SELBRegInst<R16C>;
1890 def r8: SELBRegInst<R8C>;
1892 def v16i8_cond: SELBVecCondInst<v16i8>;
1893 def v8i16_cond: SELBVecCondInst<v8i16>;
1894 def v4i32_cond: SELBVecCondInst<v4i32>;
1895 def v2i64_cond: SELBVecCondInst<v2i64>;
1897 def v16i8_vcond: SELBVecCondInst<v16i8>;
1898 def v8i16_vcond: SELBVecCondInst<v8i16>;
1899 def v4i32_vcond: SELBVecCondInst<v4i32>;
1900 def v2i64_vcond: SELBVecCondInst<v2i64>;
1903 SELBInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
1904 [(set (v4f32 VECREG:$rT),
1905 (select (v4i32 VECREG:$rC),
1907 (v4f32 VECREG:$rA)))]>;
1909 // SELBr64_cond is defined further down, look for i64 comparisons
1910 def r32_cond: SELBRegCondInst<R32C, R32C>;
1911 def f32_cond: SELBRegCondInst<R32C, R32FP>;
1912 def r16_cond: SELBRegCondInst<R16C, R16C>;
1913 def r8_cond: SELBRegCondInst<R8C, R8C>;
1916 defm SELB : SelectBits;
1918 class SPUselbPatVec<ValueType vectype, SPUInstr inst>:
1919 Pat<(SPUselb (vectype VECREG:$rA), (vectype VECREG:$rB), (vectype VECREG:$rC)),
1920 (inst VECREG:$rA, VECREG:$rB, VECREG:$rC)>;
1922 def : SPUselbPatVec<v16i8, SELBv16i8>;
1923 def : SPUselbPatVec<v8i16, SELBv8i16>;
1924 def : SPUselbPatVec<v4i32, SELBv4i32>;
1925 def : SPUselbPatVec<v2i64, SELBv2i64>;
1927 class SPUselbPatReg<RegisterClass rclass, SPUInstr inst>:
1928 Pat<(SPUselb rclass:$rA, rclass:$rB, rclass:$rC),
1929 (inst rclass:$rA, rclass:$rB, rclass:$rC)>;
1931 def : SPUselbPatReg<R8C, SELBr8>;
1932 def : SPUselbPatReg<R16C, SELBr16>;
1933 def : SPUselbPatReg<R32C, SELBr32>;
1934 def : SPUselbPatReg<R64C, SELBr64>;
1936 // EQV: Equivalence (1 for each same bit, otherwise 0)
1938 // Note: There are a lot of ways to match this bit operator and these patterns
1939 // attempt to be as exhaustive as possible.
1941 class EQVInst<dag OOL, dag IOL, list<dag> pattern>:
1942 RRForm<0b10010010000, OOL, IOL, "eqv\t$rT, $rA, $rB",
1943 IntegerOp, pattern>;
1945 class EQVVecInst<ValueType vectype>:
1946 EQVInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1947 [(set (vectype VECREG:$rT),
1948 (or (and (vectype VECREG:$rA), (vectype VECREG:$rB)),
1949 (and (vnot (vectype VECREG:$rA)),
1950 (vnot (vectype VECREG:$rB)))))]>;
1952 class EQVRegInst<RegisterClass rclass>:
1953 EQVInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
1954 [(set rclass:$rT, (or (and rclass:$rA, rclass:$rB),
1955 (and (not rclass:$rA), (not rclass:$rB))))]>;
1957 class EQVVecPattern1<ValueType vectype>:
1958 EQVInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1959 [(set (vectype VECREG:$rT),
1960 (xor (vectype VECREG:$rA), (vnot (vectype VECREG:$rB))))]>;
1962 class EQVRegPattern1<RegisterClass rclass>:
1963 EQVInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
1964 [(set rclass:$rT, (xor rclass:$rA, (not rclass:$rB)))]>;
1966 class EQVVecPattern2<ValueType vectype>:
1967 EQVInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1968 [(set (vectype VECREG:$rT),
1969 (or (and (vectype VECREG:$rA), (vectype VECREG:$rB)),
1970 (vnot (or (vectype VECREG:$rA), (vectype VECREG:$rB)))))]>;
1972 class EQVRegPattern2<RegisterClass rclass>:
1973 EQVInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
1975 (or (and rclass:$rA, rclass:$rB),
1976 (not (or rclass:$rA, rclass:$rB))))]>;
1978 class EQVVecPattern3<ValueType vectype>:
1979 EQVInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1980 [(set (vectype VECREG:$rT),
1981 (not (xor (vectype VECREG:$rA), (vectype VECREG:$rB))))]>;
1983 class EQVRegPattern3<RegisterClass rclass>:
1984 EQVInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
1985 [(set rclass:$rT, (not (xor rclass:$rA, rclass:$rB)))]>;
1987 multiclass BitEquivalence
1989 def v16i8: EQVVecInst<v16i8>;
1990 def v8i16: EQVVecInst<v8i16>;
1991 def v4i32: EQVVecInst<v4i32>;
1992 def v2i64: EQVVecInst<v2i64>;
1994 def v16i8_1: EQVVecPattern1<v16i8>;
1995 def v8i16_1: EQVVecPattern1<v8i16>;
1996 def v4i32_1: EQVVecPattern1<v4i32>;
1997 def v2i64_1: EQVVecPattern1<v2i64>;
1999 def v16i8_2: EQVVecPattern2<v16i8>;
2000 def v8i16_2: EQVVecPattern2<v8i16>;
2001 def v4i32_2: EQVVecPattern2<v4i32>;
2002 def v2i64_2: EQVVecPattern2<v2i64>;
2004 def v16i8_3: EQVVecPattern3<v16i8>;
2005 def v8i16_3: EQVVecPattern3<v8i16>;
2006 def v4i32_3: EQVVecPattern3<v4i32>;
2007 def v2i64_3: EQVVecPattern3<v2i64>;
2009 def r128: EQVRegInst<GPRC>;
2010 def r64: EQVRegInst<R64C>;
2011 def r32: EQVRegInst<R32C>;
2012 def r16: EQVRegInst<R16C>;
2013 def r8: EQVRegInst<R8C>;
2015 def r128_1: EQVRegPattern1<GPRC>;
2016 def r64_1: EQVRegPattern1<R64C>;
2017 def r32_1: EQVRegPattern1<R32C>;
2018 def r16_1: EQVRegPattern1<R16C>;
2019 def r8_1: EQVRegPattern1<R8C>;
2021 def r128_2: EQVRegPattern2<GPRC>;
2022 def r64_2: EQVRegPattern2<R64C>;
2023 def r32_2: EQVRegPattern2<R32C>;
2024 def r16_2: EQVRegPattern2<R16C>;
2025 def r8_2: EQVRegPattern2<R8C>;
2027 def r128_3: EQVRegPattern3<GPRC>;
2028 def r64_3: EQVRegPattern3<R64C>;
2029 def r32_3: EQVRegPattern3<R32C>;
2030 def r16_3: EQVRegPattern3<R16C>;
2031 def r8_3: EQVRegPattern3<R8C>;
2034 defm EQV: BitEquivalence;
2036 //===----------------------------------------------------------------------===//
2037 // Vector shuffle...
2038 //===----------------------------------------------------------------------===//
2039 // SPUshuffle is generated in LowerVECTOR_SHUFFLE and gets replaced with SHUFB.
2040 // See the SPUshuffle SDNode operand above, which sets up the DAG pattern
2041 // matcher to emit something when the LowerVECTOR_SHUFFLE generates a node with
2042 // the SPUISD::SHUFB opcode.
2043 //===----------------------------------------------------------------------===//
2045 class SHUFBInst<dag OOL, dag IOL, list<dag> pattern>:
2046 RRRForm<0b1000, OOL, IOL, "shufb\t$rT, $rA, $rB, $rC",
2047 IntegerOp, pattern>;
2049 class SHUFBVecInst<ValueType resultvec, ValueType maskvec>:
2050 SHUFBInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
2051 [(set (resultvec VECREG:$rT),
2052 (SPUshuffle (resultvec VECREG:$rA),
2053 (resultvec VECREG:$rB),
2054 (maskvec VECREG:$rC)))]>;
2056 class SHUFBGPRCInst:
2057 SHUFBInst<(outs VECREG:$rT), (ins GPRC:$rA, GPRC:$rB, VECREG:$rC),
2058 [/* no pattern */]>;
2060 multiclass ShuffleBytes
2062 def v16i8 : SHUFBVecInst<v16i8, v16i8>;
2063 def v16i8_m32 : SHUFBVecInst<v16i8, v4i32>;
2064 def v8i16 : SHUFBVecInst<v8i16, v16i8>;
2065 def v8i16_m32 : SHUFBVecInst<v8i16, v4i32>;
2066 def v4i32 : SHUFBVecInst<v4i32, v16i8>;
2067 def v4i32_m32 : SHUFBVecInst<v4i32, v4i32>;
2068 def v2i64 : SHUFBVecInst<v2i64, v16i8>;
2069 def v2i64_m32 : SHUFBVecInst<v2i64, v4i32>;
2071 def v4f32 : SHUFBVecInst<v4f32, v16i8>;
2072 def v4f32_m32 : SHUFBVecInst<v4f32, v4i32>;
2074 def v2f64 : SHUFBVecInst<v2f64, v16i8>;
2075 def v2f64_m32 : SHUFBVecInst<v2f64, v4i32>;
2077 def gprc : SHUFBGPRCInst;
2080 defm SHUFB : ShuffleBytes;
2082 //===----------------------------------------------------------------------===//
2083 // Shift and rotate group:
2084 //===----------------------------------------------------------------------===//
2086 class SHLHInst<dag OOL, dag IOL, list<dag> pattern>:
2087 RRForm<0b11111010000, OOL, IOL, "shlh\t$rT, $rA, $rB",
2088 RotateShift, pattern>;
2090 class SHLHVecInst<ValueType vectype>:
2091 SHLHInst<(outs VECREG:$rT), (ins VECREG:$rA, R16C:$rB),
2092 [(set (vectype VECREG:$rT),
2093 (SPUvec_shl (vectype VECREG:$rA), R16C:$rB))]>;
2095 // $rB gets promoted to 32-bit register type when confronted with
2096 // this llvm assembly code:
2098 // define i16 @shlh_i16_1(i16 %arg1, i16 %arg2) {
2099 // %A = shl i16 %arg1, %arg2
2103 multiclass ShiftLeftHalfword
2105 def v8i16: SHLHVecInst<v8i16>;
2106 def r16: SHLHInst<(outs R16C:$rT), (ins R16C:$rA, R16C:$rB),
2107 [(set R16C:$rT, (shl R16C:$rA, R16C:$rB))]>;
2108 def r16_r32: SHLHInst<(outs R16C:$rT), (ins R16C:$rA, R32C:$rB),
2109 [(set R16C:$rT, (shl R16C:$rA, R32C:$rB))]>;
2112 defm SHLH : ShiftLeftHalfword;
2114 //===----------------------------------------------------------------------===//
2116 class SHLHIInst<dag OOL, dag IOL, list<dag> pattern>:
2117 RI7Form<0b11111010000, OOL, IOL, "shlhi\t$rT, $rA, $val",
2118 RotateShift, pattern>;
2120 class SHLHIVecInst<ValueType vectype>:
2121 SHLHIInst<(outs VECREG:$rT), (ins VECREG:$rA, u7imm:$val),
2122 [(set (vectype VECREG:$rT),
2123 (SPUvec_shl (vectype VECREG:$rA), (i16 uimm7:$val)))]>;
2125 multiclass ShiftLeftHalfwordImm
2127 def v8i16: SHLHIVecInst<v8i16>;
2128 def r16: SHLHIInst<(outs R16C:$rT), (ins R16C:$rA, u7imm:$val),
2129 [(set R16C:$rT, (shl R16C:$rA, (i16 uimm7:$val)))]>;
2132 defm SHLHI : ShiftLeftHalfwordImm;
2134 def : Pat<(SPUvec_shl (v8i16 VECREG:$rA), (i32 uimm7:$val)),
2135 (SHLHIv8i16 VECREG:$rA, uimm7:$val)>;
2137 def : Pat<(shl R16C:$rA, (i32 uimm7:$val)),
2138 (SHLHIr16 R16C:$rA, uimm7:$val)>;
2140 //===----------------------------------------------------------------------===//
2142 class SHLInst<dag OOL, dag IOL, list<dag> pattern>:
2143 RRForm<0b11111010000, OOL, IOL, "shl\t$rT, $rA, $rB",
2144 RotateShift, pattern>;
2146 multiclass ShiftLeftWord
2149 SHLInst<(outs VECREG:$rT), (ins VECREG:$rA, R16C:$rB),
2150 [(set (v4i32 VECREG:$rT),
2151 (SPUvec_shl (v4i32 VECREG:$rA), R16C:$rB))]>;
2153 SHLInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
2154 [(set R32C:$rT, (shl R32C:$rA, R32C:$rB))]>;
2157 defm SHL: ShiftLeftWord;
2159 //===----------------------------------------------------------------------===//
2161 class SHLIInst<dag OOL, dag IOL, list<dag> pattern>:
2162 RI7Form<0b11111010000, OOL, IOL, "shli\t$rT, $rA, $val",
2163 RotateShift, pattern>;
2165 multiclass ShiftLeftWordImm
2168 SHLIInst<(outs VECREG:$rT), (ins VECREG:$rA, u7imm_i32:$val),
2169 [(set (v4i32 VECREG:$rT),
2170 (SPUvec_shl (v4i32 VECREG:$rA), (i32 uimm7:$val)))]>;
2173 SHLIInst<(outs R32C:$rT), (ins R32C:$rA, u7imm_i32:$val),
2174 [(set R32C:$rT, (shl R32C:$rA, (i32 uimm7:$val)))]>;
2177 defm SHLI : ShiftLeftWordImm;
2179 //===----------------------------------------------------------------------===//
2180 // SHLQBI vec form: Note that this will shift the entire vector (the 128-bit
2181 // register) to the left. Vector form is here to ensure type correctness.
2183 // The shift count is in the lowest 3 bits (29-31) of $rB, so only a bit shift
2184 // of 7 bits is actually possible.
2186 // Note also that SHLQBI/SHLQBII are used in conjunction with SHLQBY/SHLQBYI
2187 // to shift i64 and i128. SHLQBI is the residual left over after shifting by
2188 // bytes with SHLQBY.
2190 class SHLQBIInst<dag OOL, dag IOL, list<dag> pattern>:
2191 RRForm<0b11011011100, OOL, IOL, "shlqbi\t$rT, $rA, $rB",
2192 RotateShift, pattern>;
2194 class SHLQBIVecInst<ValueType vectype>:
2195 SHLQBIInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2196 [(set (vectype VECREG:$rT),
2197 (SPUshlquad_l_bits (vectype VECREG:$rA), R32C:$rB))]>;
2199 multiclass ShiftLeftQuadByBits
2201 def v16i8: SHLQBIVecInst<v16i8>;
2202 def v8i16: SHLQBIVecInst<v8i16>;
2203 def v4i32: SHLQBIVecInst<v4i32>;
2204 def v4f32: SHLQBIVecInst<v4f32>;
2205 def v2i64: SHLQBIVecInst<v2i64>;
2206 def v2f64: SHLQBIVecInst<v2f64>;
2209 defm SHLQBI : ShiftLeftQuadByBits;
2211 // See note above on SHLQBI. In this case, the predicate actually does then
2212 // enforcement, whereas with SHLQBI, we have to "take it on faith."
2213 class SHLQBIIInst<dag OOL, dag IOL, list<dag> pattern>:
2214 RI7Form<0b11011111100, OOL, IOL, "shlqbii\t$rT, $rA, $val",
2215 RotateShift, pattern>;
2217 class SHLQBIIVecInst<ValueType vectype>:
2218 SHLQBIIInst<(outs VECREG:$rT), (ins VECREG:$rA, u7imm_i32:$val),
2219 [(set (vectype VECREG:$rT),
2220 (SPUshlquad_l_bits (vectype VECREG:$rA), (i32 bitshift:$val)))]>;
2222 multiclass ShiftLeftQuadByBitsImm
2224 def v16i8 : SHLQBIIVecInst<v16i8>;
2225 def v8i16 : SHLQBIIVecInst<v8i16>;
2226 def v4i32 : SHLQBIIVecInst<v4i32>;
2227 def v4f32 : SHLQBIIVecInst<v4f32>;
2228 def v2i64 : SHLQBIIVecInst<v2i64>;
2229 def v2f64 : SHLQBIIVecInst<v2f64>;
2232 defm SHLQBII : ShiftLeftQuadByBitsImm;
2234 // SHLQBY, SHLQBYI vector forms: Shift the entire vector to the left by bytes,
2235 // not by bits. See notes above on SHLQBI.
2237 class SHLQBYInst<dag OOL, dag IOL, list<dag> pattern>:
2238 RI7Form<0b11111011100, OOL, IOL, "shlqby\t$rT, $rA, $rB",
2239 RotateShift, pattern>;
2241 class SHLQBYVecInst<ValueType vectype>:
2242 SHLQBYInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2243 [(set (vectype VECREG:$rT),
2244 (SPUshlquad_l_bytes (vectype VECREG:$rA), R32C:$rB))]>;
2246 multiclass ShiftLeftQuadBytes
2248 def v16i8: SHLQBYVecInst<v16i8>;
2249 def v8i16: SHLQBYVecInst<v8i16>;
2250 def v4i32: SHLQBYVecInst<v4i32>;
2251 def v4f32: SHLQBYVecInst<v4f32>;
2252 def v2i64: SHLQBYVecInst<v2i64>;
2253 def v2f64: SHLQBYVecInst<v2f64>;
2254 def r128: SHLQBYInst<(outs GPRC:$rT), (ins GPRC:$rA, R32C:$rB),
2255 [(set GPRC:$rT, (SPUshlquad_l_bytes GPRC:$rA, R32C:$rB))]>;
2258 defm SHLQBY: ShiftLeftQuadBytes;
2260 class SHLQBYIInst<dag OOL, dag IOL, list<dag> pattern>:
2261 RI7Form<0b11111111100, OOL, IOL, "shlqbyi\t$rT, $rA, $val",
2262 RotateShift, pattern>;
2264 class SHLQBYIVecInst<ValueType vectype>:
2265 SHLQBYIInst<(outs VECREG:$rT), (ins VECREG:$rA, u7imm_i32:$val),
2266 [(set (vectype VECREG:$rT),
2267 (SPUshlquad_l_bytes (vectype VECREG:$rA), (i32 uimm7:$val)))]>;
2269 multiclass ShiftLeftQuadBytesImm
2271 def v16i8: SHLQBYIVecInst<v16i8>;
2272 def v8i16: SHLQBYIVecInst<v8i16>;
2273 def v4i32: SHLQBYIVecInst<v4i32>;
2274 def v4f32: SHLQBYIVecInst<v4f32>;
2275 def v2i64: SHLQBYIVecInst<v2i64>;
2276 def v2f64: SHLQBYIVecInst<v2f64>;
2277 def r128: SHLQBYIInst<(outs GPRC:$rT), (ins GPRC:$rA, u7imm_i32:$val),
2279 (SPUshlquad_l_bytes GPRC:$rA, (i32 uimm7:$val)))]>;
2282 defm SHLQBYI : ShiftLeftQuadBytesImm;
2284 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2286 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2287 class ROTHInst<dag OOL, dag IOL, list<dag> pattern>:
2288 RRForm<0b00111010000, OOL, IOL, "roth\t$rT, $rA, $rB",
2289 RotateShift, pattern>;
2291 class ROTHVecInst<ValueType vectype>:
2292 ROTHInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
2293 [(set (vectype VECREG:$rT),
2294 (SPUvec_rotl VECREG:$rA, VECREG:$rB))]>;
2296 class ROTHRegInst<RegisterClass rclass>:
2297 ROTHInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
2298 [(set rclass:$rT, (rotl rclass:$rA, rclass:$rB))]>;
2300 multiclass RotateLeftHalfword
2302 def v8i16: ROTHVecInst<v8i16>;
2303 def r16: ROTHRegInst<R16C>;
2306 defm ROTH: RotateLeftHalfword;
2308 def ROTHr16_r32: ROTHInst<(outs R16C:$rT), (ins R16C:$rA, R32C:$rB),
2309 [(set R16C:$rT, (rotl R16C:$rA, R32C:$rB))]>;
2311 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2312 // Rotate halfword, immediate:
2313 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2314 class ROTHIInst<dag OOL, dag IOL, list<dag> pattern>:
2315 RI7Form<0b00111110000, OOL, IOL, "rothi\t$rT, $rA, $val",
2316 RotateShift, pattern>;
2318 class ROTHIVecInst<ValueType vectype>:
2319 ROTHIInst<(outs VECREG:$rT), (ins VECREG:$rA, u7imm:$val),
2320 [(set (vectype VECREG:$rT),
2321 (SPUvec_rotl VECREG:$rA, (i16 uimm7:$val)))]>;
2323 multiclass RotateLeftHalfwordImm
2325 def v8i16: ROTHIVecInst<v8i16>;
2326 def r16: ROTHIInst<(outs R16C:$rT), (ins R16C:$rA, u7imm:$val),
2327 [(set R16C:$rT, (rotl R16C:$rA, (i16 uimm7:$val)))]>;
2328 def r16_r32: ROTHIInst<(outs R16C:$rT), (ins R16C:$rA, u7imm_i32:$val),
2329 [(set R16C:$rT, (rotl R16C:$rA, (i32 uimm7:$val)))]>;
2332 defm ROTHI: RotateLeftHalfwordImm;
2334 def : Pat<(SPUvec_rotl VECREG:$rA, (i32 uimm7:$val)),
2335 (ROTHIv8i16 VECREG:$rA, imm:$val)>;
2337 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2339 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2341 class ROTInst<dag OOL, dag IOL, list<dag> pattern>:
2342 RRForm<0b00011010000, OOL, IOL, "rot\t$rT, $rA, $rB",
2343 RotateShift, pattern>;
2345 class ROTVecInst<ValueType vectype>:
2346 ROTInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2347 [(set (vectype VECREG:$rT),
2348 (SPUvec_rotl (vectype VECREG:$rA), R32C:$rB))]>;
2350 class ROTRegInst<RegisterClass rclass>:
2351 ROTInst<(outs rclass:$rT), (ins rclass:$rA, R32C:$rB),
2353 (rotl rclass:$rA, R32C:$rB))]>;
2355 multiclass RotateLeftWord
2357 def v4i32: ROTVecInst<v4i32>;
2358 def r32: ROTRegInst<R32C>;
2361 defm ROT: RotateLeftWord;
2363 // The rotate amount is in the same bits whether we've got an 8-bit, 16-bit or
2365 def ROTr32_r16_anyext:
2366 ROTInst<(outs R32C:$rT), (ins R32C:$rA, R16C:$rB),
2367 [(set R32C:$rT, (rotl R32C:$rA, (i32 (anyext R16C:$rB))))]>;
2369 def : Pat<(rotl R32C:$rA, (i32 (zext R16C:$rB))),
2370 (ROTr32_r16_anyext R32C:$rA, R16C:$rB)>;
2372 def : Pat<(rotl R32C:$rA, (i32 (sext R16C:$rB))),
2373 (ROTr32_r16_anyext R32C:$rA, R16C:$rB)>;
2375 def ROTr32_r8_anyext:
2376 ROTInst<(outs R32C:$rT), (ins R32C:$rA, R8C:$rB),
2377 [(set R32C:$rT, (rotl R32C:$rA, (i32 (anyext R8C:$rB))))]>;
2379 def : Pat<(rotl R32C:$rA, (i32 (zext R8C:$rB))),
2380 (ROTr32_r8_anyext R32C:$rA, R8C:$rB)>;
2382 def : Pat<(rotl R32C:$rA, (i32 (sext R8C:$rB))),
2383 (ROTr32_r8_anyext R32C:$rA, R8C:$rB)>;
2385 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2386 // Rotate word, immediate
2387 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2389 class ROTIInst<dag OOL, dag IOL, list<dag> pattern>:
2390 RI7Form<0b00011110000, OOL, IOL, "roti\t$rT, $rA, $val",
2391 RotateShift, pattern>;
2393 class ROTIVecInst<ValueType vectype, Operand optype, ValueType inttype, PatLeaf pred>:
2394 ROTIInst<(outs VECREG:$rT), (ins VECREG:$rA, optype:$val),
2395 [(set (vectype VECREG:$rT),
2396 (SPUvec_rotl (vectype VECREG:$rA), (inttype pred:$val)))]>;
2398 class ROTIRegInst<RegisterClass rclass, Operand optype, ValueType inttype, PatLeaf pred>:
2399 ROTIInst<(outs rclass:$rT), (ins rclass:$rA, optype:$val),
2400 [(set rclass:$rT, (rotl rclass:$rA, (inttype pred:$val)))]>;
2402 multiclass RotateLeftWordImm
2404 def v4i32: ROTIVecInst<v4i32, u7imm_i32, i32, uimm7>;
2405 def v4i32_i16: ROTIVecInst<v4i32, u7imm, i16, uimm7>;
2406 def v4i32_i8: ROTIVecInst<v4i32, u7imm_i8, i8, uimm7>;
2408 def r32: ROTIRegInst<R32C, u7imm_i32, i32, uimm7>;
2409 def r32_i16: ROTIRegInst<R32C, u7imm, i16, uimm7>;
2410 def r32_i8: ROTIRegInst<R32C, u7imm_i8, i8, uimm7>;
2413 defm ROTI : RotateLeftWordImm;
2415 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2416 // Rotate quad by byte (count)
2417 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2419 class ROTQBYInst<dag OOL, dag IOL, list<dag> pattern>:
2420 RRForm<0b00111011100, OOL, IOL, "rotqby\t$rT, $rA, $rB",
2421 RotateShift, pattern>;
2423 class ROTQBYVecInst<ValueType vectype>:
2424 ROTQBYInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2425 [(set (vectype VECREG:$rT),
2426 (SPUrotbytes_left (vectype VECREG:$rA), R32C:$rB))]>;
2428 multiclass RotateQuadLeftByBytes
2430 def v16i8: ROTQBYVecInst<v16i8>;
2431 def v8i16: ROTQBYVecInst<v8i16>;
2432 def v4i32: ROTQBYVecInst<v4i32>;
2433 def v4f32: ROTQBYVecInst<v4f32>;
2434 def v2i64: ROTQBYVecInst<v2i64>;
2435 def v2f64: ROTQBYVecInst<v2f64>;
2438 defm ROTQBY: RotateQuadLeftByBytes;
2440 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2441 // Rotate quad by byte (count), immediate
2442 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2444 class ROTQBYIInst<dag OOL, dag IOL, list<dag> pattern>:
2445 RI7Form<0b00111111100, OOL, IOL, "rotqbyi\t$rT, $rA, $val",
2446 RotateShift, pattern>;
2448 class ROTQBYIVecInst<ValueType vectype>:
2449 ROTQBYIInst<(outs VECREG:$rT), (ins VECREG:$rA, u7imm:$val),
2450 [(set (vectype VECREG:$rT),
2451 (SPUrotbytes_left (vectype VECREG:$rA), (i16 uimm7:$val)))]>;
2453 multiclass RotateQuadByBytesImm
2455 def v16i8: ROTQBYIVecInst<v16i8>;
2456 def v8i16: ROTQBYIVecInst<v8i16>;
2457 def v4i32: ROTQBYIVecInst<v4i32>;
2458 def v4f32: ROTQBYIVecInst<v4f32>;
2459 def v2i64: ROTQBYIVecInst<v2i64>;
2460 def vfi64: ROTQBYIVecInst<v2f64>;
2463 defm ROTQBYI: RotateQuadByBytesImm;
2465 // See ROTQBY note above.
2466 class ROTQBYBIInst<dag OOL, dag IOL, list<dag> pattern>:
2467 RI7Form<0b00110011100, OOL, IOL,
2468 "rotqbybi\t$rT, $rA, $shift",
2469 RotateShift, pattern>;
2471 class ROTQBYBIVecInst<ValueType vectype, RegisterClass rclass>:
2472 ROTQBYBIInst<(outs VECREG:$rT), (ins VECREG:$rA, rclass:$shift),
2473 [(set (vectype VECREG:$rT),
2474 (SPUrotbytes_left_bits (vectype VECREG:$rA), rclass:$shift))]>;
2476 multiclass RotateQuadByBytesByBitshift {
2477 def v16i8_r32: ROTQBYBIVecInst<v16i8, R32C>;
2478 def v8i16_r32: ROTQBYBIVecInst<v8i16, R32C>;
2479 def v4i32_r32: ROTQBYBIVecInst<v4i32, R32C>;
2480 def v2i64_r32: ROTQBYBIVecInst<v2i64, R32C>;
2483 defm ROTQBYBI : RotateQuadByBytesByBitshift;
2485 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2486 // See ROTQBY note above.
2488 // Assume that the user of this instruction knows to shift the rotate count
2490 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2492 class ROTQBIInst<dag OOL, dag IOL, list<dag> pattern>:
2493 RRForm<0b00011011100, OOL, IOL, "rotqbi\t$rT, $rA, $rB",
2494 RotateShift, pattern>;
2496 class ROTQBIVecInst<ValueType vectype>:
2497 ROTQBIInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2498 [/* no pattern yet */]>;
2500 class ROTQBIRegInst<RegisterClass rclass>:
2501 ROTQBIInst<(outs rclass:$rT), (ins rclass:$rA, R32C:$rB),
2502 [/* no pattern yet */]>;
2504 multiclass RotateQuadByBitCount
2506 def v16i8: ROTQBIVecInst<v16i8>;
2507 def v8i16: ROTQBIVecInst<v8i16>;
2508 def v4i32: ROTQBIVecInst<v4i32>;
2509 def v2i64: ROTQBIVecInst<v2i64>;
2511 def r128: ROTQBIRegInst<GPRC>;
2512 def r64: ROTQBIRegInst<R64C>;
2515 defm ROTQBI: RotateQuadByBitCount;
2517 class ROTQBIIInst<dag OOL, dag IOL, list<dag> pattern>:
2518 RI7Form<0b00011111100, OOL, IOL, "rotqbii\t$rT, $rA, $val",
2519 RotateShift, pattern>;
2521 class ROTQBIIVecInst<ValueType vectype, Operand optype, ValueType inttype,
2523 ROTQBIIInst<(outs VECREG:$rT), (ins VECREG:$rA, optype:$val),
2524 [/* no pattern yet */]>;
2526 class ROTQBIIRegInst<RegisterClass rclass, Operand optype, ValueType inttype,
2528 ROTQBIIInst<(outs rclass:$rT), (ins rclass:$rA, optype:$val),
2529 [/* no pattern yet */]>;
2531 multiclass RotateQuadByBitCountImm
2533 def v16i8: ROTQBIIVecInst<v16i8, u7imm_i32, i32, uimm7>;
2534 def v8i16: ROTQBIIVecInst<v8i16, u7imm_i32, i32, uimm7>;
2535 def v4i32: ROTQBIIVecInst<v4i32, u7imm_i32, i32, uimm7>;
2536 def v2i64: ROTQBIIVecInst<v2i64, u7imm_i32, i32, uimm7>;
2538 def r128: ROTQBIIRegInst<GPRC, u7imm_i32, i32, uimm7>;
2539 def r64: ROTQBIIRegInst<R64C, u7imm_i32, i32, uimm7>;
2542 defm ROTQBII : RotateQuadByBitCountImm;
2544 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2545 // ROTHM v8i16 form:
2546 // NOTE(1): No vector rotate is generated by the C/C++ frontend (today),
2547 // so this only matches a synthetically generated/lowered code
2549 // NOTE(2): $rB must be negated before the right rotate!
2550 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2552 class ROTHMInst<dag OOL, dag IOL, list<dag> pattern>:
2553 RRForm<0b10111010000, OOL, IOL, "rothm\t$rT, $rA, $rB",
2554 RotateShift, pattern>;
2557 ROTHMInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2558 [/* see patterns below - $rB must be negated */]>;
2560 def : Pat<(SPUvec_srl (v8i16 VECREG:$rA), R32C:$rB),
2561 (ROTHMv8i16 VECREG:$rA, (SFIr32 R32C:$rB, 0))>;
2563 def : Pat<(SPUvec_srl (v8i16 VECREG:$rA), R16C:$rB),
2564 (ROTHMv8i16 VECREG:$rA,
2565 (SFIr32 (XSHWr16 R16C:$rB), 0))>;
2567 def : Pat<(SPUvec_srl (v8i16 VECREG:$rA), R8C:$rB),
2568 (ROTHMv8i16 VECREG:$rA,
2569 (SFIr32 (XSHWr16 (XSBHr8 R8C:$rB) ), 0))>;
2571 // ROTHM r16 form: Rotate 16-bit quantity to right, zero fill at the left
2572 // Note: This instruction doesn't match a pattern because rB must be negated
2573 // for the instruction to work. Thus, the pattern below the instruction!
2576 ROTHMInst<(outs R16C:$rT), (ins R16C:$rA, R32C:$rB),
2577 [/* see patterns below - $rB must be negated! */]>;
2579 def : Pat<(srl R16C:$rA, R32C:$rB),
2580 (ROTHMr16 R16C:$rA, (SFIr32 R32C:$rB, 0))>;
2582 def : Pat<(srl R16C:$rA, R16C:$rB),
2584 (SFIr32 (XSHWr16 R16C:$rB), 0))>;
2586 def : Pat<(srl R16C:$rA, R8C:$rB),
2588 (SFIr32 (XSHWr16 (XSBHr8 R8C:$rB) ), 0))>;
2590 // ROTHMI v8i16 form: See the comment for ROTHM v8i16. The difference here is
2591 // that the immediate can be complemented, so that the user doesn't have to
2594 class ROTHMIInst<dag OOL, dag IOL, list<dag> pattern>:
2595 RI7Form<0b10111110000, OOL, IOL, "rothmi\t$rT, $rA, $val",
2596 RotateShift, pattern>;
2599 ROTHMIInst<(outs VECREG:$rT), (ins VECREG:$rA, rothNeg7imm:$val),
2600 [/* no pattern */]>;
2602 def : Pat<(SPUvec_srl (v8i16 VECREG:$rA), (i32 imm:$val)),
2603 (ROTHMIv8i16 VECREG:$rA, imm:$val)>;
2605 def: Pat<(SPUvec_srl (v8i16 VECREG:$rA), (i16 imm:$val)),
2606 (ROTHMIv8i16 VECREG:$rA, imm:$val)>;
2608 def: Pat<(SPUvec_srl (v8i16 VECREG:$rA), (i8 imm:$val)),
2609 (ROTHMIv8i16 VECREG:$rA, imm:$val)>;
2612 ROTHMIInst<(outs R16C:$rT), (ins R16C:$rA, rothNeg7imm:$val),
2613 [/* no pattern */]>;
2615 def: Pat<(srl R16C:$rA, (i32 uimm7:$val)),
2616 (ROTHMIr16 R16C:$rA, uimm7:$val)>;
2618 def: Pat<(srl R16C:$rA, (i16 uimm7:$val)),
2619 (ROTHMIr16 R16C:$rA, uimm7:$val)>;
2621 def: Pat<(srl R16C:$rA, (i8 uimm7:$val)),
2622 (ROTHMIr16 R16C:$rA, uimm7:$val)>;
2624 // ROTM v4i32 form: See the ROTHM v8i16 comments.
2625 class ROTMInst<dag OOL, dag IOL, list<dag> pattern>:
2626 RRForm<0b10011010000, OOL, IOL, "rotm\t$rT, $rA, $rB",
2627 RotateShift, pattern>;
2630 ROTMInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2631 [/* see patterns below - $rB must be negated */]>;
2633 def : Pat<(SPUvec_srl VECREG:$rA, R32C:$rB),
2634 (ROTMv4i32 VECREG:$rA, (SFIr32 R32C:$rB, 0))>;
2636 def : Pat<(SPUvec_srl VECREG:$rA, R16C:$rB),
2637 (ROTMv4i32 VECREG:$rA,
2638 (SFIr32 (XSHWr16 R16C:$rB), 0))>;
2640 def : Pat<(SPUvec_srl VECREG:$rA, R8C:$rB),
2641 (ROTMv4i32 VECREG:$rA,
2642 (SFIr32 (XSHWr16 (XSBHr8 R8C:$rB)), 0))>;
2645 ROTMInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
2646 [/* see patterns below - $rB must be negated */]>;
2648 def : Pat<(srl R32C:$rA, R32C:$rB),
2649 (ROTMr32 R32C:$rA, (SFIr32 R32C:$rB, 0))>;
2651 def : Pat<(srl R32C:$rA, R16C:$rB),
2653 (SFIr32 (XSHWr16 R16C:$rB), 0))>;
2655 def : Pat<(srl R32C:$rA, R8C:$rB),
2657 (SFIr32 (XSHWr16 (XSBHr8 R8C:$rB)), 0))>;
2659 // ROTMI v4i32 form: See the comment for ROTHM v8i16.
2661 RI7Form<0b10011110000, (outs VECREG:$rT), (ins VECREG:$rA, rotNeg7imm:$val),
2662 "rotmi\t$rT, $rA, $val", RotateShift,
2663 [(set (v4i32 VECREG:$rT),
2664 (SPUvec_srl VECREG:$rA, (i32 uimm7:$val)))]>;
2666 def : Pat<(SPUvec_srl VECREG:$rA, (i16 uimm7:$val)),
2667 (ROTMIv4i32 VECREG:$rA, uimm7:$val)>;
2669 def : Pat<(SPUvec_srl VECREG:$rA, (i8 uimm7:$val)),
2670 (ROTMIv4i32 VECREG:$rA, uimm7:$val)>;
2672 // ROTMI r32 form: know how to complement the immediate value.
2674 RI7Form<0b10011110000, (outs R32C:$rT), (ins R32C:$rA, rotNeg7imm:$val),
2675 "rotmi\t$rT, $rA, $val", RotateShift,
2676 [(set R32C:$rT, (srl R32C:$rA, (i32 uimm7:$val)))]>;
2678 def : Pat<(srl R32C:$rA, (i16 imm:$val)),
2679 (ROTMIr32 R32C:$rA, uimm7:$val)>;
2681 def : Pat<(srl R32C:$rA, (i8 imm:$val)),
2682 (ROTMIr32 R32C:$rA, uimm7:$val)>;
2684 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2685 // ROTQMBYvec: This is a vector form merely so that when used in an
2686 // instruction pattern, type checking will succeed. This instruction assumes
2687 // that the user knew to negate $rB.
2688 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2690 class ROTQMBYInst<dag OOL, dag IOL, list<dag> pattern>:
2691 RRForm<0b10111011100, OOL, IOL, "rotqmby\t$rT, $rA, $rB",
2692 RotateShift, pattern>;
2694 class ROTQMBYVecInst<ValueType vectype>:
2695 ROTQMBYInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2696 [/* no pattern, $rB must be negated */]>;
2698 class ROTQMBYRegInst<RegisterClass rclass>:
2699 ROTQMBYInst<(outs rclass:$rT), (ins rclass:$rA, R32C:$rB),
2700 [/* no pattern */]>;
2702 multiclass RotateQuadBytes
2704 def v16i8: ROTQMBYVecInst<v16i8>;
2705 def v8i16: ROTQMBYVecInst<v8i16>;
2706 def v4i32: ROTQMBYVecInst<v4i32>;
2707 def v2i64: ROTQMBYVecInst<v2i64>;
2709 def r128: ROTQMBYRegInst<GPRC>;
2710 def r64: ROTQMBYRegInst<R64C>;
2713 defm ROTQMBY : RotateQuadBytes;
2715 class ROTQMBYIInst<dag OOL, dag IOL, list<dag> pattern>:
2716 RI7Form<0b10111111100, OOL, IOL, "rotqmbyi\t$rT, $rA, $val",
2717 RotateShift, pattern>;
2719 class ROTQMBYIVecInst<ValueType vectype>:
2720 ROTQMBYIInst<(outs VECREG:$rT), (ins VECREG:$rA, rotNeg7imm:$val),
2721 [/* no pattern */]>;
2723 class ROTQMBYIRegInst<RegisterClass rclass, Operand optype, ValueType inttype, PatLeaf pred>:
2724 ROTQMBYIInst<(outs rclass:$rT), (ins rclass:$rA, optype:$val),
2725 [/* no pattern */]>;
2727 multiclass RotateQuadBytesImm
2729 def v16i8: ROTQMBYIVecInst<v16i8>;
2730 def v8i16: ROTQMBYIVecInst<v8i16>;
2731 def v4i32: ROTQMBYIVecInst<v4i32>;
2732 def v2i64: ROTQMBYIVecInst<v2i64>;
2734 def r128: ROTQMBYIRegInst<GPRC, rotNeg7imm, i32, uimm7>;
2735 def r64: ROTQMBYIRegInst<R64C, rotNeg7imm, i32, uimm7>;
2738 defm ROTQMBYI : RotateQuadBytesImm;
2740 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2741 // Rotate right and mask by bit count
2742 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2744 class ROTQMBYBIInst<dag OOL, dag IOL, list<dag> pattern>:
2745 RRForm<0b10110011100, OOL, IOL, "rotqmbybi\t$rT, $rA, $rB",
2746 RotateShift, pattern>;
2748 class ROTQMBYBIVecInst<ValueType vectype>:
2749 ROTQMBYBIInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2750 [/* no pattern, */]>;
2752 multiclass RotateMaskQuadByBitCount
2754 def v16i8: ROTQMBYBIVecInst<v16i8>;
2755 def v8i16: ROTQMBYBIVecInst<v8i16>;
2756 def v4i32: ROTQMBYBIVecInst<v4i32>;
2757 def v2i64: ROTQMBYBIVecInst<v2i64>;
2760 defm ROTQMBYBI: RotateMaskQuadByBitCount;
2762 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2763 // Rotate quad and mask by bits
2764 // Note that the rotate amount has to be negated
2765 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2767 class ROTQMBIInst<dag OOL, dag IOL, list<dag> pattern>:
2768 RRForm<0b10011011100, OOL, IOL, "rotqmbi\t$rT, $rA, $rB",
2769 RotateShift, pattern>;
2771 class ROTQMBIVecInst<ValueType vectype>:
2772 ROTQMBIInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2773 [/* no pattern */]>;
2775 class ROTQMBIRegInst<RegisterClass rclass>:
2776 ROTQMBIInst<(outs rclass:$rT), (ins rclass:$rA, R32C:$rB),
2777 [/* no pattern */]>;
2779 multiclass RotateMaskQuadByBits
2781 def v16i8: ROTQMBIVecInst<v16i8>;
2782 def v8i16: ROTQMBIVecInst<v8i16>;
2783 def v4i32: ROTQMBIVecInst<v4i32>;
2784 def v2i64: ROTQMBIVecInst<v2i64>;
2786 def r128: ROTQMBIRegInst<GPRC>;
2787 def r64: ROTQMBIRegInst<R64C>;
2790 defm ROTQMBI: RotateMaskQuadByBits;
2792 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2793 // Rotate quad and mask by bits, immediate
2794 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2796 class ROTQMBIIInst<dag OOL, dag IOL, list<dag> pattern>:
2797 RI7Form<0b10011111100, OOL, IOL, "rotqmbii\t$rT, $rA, $val",
2798 RotateShift, pattern>;
2800 class ROTQMBIIVecInst<ValueType vectype>:
2801 ROTQMBIIInst<(outs VECREG:$rT), (ins VECREG:$rA, rotNeg7imm:$val),
2802 [/* no pattern */]>;
2804 class ROTQMBIIRegInst<RegisterClass rclass>:
2805 ROTQMBIIInst<(outs rclass:$rT), (ins rclass:$rA, rotNeg7imm:$val),
2806 [/* no pattern */]>;
2808 multiclass RotateMaskQuadByBitsImm
2810 def v16i8: ROTQMBIIVecInst<v16i8>;
2811 def v8i16: ROTQMBIIVecInst<v8i16>;
2812 def v4i32: ROTQMBIIVecInst<v4i32>;
2813 def v2i64: ROTQMBIIVecInst<v2i64>;
2815 def r128: ROTQMBIIRegInst<GPRC>;
2816 def r64: ROTQMBIIRegInst<R64C>;
2819 defm ROTQMBII: RotateMaskQuadByBitsImm;
2821 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2822 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2825 RRForm<0b01111010000, (outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2826 "rotmah\t$rT, $rA, $rB", RotateShift,
2827 [/* see patterns below - $rB must be negated */]>;
2829 def : Pat<(SPUvec_sra VECREG:$rA, R32C:$rB),
2830 (ROTMAHv8i16 VECREG:$rA, (SFIr32 R32C:$rB, 0))>;
2832 def : Pat<(SPUvec_sra VECREG:$rA, R16C:$rB),
2833 (ROTMAHv8i16 VECREG:$rA,
2834 (SFIr32 (XSHWr16 R16C:$rB), 0))>;
2836 def : Pat<(SPUvec_sra VECREG:$rA, R8C:$rB),
2837 (ROTMAHv8i16 VECREG:$rA,
2838 (SFIr32 (XSHWr16 (XSBHr8 R8C:$rB)), 0))>;
2841 RRForm<0b01111010000, (outs R16C:$rT), (ins R16C:$rA, R32C:$rB),
2842 "rotmah\t$rT, $rA, $rB", RotateShift,
2843 [/* see patterns below - $rB must be negated */]>;
2845 def : Pat<(sra R16C:$rA, R32C:$rB),
2846 (ROTMAHr16 R16C:$rA, (SFIr32 R32C:$rB, 0))>;
2848 def : Pat<(sra R16C:$rA, R16C:$rB),
2849 (ROTMAHr16 R16C:$rA,
2850 (SFIr32 (XSHWr16 R16C:$rB), 0))>;
2852 def : Pat<(sra R16C:$rA, R8C:$rB),
2853 (ROTMAHr16 R16C:$rA,
2854 (SFIr32 (XSHWr16 (XSBHr8 R8C:$rB)), 0))>;
2857 RRForm<0b01111110000, (outs VECREG:$rT), (ins VECREG:$rA, rothNeg7imm:$val),
2858 "rotmahi\t$rT, $rA, $val", RotateShift,
2859 [(set (v8i16 VECREG:$rT),
2860 (SPUvec_sra (v8i16 VECREG:$rA), (i32 uimm7:$val)))]>;
2862 def : Pat<(SPUvec_sra (v8i16 VECREG:$rA), (i16 uimm7:$val)),
2863 (ROTMAHIv8i16 (v8i16 VECREG:$rA), (i32 uimm7:$val))>;
2865 def : Pat<(SPUvec_sra (v8i16 VECREG:$rA), (i8 uimm7:$val)),
2866 (ROTMAHIv8i16 (v8i16 VECREG:$rA), (i32 uimm7:$val))>;
2869 RRForm<0b01111110000, (outs R16C:$rT), (ins R16C:$rA, rothNeg7imm_i16:$val),
2870 "rotmahi\t$rT, $rA, $val", RotateShift,
2871 [(set R16C:$rT, (sra R16C:$rA, (i16 uimm7:$val)))]>;
2873 def : Pat<(sra R16C:$rA, (i32 imm:$val)),
2874 (ROTMAHIr16 R16C:$rA, uimm7:$val)>;
2876 def : Pat<(sra R16C:$rA, (i8 imm:$val)),
2877 (ROTMAHIr16 R16C:$rA, uimm7:$val)>;
2880 RRForm<0b01011010000, (outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2881 "rotma\t$rT, $rA, $rB", RotateShift,
2882 [/* see patterns below - $rB must be negated */]>;
2884 def : Pat<(SPUvec_sra VECREG:$rA, R32C:$rB),
2885 (ROTMAv4i32 (v4i32 VECREG:$rA), (SFIr32 R32C:$rB, 0))>;
2887 def : Pat<(SPUvec_sra VECREG:$rA, R16C:$rB),
2888 (ROTMAv4i32 (v4i32 VECREG:$rA),
2889 (SFIr32 (XSHWr16 R16C:$rB), 0))>;
2891 def : Pat<(SPUvec_sra VECREG:$rA, R8C:$rB),
2892 (ROTMAv4i32 (v4i32 VECREG:$rA),
2893 (SFIr32 (XSHWr16 (XSBHr8 R8C:$rB)), 0))>;
2896 RRForm<0b01011010000, (outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
2897 "rotma\t$rT, $rA, $rB", RotateShift,
2898 [/* see patterns below - $rB must be negated */]>;
2900 def : Pat<(sra R32C:$rA, R32C:$rB),
2901 (ROTMAr32 R32C:$rA, (SFIr32 R32C:$rB, 0))>;
2903 def : Pat<(sra R32C:$rA, R16C:$rB),
2905 (SFIr32 (XSHWr16 R16C:$rB), 0))>;
2907 def : Pat<(sra R32C:$rA, R8C:$rB),
2909 (SFIr32 (XSHWr16 (XSBHr8 R8C:$rB)), 0))>;
2911 class ROTMAIInst<dag OOL, dag IOL, list<dag> pattern>:
2912 RRForm<0b01011110000, OOL, IOL,
2913 "rotmai\t$rT, $rA, $val",
2914 RotateShift, pattern>;
2916 class ROTMAIVecInst<ValueType vectype, Operand intop, ValueType inttype>:
2917 ROTMAIInst<(outs VECREG:$rT), (ins VECREG:$rA, intop:$val),
2918 [(set (vectype VECREG:$rT),
2919 (SPUvec_sra VECREG:$rA, (inttype uimm7:$val)))]>;
2921 class ROTMAIRegInst<RegisterClass rclass, Operand intop, ValueType inttype>:
2922 ROTMAIInst<(outs rclass:$rT), (ins rclass:$rA, intop:$val),
2923 [(set rclass:$rT, (sra rclass:$rA, (inttype uimm7:$val)))]>;
2925 multiclass RotateMaskAlgebraicImm {
2926 def v2i64_i32 : ROTMAIVecInst<v2i64, rotNeg7imm, i32>;
2927 def v4i32_i32 : ROTMAIVecInst<v4i32, rotNeg7imm, i32>;
2928 def r64_i32 : ROTMAIRegInst<R64C, rotNeg7imm, i32>;
2929 def r32_i32 : ROTMAIRegInst<R32C, rotNeg7imm, i32>;
2932 defm ROTMAI : RotateMaskAlgebraicImm;
2934 //===----------------------------------------------------------------------===//
2935 // Branch and conditionals:
2936 //===----------------------------------------------------------------------===//
2938 let isTerminator = 1, isBarrier = 1 in {
2939 // Halt If Equal (r32 preferred slot only, no vector form)
2941 RRForm_3<0b00011011110, (outs), (ins R32C:$rA, R32C:$rB),
2942 "heq\t$rA, $rB", BranchResolv,
2943 [/* no pattern to match */]>;
2946 RI10Form_2<0b11111110, (outs), (ins R32C:$rA, s10imm:$val),
2947 "heqi\t$rA, $val", BranchResolv,
2948 [/* no pattern to match */]>;
2950 // HGT/HGTI: These instructions use signed arithmetic for the comparison,
2951 // contrasting with HLGT/HLGTI, which use unsigned comparison:
2953 RRForm_3<0b00011010010, (outs), (ins R32C:$rA, R32C:$rB),
2954 "hgt\t$rA, $rB", BranchResolv,
2955 [/* no pattern to match */]>;
2958 RI10Form_2<0b11110010, (outs), (ins R32C:$rA, s10imm:$val),
2959 "hgti\t$rA, $val", BranchResolv,
2960 [/* no pattern to match */]>;
2963 RRForm_3<0b00011011010, (outs), (ins R32C:$rA, R32C:$rB),
2964 "hlgt\t$rA, $rB", BranchResolv,
2965 [/* no pattern to match */]>;
2968 RI10Form_2<0b11111010, (outs), (ins R32C:$rA, s10imm:$val),
2969 "hlgti\t$rA, $val", BranchResolv,
2970 [/* no pattern to match */]>;
2973 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2974 // Comparison operators for i8, i16 and i32:
2975 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2977 class CEQBInst<dag OOL, dag IOL, list<dag> pattern> :
2978 RRForm<0b00001011110, OOL, IOL, "ceqb\t$rT, $rA, $rB",
2981 multiclass CmpEqualByte
2984 CEQBInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
2985 [(set (v16i8 VECREG:$rT), (seteq (v8i16 VECREG:$rA),
2986 (v8i16 VECREG:$rB)))]>;
2989 CEQBInst<(outs R8C:$rT), (ins R8C:$rA, R8C:$rB),
2990 [(set R8C:$rT, (seteq R8C:$rA, R8C:$rB))]>;
2993 class CEQBIInst<dag OOL, dag IOL, list<dag> pattern> :
2994 RI10Form<0b01111110, OOL, IOL, "ceqbi\t$rT, $rA, $val",
2997 multiclass CmpEqualByteImm
3000 CEQBIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm_i8:$val),
3001 [(set (v16i8 VECREG:$rT), (seteq (v16i8 VECREG:$rA),
3002 v16i8SExt8Imm:$val))]>;
3004 CEQBIInst<(outs R8C:$rT), (ins R8C:$rA, s10imm_i8:$val),
3005 [(set R8C:$rT, (seteq R8C:$rA, immSExt8:$val))]>;
3008 class CEQHInst<dag OOL, dag IOL, list<dag> pattern> :
3009 RRForm<0b00010011110, OOL, IOL, "ceqh\t$rT, $rA, $rB",
3012 multiclass CmpEqualHalfword
3014 def v8i16 : CEQHInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3015 [(set (v8i16 VECREG:$rT), (seteq (v8i16 VECREG:$rA),
3016 (v8i16 VECREG:$rB)))]>;
3018 def r16 : CEQHInst<(outs R16C:$rT), (ins R16C:$rA, R16C:$rB),
3019 [(set R16C:$rT, (seteq R16C:$rA, R16C:$rB))]>;
3022 class CEQHIInst<dag OOL, dag IOL, list<dag> pattern> :
3023 RI10Form<0b10111110, OOL, IOL, "ceqhi\t$rT, $rA, $val",
3026 multiclass CmpEqualHalfwordImm
3028 def v8i16 : CEQHIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
3029 [(set (v8i16 VECREG:$rT),
3030 (seteq (v8i16 VECREG:$rA),
3031 (v8i16 v8i16SExt10Imm:$val)))]>;
3032 def r16 : CEQHIInst<(outs R16C:$rT), (ins R16C:$rA, s10imm:$val),
3033 [(set R16C:$rT, (seteq R16C:$rA, i16ImmSExt10:$val))]>;
3036 class CEQInst<dag OOL, dag IOL, list<dag> pattern> :
3037 RRForm<0b00000011110, OOL, IOL, "ceq\t$rT, $rA, $rB",
3040 multiclass CmpEqualWord
3042 def v4i32 : CEQInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3043 [(set (v4i32 VECREG:$rT),
3044 (seteq (v4i32 VECREG:$rA), (v4i32 VECREG:$rB)))]>;
3046 def r32 : CEQInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
3047 [(set R32C:$rT, (seteq R32C:$rA, R32C:$rB))]>;
3050 class CEQIInst<dag OOL, dag IOL, list<dag> pattern> :
3051 RI10Form<0b00111110, OOL, IOL, "ceqi\t$rT, $rA, $val",
3054 multiclass CmpEqualWordImm
3056 def v4i32 : CEQIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
3057 [(set (v4i32 VECREG:$rT),
3058 (seteq (v4i32 VECREG:$rA),
3059 (v4i32 v4i32SExt16Imm:$val)))]>;
3061 def r32: CEQIInst<(outs R32C:$rT), (ins R32C:$rA, s10imm_i32:$val),
3062 [(set R32C:$rT, (seteq R32C:$rA, i32ImmSExt10:$val))]>;
3065 class CGTBInst<dag OOL, dag IOL, list<dag> pattern> :
3066 RRForm<0b00001010010, OOL, IOL, "cgtb\t$rT, $rA, $rB",
3069 multiclass CmpGtrByte
3072 CGTBInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3073 [(set (v16i8 VECREG:$rT), (setgt (v8i16 VECREG:$rA),
3074 (v8i16 VECREG:$rB)))]>;
3077 CGTBInst<(outs R8C:$rT), (ins R8C:$rA, R8C:$rB),
3078 [(set R8C:$rT, (setgt R8C:$rA, R8C:$rB))]>;
3081 class CGTBIInst<dag OOL, dag IOL, list<dag> pattern> :
3082 RI10Form<0b01110010, OOL, IOL, "cgtbi\t$rT, $rA, $val",
3085 multiclass CmpGtrByteImm
3088 CGTBIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm_i8:$val),
3089 [(set (v16i8 VECREG:$rT), (setgt (v16i8 VECREG:$rA),
3090 v16i8SExt8Imm:$val))]>;
3092 CGTBIInst<(outs R8C:$rT), (ins R8C:$rA, s10imm_i8:$val),
3093 [(set R8C:$rT, (setgt R8C:$rA, immSExt8:$val))]>;
3096 class CGTHInst<dag OOL, dag IOL, list<dag> pattern> :
3097 RRForm<0b00010010010, OOL, IOL, "cgth\t$rT, $rA, $rB",
3100 multiclass CmpGtrHalfword
3102 def v8i16 : CGTHInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3103 [(set (v8i16 VECREG:$rT), (setgt (v8i16 VECREG:$rA),
3104 (v8i16 VECREG:$rB)))]>;
3106 def r16 : CGTHInst<(outs R16C:$rT), (ins R16C:$rA, R16C:$rB),
3107 [(set R16C:$rT, (setgt R16C:$rA, R16C:$rB))]>;
3110 class CGTHIInst<dag OOL, dag IOL, list<dag> pattern> :
3111 RI10Form<0b10110010, OOL, IOL, "cgthi\t$rT, $rA, $val",
3114 multiclass CmpGtrHalfwordImm
3116 def v8i16 : CGTHIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
3117 [(set (v8i16 VECREG:$rT),
3118 (setgt (v8i16 VECREG:$rA),
3119 (v8i16 v8i16SExt10Imm:$val)))]>;
3120 def r16 : CGTHIInst<(outs R16C:$rT), (ins R16C:$rA, s10imm:$val),
3121 [(set R16C:$rT, (setgt R16C:$rA, i16ImmSExt10:$val))]>;
3124 class CGTInst<dag OOL, dag IOL, list<dag> pattern> :
3125 RRForm<0b00000010010, OOL, IOL, "cgt\t$rT, $rA, $rB",
3128 multiclass CmpGtrWord
3130 def v4i32 : CGTInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3131 [(set (v4i32 VECREG:$rT),
3132 (setgt (v4i32 VECREG:$rA), (v4i32 VECREG:$rB)))]>;
3134 def r32 : CGTInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
3135 [(set R32C:$rT, (setgt R32C:$rA, R32C:$rB))]>;
3138 class CGTIInst<dag OOL, dag IOL, list<dag> pattern> :
3139 RI10Form<0b00110010, OOL, IOL, "cgti\t$rT, $rA, $val",
3142 multiclass CmpGtrWordImm
3144 def v4i32 : CGTIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
3145 [(set (v4i32 VECREG:$rT),
3146 (setgt (v4i32 VECREG:$rA),
3147 (v4i32 v4i32SExt16Imm:$val)))]>;
3149 def r32: CGTIInst<(outs R32C:$rT), (ins R32C:$rA, s10imm_i32:$val),
3150 [(set R32C:$rT, (setgt R32C:$rA, i32ImmSExt10:$val))]>;
3152 // CGTIv4f32, CGTIf32: These are used in the f32 fdiv instruction sequence:
3153 def v4f32: CGTIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
3154 [(set (v4i32 VECREG:$rT),
3155 (setgt (v4i32 (bitconvert (v4f32 VECREG:$rA))),
3156 (v4i32 v4i32SExt16Imm:$val)))]>;
3158 def f32: CGTIInst<(outs R32C:$rT), (ins R32FP:$rA, s10imm_i32:$val),
3159 [/* no pattern */]>;
3162 class CLGTBInst<dag OOL, dag IOL, list<dag> pattern> :
3163 RRForm<0b00001011010, OOL, IOL, "clgtb\t$rT, $rA, $rB",
3166 multiclass CmpLGtrByte
3169 CLGTBInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3170 [(set (v16i8 VECREG:$rT), (setugt (v8i16 VECREG:$rA),
3171 (v8i16 VECREG:$rB)))]>;
3174 CLGTBInst<(outs R8C:$rT), (ins R8C:$rA, R8C:$rB),
3175 [(set R8C:$rT, (setugt R8C:$rA, R8C:$rB))]>;
3178 class CLGTBIInst<dag OOL, dag IOL, list<dag> pattern> :
3179 RI10Form<0b01111010, OOL, IOL, "clgtbi\t$rT, $rA, $val",
3182 multiclass CmpLGtrByteImm
3185 CLGTBIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm_i8:$val),
3186 [(set (v16i8 VECREG:$rT), (setugt (v16i8 VECREG:$rA),
3187 v16i8SExt8Imm:$val))]>;
3189 CLGTBIInst<(outs R8C:$rT), (ins R8C:$rA, s10imm_i8:$val),
3190 [(set R8C:$rT, (setugt R8C:$rA, immSExt8:$val))]>;
3193 class CLGTHInst<dag OOL, dag IOL, list<dag> pattern> :
3194 RRForm<0b00010011010, OOL, IOL, "clgth\t$rT, $rA, $rB",
3197 multiclass CmpLGtrHalfword
3199 def v8i16 : CLGTHInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3200 [(set (v8i16 VECREG:$rT), (setugt (v8i16 VECREG:$rA),
3201 (v8i16 VECREG:$rB)))]>;
3203 def r16 : CLGTHInst<(outs R16C:$rT), (ins R16C:$rA, R16C:$rB),
3204 [(set R16C:$rT, (setugt R16C:$rA, R16C:$rB))]>;
3207 class CLGTHIInst<dag OOL, dag IOL, list<dag> pattern> :
3208 RI10Form<0b10111010, OOL, IOL, "clgthi\t$rT, $rA, $val",
3211 multiclass CmpLGtrHalfwordImm
3213 def v8i16 : CLGTHIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
3214 [(set (v8i16 VECREG:$rT),
3215 (setugt (v8i16 VECREG:$rA),
3216 (v8i16 v8i16SExt10Imm:$val)))]>;
3217 def r16 : CLGTHIInst<(outs R16C:$rT), (ins R16C:$rA, s10imm:$val),
3218 [(set R16C:$rT, (setugt R16C:$rA, i16ImmSExt10:$val))]>;
3221 class CLGTInst<dag OOL, dag IOL, list<dag> pattern> :
3222 RRForm<0b00000011010, OOL, IOL, "clgt\t$rT, $rA, $rB",
3225 multiclass CmpLGtrWord
3227 def v4i32 : CLGTInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3228 [(set (v4i32 VECREG:$rT),
3229 (setugt (v4i32 VECREG:$rA), (v4i32 VECREG:$rB)))]>;
3231 def r32 : CLGTInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
3232 [(set R32C:$rT, (setugt R32C:$rA, R32C:$rB))]>;
3235 class CLGTIInst<dag OOL, dag IOL, list<dag> pattern> :
3236 RI10Form<0b00111010, OOL, IOL, "clgti\t$rT, $rA, $val",
3239 multiclass CmpLGtrWordImm
3241 def v4i32 : CLGTIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
3242 [(set (v4i32 VECREG:$rT),
3243 (setugt (v4i32 VECREG:$rA),
3244 (v4i32 v4i32SExt16Imm:$val)))]>;
3246 def r32: CLGTIInst<(outs R32C:$rT), (ins R32C:$rA, s10imm_i32:$val),
3247 [(set R32C:$rT, (setugt R32C:$rA, i32ImmSExt10:$val))]>;
3250 defm CEQB : CmpEqualByte;
3251 defm CEQBI : CmpEqualByteImm;
3252 defm CEQH : CmpEqualHalfword;
3253 defm CEQHI : CmpEqualHalfwordImm;
3254 defm CEQ : CmpEqualWord;
3255 defm CEQI : CmpEqualWordImm;
3256 defm CGTB : CmpGtrByte;
3257 defm CGTBI : CmpGtrByteImm;
3258 defm CGTH : CmpGtrHalfword;
3259 defm CGTHI : CmpGtrHalfwordImm;
3260 defm CGT : CmpGtrWord;
3261 defm CGTI : CmpGtrWordImm;
3262 defm CLGTB : CmpLGtrByte;
3263 defm CLGTBI : CmpLGtrByteImm;
3264 defm CLGTH : CmpLGtrHalfword;
3265 defm CLGTHI : CmpLGtrHalfwordImm;
3266 defm CLGT : CmpLGtrWord;
3267 defm CLGTI : CmpLGtrWordImm;
3269 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
3270 // For SETCC primitives not supported above (setlt, setle, setge, etc.)
3271 // define a pattern to generate the right code, as a binary operator
3272 // (in a manner of speaking.)
3275 // 1. This only matches the setcc set of conditionals. Special pattern
3276 // matching is used for select conditionals.
3278 // 2. The "DAG" versions of these classes is almost exclusively used for
3279 // i64 comparisons. See the tblgen fundamentals documentation for what
3280 // ".ResultInstrs[0]" means; see TargetSelectionDAG.td and the Pattern
3281 // class for where ResultInstrs originates.
3282 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
3284 class SETCCNegCondReg<PatFrag cond, RegisterClass rclass, ValueType inttype,
3285 SPUInstr xorinst, SPUInstr cmpare>:
3286 Pat<(cond rclass:$rA, rclass:$rB),
3287 (xorinst (cmpare rclass:$rA, rclass:$rB), (inttype -1))>;
3289 class SETCCNegCondImm<PatFrag cond, RegisterClass rclass, ValueType inttype,
3290 PatLeaf immpred, SPUInstr xorinst, SPUInstr cmpare>:
3291 Pat<(cond rclass:$rA, (inttype immpred:$imm)),
3292 (xorinst (cmpare rclass:$rA, (inttype immpred:$imm)), (inttype -1))>;
3294 def : SETCCNegCondReg<setne, R8C, i8, XORBIr8, CEQBr8>;
3295 def : SETCCNegCondImm<setne, R8C, i8, immSExt8, XORBIr8, CEQBIr8>;
3297 def : SETCCNegCondReg<setne, R16C, i16, XORHIr16, CEQHr16>;
3298 def : SETCCNegCondImm<setne, R16C, i16, i16ImmSExt10, XORHIr16, CEQHIr16>;
3300 def : SETCCNegCondReg<setne, R32C, i32, XORIr32, CEQr32>;
3301 def : SETCCNegCondImm<setne, R32C, i32, i32ImmSExt10, XORIr32, CEQIr32>;
3303 class SETCCBinOpReg<PatFrag cond, RegisterClass rclass,
3304 SPUInstr binop, SPUInstr cmpOp1, SPUInstr cmpOp2>:
3305 Pat<(cond rclass:$rA, rclass:$rB),
3306 (binop (cmpOp1 rclass:$rA, rclass:$rB),
3307 (cmpOp2 rclass:$rA, rclass:$rB))>;
3309 class SETCCBinOpImm<PatFrag cond, RegisterClass rclass, PatLeaf immpred,
3311 SPUInstr binop, SPUInstr cmpOp1, SPUInstr cmpOp2>:
3312 Pat<(cond rclass:$rA, (immtype immpred:$imm)),
3313 (binop (cmpOp1 rclass:$rA, (immtype immpred:$imm)),
3314 (cmpOp2 rclass:$rA, (immtype immpred:$imm)))>;
3316 def : SETCCBinOpReg<setge, R8C, ORr8, CGTBr8, CEQBr8>;
3317 def : SETCCBinOpImm<setge, R8C, immSExt8, i8, ORr8, CGTBIr8, CEQBIr8>;
3318 def : SETCCBinOpReg<setlt, R8C, NORr8, CGTBr8, CEQBr8>;
3319 def : SETCCBinOpImm<setlt, R8C, immSExt8, i8, NORr8, CGTBIr8, CEQBIr8>;
3320 def : Pat<(setle R8C:$rA, R8C:$rB),
3321 (XORBIr8 (CGTBr8 R8C:$rA, R8C:$rB), 0xff)>;
3322 def : Pat<(setle R8C:$rA, immU8:$imm),
3323 (XORBIr8 (CGTBIr8 R8C:$rA, immU8:$imm), 0xff)>;
3325 def : SETCCBinOpReg<setge, R16C, ORr16, CGTHr16, CEQHr16>;
3326 def : SETCCBinOpImm<setge, R16C, i16ImmSExt10, i16,
3327 ORr16, CGTHIr16, CEQHIr16>;
3328 def : SETCCBinOpReg<setlt, R16C, NORr16, CGTHr16, CEQHr16>;
3329 def : SETCCBinOpImm<setlt, R16C, i16ImmSExt10, i16, NORr16, CGTHIr16, CEQHIr16>;
3330 def : Pat<(setle R16C:$rA, R16C:$rB),
3331 (XORHIr16 (CGTHr16 R16C:$rA, R16C:$rB), 0xffff)>;
3332 def : Pat<(setle R16C:$rA, i16ImmSExt10:$imm),
3333 (XORHIr16 (CGTHIr16 R16C:$rA, i16ImmSExt10:$imm), 0xffff)>;
3335 def : SETCCBinOpReg<setge, R32C, ORr32, CGTr32, CEQr32>;
3336 def : SETCCBinOpImm<setge, R32C, i32ImmSExt10, i32,
3337 ORr32, CGTIr32, CEQIr32>;
3338 def : SETCCBinOpReg<setlt, R32C, NORr32, CGTr32, CEQr32>;
3339 def : SETCCBinOpImm<setlt, R32C, i32ImmSExt10, i32, NORr32, CGTIr32, CEQIr32>;
3340 def : Pat<(setle R32C:$rA, R32C:$rB),
3341 (XORIr32 (CGTr32 R32C:$rA, R32C:$rB), 0xffffffff)>;
3342 def : Pat<(setle R32C:$rA, i32ImmSExt10:$imm),
3343 (XORIr32 (CGTIr32 R32C:$rA, i32ImmSExt10:$imm), 0xffffffff)>;
3345 def : SETCCBinOpReg<setuge, R8C, ORr8, CLGTBr8, CEQBr8>;
3346 def : SETCCBinOpImm<setuge, R8C, immSExt8, i8, ORr8, CLGTBIr8, CEQBIr8>;
3347 def : SETCCBinOpReg<setult, R8C, NORr8, CLGTBr8, CEQBr8>;
3348 def : SETCCBinOpImm<setult, R8C, immSExt8, i8, NORr8, CLGTBIr8, CEQBIr8>;
3349 def : Pat<(setule R8C:$rA, R8C:$rB),
3350 (XORBIr8 (CLGTBr8 R8C:$rA, R8C:$rB), 0xff)>;
3351 def : Pat<(setule R8C:$rA, immU8:$imm),
3352 (XORBIr8 (CLGTBIr8 R8C:$rA, immU8:$imm), 0xff)>;
3354 def : SETCCBinOpReg<setuge, R16C, ORr16, CLGTHr16, CEQHr16>;
3355 def : SETCCBinOpImm<setuge, R16C, i16ImmSExt10, i16,
3356 ORr16, CLGTHIr16, CEQHIr16>;
3357 def : SETCCBinOpReg<setult, R16C, NORr16, CLGTHr16, CEQHr16>;
3358 def : SETCCBinOpImm<setult, R16C, i16ImmSExt10, i16, NORr16,
3359 CLGTHIr16, CEQHIr16>;
3360 def : Pat<(setule R16C:$rA, R16C:$rB),
3361 (XORHIr16 (CLGTHr16 R16C:$rA, R16C:$rB), 0xffff)>;
3362 def : Pat<(setule R16C:$rA, i16ImmSExt10:$imm),
3363 (XORHIr16 (CLGTHIr16 R16C:$rA, i16ImmSExt10:$imm), 0xffff)>;
3365 def : SETCCBinOpReg<setuge, R32C, ORr32, CLGTr32, CEQr32>;
3366 def : SETCCBinOpImm<setuge, R32C, i32ImmSExt10, i32,
3367 ORr32, CLGTIr32, CEQIr32>;
3368 def : SETCCBinOpReg<setult, R32C, NORr32, CLGTr32, CEQr32>;
3369 def : SETCCBinOpImm<setult, R32C, i32ImmSExt10, i32, NORr32, CLGTIr32, CEQIr32>;
3370 def : Pat<(setule R32C:$rA, R32C:$rB),
3371 (XORIr32 (CLGTr32 R32C:$rA, R32C:$rB), 0xffffffff)>;
3372 def : Pat<(setule R32C:$rA, i32ImmSExt10:$imm),
3373 (XORIr32 (CLGTIr32 R32C:$rA, i32ImmSExt10:$imm), 0xffffffff)>;
3375 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
3376 // select conditional patterns:
3377 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
3379 class SELECTNegCondReg<PatFrag cond, RegisterClass rclass, ValueType inttype,
3380 SPUInstr selinstr, SPUInstr cmpare>:
3381 Pat<(select (inttype (cond rclass:$rA, rclass:$rB)),
3382 rclass:$rTrue, rclass:$rFalse),
3383 (selinstr rclass:$rTrue, rclass:$rFalse,
3384 (cmpare rclass:$rA, rclass:$rB))>;
3386 class SELECTNegCondImm<PatFrag cond, RegisterClass rclass, ValueType inttype,
3387 PatLeaf immpred, SPUInstr selinstr, SPUInstr cmpare>:
3388 Pat<(select (inttype (cond rclass:$rA, immpred:$imm)),
3389 rclass:$rTrue, rclass:$rFalse),
3390 (selinstr rclass:$rTrue, rclass:$rFalse,
3391 (cmpare rclass:$rA, immpred:$imm))>;
3393 def : SELECTNegCondReg<setne, R8C, i8, SELBr8, CEQBr8>;
3394 def : SELECTNegCondImm<setne, R8C, i8, immSExt8, SELBr8, CEQBIr8>;
3395 def : SELECTNegCondReg<setle, R8C, i8, SELBr8, CGTBr8>;
3396 def : SELECTNegCondImm<setle, R8C, i8, immSExt8, SELBr8, CGTBr8>;
3397 def : SELECTNegCondReg<setule, R8C, i8, SELBr8, CLGTBr8>;
3398 def : SELECTNegCondImm<setule, R8C, i8, immU8, SELBr8, CLGTBIr8>;
3400 def : SELECTNegCondReg<setne, R16C, i16, SELBr16, CEQHr16>;
3401 def : SELECTNegCondImm<setne, R16C, i16, i16ImmSExt10, SELBr16, CEQHIr16>;
3402 def : SELECTNegCondReg<setle, R16C, i16, SELBr16, CGTHr16>;
3403 def : SELECTNegCondImm<setle, R16C, i16, i16ImmSExt10, SELBr16, CGTHIr16>;
3404 def : SELECTNegCondReg<setule, R16C, i16, SELBr16, CLGTHr16>;
3405 def : SELECTNegCondImm<setule, R16C, i16, i16ImmSExt10, SELBr16, CLGTHIr16>;
3407 def : SELECTNegCondReg<setne, R32C, i32, SELBr32, CEQr32>;
3408 def : SELECTNegCondImm<setne, R32C, i32, i32ImmSExt10, SELBr32, CEQIr32>;
3409 def : SELECTNegCondReg<setle, R32C, i32, SELBr32, CGTr32>;
3410 def : SELECTNegCondImm<setle, R32C, i32, i32ImmSExt10, SELBr32, CGTIr32>;
3411 def : SELECTNegCondReg<setule, R32C, i32, SELBr32, CLGTr32>;
3412 def : SELECTNegCondImm<setule, R32C, i32, i32ImmSExt10, SELBr32, CLGTIr32>;
3414 class SELECTBinOpReg<PatFrag cond, RegisterClass rclass, ValueType inttype,
3415 SPUInstr selinstr, SPUInstr binop, SPUInstr cmpOp1,
3417 Pat<(select (inttype (cond rclass:$rA, rclass:$rB)),
3418 rclass:$rTrue, rclass:$rFalse),
3419 (selinstr rclass:$rFalse, rclass:$rTrue,
3420 (binop (cmpOp1 rclass:$rA, rclass:$rB),
3421 (cmpOp2 rclass:$rA, rclass:$rB)))>;
3423 class SELECTBinOpImm<PatFrag cond, RegisterClass rclass, PatLeaf immpred,
3425 SPUInstr selinstr, SPUInstr binop, SPUInstr cmpOp1,
3427 Pat<(select (inttype (cond rclass:$rA, (inttype immpred:$imm))),
3428 rclass:$rTrue, rclass:$rFalse),
3429 (selinstr rclass:$rFalse, rclass:$rTrue,
3430 (binop (cmpOp1 rclass:$rA, (inttype immpred:$imm)),
3431 (cmpOp2 rclass:$rA, (inttype immpred:$imm))))>;
3433 def : SELECTBinOpReg<setge, R8C, i8, SELBr8, ORr8, CGTBr8, CEQBr8>;
3434 def : SELECTBinOpImm<setge, R8C, immSExt8, i8,
3435 SELBr8, ORr8, CGTBIr8, CEQBIr8>;
3437 def : SELECTBinOpReg<setge, R16C, i16, SELBr16, ORr16, CGTHr16, CEQHr16>;
3438 def : SELECTBinOpImm<setge, R16C, i16ImmSExt10, i16,
3439 SELBr16, ORr16, CGTHIr16, CEQHIr16>;
3441 def : SELECTBinOpReg<setge, R32C, i32, SELBr32, ORr32, CGTr32, CEQr32>;
3442 def : SELECTBinOpImm<setge, R32C, i32ImmSExt10, i32,
3443 SELBr32, ORr32, CGTIr32, CEQIr32>;
3445 def : SELECTBinOpReg<setuge, R8C, i8, SELBr8, ORr8, CLGTBr8, CEQBr8>;
3446 def : SELECTBinOpImm<setuge, R8C, immSExt8, i8,
3447 SELBr8, ORr8, CLGTBIr8, CEQBIr8>;
3449 def : SELECTBinOpReg<setuge, R16C, i16, SELBr16, ORr16, CLGTHr16, CEQHr16>;
3450 def : SELECTBinOpImm<setuge, R16C, i16ImmUns10, i16,
3451 SELBr16, ORr16, CLGTHIr16, CEQHIr16>;
3453 def : SELECTBinOpReg<setuge, R32C, i32, SELBr32, ORr32, CLGTr32, CEQr32>;
3454 def : SELECTBinOpImm<setuge, R32C, i32ImmUns10, i32,
3455 SELBr32, ORr32, CLGTIr32, CEQIr32>;
3457 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
3460 // All calls clobber the non-callee-saved registers:
3461 Defs = [R0, R1, R2, R3, R4, R5, R6, R7, R8, R9,
3462 R10,R11,R12,R13,R14,R15,R16,R17,R18,R19,
3463 R20,R21,R22,R23,R24,R25,R26,R27,R28,R29,
3464 R30,R31,R32,R33,R34,R35,R36,R37,R38,R39,
3465 R40,R41,R42,R43,R44,R45,R46,R47,R48,R49,
3466 R50,R51,R52,R53,R54,R55,R56,R57,R58,R59,
3467 R60,R61,R62,R63,R64,R65,R66,R67,R68,R69,
3468 R70,R71,R72,R73,R74,R75,R76,R77,R78,R79],
3469 // All of these instructions use $lr (aka $0)
3471 // Branch relative and set link: Used if we actually know that the target
3472 // is within [-32768, 32767] bytes of the target
3474 BranchSetLink<0b011001100, (outs), (ins relcalltarget:$func, variable_ops),
3475 "brsl\t$$lr, $func",
3476 [(SPUcall (SPUpcrel tglobaladdr:$func, 0))]>;
3478 // Branch absolute and set link: Used if we actually know that the target
3479 // is an absolute address
3481 BranchSetLink<0b011001100, (outs), (ins calltarget:$func, variable_ops),
3482 "brasl\t$$lr, $func",
3483 [(SPUcall (SPUaform tglobaladdr:$func, 0))]>;
3485 // Branch indirect and set link if external data. These instructions are not
3486 // actually generated, matched by an intrinsic:
3487 def BISLED_00: BISLEDForm<0b11, "bisled\t$$lr, $func", [/* empty pattern */]>;
3488 def BISLED_E0: BISLEDForm<0b10, "bisled\t$$lr, $func", [/* empty pattern */]>;
3489 def BISLED_0D: BISLEDForm<0b01, "bisled\t$$lr, $func", [/* empty pattern */]>;
3490 def BISLED_ED: BISLEDForm<0b00, "bisled\t$$lr, $func", [/* empty pattern */]>;
3492 // Branch indirect and set link. This is the "X-form" address version of a
3495 BIForm<0b10010101100, "bisl\t$$lr, $func", [(SPUcall R32C:$func)]>;
3498 // Support calls to external symbols:
3499 def : Pat<(SPUcall (SPUpcrel texternalsym:$func, 0)),
3500 (BRSL texternalsym:$func)>;
3502 def : Pat<(SPUcall (SPUaform texternalsym:$func, 0)),
3503 (BRASL texternalsym:$func)>;
3505 // Unconditional branches:
3506 let isBranch = 1, isTerminator = 1, hasCtrlDep = 1, isBarrier = 1 in {
3508 UncondBranch<0b001001100, (outs), (ins brtarget:$dest),
3512 // Unconditional, absolute address branch
3514 UncondBranch<0b001100000, (outs), (ins brtarget:$dest),
3516 [/* no pattern */]>;
3520 BIForm<0b00010101100, "bi\t$func", [(brind R32C:$func)]>;
3522 // Conditional branches:
3523 class BRNZInst<dag IOL, list<dag> pattern>:
3524 RI16Form<0b010000100, (outs), IOL, "brnz\t$rCond,$dest",
3525 BranchResolv, pattern>;
3527 class BRNZRegInst<RegisterClass rclass>:
3528 BRNZInst<(ins rclass:$rCond, brtarget:$dest),
3529 [(brcond rclass:$rCond, bb:$dest)]>;
3531 class BRNZVecInst<ValueType vectype>:
3532 BRNZInst<(ins VECREG:$rCond, brtarget:$dest),
3533 [(brcond (vectype VECREG:$rCond), bb:$dest)]>;
3535 multiclass BranchNotZero {
3536 def v4i32 : BRNZVecInst<v4i32>;
3537 def r32 : BRNZRegInst<R32C>;
3540 defm BRNZ : BranchNotZero;
3542 class BRZInst<dag IOL, list<dag> pattern>:
3543 RI16Form<0b000000100, (outs), IOL, "brz\t$rT,$dest",
3544 BranchResolv, pattern>;
3546 class BRZRegInst<RegisterClass rclass>:
3547 BRZInst<(ins rclass:$rT, brtarget:$dest), [/* no pattern */]>;
3549 class BRZVecInst<ValueType vectype>:
3550 BRZInst<(ins VECREG:$rT, brtarget:$dest), [/* no pattern */]>;
3552 multiclass BranchZero {
3553 def v4i32: BRZVecInst<v4i32>;
3554 def r32: BRZRegInst<R32C>;
3557 defm BRZ: BranchZero;
3559 // Note: LLVM doesn't do branch conditional, indirect. Otherwise these would
3562 class BINZInst<dag IOL, list<dag> pattern>:
3563 BICondForm<0b10010100100, (outs), IOL, "binz\t$rA, $dest", pattern>;
3565 class BINZRegInst<RegisterClass rclass>:
3566 BINZInst<(ins rclass:$rA, brtarget:$dest),
3567 [(brcond rclass:$rA, R32C:$dest)]>;
3569 class BINZVecInst<ValueType vectype>:
3570 BINZInst<(ins VECREG:$rA, R32C:$dest),
3571 [(brcond (vectype VECREG:$rA), R32C:$dest)]>;
3573 multiclass BranchNotZeroIndirect {
3574 def v4i32: BINZVecInst<v4i32>;
3575 def r32: BINZRegInst<R32C>;
3578 defm BINZ: BranchNotZeroIndirect;
3580 class BIZInst<dag IOL, list<dag> pattern>:
3581 BICondForm<0b00010100100, (outs), IOL, "biz\t$rA, $func", pattern>;
3583 class BIZRegInst<RegisterClass rclass>:
3584 BIZInst<(ins rclass:$rA, R32C:$func), [/* no pattern */]>;
3586 class BIZVecInst<ValueType vectype>:
3587 BIZInst<(ins VECREG:$rA, R32C:$func), [/* no pattern */]>;
3589 multiclass BranchZeroIndirect {
3590 def v4i32: BIZVecInst<v4i32>;
3591 def r32: BIZRegInst<R32C>;
3594 defm BIZ: BranchZeroIndirect;
3597 class BRHNZInst<dag IOL, list<dag> pattern>:
3598 RI16Form<0b011000100, (outs), IOL, "brhnz\t$rCond,$dest", BranchResolv,
3601 class BRHNZRegInst<RegisterClass rclass>:
3602 BRHNZInst<(ins rclass:$rCond, brtarget:$dest),
3603 [(brcond rclass:$rCond, bb:$dest)]>;
3605 class BRHNZVecInst<ValueType vectype>:
3606 BRHNZInst<(ins VECREG:$rCond, brtarget:$dest), [/* no pattern */]>;
3608 multiclass BranchNotZeroHalfword {
3609 def v8i16: BRHNZVecInst<v8i16>;
3610 def r16: BRHNZRegInst<R16C>;
3613 defm BRHNZ: BranchNotZeroHalfword;
3615 class BRHZInst<dag IOL, list<dag> pattern>:
3616 RI16Form<0b001000100, (outs), IOL, "brhz\t$rT,$dest", BranchResolv,
3619 class BRHZRegInst<RegisterClass rclass>:
3620 BRHZInst<(ins rclass:$rT, brtarget:$dest), [/* no pattern */]>;
3622 class BRHZVecInst<ValueType vectype>:
3623 BRHZInst<(ins VECREG:$rT, brtarget:$dest), [/* no pattern */]>;
3625 multiclass BranchZeroHalfword {
3626 def v8i16: BRHZVecInst<v8i16>;
3627 def r16: BRHZRegInst<R16C>;
3630 defm BRHZ: BranchZeroHalfword;
3633 //===----------------------------------------------------------------------===//
3634 // setcc and brcond patterns:
3635 //===----------------------------------------------------------------------===//
3637 def : Pat<(brcond (i16 (seteq R16C:$rA, 0)), bb:$dest),
3638 (BRHZr16 R16C:$rA, bb:$dest)>;
3639 def : Pat<(brcond (i16 (setne R16C:$rA, 0)), bb:$dest),
3640 (BRHNZr16 R16C:$rA, bb:$dest)>;
3642 def : Pat<(brcond (i32 (seteq R32C:$rA, 0)), bb:$dest),
3643 (BRZr32 R32C:$rA, bb:$dest)>;
3644 def : Pat<(brcond (i32 (setne R32C:$rA, 0)), bb:$dest),
3645 (BRNZr32 R32C:$rA, bb:$dest)>;
3647 multiclass BranchCondEQ<PatFrag cond, SPUInstr brinst16, SPUInstr brinst32>
3649 def r16imm: Pat<(brcond (i16 (cond R16C:$rA, i16ImmSExt10:$val)), bb:$dest),
3650 (brinst16 (CEQHIr16 R16C:$rA, i16ImmSExt10:$val), bb:$dest)>;
3652 def r16 : Pat<(brcond (i16 (cond R16C:$rA, R16C:$rB)), bb:$dest),
3653 (brinst16 (CEQHr16 R16C:$rA, R16:$rB), bb:$dest)>;
3655 def r32imm : Pat<(brcond (i32 (cond R32C:$rA, i32ImmSExt10:$val)), bb:$dest),
3656 (brinst32 (CEQIr32 R32C:$rA, i32ImmSExt10:$val), bb:$dest)>;
3658 def r32 : Pat<(brcond (i32 (cond R32C:$rA, R32C:$rB)), bb:$dest),
3659 (brinst32 (CEQr32 R32C:$rA, R32C:$rB), bb:$dest)>;
3662 defm BRCONDeq : BranchCondEQ<seteq, BRHNZr16, BRNZr32>;
3663 defm BRCONDne : BranchCondEQ<setne, BRHZr16, BRZr32>;
3665 multiclass BranchCondLGT<PatFrag cond, SPUInstr brinst16, SPUInstr brinst32>
3667 def r16imm : Pat<(brcond (i16 (cond R16C:$rA, i16ImmSExt10:$val)), bb:$dest),
3668 (brinst16 (CLGTHIr16 R16C:$rA, i16ImmSExt10:$val), bb:$dest)>;
3670 def r16 : Pat<(brcond (i16 (cond R16C:$rA, R16C:$rB)), bb:$dest),
3671 (brinst16 (CLGTHr16 R16C:$rA, R16:$rB), bb:$dest)>;
3673 def r32imm : Pat<(brcond (i32 (cond R32C:$rA, i32ImmSExt10:$val)), bb:$dest),
3674 (brinst32 (CLGTIr32 R32C:$rA, i32ImmSExt10:$val), bb:$dest)>;
3676 def r32 : Pat<(brcond (i32 (cond R32C:$rA, R32C:$rB)), bb:$dest),
3677 (brinst32 (CLGTr32 R32C:$rA, R32C:$rB), bb:$dest)>;
3680 defm BRCONDugt : BranchCondLGT<setugt, BRHNZr16, BRNZr32>;
3681 defm BRCONDule : BranchCondLGT<setule, BRHZr16, BRZr32>;
3683 multiclass BranchCondLGTEQ<PatFrag cond, SPUInstr orinst16, SPUInstr brinst16,
3684 SPUInstr orinst32, SPUInstr brinst32>
3686 def r16imm: Pat<(brcond (i16 (cond R16C:$rA, i16ImmSExt10:$val)), bb:$dest),
3687 (brinst16 (orinst16 (CLGTHIr16 R16C:$rA, i16ImmSExt10:$val),
3688 (CEQHIr16 R16C:$rA, i16ImmSExt10:$val)),
3691 def r16: Pat<(brcond (i16 (cond R16C:$rA, R16C:$rB)), bb:$dest),
3692 (brinst16 (orinst16 (CLGTHr16 R16C:$rA, R16:$rB),
3693 (CEQHr16 R16C:$rA, R16:$rB)),
3696 def r32imm : Pat<(brcond (i32 (cond R32C:$rA, i32ImmSExt10:$val)), bb:$dest),
3697 (brinst32 (orinst32 (CLGTIr32 R32C:$rA, i32ImmSExt10:$val),
3698 (CEQIr32 R32C:$rA, i32ImmSExt10:$val)),
3701 def r32 : Pat<(brcond (i32 (cond R32C:$rA, R32C:$rB)), bb:$dest),
3702 (brinst32 (orinst32 (CLGTr32 R32C:$rA, R32C:$rB),
3703 (CEQr32 R32C:$rA, R32C:$rB)),
3707 defm BRCONDuge : BranchCondLGTEQ<setuge, ORr16, BRHNZr16, ORr32, BRNZr32>;
3708 defm BRCONDult : BranchCondLGTEQ<setult, ORr16, BRHZr16, ORr32, BRZr32>;
3710 multiclass BranchCondGT<PatFrag cond, SPUInstr brinst16, SPUInstr brinst32>
3712 def r16imm : Pat<(brcond (i16 (cond R16C:$rA, i16ImmSExt10:$val)), bb:$dest),
3713 (brinst16 (CGTHIr16 R16C:$rA, i16ImmSExt10:$val), bb:$dest)>;
3715 def r16 : Pat<(brcond (i16 (cond R16C:$rA, R16C:$rB)), bb:$dest),
3716 (brinst16 (CGTHr16 R16C:$rA, R16:$rB), bb:$dest)>;
3718 def r32imm : Pat<(brcond (i32 (cond R32C:$rA, i32ImmSExt10:$val)), bb:$dest),
3719 (brinst32 (CGTIr32 R32C:$rA, i32ImmSExt10:$val), bb:$dest)>;
3721 def r32 : Pat<(brcond (i32 (cond R32C:$rA, R32C:$rB)), bb:$dest),
3722 (brinst32 (CGTr32 R32C:$rA, R32C:$rB), bb:$dest)>;
3725 defm BRCONDgt : BranchCondGT<setgt, BRHNZr16, BRNZr32>;
3726 defm BRCONDle : BranchCondGT<setle, BRHZr16, BRZr32>;
3728 multiclass BranchCondGTEQ<PatFrag cond, SPUInstr orinst16, SPUInstr brinst16,
3729 SPUInstr orinst32, SPUInstr brinst32>
3731 def r16imm: Pat<(brcond (i16 (cond R16C:$rA, i16ImmSExt10:$val)), bb:$dest),
3732 (brinst16 (orinst16 (CGTHIr16 R16C:$rA, i16ImmSExt10:$val),
3733 (CEQHIr16 R16C:$rA, i16ImmSExt10:$val)),
3736 def r16: Pat<(brcond (i16 (cond R16C:$rA, R16C:$rB)), bb:$dest),
3737 (brinst16 (orinst16 (CGTHr16 R16C:$rA, R16:$rB),
3738 (CEQHr16 R16C:$rA, R16:$rB)),
3741 def r32imm : Pat<(brcond (i32 (cond R32C:$rA, i32ImmSExt10:$val)), bb:$dest),
3742 (brinst32 (orinst32 (CGTIr32 R32C:$rA, i32ImmSExt10:$val),
3743 (CEQIr32 R32C:$rA, i32ImmSExt10:$val)),
3746 def r32 : Pat<(brcond (i32 (cond R32C:$rA, R32C:$rB)), bb:$dest),
3747 (brinst32 (orinst32 (CGTr32 R32C:$rA, R32C:$rB),
3748 (CEQr32 R32C:$rA, R32C:$rB)),
3752 defm BRCONDge : BranchCondGTEQ<setge, ORr16, BRHNZr16, ORr32, BRNZr32>;
3753 defm BRCONDlt : BranchCondGTEQ<setlt, ORr16, BRHZr16, ORr32, BRZr32>;
3755 let isTerminator = 1, isBarrier = 1 in {
3756 let isReturn = 1 in {
3758 RETForm<"bi\t$$lr", [(retflag)]>;
3762 //===----------------------------------------------------------------------===//
3763 // Single precision floating point instructions
3764 //===----------------------------------------------------------------------===//
3766 class FAInst<dag OOL, dag IOL, list<dag> pattern>:
3767 RRForm<0b01011000100, OOL, IOL, "fa\t$rT, $rA, $rB",
3770 class FAVecInst<ValueType vectype>:
3771 FAInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3772 [(set (vectype VECREG:$rT),
3773 (fadd (vectype VECREG:$rA), (vectype VECREG:$rB)))]>;
3777 def v4f32: FAVecInst<v4f32>;
3778 def f32: FAInst<(outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB),
3779 [(set R32FP:$rT, (fadd R32FP:$rA, R32FP:$rB))]>;
3784 class FSInst<dag OOL, dag IOL, list<dag> pattern>:
3785 RRForm<0b01011000100, OOL, IOL, "fs\t$rT, $rA, $rB",
3788 class FSVecInst<ValueType vectype>:
3789 FSInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3790 [(set (vectype VECREG:$rT),
3791 (fsub (vectype VECREG:$rA), (vectype VECREG:$rB)))]>;
3795 def v4f32: FSVecInst<v4f32>;
3796 def f32: FSInst<(outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB),
3797 [(set R32FP:$rT, (fsub R32FP:$rA, R32FP:$rB))]>;
3802 // Floating point reciprocal estimate
3804 class FRESTInst<dag OOL, dag IOL>:
3805 RRForm_1<0b00110111000, OOL, IOL,
3806 "frest\t$rT, $rA", SPrecFP,
3807 [/* no pattern */]>;
3810 FRESTInst<(outs VECREG:$rT), (ins VECREG:$rA)>;
3813 FRESTInst<(outs R32FP:$rT), (ins R32FP:$rA)>;
3815 // Floating point interpolate (used in conjunction with reciprocal estimate)
3817 RRForm<0b00101011110, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3818 "fi\t$rT, $rA, $rB", SPrecFP,
3819 [/* no pattern */]>;
3822 RRForm<0b00101011110, (outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB),
3823 "fi\t$rT, $rA, $rB", SPrecFP,
3824 [/* no pattern */]>;
3826 //--------------------------------------------------------------------------
3827 // Basic single precision floating point comparisons:
3829 // Note: There is no support on SPU for single precision NaN. Consequently,
3830 // ordered and unordered comparisons are the same.
3831 //--------------------------------------------------------------------------
3834 RRForm<0b01000011110, (outs R32C:$rT), (ins R32FP:$rA, R32FP:$rB),
3835 "fceq\t$rT, $rA, $rB", SPrecFP,
3836 [(set R32C:$rT, (setueq R32FP:$rA, R32FP:$rB))]>;
3838 def : Pat<(setoeq R32FP:$rA, R32FP:$rB),
3839 (FCEQf32 R32FP:$rA, R32FP:$rB)>;
3842 RRForm<0b01010011110, (outs R32C:$rT), (ins R32FP:$rA, R32FP:$rB),
3843 "fcmeq\t$rT, $rA, $rB", SPrecFP,
3844 [(set R32C:$rT, (setueq (fabs R32FP:$rA), (fabs R32FP:$rB)))]>;
3846 def : Pat<(setoeq (fabs R32FP:$rA), (fabs R32FP:$rB)),
3847 (FCMEQf32 R32FP:$rA, R32FP:$rB)>;
3850 RRForm<0b01000011010, (outs R32C:$rT), (ins R32FP:$rA, R32FP:$rB),
3851 "fcgt\t$rT, $rA, $rB", SPrecFP,
3852 [(set R32C:$rT, (setugt R32FP:$rA, R32FP:$rB))]>;
3854 def : Pat<(setugt R32FP:$rA, R32FP:$rB),
3855 (FCGTf32 R32FP:$rA, R32FP:$rB)>;
3858 RRForm<0b01010011010, (outs R32C:$rT), (ins R32FP:$rA, R32FP:$rB),
3859 "fcmgt\t$rT, $rA, $rB", SPrecFP,
3860 [(set R32C:$rT, (setugt (fabs R32FP:$rA), (fabs R32FP:$rB)))]>;
3862 def : Pat<(setugt (fabs R32FP:$rA), (fabs R32FP:$rB)),
3863 (FCMGTf32 R32FP:$rA, R32FP:$rB)>;
3865 //--------------------------------------------------------------------------
3866 // Single precision floating point comparisons and SETCC equivalents:
3867 //--------------------------------------------------------------------------
3869 def : SETCCNegCondReg<setune, R32FP, i32, XORIr32, FCEQf32>;
3870 def : SETCCNegCondReg<setone, R32FP, i32, XORIr32, FCEQf32>;
3872 def : SETCCBinOpReg<setuge, R32FP, ORr32, FCGTf32, FCEQf32>;
3873 def : SETCCBinOpReg<setoge, R32FP, ORr32, FCGTf32, FCEQf32>;
3875 def : SETCCBinOpReg<setult, R32FP, NORr32, FCGTf32, FCEQf32>;
3876 def : SETCCBinOpReg<setolt, R32FP, NORr32, FCGTf32, FCEQf32>;
3878 def : Pat<(setule R32FP:$rA, R32FP:$rB),
3879 (XORIr32 (FCGTf32 R32FP:$rA, R32FP:$rB), 0xffffffff)>;
3880 def : Pat<(setole R32FP:$rA, R32FP:$rB),
3881 (XORIr32 (FCGTf32 R32FP:$rA, R32FP:$rB), 0xffffffff)>;
3883 // FP Status and Control Register Write
3884 // Why isn't rT a don't care in the ISA?
3885 // Should we create a special RRForm_3 for this guy and zero out the rT?
3887 RRForm_1<0b01011101110, (outs R32FP:$rT), (ins R32FP:$rA),
3888 "fscrwr\t$rA", SPrecFP,
3889 [/* This instruction requires an intrinsic. Note: rT is unused. */]>;
3891 // FP Status and Control Register Read
3893 RRForm_2<0b01011101110, (outs R32FP:$rT), (ins),
3894 "fscrrd\t$rT", SPrecFP,
3895 [/* This instruction requires an intrinsic */]>;
3897 // llvm instruction space
3898 // How do these map onto cell instructions?
3900 // frest rC rB # c = 1/b (both lines)
3902 // fm rD rA rC # d = a * 1/b
3903 // fnms rB rD rB rA # b = - (d * b - a) --should == 0 in a perfect world
3904 // fma rB rB rC rD # b = b * c + d
3905 // = -(d *b -a) * c + d
3906 // = a * c - c ( a *b *c - a)
3911 // These llvm instructions will actually map to library calls.
3912 // All that's needed, then, is to check that the appropriate library is
3913 // imported and do a brsl to the proper function name.
3914 // frem # fmod(x, y): x - (x/y) * y
3915 // (Note: fmod(double, double), fmodf(float,float)
3919 // Unimplemented SPU instruction space
3920 // floating reciprocal absolute square root estimate (frsqest)
3922 // The following are probably just intrinsics
3923 // status and control register write
3924 // status and control register read
3926 //--------------------------------------
3927 // Floating point multiply instructions
3928 //--------------------------------------
3931 RRForm<0b00100011010, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3932 "fm\t$rT, $rA, $rB", SPrecFP,
3933 [(set (v4f32 VECREG:$rT), (fmul (v4f32 VECREG:$rA),
3934 (v4f32 VECREG:$rB)))]>;
3937 RRForm<0b01100011010, (outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB),
3938 "fm\t$rT, $rA, $rB", SPrecFP,
3939 [(set R32FP:$rT, (fmul R32FP:$rA, R32FP:$rB))]>;
3941 // Floating point multiply and add
3942 // e.g. d = c + (a * b)
3944 RRRForm<0b0111, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
3945 "fma\t$rT, $rA, $rB, $rC", SPrecFP,
3946 [(set (v4f32 VECREG:$rT),
3947 (fadd (v4f32 VECREG:$rC),
3948 (fmul (v4f32 VECREG:$rA), (v4f32 VECREG:$rB))))]>;
3951 RRRForm<0b0111, (outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB, R32FP:$rC),
3952 "fma\t$rT, $rA, $rB, $rC", SPrecFP,
3953 [(set R32FP:$rT, (fadd R32FP:$rC, (fmul R32FP:$rA, R32FP:$rB)))]>;
3955 // FP multiply and subtract
3956 // Subtracts value in rC from product
3959 RRRForm<0b0111, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
3960 "fms\t$rT, $rA, $rB, $rC", SPrecFP,
3961 [(set (v4f32 VECREG:$rT),
3962 (fsub (fmul (v4f32 VECREG:$rA), (v4f32 VECREG:$rB)),
3963 (v4f32 VECREG:$rC)))]>;
3966 RRRForm<0b0111, (outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB, R32FP:$rC),
3967 "fms\t$rT, $rA, $rB, $rC", SPrecFP,
3969 (fsub (fmul R32FP:$rA, R32FP:$rB), R32FP:$rC))]>;
3971 // Floating Negative Mulitply and Subtract
3972 // Subtracts product from value in rC
3973 // res = fneg(fms a b c)
3976 // NOTE: subtraction order
3980 RRRForm<0b1101, (outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB, R32FP:$rC),
3981 "fnms\t$rT, $rA, $rB, $rC", SPrecFP,
3982 [(set R32FP:$rT, (fsub R32FP:$rC, (fmul R32FP:$rA, R32FP:$rB)))]>;
3985 RRRForm<0b1101, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
3986 "fnms\t$rT, $rA, $rB, $rC", SPrecFP,
3987 [(set (v4f32 VECREG:$rT),
3988 (fsub (v4f32 VECREG:$rC),
3989 (fmul (v4f32 VECREG:$rA),
3990 (v4f32 VECREG:$rB))))]>;
3992 //--------------------------------------
3993 // Floating Point Conversions
3994 // Signed conversions:
3996 CVTIntFPForm<0b0101101110, (outs VECREG:$rT), (ins VECREG:$rA),
3997 "csflt\t$rT, $rA, 0", SPrecFP,
3998 [(set (v4f32 VECREG:$rT), (sint_to_fp (v4i32 VECREG:$rA)))]>;
4000 // Convert signed integer to floating point
4002 CVTIntFPForm<0b0101101110, (outs R32FP:$rT), (ins R32C:$rA),
4003 "csflt\t$rT, $rA, 0", SPrecFP,
4004 [(set R32FP:$rT, (sint_to_fp R32C:$rA))]>;
4006 // Convert unsigned into to float
4008 CVTIntFPForm<0b1101101110, (outs VECREG:$rT), (ins VECREG:$rA),
4009 "cuflt\t$rT, $rA, 0", SPrecFP,
4010 [(set (v4f32 VECREG:$rT), (uint_to_fp (v4i32 VECREG:$rA)))]>;
4013 CVTIntFPForm<0b1101101110, (outs R32FP:$rT), (ins R32C:$rA),
4014 "cuflt\t$rT, $rA, 0", SPrecFP,
4015 [(set R32FP:$rT, (uint_to_fp R32C:$rA))]>;
4017 // Convert float to unsigned int
4018 // Assume that scale = 0
4021 CVTIntFPForm<0b1101101110, (outs VECREG:$rT), (ins VECREG:$rA),
4022 "cfltu\t$rT, $rA, 0", SPrecFP,
4023 [(set (v4i32 VECREG:$rT), (fp_to_uint (v4f32 VECREG:$rA)))]>;
4026 CVTIntFPForm<0b1101101110, (outs R32C:$rT), (ins R32FP:$rA),
4027 "cfltu\t$rT, $rA, 0", SPrecFP,
4028 [(set R32C:$rT, (fp_to_uint R32FP:$rA))]>;
4030 // Convert float to signed int
4031 // Assume that scale = 0
4034 CVTIntFPForm<0b1101101110, (outs VECREG:$rT), (ins VECREG:$rA),
4035 "cflts\t$rT, $rA, 0", SPrecFP,
4036 [(set (v4i32 VECREG:$rT), (fp_to_sint (v4f32 VECREG:$rA)))]>;
4039 CVTIntFPForm<0b1101101110, (outs R32C:$rT), (ins R32FP:$rA),
4040 "cflts\t$rT, $rA, 0", SPrecFP,
4041 [(set R32C:$rT, (fp_to_sint R32FP:$rA))]>;
4043 //===----------------------------------------------------------------------==//
4044 // Single<->Double precision conversions
4045 //===----------------------------------------------------------------------==//
4047 // NOTE: We use "vec" name suffix here to avoid confusion (e.g. input is a
4048 // v4f32, output is v2f64--which goes in the name?)
4050 // Floating point extend single to double
4051 // NOTE: Not sure if passing in v4f32 to FESDvec is correct since it
4052 // operates on two double-word slots (i.e. 1st and 3rd fp numbers
4055 RRForm_1<0b00011101110, (outs VECREG:$rT), (ins VECREG:$rA),
4056 "fesd\t$rT, $rA", SPrecFP,
4057 [(set (v2f64 VECREG:$rT), (fextend (v4f32 VECREG:$rA)))]>;
4060 RRForm_1<0b00011101110, (outs R64FP:$rT), (ins R32FP:$rA),
4061 "fesd\t$rT, $rA", SPrecFP,
4062 [(set R64FP:$rT, (fextend R32FP:$rA))]>;
4064 // Floating point round double to single
4066 // RRForm_1<0b10011101110, (outs VECREG:$rT), (ins VECREG:$rA),
4067 // "frds\t$rT, $rA,", SPrecFP,
4068 // [(set (v4f32 R32FP:$rT), (fround (v2f64 R64FP:$rA)))]>;
4071 RRForm_1<0b10011101110, (outs R32FP:$rT), (ins R64FP:$rA),
4072 "frds\t$rT, $rA", SPrecFP,
4073 [(set R32FP:$rT, (fround R64FP:$rA))]>;
4075 //ToDo include anyextend?
4077 //===----------------------------------------------------------------------==//
4078 // Double precision floating point instructions
4079 //===----------------------------------------------------------------------==//
4081 RRForm<0b00110011010, (outs R64FP:$rT), (ins R64FP:$rA, R64FP:$rB),
4082 "dfa\t$rT, $rA, $rB", DPrecFP,
4083 [(set R64FP:$rT, (fadd R64FP:$rA, R64FP:$rB))]>;
4086 RRForm<0b00110011010, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
4087 "dfa\t$rT, $rA, $rB", DPrecFP,
4088 [(set (v2f64 VECREG:$rT), (fadd (v2f64 VECREG:$rA), (v2f64 VECREG:$rB)))]>;
4091 RRForm<0b10100011010, (outs R64FP:$rT), (ins R64FP:$rA, R64FP:$rB),
4092 "dfs\t$rT, $rA, $rB", DPrecFP,
4093 [(set R64FP:$rT, (fsub R64FP:$rA, R64FP:$rB))]>;
4096 RRForm<0b10100011010, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
4097 "dfs\t$rT, $rA, $rB", DPrecFP,
4098 [(set (v2f64 VECREG:$rT),
4099 (fsub (v2f64 VECREG:$rA), (v2f64 VECREG:$rB)))]>;
4102 RRForm<0b01100011010, (outs R64FP:$rT), (ins R64FP:$rA, R64FP:$rB),
4103 "dfm\t$rT, $rA, $rB", DPrecFP,
4104 [(set R64FP:$rT, (fmul R64FP:$rA, R64FP:$rB))]>;
4107 RRForm<0b00100011010, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
4108 "dfm\t$rT, $rA, $rB", DPrecFP,
4109 [(set (v2f64 VECREG:$rT),
4110 (fmul (v2f64 VECREG:$rA), (v2f64 VECREG:$rB)))]>;
4113 RRForm<0b00111010110, (outs R64FP:$rT),
4114 (ins R64FP:$rA, R64FP:$rB, R64FP:$rC),
4115 "dfma\t$rT, $rA, $rB", DPrecFP,
4116 [(set R64FP:$rT, (fadd R64FP:$rC, (fmul R64FP:$rA, R64FP:$rB)))]>,
4117 RegConstraint<"$rC = $rT">,
4121 RRForm<0b00111010110, (outs VECREG:$rT),
4122 (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
4123 "dfma\t$rT, $rA, $rB", DPrecFP,
4124 [(set (v2f64 VECREG:$rT),
4125 (fadd (v2f64 VECREG:$rC),
4126 (fmul (v2f64 VECREG:$rA), (v2f64 VECREG:$rB))))]>,
4127 RegConstraint<"$rC = $rT">,
4131 RRForm<0b10111010110, (outs R64FP:$rT),
4132 (ins R64FP:$rA, R64FP:$rB, R64FP:$rC),
4133 "dfms\t$rT, $rA, $rB", DPrecFP,
4134 [(set R64FP:$rT, (fsub (fmul R64FP:$rA, R64FP:$rB), R64FP:$rC))]>,
4135 RegConstraint<"$rC = $rT">,
4139 RRForm<0b10111010110, (outs VECREG:$rT),
4140 (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
4141 "dfms\t$rT, $rA, $rB", DPrecFP,
4142 [(set (v2f64 VECREG:$rT),
4143 (fsub (fmul (v2f64 VECREG:$rA), (v2f64 VECREG:$rB)),
4144 (v2f64 VECREG:$rC)))]>;
4146 // FNMS: - (a * b - c)
4147 // - (a * b) + c => c - (a * b)
4149 RRForm<0b01111010110, (outs R64FP:$rT),
4150 (ins R64FP:$rA, R64FP:$rB, R64FP:$rC),
4151 "dfnms\t$rT, $rA, $rB", DPrecFP,
4152 [(set R64FP:$rT, (fsub R64FP:$rC, (fmul R64FP:$rA, R64FP:$rB)))]>,
4153 RegConstraint<"$rC = $rT">,
4156 def : Pat<(fneg (fsub (fmul R64FP:$rA, R64FP:$rB), R64FP:$rC)),
4157 (FNMSf64 R64FP:$rA, R64FP:$rB, R64FP:$rC)>;
4160 RRForm<0b01111010110, (outs VECREG:$rT),
4161 (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
4162 "dfnms\t$rT, $rA, $rB", DPrecFP,
4163 [(set (v2f64 VECREG:$rT),
4164 (fsub (v2f64 VECREG:$rC),
4165 (fmul (v2f64 VECREG:$rA),
4166 (v2f64 VECREG:$rB))))]>,
4167 RegConstraint<"$rC = $rT">,
4170 def : Pat<(fneg (fsub (fmul (v2f64 VECREG:$rA), (v2f64 VECREG:$rB)),
4171 (v2f64 VECREG:$rC))),
4172 (FNMSv2f64 VECREG:$rA, VECREG:$rB, VECREG:$rC)>;
4177 RRForm<0b11111010110, (outs R64FP:$rT),
4178 (ins R64FP:$rA, R64FP:$rB, R64FP:$rC),
4179 "dfnma\t$rT, $rA, $rB", DPrecFP,
4180 [(set R64FP:$rT, (fneg (fadd R64FP:$rC, (fmul R64FP:$rA, R64FP:$rB))))]>,
4181 RegConstraint<"$rC = $rT">,
4185 RRForm<0b11111010110, (outs VECREG:$rT),
4186 (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
4187 "dfnma\t$rT, $rA, $rB", DPrecFP,
4188 [(set (v2f64 VECREG:$rT),
4189 (fneg (fadd (v2f64 VECREG:$rC),
4190 (fmul (v2f64 VECREG:$rA),
4191 (v2f64 VECREG:$rB)))))]>,
4192 RegConstraint<"$rC = $rT">,
4195 //===----------------------------------------------------------------------==//
4196 // Floating point negation and absolute value
4197 //===----------------------------------------------------------------------==//
4199 def : Pat<(fneg (v4f32 VECREG:$rA)),
4200 (XORfnegvec (v4f32 VECREG:$rA),
4201 (v4f32 (ILHUv4i32 0x8000)))>;
4203 def : Pat<(fneg R32FP:$rA),
4204 (XORfneg32 R32FP:$rA, (ILHUr32 0x8000))>;
4206 def : Pat<(fneg (v2f64 VECREG:$rA)),
4207 (XORfnegvec (v2f64 VECREG:$rA),
4208 (v2f64 (ANDBIv16i8 (FSMBIv16i8 0x8080), 0x80)))>;
4210 def : Pat<(fneg R64FP:$rA),
4211 (XORfneg64 R64FP:$rA,
4212 (ANDBIv16i8 (FSMBIv16i8 0x8080), 0x80))>;
4214 // Floating point absolute value
4216 def : Pat<(fabs R32FP:$rA),
4217 (ANDfabs32 R32FP:$rA, (IOHLr32 (ILHUr32 0x7fff), 0xffff))>;
4219 def : Pat<(fabs (v4f32 VECREG:$rA)),
4220 (ANDfabsvec (v4f32 VECREG:$rA),
4221 (v4f32 (ANDBIv16i8 (FSMBIv16i8 0xffff), 0x7f)))>;
4223 def : Pat<(fabs R64FP:$rA),
4224 (ANDfabs64 R64FP:$rA, (ANDBIv16i8 (FSMBIv16i8 0xffff), 0x7f))>;
4226 def : Pat<(fabs (v2f64 VECREG:$rA)),
4227 (ANDfabsvec (v2f64 VECREG:$rA),
4228 (v2f64 (ANDBIv16i8 (FSMBIv16i8 0xffff), 0x7f)))>;
4230 //===----------------------------------------------------------------------===//
4231 // Hint for branch instructions:
4232 //===----------------------------------------------------------------------===//
4234 /* def HBR : SPUInstr<(outs), (ins), "hbr\t" */
4236 //===----------------------------------------------------------------------===//
4237 // Execution, Load NOP (execute NOPs belong in even pipeline, load NOPs belong
4238 // in the odd pipeline)
4239 //===----------------------------------------------------------------------===//
4241 def ENOP : SPUInstr<(outs), (ins), "enop", ExecNOP> {
4244 let Inst{0-10} = 0b10000000010;
4245 let Inst{11-17} = 0;
4246 let Inst{18-24} = 0;
4247 let Inst{25-31} = 0;
4250 def LNOP : SPUInstr<(outs), (ins), "lnop", LoadNOP> {
4253 let Inst{0-10} = 0b10000000000;
4254 let Inst{11-17} = 0;
4255 let Inst{18-24} = 0;
4256 let Inst{25-31} = 0;
4259 //===----------------------------------------------------------------------===//
4260 // Bit conversions (type conversions between vector/packed types)
4261 // NOTE: Promotions are handled using the XS* instructions. Truncation
4263 //===----------------------------------------------------------------------===//
4264 def : Pat<(v16i8 (bitconvert (v8i16 VECREG:$src))), (v16i8 VECREG:$src)>;
4265 def : Pat<(v16i8 (bitconvert (v4i32 VECREG:$src))), (v16i8 VECREG:$src)>;
4266 def : Pat<(v16i8 (bitconvert (v2i64 VECREG:$src))), (v16i8 VECREG:$src)>;
4267 def : Pat<(v16i8 (bitconvert (v4f32 VECREG:$src))), (v16i8 VECREG:$src)>;
4268 def : Pat<(v16i8 (bitconvert (v2f64 VECREG:$src))), (v16i8 VECREG:$src)>;
4270 def : Pat<(v8i16 (bitconvert (v16i8 VECREG:$src))), (v8i16 VECREG:$src)>;
4271 def : Pat<(v8i16 (bitconvert (v4i32 VECREG:$src))), (v8i16 VECREG:$src)>;
4272 def : Pat<(v8i16 (bitconvert (v2i64 VECREG:$src))), (v8i16 VECREG:$src)>;
4273 def : Pat<(v8i16 (bitconvert (v4f32 VECREG:$src))), (v8i16 VECREG:$src)>;
4274 def : Pat<(v8i16 (bitconvert (v2f64 VECREG:$src))), (v8i16 VECREG:$src)>;
4276 def : Pat<(v4i32 (bitconvert (v16i8 VECREG:$src))), (v4i32 VECREG:$src)>;
4277 def : Pat<(v4i32 (bitconvert (v8i16 VECREG:$src))), (v4i32 VECREG:$src)>;
4278 def : Pat<(v4i32 (bitconvert (v2i64 VECREG:$src))), (v4i32 VECREG:$src)>;
4279 def : Pat<(v4i32 (bitconvert (v4f32 VECREG:$src))), (v4i32 VECREG:$src)>;
4280 def : Pat<(v4i32 (bitconvert (v2f64 VECREG:$src))), (v4i32 VECREG:$src)>;
4282 def : Pat<(v2i64 (bitconvert (v16i8 VECREG:$src))), (v2i64 VECREG:$src)>;
4283 def : Pat<(v2i64 (bitconvert (v8i16 VECREG:$src))), (v2i64 VECREG:$src)>;
4284 def : Pat<(v2i64 (bitconvert (v4i32 VECREG:$src))), (v2i64 VECREG:$src)>;
4285 def : Pat<(v2i64 (bitconvert (v4f32 VECREG:$src))), (v2i64 VECREG:$src)>;
4286 def : Pat<(v2i64 (bitconvert (v2f64 VECREG:$src))), (v2i64 VECREG:$src)>;
4288 def : Pat<(v4f32 (bitconvert (v16i8 VECREG:$src))), (v4f32 VECREG:$src)>;
4289 def : Pat<(v4f32 (bitconvert (v8i16 VECREG:$src))), (v4f32 VECREG:$src)>;
4290 def : Pat<(v4f32 (bitconvert (v2i64 VECREG:$src))), (v4f32 VECREG:$src)>;
4291 def : Pat<(v4f32 (bitconvert (v4i32 VECREG:$src))), (v4f32 VECREG:$src)>;
4292 def : Pat<(v4f32 (bitconvert (v2f64 VECREG:$src))), (v4f32 VECREG:$src)>;
4294 def : Pat<(v2f64 (bitconvert (v16i8 VECREG:$src))), (v2f64 VECREG:$src)>;
4295 def : Pat<(v2f64 (bitconvert (v8i16 VECREG:$src))), (v2f64 VECREG:$src)>;
4296 def : Pat<(v2f64 (bitconvert (v4i32 VECREG:$src))), (v2f64 VECREG:$src)>;
4297 def : Pat<(v2f64 (bitconvert (v2i64 VECREG:$src))), (v2f64 VECREG:$src)>;
4298 def : Pat<(v2f64 (bitconvert (v2f64 VECREG:$src))), (v2f64 VECREG:$src)>;
4300 def : Pat<(f32 (bitconvert (i32 R32C:$src))), (f32 R32FP:$src)>;
4301 def : Pat<(f64 (bitconvert (i64 R64C:$src))), (f64 R64FP:$src)>;
4303 //===----------------------------------------------------------------------===//
4304 // Instruction patterns:
4305 //===----------------------------------------------------------------------===//
4307 // General 32-bit constants:
4308 def : Pat<(i32 imm:$imm),
4309 (IOHLr32 (ILHUr32 (HI16 imm:$imm)), (LO16 imm:$imm))>;
4311 // Single precision float constants:
4312 def : Pat<(f32 fpimm:$imm),
4313 (IOHLf32 (ILHUf32 (HI16_f32 fpimm:$imm)), (LO16_f32 fpimm:$imm))>;
4315 // General constant 32-bit vectors
4316 def : Pat<(v4i32 v4i32Imm:$imm),
4317 (IOHLv4i32 (v4i32 (ILHUv4i32 (HI16_vec v4i32Imm:$imm))),
4318 (LO16_vec v4i32Imm:$imm))>;
4321 def : Pat<(i8 imm:$imm),
4324 //===----------------------------------------------------------------------===//
4325 // Call instruction patterns:
4326 //===----------------------------------------------------------------------===//
4331 //===----------------------------------------------------------------------===//
4332 // Zero/Any/Sign extensions
4333 //===----------------------------------------------------------------------===//
4335 // sext 8->32: Sign extend bytes to words
4336 def : Pat<(sext_inreg R32C:$rSrc, i8),
4337 (XSHWr32 (XSBHr32 R32C:$rSrc))>;
4339 def : Pat<(i32 (sext R8C:$rSrc)),
4340 (XSHWr16 (XSBHr8 R8C:$rSrc))>;
4342 // zext 8->16: Zero extend bytes to halfwords
4343 def : Pat<(i16 (zext R8C:$rSrc)),
4344 (ANDHIi8i16 R8C:$rSrc, 0xff)>;
4346 // zext 8->32: Zero extend bytes to words
4347 def : Pat<(i32 (zext R8C:$rSrc)),
4348 (ANDIi8i32 R8C:$rSrc, 0xff)>;
4350 // anyext 8->16: Extend 8->16 bits, irrespective of sign
4351 def : Pat<(i16 (anyext R8C:$rSrc)),
4352 (ORHIi8i16 R8C:$rSrc, 0)>;
4354 // anyext 8->32: Extend 8->32 bits, irrespective of sign
4355 def : Pat<(i32 (anyext R8C:$rSrc)),
4356 (ORIi8i32 R8C:$rSrc, 0)>;
4358 // zext 16->32: Zero extend halfwords to words
4359 def : Pat<(i32 (zext R16C:$rSrc)),
4360 (ANDi16i32 R16C:$rSrc, (ILAr32 0xffff))>;
4362 def : Pat<(i32 (zext (and R16C:$rSrc, 0xf))),
4363 (ANDIi16i32 R16C:$rSrc, 0xf)>;
4365 def : Pat<(i32 (zext (and R16C:$rSrc, 0xff))),
4366 (ANDIi16i32 R16C:$rSrc, 0xff)>;
4368 def : Pat<(i32 (zext (and R16C:$rSrc, 0xfff))),
4369 (ANDIi16i32 R16C:$rSrc, 0xfff)>;
4371 // anyext 16->32: Extend 16->32 bits, irrespective of sign
4372 def : Pat<(i32 (anyext R16C:$rSrc)),
4373 (ORIi16i32 R16C:$rSrc, 0)>;
4375 //===----------------------------------------------------------------------===//
4377 // These truncates are for the SPU's supported types (i8, i16, i32). i64 and
4378 // above are custom lowered.
4379 //===----------------------------------------------------------------------===//
4381 def : Pat<(i8 (trunc GPRC:$src)),
4383 (SHUFBgprc GPRC:$src, GPRC:$src,
4384 (IOHLv4i32 (ILHUv4i32 0x0f0f), 0x0f0f)))>;
4386 def : Pat<(i8 (trunc R64C:$src)),
4389 (ORv2i64_i64 R64C:$src),
4390 (ORv2i64_i64 R64C:$src),
4391 (IOHLv4i32 (ILHUv4i32 0x0707), 0x0707)))>;
4393 def : Pat<(i8 (trunc R32C:$src)),
4396 (ORv4i32_i32 R32C:$src),
4397 (ORv4i32_i32 R32C:$src),
4398 (IOHLv4i32 (ILHUv4i32 0x0303), 0x0303)))>;
4400 def : Pat<(i8 (trunc R16C:$src)),
4403 (ORv8i16_i16 R16C:$src),
4404 (ORv8i16_i16 R16C:$src),
4405 (IOHLv4i32 (ILHUv4i32 0x0303), 0x0303)))>;
4407 def : Pat<(i16 (trunc GPRC:$src)),
4409 (SHUFBgprc GPRC:$src, GPRC:$src,
4410 (IOHLv4i32 (ILHUv4i32 0x0e0f), 0x0e0f)))>;
4412 def : Pat<(i16 (trunc R64C:$src)),
4415 (ORv2i64_i64 R64C:$src),
4416 (ORv2i64_i64 R64C:$src),
4417 (IOHLv4i32 (ILHUv4i32 0x0607), 0x0607)))>;
4419 def : Pat<(i16 (trunc R32C:$src)),
4422 (ORv4i32_i32 R32C:$src),
4423 (ORv4i32_i32 R32C:$src),
4424 (IOHLv4i32 (ILHUv4i32 0x0203), 0x0203)))>;
4426 def : Pat<(i32 (trunc GPRC:$src)),
4428 (SHUFBgprc GPRC:$src, GPRC:$src,
4429 (IOHLv4i32 (ILHUv4i32 0x0c0d), 0x0e0f)))>;
4431 def : Pat<(i32 (trunc R64C:$src)),
4434 (ORv2i64_i64 R64C:$src),
4435 (ORv2i64_i64 R64C:$src),
4436 (IOHLv4i32 (ILHUv4i32 0x0405), 0x0607)))>;
4438 //===----------------------------------------------------------------------===//
4439 // Address generation: SPU, like PPC, has to split addresses into high and
4440 // low parts in order to load them into a register.
4441 //===----------------------------------------------------------------------===//
4443 def : Pat<(SPUaform tglobaladdr:$in, 0), (ILAlsa tglobaladdr:$in)>;
4444 def : Pat<(SPUaform texternalsym:$in, 0), (ILAlsa texternalsym:$in)>;
4445 def : Pat<(SPUaform tjumptable:$in, 0), (ILAlsa tjumptable:$in)>;
4446 def : Pat<(SPUaform tconstpool:$in, 0), (ILAlsa tconstpool:$in)>;
4448 def : Pat<(SPUindirect (SPUhi tglobaladdr:$in, 0),
4449 (SPUlo tglobaladdr:$in, 0)),
4450 (IOHLlo (ILHUhi tglobaladdr:$in), tglobaladdr:$in)>;
4452 def : Pat<(SPUindirect (SPUhi texternalsym:$in, 0),
4453 (SPUlo texternalsym:$in, 0)),
4454 (IOHLlo (ILHUhi texternalsym:$in), texternalsym:$in)>;
4456 def : Pat<(SPUindirect (SPUhi tjumptable:$in, 0),
4457 (SPUlo tjumptable:$in, 0)),
4458 (IOHLlo (ILHUhi tjumptable:$in), tjumptable:$in)>;
4460 def : Pat<(SPUindirect (SPUhi tconstpool:$in, 0),
4461 (SPUlo tconstpool:$in, 0)),
4462 (IOHLlo (ILHUhi tconstpool:$in), tconstpool:$in)>;
4465 def : Pat<(SPUindirect R32C:$sp, i32ImmSExt10:$imm),
4466 (AIr32 R32C:$sp, i32ImmSExt10:$imm)>;
4468 def : Pat<(SPUindirect R32C:$sp, imm:$imm),
4470 (IOHLr32 (ILHUr32 (HI16 imm:$imm)), (LO16 imm:$imm)))>;
4473 def : Pat<(add (SPUhi tglobaladdr:$in, 0), (SPUlo tglobaladdr:$in, 0)),
4474 (IOHLlo (ILHUhi tglobaladdr:$in), tglobaladdr:$in)>;
4476 def : Pat<(add (SPUhi texternalsym:$in, 0), (SPUlo texternalsym:$in, 0)),
4477 (IOHLlo (ILHUhi texternalsym:$in), texternalsym:$in)>;
4479 def : Pat<(add (SPUhi tjumptable:$in, 0), (SPUlo tjumptable:$in, 0)),
4480 (IOHLlo (ILHUhi tjumptable:$in), tjumptable:$in)>;
4482 def : Pat<(add (SPUhi tconstpool:$in, 0), (SPUlo tconstpool:$in, 0)),
4483 (IOHLlo (ILHUhi tconstpool:$in), tconstpool:$in)>;
4486 include "CellSDKIntrinsics.td"
4487 // Various math operator instruction sequences
4488 include "SPUMathInstr.td"
4489 // 64-bit "instructions"/support
4490 include "SPU64InstrInfo.td"