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:$src))]>
67 def v16i8: LoadDFormVec<v16i8>;
68 def v8i16: LoadDFormVec<v8i16>;
69 def v4i32: LoadDFormVec<v4i32>;
70 def v2i64: LoadDFormVec<v2i64>;
71 def v4f32: LoadDFormVec<v4f32>;
72 def v2f64: LoadDFormVec<v2f64>;
74 def v2i32: LoadDFormVec<v2i32>;
76 def r128: LoadDForm<GPRC>;
77 def r64: LoadDForm<R64C>;
78 def r32: LoadDForm<R32C>;
79 def f32: LoadDForm<R32FP>;
80 def f64: LoadDForm<R64FP>;
81 def r16: LoadDForm<R16C>;
82 def r8: LoadDForm<R8C>;
85 class LoadAFormVec<ValueType vectype>
86 : RI16Form<0b100001100, (outs VECREG:$rT), (ins addr256k:$src),
89 [(set (vectype VECREG:$rT), (load aform_addr:$src))]>
92 class LoadAForm<RegisterClass rclass>
93 : RI16Form<0b100001100, (outs rclass:$rT), (ins addr256k:$src),
96 [(set rclass:$rT, (load aform_addr:$src))]>
101 def v16i8: LoadAFormVec<v16i8>;
102 def v8i16: LoadAFormVec<v8i16>;
103 def v4i32: LoadAFormVec<v4i32>;
104 def v2i64: LoadAFormVec<v2i64>;
105 def v4f32: LoadAFormVec<v4f32>;
106 def v2f64: LoadAFormVec<v2f64>;
108 def v2i32: LoadAFormVec<v2i32>;
110 def r128: LoadAForm<GPRC>;
111 def r64: LoadAForm<R64C>;
112 def r32: LoadAForm<R32C>;
113 def f32: LoadAForm<R32FP>;
114 def f64: LoadAForm<R64FP>;
115 def r16: LoadAForm<R16C>;
116 def r8: LoadAForm<R8C>;
119 class LoadXFormVec<ValueType vectype>
120 : RRForm<0b00100011100, (outs VECREG:$rT), (ins memrr:$src),
123 [(set (vectype VECREG:$rT), (load xform_addr:$src))]>
126 class LoadXForm<RegisterClass rclass>
127 : RRForm<0b00100011100, (outs rclass:$rT), (ins memrr:$src),
130 [(set rclass:$rT, (load xform_addr:$src))]>
133 multiclass LoadXForms
135 def v16i8: LoadXFormVec<v16i8>;
136 def v8i16: LoadXFormVec<v8i16>;
137 def v4i32: LoadXFormVec<v4i32>;
138 def v2i64: LoadXFormVec<v2i64>;
139 def v4f32: LoadXFormVec<v4f32>;
140 def v2f64: LoadXFormVec<v2f64>;
142 def v2i32: LoadXFormVec<v2i32>;
144 def r128: LoadXForm<GPRC>;
145 def r64: LoadXForm<R64C>;
146 def r32: LoadXForm<R32C>;
147 def f32: LoadXForm<R32FP>;
148 def f64: LoadXForm<R64FP>;
149 def r16: LoadXForm<R16C>;
150 def r8: LoadXForm<R8C>;
153 defm LQA : LoadAForms;
154 defm LQD : LoadDForms;
155 defm LQX : LoadXForms;
157 /* Load quadword, PC relative: Not much use at this point in time.
158 Might be of use later for relocatable code. It's effectively the
159 same as LQA, but uses PC-relative addressing.
160 def LQR : RI16Form<0b111001100, (outs VECREG:$rT), (ins s16imm:$disp),
161 "lqr\t$rT, $disp", LoadStore,
162 [(set VECREG:$rT, (load iaddr:$disp))]>;
166 //===----------------------------------------------------------------------===//
168 //===----------------------------------------------------------------------===//
169 class StoreDFormVec<ValueType vectype>
170 : RI10Form<0b00100100, (outs), (ins VECREG:$rT, dformaddr:$src),
173 [(store (vectype VECREG:$rT), dform_addr:$src)]>
176 class StoreDForm<RegisterClass rclass>
177 : RI10Form<0b00100100, (outs), (ins rclass:$rT, dformaddr:$src),
180 [(store rclass:$rT, dform_addr:$src)]>
183 multiclass StoreDForms
185 def v16i8: StoreDFormVec<v16i8>;
186 def v8i16: StoreDFormVec<v8i16>;
187 def v4i32: StoreDFormVec<v4i32>;
188 def v2i64: StoreDFormVec<v2i64>;
189 def v4f32: StoreDFormVec<v4f32>;
190 def v2f64: StoreDFormVec<v2f64>;
192 def v2i32: StoreDFormVec<v2i32>;
194 def r128: StoreDForm<GPRC>;
195 def r64: StoreDForm<R64C>;
196 def r32: StoreDForm<R32C>;
197 def f32: StoreDForm<R32FP>;
198 def f64: StoreDForm<R64FP>;
199 def r16: StoreDForm<R16C>;
200 def r8: StoreDForm<R8C>;
203 class StoreAFormVec<ValueType vectype>
204 : RI16Form<0b0010010, (outs), (ins VECREG:$rT, addr256k:$src),
207 [(store (vectype VECREG:$rT), aform_addr:$src)]>;
209 class StoreAForm<RegisterClass rclass>
210 : RI16Form<0b001001, (outs), (ins rclass:$rT, addr256k:$src),
213 [(store rclass:$rT, aform_addr:$src)]>;
215 multiclass StoreAForms
217 def v16i8: StoreAFormVec<v16i8>;
218 def v8i16: StoreAFormVec<v8i16>;
219 def v4i32: StoreAFormVec<v4i32>;
220 def v2i64: StoreAFormVec<v2i64>;
221 def v4f32: StoreAFormVec<v4f32>;
222 def v2f64: StoreAFormVec<v2f64>;
224 def v2i32: StoreAFormVec<v2i32>;
226 def r128: StoreAForm<GPRC>;
227 def r64: StoreAForm<R64C>;
228 def r32: StoreAForm<R32C>;
229 def f32: StoreAForm<R32FP>;
230 def f64: StoreAForm<R64FP>;
231 def r16: StoreAForm<R16C>;
232 def r8: StoreAForm<R8C>;
235 class StoreXFormVec<ValueType vectype>
236 : RRForm<0b00100100, (outs), (ins VECREG:$rT, memrr:$src),
239 [(store (vectype VECREG:$rT), xform_addr:$src)]>
242 class StoreXForm<RegisterClass rclass>
243 : RRForm<0b00100100, (outs), (ins rclass:$rT, memrr:$src),
246 [(store rclass:$rT, xform_addr:$src)]>
249 multiclass StoreXForms
251 def v16i8: StoreXFormVec<v16i8>;
252 def v8i16: StoreXFormVec<v8i16>;
253 def v4i32: StoreXFormVec<v4i32>;
254 def v2i64: StoreXFormVec<v2i64>;
255 def v4f32: StoreXFormVec<v4f32>;
256 def v2f64: StoreXFormVec<v2f64>;
258 def v2i32: StoreXFormVec<v2i32>;
260 def r128: StoreXForm<GPRC>;
261 def r64: StoreXForm<R64C>;
262 def r32: StoreXForm<R32C>;
263 def f32: StoreXForm<R32FP>;
264 def f64: StoreXForm<R64FP>;
265 def r16: StoreXForm<R16C>;
266 def r8: StoreXForm<R8C>;
269 defm STQD : StoreDForms;
270 defm STQA : StoreAForms;
271 defm STQX : StoreXForms;
273 /* Store quadword, PC relative: Not much use at this point in time. Might
274 be useful for relocatable code.
275 def STQR : RI16Form<0b111000100, (outs), (ins VECREG:$rT, s16imm:$disp),
276 "stqr\t$rT, $disp", LoadStore,
277 [(store VECREG:$rT, iaddr:$disp)]>;
280 //===----------------------------------------------------------------------===//
281 // Generate Controls for Insertion:
282 //===----------------------------------------------------------------------===//
284 def CBD: RI7Form<0b10101111100, (outs VECREG:$rT), (ins shufaddr:$src),
285 "cbd\t$rT, $src", ShuffleOp,
286 [(set (v16i8 VECREG:$rT), (SPUshufmask dform2_addr:$src))]>;
288 def CBX: RRForm<0b00101011100, (outs VECREG:$rT), (ins memrr:$src),
289 "cbx\t$rT, $src", ShuffleOp,
290 [(set (v16i8 VECREG:$rT), (SPUshufmask xform_addr:$src))]>;
292 def CHD: RI7Form<0b10101111100, (outs VECREG:$rT), (ins shufaddr:$src),
293 "chd\t$rT, $src", ShuffleOp,
294 [(set (v8i16 VECREG:$rT), (SPUshufmask dform2_addr:$src))]>;
296 def CHX: RRForm<0b10101011100, (outs VECREG:$rT), (ins memrr:$src),
297 "chx\t$rT, $src", ShuffleOp,
298 [(set (v8i16 VECREG:$rT), (SPUshufmask xform_addr:$src))]>;
300 def CWD: RI7Form<0b01101111100, (outs VECREG:$rT), (ins shufaddr:$src),
301 "cwd\t$rT, $src", ShuffleOp,
302 [(set (v4i32 VECREG:$rT), (SPUshufmask dform2_addr:$src))]>;
304 def CWX: RRForm<0b01101011100, (outs VECREG:$rT), (ins memrr:$src),
305 "cwx\t$rT, $src", ShuffleOp,
306 [(set (v4i32 VECREG:$rT), (SPUshufmask xform_addr:$src))]>;
308 def CWDf32: RI7Form<0b01101111100, (outs VECREG:$rT), (ins shufaddr:$src),
309 "cwd\t$rT, $src", ShuffleOp,
310 [(set (v4f32 VECREG:$rT), (SPUshufmask dform2_addr:$src))]>;
312 def CWXf32: RRForm<0b01101011100, (outs VECREG:$rT), (ins memrr:$src),
313 "cwx\t$rT, $src", ShuffleOp,
314 [(set (v4f32 VECREG:$rT), (SPUshufmask xform_addr:$src))]>;
316 def CDD: RI7Form<0b11101111100, (outs VECREG:$rT), (ins shufaddr:$src),
317 "cdd\t$rT, $src", ShuffleOp,
318 [(set (v2i64 VECREG:$rT), (SPUshufmask dform2_addr:$src))]>;
320 def CDX: RRForm<0b11101011100, (outs VECREG:$rT), (ins memrr:$src),
321 "cdx\t$rT, $src", ShuffleOp,
322 [(set (v2i64 VECREG:$rT), (SPUshufmask xform_addr:$src))]>;
324 def CDDf64: RI7Form<0b11101111100, (outs VECREG:$rT), (ins shufaddr:$src),
325 "cdd\t$rT, $src", ShuffleOp,
326 [(set (v2f64 VECREG:$rT), (SPUshufmask dform2_addr:$src))]>;
328 def CDXf64: RRForm<0b11101011100, (outs VECREG:$rT), (ins memrr:$src),
329 "cdx\t$rT, $src", ShuffleOp,
330 [(set (v2f64 VECREG:$rT), (SPUshufmask xform_addr:$src))]>;
332 //===----------------------------------------------------------------------===//
333 // Constant formation:
334 //===----------------------------------------------------------------------===//
337 RI16Form<0b110000010, (outs VECREG:$rT), (ins s16imm:$val),
338 "ilh\t$rT, $val", ImmLoad,
339 [(set (v8i16 VECREG:$rT), (v8i16 v8i16SExt16Imm:$val))]>;
342 RI16Form<0b110000010, (outs R16C:$rT), (ins s16imm:$val),
343 "ilh\t$rT, $val", ImmLoad,
344 [(set R16C:$rT, immSExt16:$val)]>;
346 // Cell SPU doesn't have a native 8-bit immediate load, but ILH works ("with
347 // the right constant")
349 RI16Form<0b110000010, (outs R8C:$rT), (ins s16imm_i8:$val),
350 "ilh\t$rT, $val", ImmLoad,
351 [(set R8C:$rT, immSExt8:$val)]>;
353 // IL does sign extension!
355 class ILInst<dag OOL, dag IOL, list<dag> pattern>:
356 RI16Form<0b100000010, OOL, IOL, "il\t$rT, $val",
359 class ILVecInst<ValueType vectype, Operand immtype, PatLeaf xform>:
360 ILInst<(outs VECREG:$rT), (ins immtype:$val),
361 [(set (vectype VECREG:$rT), (vectype xform:$val))]>;
363 class ILRegInst<RegisterClass rclass, Operand immtype, PatLeaf xform>:
364 ILInst<(outs rclass:$rT), (ins immtype:$val),
365 [(set rclass:$rT, xform:$val)]>;
367 multiclass ImmediateLoad
369 def v2i64: ILVecInst<v2i64, s16imm_i64, v2i64SExt16Imm>;
370 def v4i32: ILVecInst<v4i32, s16imm_i32, v4i32SExt16Imm>;
372 // TODO: Need v2f64, v4f32
374 def r64: ILRegInst<R64C, s16imm_i64, immSExt16>;
375 def r32: ILRegInst<R32C, s16imm_i32, immSExt16>;
376 def f32: ILRegInst<R32FP, s16imm_f32, fpimmSExt16>;
377 def f64: ILRegInst<R64FP, s16imm_f64, fpimmSExt16>;
380 defm IL : ImmediateLoad;
382 class ILHUInst<dag OOL, dag IOL, list<dag> pattern>:
383 RI16Form<0b010000010, OOL, IOL, "ilhu\t$rT, $val",
386 class ILHUVecInst<ValueType vectype, Operand immtype, PatLeaf xform>:
387 ILHUInst<(outs VECREG:$rT), (ins immtype:$val),
388 [(set (vectype VECREG:$rT), (vectype xform:$val))]>;
390 class ILHURegInst<RegisterClass rclass, Operand immtype, PatLeaf xform>:
391 ILHUInst<(outs rclass:$rT), (ins immtype:$val),
392 [(set rclass:$rT, xform:$val)]>;
394 multiclass ImmLoadHalfwordUpper
396 def v2i64: ILHUVecInst<v2i64, u16imm_i64, immILHUvec_i64>;
397 def v4i32: ILHUVecInst<v4i32, u16imm_i32, immILHUvec>;
399 def r64: ILHURegInst<R64C, u16imm_i64, hi16>;
400 def r32: ILHURegInst<R32C, u16imm_i32, hi16>;
402 // Loads the high portion of an address
403 def hi: ILHURegInst<R32C, symbolHi, hi16>;
405 // Used in custom lowering constant SFP loads:
406 def f32: ILHURegInst<R32FP, f16imm, hi16_f32>;
409 defm ILHU : ImmLoadHalfwordUpper;
411 // Immediate load address (can also be used to load 18-bit unsigned constants,
412 // see the zext 16->32 pattern)
414 class ILAInst<dag OOL, dag IOL, list<dag> pattern>:
415 RI18Form<0b1000010, OOL, IOL, "ila\t$rT, $val",
418 class ILAVecInst<ValueType vectype, Operand immtype, PatLeaf xform>:
419 ILAInst<(outs VECREG:$rT), (ins immtype:$val),
420 [(set (vectype VECREG:$rT), (vectype xform:$val))]>;
422 class ILARegInst<RegisterClass rclass, Operand immtype, PatLeaf xform>:
423 ILAInst<(outs rclass:$rT), (ins immtype:$val),
424 [(set rclass:$rT, xform:$val)]>;
426 multiclass ImmLoadAddress
428 def v2i64: ILAVecInst<v2i64, u18imm, v2i64Uns18Imm>;
429 def v4i32: ILAVecInst<v4i32, u18imm, v4i32Uns18Imm>;
431 def r64: ILARegInst<R64C, u18imm_i64, imm18>;
432 def r32: ILARegInst<R32C, u18imm, imm18>;
433 def f32: ILARegInst<R32FP, f18imm, fpimm18>;
434 def f64: ILARegInst<R64FP, f18imm_f64, fpimm18>;
436 def hi: ILARegInst<R32C, symbolHi, imm18>;
437 def lo: ILARegInst<R32C, symbolLo, imm18>;
439 def lsa: ILAInst<(outs R32C:$rT), (ins symbolLSA:$val),
443 defm ILA : ImmLoadAddress;
445 // Immediate OR, Halfword Lower: The "other" part of loading large constants
446 // into 32-bit registers. See the anonymous pattern Pat<(i32 imm:$imm), ...>
447 // Note that these are really two operand instructions, but they're encoded
448 // as three operands with the first two arguments tied-to each other.
450 class IOHLInst<dag OOL, dag IOL, list<dag> pattern>:
451 RI16Form<0b100000110, OOL, IOL, "iohl\t$rT, $val",
453 RegConstraint<"$rS = $rT">,
456 class IOHLVecInst<ValueType vectype, Operand immtype /* , PatLeaf xform */>:
457 IOHLInst<(outs VECREG:$rT), (ins VECREG:$rS, immtype:$val),
460 class IOHLRegInst<RegisterClass rclass, Operand immtype /* , PatLeaf xform */>:
461 IOHLInst<(outs rclass:$rT), (ins rclass:$rS, immtype:$val),
464 multiclass ImmOrHalfwordLower
466 def v2i64: IOHLVecInst<v2i64, u16imm_i64>;
467 def v4i32: IOHLVecInst<v4i32, u16imm_i32>;
469 def r32: IOHLRegInst<R32C, i32imm>;
470 def f32: IOHLRegInst<R32FP, f32imm>;
472 def lo: IOHLRegInst<R32C, symbolLo>;
475 defm IOHL: ImmOrHalfwordLower;
477 // Form select mask for bytes using immediate, used in conjunction with the
480 class FSMBIVec<ValueType vectype>:
481 RI16Form<0b101001100, (outs VECREG:$rT), (ins u16imm:$val),
484 [(set (vectype VECREG:$rT), (SPUselmask (i16 immU16:$val)))]>;
486 multiclass FormSelectMaskBytesImm
488 def v16i8: FSMBIVec<v16i8>;
489 def v8i16: FSMBIVec<v8i16>;
490 def v4i32: FSMBIVec<v4i32>;
491 def v2i64: FSMBIVec<v2i64>;
494 defm FSMBI : FormSelectMaskBytesImm;
496 // fsmb: Form select mask for bytes. N.B. Input operand, $rA, is 16-bits
497 class FSMBInst<dag OOL, dag IOL, list<dag> pattern>:
498 RRForm_1<0b01101101100, OOL, IOL, "fsmb\t$rT, $rA", SelectOp,
501 class FSMBRegInst<RegisterClass rclass, ValueType vectype>:
502 FSMBInst<(outs VECREG:$rT), (ins rclass:$rA),
503 [(set (vectype VECREG:$rT), (SPUselmask rclass:$rA))]>;
505 class FSMBVecInst<ValueType vectype>:
506 FSMBInst<(outs VECREG:$rT), (ins VECREG:$rA),
507 [(set (vectype VECREG:$rT),
508 (SPUselmask (vectype VECREG:$rA)))]>;
510 multiclass FormSelectMaskBits {
511 def v16i8_r16: FSMBRegInst<R16C, v16i8>;
512 def v16i8: FSMBVecInst<v16i8>;
515 defm FSMB: FormSelectMaskBits;
517 // fsmh: Form select mask for halfwords. N.B., Input operand, $rA, is
518 // only 8-bits wide (even though it's input as 16-bits here)
520 class FSMHInst<dag OOL, dag IOL, list<dag> pattern>:
521 RRForm_1<0b10101101100, OOL, IOL, "fsmh\t$rT, $rA", SelectOp,
524 class FSMHRegInst<RegisterClass rclass, ValueType vectype>:
525 FSMHInst<(outs VECREG:$rT), (ins rclass:$rA),
526 [(set (vectype VECREG:$rT), (SPUselmask rclass:$rA))]>;
528 class FSMHVecInst<ValueType vectype>:
529 FSMHInst<(outs VECREG:$rT), (ins VECREG:$rA),
530 [(set (vectype VECREG:$rT),
531 (SPUselmask (vectype VECREG:$rA)))]>;
533 multiclass FormSelectMaskHalfword {
534 def v8i16_r16: FSMHRegInst<R16C, v8i16>;
535 def v8i16: FSMHVecInst<v8i16>;
538 defm FSMH: FormSelectMaskHalfword;
540 // fsm: Form select mask for words. Like the other fsm* instructions,
541 // only the lower 4 bits of $rA are significant.
543 class FSMInst<dag OOL, dag IOL, list<dag> pattern>:
544 RRForm_1<0b00101101100, OOL, IOL, "fsm\t$rT, $rA", SelectOp,
547 class FSMRegInst<ValueType vectype, RegisterClass rclass>:
548 FSMInst<(outs VECREG:$rT), (ins rclass:$rA),
549 [(set (vectype VECREG:$rT), (SPUselmask rclass:$rA))]>;
551 class FSMVecInst<ValueType vectype>:
552 FSMInst<(outs VECREG:$rT), (ins VECREG:$rA),
553 [(set (vectype VECREG:$rT), (SPUselmask (vectype VECREG:$rA)))]>;
555 multiclass FormSelectMaskWord {
556 def v4i32: FSMVecInst<v4i32>;
558 def r32 : FSMRegInst<v4i32, R32C>;
559 def r16 : FSMRegInst<v4i32, R16C>;
562 defm FSM : FormSelectMaskWord;
564 // Special case when used for i64 math operations
565 multiclass FormSelectMaskWord64 {
566 def r32 : FSMRegInst<v2i64, R32C>;
567 def r16 : FSMRegInst<v2i64, R16C>;
570 defm FSM64 : FormSelectMaskWord64;
572 //===----------------------------------------------------------------------===//
573 // Integer and Logical Operations:
574 //===----------------------------------------------------------------------===//
577 RRForm<0b00010011000, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
578 "ah\t$rT, $rA, $rB", IntegerOp,
579 [(set (v8i16 VECREG:$rT), (int_spu_si_ah VECREG:$rA, VECREG:$rB))]>;
581 def : Pat<(add (v8i16 VECREG:$rA), (v8i16 VECREG:$rB)),
582 (AHv8i16 VECREG:$rA, VECREG:$rB)>;
585 RRForm<0b00010011000, (outs R16C:$rT), (ins R16C:$rA, R16C:$rB),
586 "ah\t$rT, $rA, $rB", IntegerOp,
587 [(set R16C:$rT, (add R16C:$rA, R16C:$rB))]>;
590 RI10Form<0b10111000, (outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
591 "ahi\t$rT, $rA, $val", IntegerOp,
592 [(set (v8i16 VECREG:$rT), (add (v8i16 VECREG:$rA),
593 v8i16SExt10Imm:$val))]>;
596 RI10Form<0b10111000, (outs R16C:$rT), (ins R16C:$rA, s10imm:$val),
597 "ahi\t$rT, $rA, $val", IntegerOp,
598 [(set R16C:$rT, (add R16C:$rA, i16ImmSExt10:$val))]>;
600 // v4i32, i32 add instruction:
602 class AInst<dag OOL, dag IOL, list<dag> pattern>:
603 RRForm<0b00000011000, OOL, IOL,
604 "a\t$rT, $rA, $rB", IntegerOp,
607 class AVecInst<ValueType vectype>:
608 AInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
609 [(set (vectype VECREG:$rT), (add (vectype VECREG:$rA),
610 (vectype VECREG:$rB)))]>;
612 class ARegInst<RegisterClass rclass>:
613 AInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
614 [(set rclass:$rT, (add rclass:$rA, rclass:$rB))]>;
616 multiclass AddInstruction {
617 def v4i32: AVecInst<v4i32>;
618 def v16i8: AVecInst<v16i8>;
620 def r32: ARegInst<R32C>;
623 defm A : AddInstruction;
625 class AIInst<dag OOL, dag IOL, list<dag> pattern>:
626 RI10Form<0b00111000, OOL, IOL,
627 "ai\t$rT, $rA, $val", IntegerOp,
630 class AIVecInst<ValueType vectype, PatLeaf immpred>:
631 AIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
632 [(set (vectype VECREG:$rT), (add (vectype VECREG:$rA), immpred:$val))]>;
634 class AIFPVecInst<ValueType vectype, PatLeaf immpred>:
635 AIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
638 class AIRegInst<RegisterClass rclass, PatLeaf immpred>:
639 AIInst<(outs rclass:$rT), (ins rclass:$rA, s10imm_i32:$val),
640 [(set rclass:$rT, (add rclass:$rA, immpred:$val))]>;
642 // This is used to add epsilons to floating point numbers in the f32 fdiv code:
643 class AIFPInst<RegisterClass rclass, PatLeaf immpred>:
644 AIInst<(outs rclass:$rT), (ins rclass:$rA, s10imm_i32:$val),
647 multiclass AddImmediate {
648 def v4i32: AIVecInst<v4i32, v4i32SExt10Imm>;
650 def r32: AIRegInst<R32C, i32ImmSExt10>;
652 def v4f32: AIFPVecInst<v4f32, v4i32SExt10Imm>;
653 def f32: AIFPInst<R32FP, i32ImmSExt10>;
656 defm AI : AddImmediate;
659 RRForm<0b00010010000, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
660 "sfh\t$rT, $rA, $rB", IntegerOp,
661 [(set (v8i16 VECREG:$rT), (sub (v8i16 VECREG:$rA),
662 (v8i16 VECREG:$rB)))]>;
665 RRForm<0b00010010000, (outs R16C:$rT), (ins R16C:$rA, R16C:$rB),
666 "sfh\t$rT, $rA, $rB", IntegerOp,
667 [(set R16C:$rT, (sub R16C:$rA, R16C:$rB))]>;
670 RI10Form<0b10110000, (outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
671 "sfhi\t$rT, $rA, $val", IntegerOp,
672 [(set (v8i16 VECREG:$rT), (sub v8i16SExt10Imm:$val,
673 (v8i16 VECREG:$rA)))]>;
675 def SFHIr16 : RI10Form<0b10110000, (outs R16C:$rT), (ins R16C:$rA, s10imm:$val),
676 "sfhi\t$rT, $rA, $val", IntegerOp,
677 [(set R16C:$rT, (sub i16ImmSExt10:$val, R16C:$rA))]>;
679 def SFvec : RRForm<0b00000010000, (outs VECREG:$rT),
680 (ins VECREG:$rA, VECREG:$rB),
681 "sf\t$rT, $rA, $rB", IntegerOp,
682 [(set (v4i32 VECREG:$rT), (sub (v4i32 VECREG:$rA), (v4i32 VECREG:$rB)))]>;
684 def SFr32 : RRForm<0b00000010000, (outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
685 "sf\t$rT, $rA, $rB", IntegerOp,
686 [(set R32C:$rT, (sub R32C:$rA, R32C:$rB))]>;
689 RI10Form<0b00110000, (outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
690 "sfi\t$rT, $rA, $val", IntegerOp,
691 [(set (v4i32 VECREG:$rT), (sub v4i32SExt10Imm:$val,
692 (v4i32 VECREG:$rA)))]>;
694 def SFIr32 : RI10Form<0b00110000, (outs R32C:$rT),
695 (ins R32C:$rA, s10imm_i32:$val),
696 "sfi\t$rT, $rA, $val", IntegerOp,
697 [(set R32C:$rT, (sub i32ImmSExt10:$val, R32C:$rA))]>;
699 // ADDX: only available in vector form, doesn't match a pattern.
700 class ADDXInst<dag OOL, dag IOL, list<dag> pattern>:
701 RRForm<0b00000010110, OOL, IOL,
702 "addx\t$rT, $rA, $rB",
705 class ADDXVecInst<ValueType vectype>:
706 ADDXInst<(outs VECREG:$rT),
707 (ins VECREG:$rA, VECREG:$rB, VECREG:$rCarry),
708 [(set (vectype VECREG:$rT),
709 (SPUaddx (vectype VECREG:$rA), (vectype VECREG:$rB),
710 (vectype VECREG:$rCarry)))]>,
711 RegConstraint<"$rCarry = $rT">,
714 class ADDXRegInst<RegisterClass rclass>:
715 ADDXInst<(outs rclass:$rT),
716 (ins rclass:$rA, rclass:$rB, rclass:$rCarry),
718 (SPUaddx rclass:$rA, rclass:$rB, rclass:$rCarry))]>,
719 RegConstraint<"$rCarry = $rT">,
722 multiclass AddExtended {
723 def v2i64 : ADDXVecInst<v2i64>;
724 def v4i32 : ADDXVecInst<v4i32>;
725 def r64 : ADDXRegInst<R64C>;
726 def r32 : ADDXRegInst<R32C>;
729 defm ADDX : AddExtended;
731 // CG: Generate carry for add
732 class CGInst<dag OOL, dag IOL, list<dag> pattern>:
733 RRForm<0b01000011000, OOL, IOL,
737 class CGVecInst<ValueType vectype>:
738 CGInst<(outs VECREG:$rT),
739 (ins VECREG:$rA, VECREG:$rB),
740 [(set (vectype VECREG:$rT),
741 (SPUcarry_gen (vectype VECREG:$rA), (vectype VECREG:$rB)))]>;
743 class CGRegInst<RegisterClass rclass>:
744 CGInst<(outs rclass:$rT),
745 (ins rclass:$rA, rclass:$rB),
747 (SPUcarry_gen rclass:$rA, rclass:$rB))]>;
749 multiclass CarryGenerate {
750 def v2i64 : CGVecInst<v2i64>;
751 def v4i32 : CGVecInst<v4i32>;
752 def r64 : CGRegInst<R64C>;
753 def r32 : CGRegInst<R32C>;
756 defm CG : CarryGenerate;
758 // SFX: Subract from, extended. This is used in conjunction with BG to subtract
759 // with carry (borrow, in this case)
760 class SFXInst<dag OOL, dag IOL, list<dag> pattern>:
761 RRForm<0b10000010110, OOL, IOL,
762 "sfx\t$rT, $rA, $rB",
765 class SFXVecInst<ValueType vectype>:
766 SFXInst<(outs VECREG:$rT),
767 (ins VECREG:$rA, VECREG:$rB, VECREG:$rCarry),
768 [(set (vectype VECREG:$rT),
769 (SPUsubx (vectype VECREG:$rA), (vectype VECREG:$rB),
770 (vectype VECREG:$rCarry)))]>,
771 RegConstraint<"$rCarry = $rT">,
774 class SFXRegInst<RegisterClass rclass>:
775 SFXInst<(outs rclass:$rT),
776 (ins rclass:$rA, rclass:$rB, rclass:$rCarry),
778 (SPUsubx rclass:$rA, rclass:$rB, rclass:$rCarry))]>,
779 RegConstraint<"$rCarry = $rT">,
782 multiclass SubtractExtended {
783 def v2i64 : SFXVecInst<v2i64>;
784 def v4i32 : SFXVecInst<v4i32>;
785 def r64 : SFXRegInst<R64C>;
786 def r32 : SFXRegInst<R32C>;
789 defm SFX : SubtractExtended;
791 // BG: only available in vector form, doesn't match a pattern.
792 class BGInst<dag OOL, dag IOL, list<dag> pattern>:
793 RRForm<0b01000010000, OOL, IOL,
797 class BGVecInst<ValueType vectype>:
798 BGInst<(outs VECREG:$rT),
799 (ins VECREG:$rA, VECREG:$rB),
800 [(set (vectype VECREG:$rT),
801 (SPUborrow_gen (vectype VECREG:$rA), (vectype VECREG:$rB)))]>;
803 class BGRegInst<RegisterClass rclass>:
804 BGInst<(outs rclass:$rT),
805 (ins rclass:$rA, rclass:$rB),
807 (SPUborrow_gen rclass:$rA, rclass:$rB))]>;
809 multiclass BorrowGenerate {
810 def v4i32 : BGVecInst<v4i32>;
811 def v2i64 : BGVecInst<v2i64>;
812 def r64 : BGRegInst<R64C>;
813 def r32 : BGRegInst<R32C>;
816 defm BG : BorrowGenerate;
818 // BGX: Borrow generate, extended.
820 RRForm<0b11000010110, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB,
822 "bgx\t$rT, $rA, $rB", IntegerOp,
824 RegConstraint<"$rCarry = $rT">,
827 // Halfword multiply variants:
828 // N.B: These can be used to build up larger quantities (16x16 -> 32)
831 RRForm<0b00100011110, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
832 "mpy\t$rT, $rA, $rB", IntegerMulDiv,
836 RRForm<0b00100011110, (outs R16C:$rT), (ins R16C:$rA, R16C:$rB),
837 "mpy\t$rT, $rA, $rB", IntegerMulDiv,
838 [(set R16C:$rT, (mul R16C:$rA, R16C:$rB))]>;
840 // Unsigned 16-bit multiply:
842 class MPYUInst<dag OOL, dag IOL, list<dag> pattern>:
843 RRForm<0b00110011110, OOL, IOL,
844 "mpyu\t$rT, $rA, $rB", IntegerMulDiv,
848 MPYUInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
852 MPYUInst<(outs R32C:$rT), (ins R16C:$rA, R16C:$rB),
853 [(set R32C:$rT, (mul (zext R16C:$rA), (zext R16C:$rB)))]>;
856 MPYUInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
859 // mpyi: multiply 16 x s10imm -> 32 result.
861 class MPYIInst<dag OOL, dag IOL, list<dag> pattern>:
862 RI10Form<0b00101110, OOL, IOL,
863 "mpyi\t$rT, $rA, $val", IntegerMulDiv,
867 MPYIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
868 [(set (v8i16 VECREG:$rT),
869 (mul (v8i16 VECREG:$rA), v8i16SExt10Imm:$val))]>;
872 MPYIInst<(outs R16C:$rT), (ins R16C:$rA, s10imm:$val),
873 [(set R16C:$rT, (mul R16C:$rA, i16ImmSExt10:$val))]>;
875 // mpyui: same issues as other multiplies, plus, this doesn't match a
876 // pattern... but may be used during target DAG selection or lowering
878 class MPYUIInst<dag OOL, dag IOL, list<dag> pattern>:
879 RI10Form<0b10101110, OOL, IOL,
880 "mpyui\t$rT, $rA, $val", IntegerMulDiv,
884 MPYUIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
888 MPYUIInst<(outs R16C:$rT), (ins R16C:$rA, s10imm:$val),
891 // mpya: 16 x 16 + 16 -> 32 bit result
892 class MPYAInst<dag OOL, dag IOL, list<dag> pattern>:
893 RRRForm<0b0011, OOL, IOL,
894 "mpya\t$rT, $rA, $rB, $rC", IntegerMulDiv,
898 MPYAInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
899 [(set (v4i32 VECREG:$rT),
900 (add (v4i32 (bitconvert (mul (v8i16 VECREG:$rA),
901 (v8i16 VECREG:$rB)))),
902 (v4i32 VECREG:$rC)))]>;
905 MPYAInst<(outs R32C:$rT), (ins R16C:$rA, R16C:$rB, R32C:$rC),
906 [(set R32C:$rT, (add (sext (mul R16C:$rA, R16C:$rB)),
910 MPYAInst<(outs R32C:$rT), (ins R16C:$rA, R16C:$rB, R32C:$rC),
911 [(set R32C:$rT, (add (mul (sext R16C:$rA), (sext R16C:$rB)),
914 def MPYAr32_sextinreg:
915 MPYAInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB, R32C:$rC),
916 [(set R32C:$rT, (add (mul (sext_inreg R32C:$rA, i16),
917 (sext_inreg R32C:$rB, i16)),
920 // mpyh: multiply high, used to synthesize 32-bit multiplies
921 class MPYHInst<dag OOL, dag IOL, list<dag> pattern>:
922 RRForm<0b10100011110, OOL, IOL,
923 "mpyh\t$rT, $rA, $rB", IntegerMulDiv,
927 MPYHInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
931 MPYHInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
934 // mpys: multiply high and shift right (returns the top half of
935 // a 16-bit multiply, sign extended to 32 bits.)
937 class MPYSInst<dag OOL, dag IOL>:
938 RRForm<0b11100011110, OOL, IOL,
939 "mpys\t$rT, $rA, $rB", IntegerMulDiv,
943 MPYSInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB)>;
946 MPYSInst<(outs R32C:$rT), (ins R16C:$rA, R16C:$rB)>;
948 // mpyhh: multiply high-high (returns the 32-bit result from multiplying
949 // the top 16 bits of the $rA, $rB)
951 class MPYHHInst<dag OOL, dag IOL>:
952 RRForm<0b01100011110, OOL, IOL,
953 "mpyhh\t$rT, $rA, $rB", IntegerMulDiv,
957 MPYHHInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB)>;
960 MPYHHInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB)>;
962 // mpyhha: Multiply high-high, add to $rT:
964 class MPYHHAInst<dag OOL, dag IOL>:
965 RRForm<0b01100010110, OOL, IOL,
966 "mpyhha\t$rT, $rA, $rB", IntegerMulDiv,
970 MPYHHAInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB)>;
973 MPYHHAInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB)>;
975 // mpyhhu: Multiply high-high, unsigned
977 class MPYHHUInst<dag OOL, dag IOL>:
978 RRForm<0b01110011110, OOL, IOL,
979 "mpyhhu\t$rT, $rA, $rB", IntegerMulDiv,
983 MPYHHUInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB)>;
986 MPYHHUInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB)>;
988 // mpyhhau: Multiply high-high, unsigned
990 class MPYHHAUInst<dag OOL, dag IOL>:
991 RRForm<0b01110010110, OOL, IOL,
992 "mpyhhau\t$rT, $rA, $rB", IntegerMulDiv,
996 MPYHHAUInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB)>;
999 MPYHHAUInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB)>;
1001 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
1002 // clz: Count leading zeroes
1003 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
1004 class CLZInst<dag OOL, dag IOL, list<dag> pattern>:
1005 RRForm_1<0b10100101010, OOL, IOL, "clz\t$rT, $rA",
1006 IntegerOp, pattern>;
1008 class CLZRegInst<RegisterClass rclass>:
1009 CLZInst<(outs rclass:$rT), (ins rclass:$rA),
1010 [(set rclass:$rT, (ctlz rclass:$rA))]>;
1012 class CLZVecInst<ValueType vectype>:
1013 CLZInst<(outs VECREG:$rT), (ins VECREG:$rA),
1014 [(set (vectype VECREG:$rT), (ctlz (vectype VECREG:$rA)))]>;
1016 multiclass CountLeadingZeroes {
1017 def v4i32 : CLZVecInst<v4i32>;
1018 def r32 : CLZRegInst<R32C>;
1021 defm CLZ : CountLeadingZeroes;
1023 // cntb: Count ones in bytes (aka "population count")
1025 // NOTE: This instruction is really a vector instruction, but the custom
1026 // lowering code uses it in unorthodox ways to support CTPOP for other
1030 RRForm_1<0b00101101010, (outs VECREG:$rT), (ins VECREG:$rA),
1031 "cntb\t$rT, $rA", IntegerOp,
1032 [(set (v16i8 VECREG:$rT), (SPUcntb (v16i8 VECREG:$rA)))]>;
1035 RRForm_1<0b00101101010, (outs VECREG:$rT), (ins VECREG:$rA),
1036 "cntb\t$rT, $rA", IntegerOp,
1037 [(set (v8i16 VECREG:$rT), (SPUcntb (v8i16 VECREG:$rA)))]>;
1040 RRForm_1<0b00101101010, (outs VECREG:$rT), (ins VECREG:$rA),
1041 "cntb\t$rT, $rA", IntegerOp,
1042 [(set (v4i32 VECREG:$rT), (SPUcntb (v4i32 VECREG:$rA)))]>;
1044 // gbb: Gather the low order bits from each byte in $rA into a single 16-bit
1045 // quantity stored into $rT's slot 0, upper 16 bits are zeroed, as are
1048 // Note: This instruction "pairs" with the fsmb instruction for all of the
1049 // various types defined here.
1051 // Note 2: The "VecInst" and "RegInst" forms refer to the result being either
1052 // a vector or register.
1054 class GBBInst<dag OOL, dag IOL, list<dag> pattern>:
1055 RRForm_1<0b01001101100, OOL, IOL, "gbb\t$rT, $rA", GatherOp, pattern>;
1057 class GBBRegInst<RegisterClass rclass, ValueType vectype>:
1058 GBBInst<(outs rclass:$rT), (ins VECREG:$rA),
1059 [/* no pattern */]>;
1061 class GBBVecInst<ValueType vectype>:
1062 GBBInst<(outs VECREG:$rT), (ins VECREG:$rA),
1063 [/* no pattern */]>;
1065 multiclass GatherBitsFromBytes {
1066 def v16i8_r32: GBBRegInst<R32C, v16i8>;
1067 def v16i8_r16: GBBRegInst<R16C, v16i8>;
1068 def v16i8: GBBVecInst<v16i8>;
1071 defm GBB: GatherBitsFromBytes;
1073 // gbh: Gather all low order bits from each halfword in $rA into a single
1074 // 8-bit quantity stored in $rT's slot 0, with the upper bits of $rT set to 0
1075 // and slots 1-3 also set to 0.
1077 // See notes for GBBInst, above.
1079 class GBHInst<dag OOL, dag IOL, list<dag> pattern>:
1080 RRForm_1<0b10001101100, OOL, IOL, "gbh\t$rT, $rA", GatherOp,
1083 class GBHRegInst<RegisterClass rclass, ValueType vectype>:
1084 GBHInst<(outs rclass:$rT), (ins VECREG:$rA),
1085 [/* no pattern */]>;
1087 class GBHVecInst<ValueType vectype>:
1088 GBHInst<(outs VECREG:$rT), (ins VECREG:$rA),
1089 [/* no pattern */]>;
1091 multiclass GatherBitsHalfword {
1092 def v8i16_r32: GBHRegInst<R32C, v8i16>;
1093 def v8i16_r16: GBHRegInst<R16C, v8i16>;
1094 def v8i16: GBHVecInst<v8i16>;
1097 defm GBH: GatherBitsHalfword;
1099 // gb: Gather all low order bits from each word in $rA into a single
1100 // 4-bit quantity stored in $rT's slot 0, upper bits in $rT set to 0,
1101 // as well as slots 1-3.
1103 // See notes for gbb, above.
1105 class GBInst<dag OOL, dag IOL, list<dag> pattern>:
1106 RRForm_1<0b00001101100, OOL, IOL, "gb\t$rT, $rA", GatherOp,
1109 class GBRegInst<RegisterClass rclass, ValueType vectype>:
1110 GBInst<(outs rclass:$rT), (ins VECREG:$rA),
1111 [/* no pattern */]>;
1113 class GBVecInst<ValueType vectype>:
1114 GBInst<(outs VECREG:$rT), (ins VECREG:$rA),
1115 [/* no pattern */]>;
1117 multiclass GatherBitsWord {
1118 def v4i32_r32: GBRegInst<R32C, v4i32>;
1119 def v4i32_r16: GBRegInst<R16C, v4i32>;
1120 def v4i32: GBVecInst<v4i32>;
1123 defm GB: GatherBitsWord;
1125 // avgb: average bytes
1127 RRForm<0b11001011000, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1128 "avgb\t$rT, $rA, $rB", ByteOp,
1131 // absdb: absolute difference of bytes
1133 RRForm<0b11001010000, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1134 "absdb\t$rT, $rA, $rB", ByteOp,
1137 // sumb: sum bytes into halfwords
1139 RRForm<0b11001010010, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1140 "sumb\t$rT, $rA, $rB", ByteOp,
1143 // Sign extension operations:
1144 class XSBHInst<dag OOL, dag IOL, list<dag> pattern>:
1145 RRForm_1<0b01101101010, OOL, IOL,
1146 "xsbh\t$rDst, $rSrc",
1147 IntegerOp, pattern>;
1149 class XSBHVecInst<ValueType vectype>:
1150 XSBHInst<(outs VECREG:$rDst), (ins VECREG:$rSrc),
1151 [(set (v8i16 VECREG:$rDst), (sext (vectype VECREG:$rSrc)))]>;
1153 class XSBHInRegInst<RegisterClass rclass, list<dag> pattern>:
1154 XSBHInst<(outs rclass:$rDst), (ins rclass:$rSrc),
1157 multiclass ExtendByteHalfword {
1158 def v16i8: XSBHVecInst<v8i16>;
1159 def r8: XSBHInst<(outs R16C:$rDst), (ins R8C:$rSrc),
1160 [(set R16C:$rDst, (sext R8C:$rSrc))]>;
1161 def r16: XSBHInRegInst<R16C,
1162 [(set R16C:$rDst, (sext_inreg R16C:$rSrc, i8))]>;
1164 // 32-bit form for XSBH: used to sign extend 8-bit quantities to 16-bit
1165 // quantities to 32-bit quantities via a 32-bit register (see the sext 8->32
1166 // pattern below). Intentionally doesn't match a pattern because we want the
1167 // sext 8->32 pattern to do the work for us, namely because we need the extra
1169 def r32: XSBHInRegInst<R32C, [/* no pattern */]>;
1171 // Same as the 32-bit version, but for i64
1172 def r64: XSBHInRegInst<R64C, [/* no pattern */]>;
1175 defm XSBH : ExtendByteHalfword;
1177 // Sign extend halfwords to words:
1179 class XSHWInst<dag OOL, dag IOL, list<dag> pattern>:
1180 RRForm_1<0b01101101010, OOL, IOL, "xshw\t$rDest, $rSrc",
1181 IntegerOp, pattern>;
1183 class XSHWVecInst<ValueType in_vectype, ValueType out_vectype>:
1184 XSHWInst<(outs VECREG:$rDest), (ins VECREG:$rSrc),
1185 [(set (out_vectype VECREG:$rDest),
1186 (sext (in_vectype VECREG:$rSrc)))]>;
1188 class XSHWInRegInst<RegisterClass rclass, list<dag> pattern>:
1189 XSHWInst<(outs rclass:$rDest), (ins rclass:$rSrc),
1192 class XSHWRegInst<RegisterClass rclass>:
1193 XSHWInst<(outs rclass:$rDest), (ins R16C:$rSrc),
1194 [(set rclass:$rDest, (sext R16C:$rSrc))]>;
1196 multiclass ExtendHalfwordWord {
1197 def v4i32: XSHWVecInst<v4i32, v8i16>;
1199 def r16: XSHWRegInst<R32C>;
1201 def r32: XSHWInRegInst<R32C,
1202 [(set R32C:$rDest, (sext_inreg R32C:$rSrc, i16))]>;
1203 def r64: XSHWInRegInst<R64C, [/* no pattern */]>;
1206 defm XSHW : ExtendHalfwordWord;
1208 // Sign-extend words to doublewords (32->64 bits)
1210 class XSWDInst<dag OOL, dag IOL, list<dag> pattern>:
1211 RRForm_1<0b01100101010, OOL, IOL, "xswd\t$rDst, $rSrc",
1212 IntegerOp, pattern>;
1214 class XSWDVecInst<ValueType in_vectype, ValueType out_vectype>:
1215 XSWDInst<(outs VECREG:$rDst), (ins VECREG:$rSrc),
1216 [(set (out_vectype VECREG:$rDst),
1217 (sext (out_vectype VECREG:$rSrc)))]>;
1219 class XSWDRegInst<RegisterClass in_rclass, RegisterClass out_rclass>:
1220 XSWDInst<(outs out_rclass:$rDst), (ins in_rclass:$rSrc),
1221 [(set out_rclass:$rDst, (sext in_rclass:$rSrc))]>;
1223 multiclass ExtendWordToDoubleWord {
1224 def v2i64: XSWDVecInst<v4i32, v2i64>;
1225 def r64: XSWDRegInst<R32C, R64C>;
1227 def r64_inreg: XSWDInst<(outs R64C:$rDst), (ins R64C:$rSrc),
1228 [(set R64C:$rDst, (sext_inreg R64C:$rSrc, i32))]>;
1231 defm XSWD : ExtendWordToDoubleWord;
1235 class ANDInst<dag OOL, dag IOL, list<dag> pattern> :
1236 RRForm<0b10000011000, OOL, IOL, "and\t$rT, $rA, $rB",
1237 IntegerOp, pattern>;
1239 class ANDVecInst<ValueType vectype>:
1240 ANDInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1241 [(set (vectype VECREG:$rT), (and (vectype VECREG:$rA),
1242 (vectype VECREG:$rB)))]>;
1244 class ANDRegInst<RegisterClass rclass>:
1245 ANDInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
1246 [(set rclass:$rT, (and rclass:$rA, rclass:$rB))]>;
1248 multiclass BitwiseAnd
1250 def v16i8: ANDVecInst<v16i8>;
1251 def v8i16: ANDVecInst<v8i16>;
1252 def v4i32: ANDVecInst<v4i32>;
1253 def v2i64: ANDVecInst<v2i64>;
1255 def r128: ANDRegInst<GPRC>;
1256 def r64: ANDRegInst<R64C>;
1257 def r32: ANDRegInst<R32C>;
1258 def r16: ANDRegInst<R16C>;
1259 def r8: ANDRegInst<R8C>;
1261 //===---------------------------------------------
1262 // Special instructions to perform the fabs instruction
1263 def fabs32: ANDInst<(outs R32FP:$rT), (ins R32FP:$rA, R32C:$rB),
1264 [/* Intentionally does not match a pattern */]>;
1266 def fabs64: ANDInst<(outs R64FP:$rT), (ins R64FP:$rA, VECREG:$rB),
1267 [/* Intentionally does not match a pattern */]>;
1269 // Could use v4i32, but won't for clarity
1270 def fabsvec: ANDInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1271 [/* Intentionally does not match a pattern */]>;
1273 //===---------------------------------------------
1275 // Hacked form of AND to zero-extend 16-bit quantities to 32-bit
1276 // quantities -- see 16->32 zext pattern.
1278 // This pattern is somewhat artificial, since it might match some
1279 // compiler generated pattern but it is unlikely to do so.
1281 def i16i32: ANDInst<(outs R32C:$rT), (ins R16C:$rA, R32C:$rB),
1282 [(set R32C:$rT, (and (zext R16C:$rA), R32C:$rB))]>;
1285 defm AND : BitwiseAnd;
1287 // N.B.: vnot_conv is one of those special target selection pattern fragments,
1288 // in which we expect there to be a bit_convert on the constant. Bear in mind
1289 // that llvm translates "not <reg>" to "xor <reg>, -1" (or in this case, a
1290 // constant -1 vector.)
1292 class ANDCInst<dag OOL, dag IOL, list<dag> pattern>:
1293 RRForm<0b10000011010, OOL, IOL, "andc\t$rT, $rA, $rB",
1294 IntegerOp, pattern>;
1296 class ANDCVecInst<ValueType vectype>:
1297 ANDCInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1298 [(set (vectype VECREG:$rT), (and (vectype VECREG:$rA),
1299 (vnot (vectype VECREG:$rB))))]>;
1301 class ANDCRegInst<RegisterClass rclass>:
1302 ANDCInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
1303 [(set rclass:$rT, (and rclass:$rA, (not rclass:$rB)))]>;
1305 multiclass AndComplement
1307 def v16i8: ANDCVecInst<v16i8>;
1308 def v8i16: ANDCVecInst<v8i16>;
1309 def v4i32: ANDCVecInst<v4i32>;
1310 def v2i64: ANDCVecInst<v2i64>;
1312 def r128: ANDCRegInst<GPRC>;
1313 def r64: ANDCRegInst<R64C>;
1314 def r32: ANDCRegInst<R32C>;
1315 def r16: ANDCRegInst<R16C>;
1316 def r8: ANDCRegInst<R8C>;
1319 defm ANDC : AndComplement;
1321 class ANDBIInst<dag OOL, dag IOL, list<dag> pattern>:
1322 RI10Form<0b01101000, OOL, IOL, "andbi\t$rT, $rA, $val",
1325 multiclass AndByteImm
1327 def v16i8: ANDBIInst<(outs VECREG:$rT), (ins VECREG:$rA, u10imm:$val),
1328 [(set (v16i8 VECREG:$rT),
1329 (and (v16i8 VECREG:$rA),
1330 (v16i8 v16i8U8Imm:$val)))]>;
1332 def r8: ANDBIInst<(outs R8C:$rT), (ins R8C:$rA, u10imm_i8:$val),
1333 [(set R8C:$rT, (and R8C:$rA, immU8:$val))]>;
1336 defm ANDBI : AndByteImm;
1338 class ANDHIInst<dag OOL, dag IOL, list<dag> pattern> :
1339 RI10Form<0b10101000, OOL, IOL, "andhi\t$rT, $rA, $val",
1342 multiclass AndHalfwordImm
1344 def v8i16: ANDHIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
1345 [(set (v8i16 VECREG:$rT),
1346 (and (v8i16 VECREG:$rA), v8i16SExt10Imm:$val))]>;
1348 def r16: ANDHIInst<(outs R16C:$rT), (ins R16C:$rA, u10imm:$val),
1349 [(set R16C:$rT, (and R16C:$rA, i16ImmUns10:$val))]>;
1351 // Zero-extend i8 to i16:
1352 def i8i16: ANDHIInst<(outs R16C:$rT), (ins R8C:$rA, u10imm:$val),
1353 [(set R16C:$rT, (and (zext R8C:$rA), i16ImmUns10:$val))]>;
1356 defm ANDHI : AndHalfwordImm;
1358 class ANDIInst<dag OOL, dag IOL, list<dag> pattern> :
1359 RI10Form<0b00101000, OOL, IOL, "andi\t$rT, $rA, $val",
1360 IntegerOp, pattern>;
1362 multiclass AndWordImm
1364 def v4i32: ANDIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
1365 [(set (v4i32 VECREG:$rT),
1366 (and (v4i32 VECREG:$rA), v4i32SExt10Imm:$val))]>;
1368 def r32: ANDIInst<(outs R32C:$rT), (ins R32C:$rA, s10imm_i32:$val),
1369 [(set R32C:$rT, (and R32C:$rA, i32ImmSExt10:$val))]>;
1371 // Hacked form of ANDI to zero-extend i8 quantities to i32. See the zext 8->32
1373 def i8i32: ANDIInst<(outs R32C:$rT), (ins R8C:$rA, s10imm_i32:$val),
1375 (and (zext R8C:$rA), i32ImmSExt10:$val))]>;
1377 // Hacked form of ANDI to zero-extend i16 quantities to i32. See the
1378 // zext 16->32 pattern below.
1380 // Note that this pattern is somewhat artificial, since it might match
1381 // something the compiler generates but is unlikely to occur in practice.
1382 def i16i32: ANDIInst<(outs R32C:$rT), (ins R16C:$rA, s10imm_i32:$val),
1384 (and (zext R16C:$rA), i32ImmSExt10:$val))]>;
1387 defm ANDI : AndWordImm;
1389 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
1390 // Bitwise OR group:
1391 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
1393 // Bitwise "or" (N.B.: These are also register-register copy instructions...)
1394 class ORInst<dag OOL, dag IOL, list<dag> pattern>:
1395 RRForm<0b10000010000, OOL, IOL, "or\t$rT, $rA, $rB",
1396 IntegerOp, pattern>;
1398 class ORVecInst<ValueType vectype>:
1399 ORInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1400 [(set (vectype VECREG:$rT), (or (vectype VECREG:$rA),
1401 (vectype VECREG:$rB)))]>;
1403 class ORRegInst<RegisterClass rclass>:
1404 ORInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
1405 [(set rclass:$rT, (or rclass:$rA, rclass:$rB))]>;
1407 // ORCvtForm: OR conversion form
1409 // This is used to "convert" the preferred slot to its vector equivalent, as
1410 // well as convert a vector back to its preferred slot.
1412 // These are effectively no-ops, but need to exist for proper type conversion
1413 // and type coercion.
1415 class ORCvtForm<dag OOL, dag IOL>
1416 : SPUInstr<OOL, IOL, "or\t$rT, $rA, $rA", IntegerOp> {
1420 let Pattern = [/* no pattern */];
1422 let Inst{0-10} = 0b10000010000;
1423 let Inst{11-17} = RA;
1424 let Inst{18-24} = RA;
1425 let Inst{25-31} = RT;
1428 class ORPromoteScalar<RegisterClass rclass>:
1429 ORCvtForm<(outs VECREG:$rT), (ins rclass:$rA)>;
1431 class ORExtractElt<RegisterClass rclass>:
1432 ORCvtForm<(outs rclass:$rT), (ins VECREG:$rA)>;
1434 class ORCvtRegGPRC<RegisterClass rclass>:
1435 ORCvtForm<(outs GPRC:$rT), (ins rclass:$rA)>;
1438 ORCvtForm<(outs GPRC:$rT), (ins VECREG:$rA)>;
1440 class ORCvtGPRCReg<RegisterClass rclass>:
1441 ORCvtForm<(outs rclass:$rT), (ins GPRC:$rA)>;
1443 class ORCvtFormR32Reg<RegisterClass rclass>:
1444 ORCvtForm<(outs rclass:$rT), (ins R32C:$rA)>;
1446 class ORCvtFormRegR32<RegisterClass rclass>:
1447 ORCvtForm<(outs R32C:$rT), (ins rclass:$rA)>;
1449 class ORCvtFormR64Reg<RegisterClass rclass>:
1450 ORCvtForm<(outs rclass:$rT), (ins R64C:$rA)>;
1452 class ORCvtFormRegR64<RegisterClass rclass>:
1453 ORCvtForm<(outs R64C:$rT), (ins rclass:$rA)>;
1456 ORCvtForm<(outs VECREG:$rT), (ins GPRC:$rA)>;
1458 multiclass BitwiseOr
1460 def v16i8: ORVecInst<v16i8>;
1461 def v8i16: ORVecInst<v8i16>;
1462 def v4i32: ORVecInst<v4i32>;
1463 def v2i64: ORVecInst<v2i64>;
1465 def v4f32: ORInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1466 [(set (v4f32 VECREG:$rT),
1467 (v4f32 (bitconvert (or (v4i32 VECREG:$rA),
1468 (v4i32 VECREG:$rB)))))]>;
1470 def v2f64: ORInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1471 [(set (v2f64 VECREG:$rT),
1472 (v2f64 (bitconvert (or (v2i64 VECREG:$rA),
1473 (v2i64 VECREG:$rB)))))]>;
1475 def r64: ORRegInst<R64C>;
1476 def r32: ORRegInst<R32C>;
1477 def r16: ORRegInst<R16C>;
1478 def r8: ORRegInst<R8C>;
1480 // OR instructions used to copy f32 and f64 registers.
1481 def f32: ORInst<(outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB),
1482 [/* no pattern */]>;
1484 def f64: ORInst<(outs R64FP:$rT), (ins R64FP:$rA, R64FP:$rB),
1485 [/* no pattern */]>;
1487 // scalar->vector promotion, prefslot2vec:
1488 def v16i8_i8: ORPromoteScalar<R8C>;
1489 def v8i16_i16: ORPromoteScalar<R16C>;
1490 def v4i32_i32: ORPromoteScalar<R32C>;
1491 def v2i64_i64: ORPromoteScalar<R64C>;
1492 def v4f32_f32: ORPromoteScalar<R32FP>;
1493 def v2f64_f64: ORPromoteScalar<R64FP>;
1495 // vector->scalar demotion, vec2prefslot:
1496 def i8_v16i8: ORExtractElt<R8C>;
1497 def i16_v8i16: ORExtractElt<R16C>;
1498 def i32_v4i32: ORExtractElt<R32C>;
1499 def i64_v2i64: ORExtractElt<R64C>;
1500 def f32_v4f32: ORExtractElt<R32FP>;
1501 def f64_v2f64: ORExtractElt<R64FP>;
1503 // Conversion from GPRC to register
1504 def i128_r64: ORCvtRegGPRC<R64C>;
1505 def i128_f64: ORCvtRegGPRC<R64FP>;
1506 def i128_r32: ORCvtRegGPRC<R32C>;
1507 def i128_f32: ORCvtRegGPRC<R32FP>;
1508 def i128_r16: ORCvtRegGPRC<R16C>;
1509 def i128_r8: ORCvtRegGPRC<R8C>;
1511 // Conversion from GPRC to vector
1512 def i128_vec: ORCvtVecGPRC;
1514 // Conversion from register to GPRC
1515 def r64_i128: ORCvtGPRCReg<R64C>;
1516 def f64_i128: ORCvtGPRCReg<R64FP>;
1517 def r32_i128: ORCvtGPRCReg<R32C>;
1518 def f32_i128: ORCvtGPRCReg<R32FP>;
1519 def r16_i128: ORCvtGPRCReg<R16C>;
1520 def r8_i128: ORCvtGPRCReg<R8C>;
1522 // Conversion from vector to GPRC
1523 def vec_i128: ORCvtGPRCVec;
1525 // Conversion from register to R32C:
1526 def r16_r32: ORCvtFormRegR32<R16C>;
1527 def r8_r32: ORCvtFormRegR32<R8C>;
1529 // Conversion from R32C to register
1530 def r32_r16: ORCvtFormR32Reg<R16C>;
1531 def r32_r8: ORCvtFormR32Reg<R8C>;
1533 // Conversion from register to R64C:
1534 def r32_r64: ORCvtFormR64Reg<R32C>;
1535 def r16_r64: ORCvtFormR64Reg<R16C>;
1536 def r8_r64: ORCvtFormR64Reg<R8C>;
1538 // Conversion from R64C to register
1539 def r64_r32: ORCvtFormRegR64<R32C>;
1540 def r64_r16: ORCvtFormRegR64<R16C>;
1541 def r64_r8: ORCvtFormRegR64<R8C>;
1544 defm OR : BitwiseOr;
1546 // scalar->vector promotion patterns (preferred slot to vector):
1547 def : Pat<(v16i8 (SPUprefslot2vec R8C:$rA)),
1548 (ORv16i8_i8 R8C:$rA)>;
1550 def : Pat<(v8i16 (SPUprefslot2vec R16C:$rA)),
1551 (ORv8i16_i16 R16C:$rA)>;
1553 def : Pat<(v4i32 (SPUprefslot2vec R32C:$rA)),
1554 (ORv4i32_i32 R32C:$rA)>;
1556 def : Pat<(v2i64 (SPUprefslot2vec R64C:$rA)),
1557 (ORv2i64_i64 R64C:$rA)>;
1559 def : Pat<(v4f32 (SPUprefslot2vec R32FP:$rA)),
1560 (ORv4f32_f32 R32FP:$rA)>;
1562 def : Pat<(v2f64 (SPUprefslot2vec R64FP:$rA)),
1563 (ORv2f64_f64 R64FP:$rA)>;
1565 // ORi*_v*: Used to extract vector element 0 (the preferred slot), otherwise
1566 // known as converting the vector back to its preferred slot
1568 def : Pat<(SPUvec2prefslot (v16i8 VECREG:$rA)),
1569 (ORi8_v16i8 VECREG:$rA)>;
1571 def : Pat<(SPUvec2prefslot (v8i16 VECREG:$rA)),
1572 (ORi16_v8i16 VECREG:$rA)>;
1574 def : Pat<(SPUvec2prefslot (v4i32 VECREG:$rA)),
1575 (ORi32_v4i32 VECREG:$rA)>;
1577 def : Pat<(SPUvec2prefslot (v2i64 VECREG:$rA)),
1578 (ORi64_v2i64 VECREG:$rA)>;
1580 def : Pat<(SPUvec2prefslot (v4f32 VECREG:$rA)),
1581 (ORf32_v4f32 VECREG:$rA)>;
1583 def : Pat<(SPUvec2prefslot (v2f64 VECREG:$rA)),
1584 (ORf64_v2f64 VECREG:$rA)>;
1586 // Load Register: This is an assembler alias for a bitwise OR of a register
1587 // against itself. It's here because it brings some clarity to assembly
1590 let hasCtrlDep = 1 in {
1591 class LRInst<dag OOL, dag IOL>
1592 : SPUInstr<OOL, IOL, "lr\t$rT, $rA", IntegerOp> {
1596 let Pattern = [/*no pattern*/];
1598 let Inst{0-10} = 0b10000010000; /* It's an OR operation */
1599 let Inst{11-17} = RA;
1600 let Inst{18-24} = RA;
1601 let Inst{25-31} = RT;
1604 class LRVecInst<ValueType vectype>:
1605 LRInst<(outs VECREG:$rT), (ins VECREG:$rA)>;
1607 class LRRegInst<RegisterClass rclass>:
1608 LRInst<(outs rclass:$rT), (ins rclass:$rA)>;
1610 multiclass LoadRegister {
1611 def v2i64: LRVecInst<v2i64>;
1612 def v2f64: LRVecInst<v2f64>;
1613 def v4i32: LRVecInst<v4i32>;
1614 def v4f32: LRVecInst<v4f32>;
1615 def v8i16: LRVecInst<v8i16>;
1616 def v16i8: LRVecInst<v16i8>;
1618 def r128: LRRegInst<GPRC>;
1619 def r64: LRRegInst<R64C>;
1620 def f64: LRRegInst<R64FP>;
1621 def r32: LRRegInst<R32C>;
1622 def f32: LRRegInst<R32FP>;
1623 def r16: LRRegInst<R16C>;
1624 def r8: LRRegInst<R8C>;
1627 defm LR: LoadRegister;
1630 // ORC: Bitwise "or" with complement (c = a | ~b)
1632 class ORCInst<dag OOL, dag IOL, list<dag> pattern>:
1633 RRForm<0b10010010000, OOL, IOL, "orc\t$rT, $rA, $rB",
1634 IntegerOp, pattern>;
1636 class ORCVecInst<ValueType vectype>:
1637 ORCInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1638 [(set (vectype VECREG:$rT), (or (vectype VECREG:$rA),
1639 (vnot (vectype VECREG:$rB))))]>;
1641 class ORCRegInst<RegisterClass rclass>:
1642 ORCInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
1643 [(set rclass:$rT, (or rclass:$rA, (not rclass:$rB)))]>;
1645 multiclass BitwiseOrComplement
1647 def v16i8: ORCVecInst<v16i8>;
1648 def v8i16: ORCVecInst<v8i16>;
1649 def v4i32: ORCVecInst<v4i32>;
1650 def v2i64: ORCVecInst<v2i64>;
1652 def r64: ORCRegInst<R64C>;
1653 def r32: ORCRegInst<R32C>;
1654 def r16: ORCRegInst<R16C>;
1655 def r8: ORCRegInst<R8C>;
1658 defm ORC : BitwiseOrComplement;
1660 // OR byte immediate
1661 class ORBIInst<dag OOL, dag IOL, list<dag> pattern>:
1662 RI10Form<0b01100000, OOL, IOL, "orbi\t$rT, $rA, $val",
1663 IntegerOp, pattern>;
1665 class ORBIVecInst<ValueType vectype, PatLeaf immpred>:
1666 ORBIInst<(outs VECREG:$rT), (ins VECREG:$rA, u10imm:$val),
1667 [(set (v16i8 VECREG:$rT), (or (vectype VECREG:$rA),
1668 (vectype immpred:$val)))]>;
1670 multiclass BitwiseOrByteImm
1672 def v16i8: ORBIVecInst<v16i8, v16i8U8Imm>;
1674 def r8: ORBIInst<(outs R8C:$rT), (ins R8C:$rA, u10imm_i8:$val),
1675 [(set R8C:$rT, (or R8C:$rA, immU8:$val))]>;
1678 defm ORBI : BitwiseOrByteImm;
1680 // OR halfword immediate
1681 class ORHIInst<dag OOL, dag IOL, list<dag> pattern>:
1682 RI10Form<0b10100000, OOL, IOL, "orhi\t$rT, $rA, $val",
1683 IntegerOp, pattern>;
1685 class ORHIVecInst<ValueType vectype, PatLeaf immpred>:
1686 ORHIInst<(outs VECREG:$rT), (ins VECREG:$rA, u10imm:$val),
1687 [(set (vectype VECREG:$rT), (or (vectype VECREG:$rA),
1690 multiclass BitwiseOrHalfwordImm
1692 def v8i16: ORHIVecInst<v8i16, v8i16Uns10Imm>;
1694 def r16: ORHIInst<(outs R16C:$rT), (ins R16C:$rA, u10imm:$val),
1695 [(set R16C:$rT, (or R16C:$rA, i16ImmUns10:$val))]>;
1697 // Specialized ORHI form used to promote 8-bit registers to 16-bit
1698 def i8i16: ORHIInst<(outs R16C:$rT), (ins R8C:$rA, s10imm:$val),
1699 [(set R16C:$rT, (or (anyext R8C:$rA),
1700 i16ImmSExt10:$val))]>;
1703 defm ORHI : BitwiseOrHalfwordImm;
1705 class ORIInst<dag OOL, dag IOL, list<dag> pattern>:
1706 RI10Form<0b00100000, OOL, IOL, "ori\t$rT, $rA, $val",
1707 IntegerOp, pattern>;
1709 class ORIVecInst<ValueType vectype, PatLeaf immpred>:
1710 ORIInst<(outs VECREG:$rT), (ins VECREG:$rA, u10imm:$val),
1711 [(set (vectype VECREG:$rT), (or (vectype VECREG:$rA),
1714 // Bitwise "or" with immediate
1715 multiclass BitwiseOrImm
1717 def v4i32: ORIVecInst<v4i32, v4i32Uns10Imm>;
1719 def r32: ORIInst<(outs R32C:$rT), (ins R32C:$rA, u10imm_i32:$val),
1720 [(set R32C:$rT, (or R32C:$rA, i32ImmUns10:$val))]>;
1722 // i16i32: hacked version of the ori instruction to extend 16-bit quantities
1723 // to 32-bit quantities. used exclusively to match "anyext" conversions (vide
1724 // infra "anyext 16->32" pattern.)
1725 def i16i32: ORIInst<(outs R32C:$rT), (ins R16C:$rA, s10imm_i32:$val),
1726 [(set R32C:$rT, (or (anyext R16C:$rA),
1727 i32ImmSExt10:$val))]>;
1729 // i8i32: Hacked version of the ORI instruction to extend 16-bit quantities
1730 // to 32-bit quantities. Used exclusively to match "anyext" conversions (vide
1731 // infra "anyext 16->32" pattern.)
1732 def i8i32: ORIInst<(outs R32C:$rT), (ins R8C:$rA, s10imm_i32:$val),
1733 [(set R32C:$rT, (or (anyext R8C:$rA),
1734 i32ImmSExt10:$val))]>;
1737 defm ORI : BitwiseOrImm;
1739 // ORX: "or" across the vector: or's $rA's word slots leaving the result in
1740 // $rT[0], slots 1-3 are zeroed.
1742 // FIXME: Needs to match an intrinsic pattern.
1744 RRForm<0b10010010000, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1745 "orx\t$rT, $rA, $rB", IntegerOp,
1750 class XORInst<dag OOL, dag IOL, list<dag> pattern> :
1751 RRForm<0b10010010000, OOL, IOL, "xor\t$rT, $rA, $rB",
1752 IntegerOp, pattern>;
1754 class XORVecInst<ValueType vectype>:
1755 XORInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1756 [(set (vectype VECREG:$rT), (xor (vectype VECREG:$rA),
1757 (vectype VECREG:$rB)))]>;
1759 class XORRegInst<RegisterClass rclass>:
1760 XORInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
1761 [(set rclass:$rT, (xor rclass:$rA, rclass:$rB))]>;
1763 multiclass BitwiseExclusiveOr
1765 def v16i8: XORVecInst<v16i8>;
1766 def v8i16: XORVecInst<v8i16>;
1767 def v4i32: XORVecInst<v4i32>;
1768 def v2i64: XORVecInst<v2i64>;
1770 def r128: XORRegInst<GPRC>;
1771 def r64: XORRegInst<R64C>;
1772 def r32: XORRegInst<R32C>;
1773 def r16: XORRegInst<R16C>;
1774 def r8: XORRegInst<R8C>;
1776 // Special forms for floating point instructions.
1777 // fneg and fabs require bitwise logical ops to manipulate the sign bit.
1779 def fneg32: XORInst<(outs R32FP:$rT), (ins R32FP:$rA, R32C:$rB),
1780 [/* no pattern */]>;
1782 def fneg64: XORInst<(outs R64FP:$rT), (ins R64FP:$rA, VECREG:$rB),
1783 [/* no pattern */]>;
1785 def fnegvec: XORInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1786 [/* no pattern, see fneg{32,64} */]>;
1789 defm XOR : BitwiseExclusiveOr;
1791 //==----------------------------------------------------------
1793 class XORBIInst<dag OOL, dag IOL, list<dag> pattern>:
1794 RI10Form<0b01100000, OOL, IOL, "xorbi\t$rT, $rA, $val",
1795 IntegerOp, pattern>;
1797 multiclass XorByteImm
1800 XORBIInst<(outs VECREG:$rT), (ins VECREG:$rA, u10imm:$val),
1801 [(set (v16i8 VECREG:$rT), (xor (v16i8 VECREG:$rA), v16i8U8Imm:$val))]>;
1804 XORBIInst<(outs R8C:$rT), (ins R8C:$rA, u10imm_i8:$val),
1805 [(set R8C:$rT, (xor R8C:$rA, immU8:$val))]>;
1808 defm XORBI : XorByteImm;
1811 RI10Form<0b10100000, (outs VECREG:$rT), (ins VECREG:$rA, u10imm:$val),
1812 "xorhi\t$rT, $rA, $val", IntegerOp,
1813 [(set (v8i16 VECREG:$rT), (xor (v8i16 VECREG:$rA),
1814 v8i16SExt10Imm:$val))]>;
1817 RI10Form<0b10100000, (outs R16C:$rT), (ins R16C:$rA, s10imm:$val),
1818 "xorhi\t$rT, $rA, $val", IntegerOp,
1819 [(set R16C:$rT, (xor R16C:$rA, i16ImmSExt10:$val))]>;
1822 RI10Form<0b00100000, (outs VECREG:$rT), (ins VECREG:$rA, s10imm_i32:$val),
1823 "xori\t$rT, $rA, $val", IntegerOp,
1824 [(set (v4i32 VECREG:$rT), (xor (v4i32 VECREG:$rA),
1825 v4i32SExt10Imm:$val))]>;
1828 RI10Form<0b00100000, (outs R32C:$rT), (ins R32C:$rA, s10imm_i32:$val),
1829 "xori\t$rT, $rA, $val", IntegerOp,
1830 [(set R32C:$rT, (xor R32C:$rA, i32ImmSExt10:$val))]>;
1834 RRForm<0b10010010000, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1835 "nand\t$rT, $rA, $rB", IntegerOp,
1836 [(set (v16i8 VECREG:$rT), (vnot (and (v16i8 VECREG:$rA),
1837 (v16i8 VECREG:$rB))))]>;
1840 RRForm<0b10010010000, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1841 "nand\t$rT, $rA, $rB", IntegerOp,
1842 [(set (v8i16 VECREG:$rT), (vnot (and (v8i16 VECREG:$rA),
1843 (v8i16 VECREG:$rB))))]>;
1846 RRForm<0b10010010000, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1847 "nand\t$rT, $rA, $rB", IntegerOp,
1848 [(set (v4i32 VECREG:$rT), (vnot (and (v4i32 VECREG:$rA),
1849 (v4i32 VECREG:$rB))))]>;
1852 RRForm<0b10010010000, (outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
1853 "nand\t$rT, $rA, $rB", IntegerOp,
1854 [(set R32C:$rT, (not (and R32C:$rA, R32C:$rB)))]>;
1857 RRForm<0b10010010000, (outs R16C:$rT), (ins R16C:$rA, R16C:$rB),
1858 "nand\t$rT, $rA, $rB", IntegerOp,
1859 [(set R16C:$rT, (not (and R16C:$rA, R16C:$rB)))]>;
1862 RRForm<0b10010010000, (outs R8C:$rT), (ins R8C:$rA, R8C:$rB),
1863 "nand\t$rT, $rA, $rB", IntegerOp,
1864 [(set R8C:$rT, (not (and R8C:$rA, R8C:$rB)))]>;
1868 RRForm<0b10010010000, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1869 "nor\t$rT, $rA, $rB", IntegerOp,
1870 [(set (v16i8 VECREG:$rT), (vnot (or (v16i8 VECREG:$rA),
1871 (v16i8 VECREG:$rB))))]>;
1874 RRForm<0b10010010000, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1875 "nor\t$rT, $rA, $rB", IntegerOp,
1876 [(set (v8i16 VECREG:$rT), (vnot (or (v8i16 VECREG:$rA),
1877 (v8i16 VECREG:$rB))))]>;
1880 RRForm<0b10010010000, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1881 "nor\t$rT, $rA, $rB", IntegerOp,
1882 [(set (v4i32 VECREG:$rT), (vnot (or (v4i32 VECREG:$rA),
1883 (v4i32 VECREG:$rB))))]>;
1886 RRForm<0b10010010000, (outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
1887 "nor\t$rT, $rA, $rB", IntegerOp,
1888 [(set R32C:$rT, (not (or R32C:$rA, R32C:$rB)))]>;
1891 RRForm<0b10010010000, (outs R16C:$rT), (ins R16C:$rA, R16C:$rB),
1892 "nor\t$rT, $rA, $rB", IntegerOp,
1893 [(set R16C:$rT, (not (or R16C:$rA, R16C:$rB)))]>;
1896 RRForm<0b10010010000, (outs R8C:$rT), (ins R8C:$rA, R8C:$rB),
1897 "nor\t$rT, $rA, $rB", IntegerOp,
1898 [(set R8C:$rT, (not (or R8C:$rA, R8C:$rB)))]>;
1901 class SELBInst<dag OOL, dag IOL, list<dag> pattern>:
1902 RRRForm<0b1000, OOL, IOL, "selb\t$rT, $rA, $rB, $rC",
1903 IntegerOp, pattern>;
1905 class SELBVecInst<ValueType vectype>:
1906 SELBInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
1907 [(set (vectype VECREG:$rT),
1908 (or (and (vectype VECREG:$rC), (vectype VECREG:$rB)),
1909 (and (vnot (vectype VECREG:$rC)),
1910 (vectype VECREG:$rA))))]>;
1912 class SELBVecVCondInst<ValueType vectype>:
1913 SELBInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
1914 [(set (vectype VECREG:$rT),
1915 (select (vectype VECREG:$rC),
1916 (vectype VECREG:$rB),
1917 (vectype VECREG:$rA)))]>;
1919 class SELBVecCondInst<ValueType vectype>:
1920 SELBInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, R32C:$rC),
1921 [(set (vectype VECREG:$rT),
1923 (vectype VECREG:$rB),
1924 (vectype VECREG:$rA)))]>;
1926 class SELBRegInst<RegisterClass rclass>:
1927 SELBInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB, rclass:$rC),
1929 (or (and rclass:$rB, rclass:$rC),
1930 (and rclass:$rA, (not rclass:$rC))))]>;
1932 class SELBRegCondInst<RegisterClass rcond, RegisterClass rclass>:
1933 SELBInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB, rcond:$rC),
1935 (select rcond:$rC, rclass:$rB, rclass:$rA))]>;
1937 multiclass SelectBits
1939 def v16i8: SELBVecInst<v16i8>;
1940 def v8i16: SELBVecInst<v8i16>;
1941 def v4i32: SELBVecInst<v4i32>;
1942 def v2i64: SELBVecInst<v2i64>;
1944 def r128: SELBRegInst<GPRC>;
1945 def r64: SELBRegInst<R64C>;
1946 def r32: SELBRegInst<R32C>;
1947 def r16: SELBRegInst<R16C>;
1948 def r8: SELBRegInst<R8C>;
1950 def v16i8_cond: SELBVecCondInst<v16i8>;
1951 def v8i16_cond: SELBVecCondInst<v8i16>;
1952 def v4i32_cond: SELBVecCondInst<v4i32>;
1953 def v2i64_cond: SELBVecCondInst<v2i64>;
1955 def v16i8_vcond: SELBVecCondInst<v16i8>;
1956 def v8i16_vcond: SELBVecCondInst<v8i16>;
1957 def v4i32_vcond: SELBVecCondInst<v4i32>;
1958 def v2i64_vcond: SELBVecCondInst<v2i64>;
1961 SELBInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
1962 [(set (v4f32 VECREG:$rT),
1963 (select (v4i32 VECREG:$rC),
1965 (v4f32 VECREG:$rA)))]>;
1967 // SELBr64_cond is defined further down, look for i64 comparisons
1968 def r32_cond: SELBRegCondInst<R32C, R32C>;
1969 def f32_cond: SELBRegCondInst<R32C, R32FP>;
1970 def r16_cond: SELBRegCondInst<R16C, R16C>;
1971 def r8_cond: SELBRegCondInst<R8C, R8C>;
1974 defm SELB : SelectBits;
1976 class SPUselbPatVec<ValueType vectype, SPUInstr inst>:
1977 Pat<(SPUselb (vectype VECREG:$rA), (vectype VECREG:$rB), (vectype VECREG:$rC)),
1978 (inst VECREG:$rA, VECREG:$rB, VECREG:$rC)>;
1980 def : SPUselbPatVec<v16i8, SELBv16i8>;
1981 def : SPUselbPatVec<v8i16, SELBv8i16>;
1982 def : SPUselbPatVec<v4i32, SELBv4i32>;
1983 def : SPUselbPatVec<v2i64, SELBv2i64>;
1985 class SPUselbPatReg<RegisterClass rclass, SPUInstr inst>:
1986 Pat<(SPUselb rclass:$rA, rclass:$rB, rclass:$rC),
1987 (inst rclass:$rA, rclass:$rB, rclass:$rC)>;
1989 def : SPUselbPatReg<R8C, SELBr8>;
1990 def : SPUselbPatReg<R16C, SELBr16>;
1991 def : SPUselbPatReg<R32C, SELBr32>;
1992 def : SPUselbPatReg<R64C, SELBr64>;
1994 // EQV: Equivalence (1 for each same bit, otherwise 0)
1996 // Note: There are a lot of ways to match this bit operator and these patterns
1997 // attempt to be as exhaustive as possible.
1999 class EQVInst<dag OOL, dag IOL, list<dag> pattern>:
2000 RRForm<0b10010010000, OOL, IOL, "eqv\t$rT, $rA, $rB",
2001 IntegerOp, pattern>;
2003 class EQVVecInst<ValueType vectype>:
2004 EQVInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
2005 [(set (vectype VECREG:$rT),
2006 (or (and (vectype VECREG:$rA), (vectype VECREG:$rB)),
2007 (and (vnot (vectype VECREG:$rA)),
2008 (vnot (vectype VECREG:$rB)))))]>;
2010 class EQVRegInst<RegisterClass rclass>:
2011 EQVInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
2012 [(set rclass:$rT, (or (and rclass:$rA, rclass:$rB),
2013 (and (not rclass:$rA), (not rclass:$rB))))]>;
2015 class EQVVecPattern1<ValueType vectype>:
2016 EQVInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
2017 [(set (vectype VECREG:$rT),
2018 (xor (vectype VECREG:$rA), (vnot (vectype VECREG:$rB))))]>;
2020 class EQVRegPattern1<RegisterClass rclass>:
2021 EQVInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
2022 [(set rclass:$rT, (xor rclass:$rA, (not rclass:$rB)))]>;
2024 class EQVVecPattern2<ValueType vectype>:
2025 EQVInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
2026 [(set (vectype VECREG:$rT),
2027 (or (and (vectype VECREG:$rA), (vectype VECREG:$rB)),
2028 (vnot (or (vectype VECREG:$rA), (vectype VECREG:$rB)))))]>;
2030 class EQVRegPattern2<RegisterClass rclass>:
2031 EQVInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
2033 (or (and rclass:$rA, rclass:$rB),
2034 (not (or rclass:$rA, rclass:$rB))))]>;
2036 class EQVVecPattern3<ValueType vectype>:
2037 EQVInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
2038 [(set (vectype VECREG:$rT),
2039 (not (xor (vectype VECREG:$rA), (vectype VECREG:$rB))))]>;
2041 class EQVRegPattern3<RegisterClass rclass>:
2042 EQVInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
2043 [(set rclass:$rT, (not (xor rclass:$rA, rclass:$rB)))]>;
2045 multiclass BitEquivalence
2047 def v16i8: EQVVecInst<v16i8>;
2048 def v8i16: EQVVecInst<v8i16>;
2049 def v4i32: EQVVecInst<v4i32>;
2050 def v2i64: EQVVecInst<v2i64>;
2052 def v16i8_1: EQVVecPattern1<v16i8>;
2053 def v8i16_1: EQVVecPattern1<v8i16>;
2054 def v4i32_1: EQVVecPattern1<v4i32>;
2055 def v2i64_1: EQVVecPattern1<v2i64>;
2057 def v16i8_2: EQVVecPattern2<v16i8>;
2058 def v8i16_2: EQVVecPattern2<v8i16>;
2059 def v4i32_2: EQVVecPattern2<v4i32>;
2060 def v2i64_2: EQVVecPattern2<v2i64>;
2062 def v16i8_3: EQVVecPattern3<v16i8>;
2063 def v8i16_3: EQVVecPattern3<v8i16>;
2064 def v4i32_3: EQVVecPattern3<v4i32>;
2065 def v2i64_3: EQVVecPattern3<v2i64>;
2067 def r128: EQVRegInst<GPRC>;
2068 def r64: EQVRegInst<R64C>;
2069 def r32: EQVRegInst<R32C>;
2070 def r16: EQVRegInst<R16C>;
2071 def r8: EQVRegInst<R8C>;
2073 def r128_1: EQVRegPattern1<GPRC>;
2074 def r64_1: EQVRegPattern1<R64C>;
2075 def r32_1: EQVRegPattern1<R32C>;
2076 def r16_1: EQVRegPattern1<R16C>;
2077 def r8_1: EQVRegPattern1<R8C>;
2079 def r128_2: EQVRegPattern2<GPRC>;
2080 def r64_2: EQVRegPattern2<R64C>;
2081 def r32_2: EQVRegPattern2<R32C>;
2082 def r16_2: EQVRegPattern2<R16C>;
2083 def r8_2: EQVRegPattern2<R8C>;
2085 def r128_3: EQVRegPattern3<GPRC>;
2086 def r64_3: EQVRegPattern3<R64C>;
2087 def r32_3: EQVRegPattern3<R32C>;
2088 def r16_3: EQVRegPattern3<R16C>;
2089 def r8_3: EQVRegPattern3<R8C>;
2092 defm EQV: BitEquivalence;
2094 //===----------------------------------------------------------------------===//
2095 // Vector shuffle...
2096 //===----------------------------------------------------------------------===//
2097 // SPUshuffle is generated in LowerVECTOR_SHUFFLE and gets replaced with SHUFB.
2098 // See the SPUshuffle SDNode operand above, which sets up the DAG pattern
2099 // matcher to emit something when the LowerVECTOR_SHUFFLE generates a node with
2100 // the SPUISD::SHUFB opcode.
2101 //===----------------------------------------------------------------------===//
2103 class SHUFBInst<dag OOL, dag IOL, list<dag> pattern>:
2104 RRRForm<0b1000, OOL, IOL, "shufb\t$rT, $rA, $rB, $rC",
2105 IntegerOp, pattern>;
2107 class SHUFBVecInst<ValueType resultvec, ValueType maskvec>:
2108 SHUFBInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
2109 [(set (resultvec VECREG:$rT),
2110 (SPUshuffle (resultvec VECREG:$rA),
2111 (resultvec VECREG:$rB),
2112 (maskvec VECREG:$rC)))]>;
2114 class SHUFBGPRCInst:
2115 SHUFBInst<(outs VECREG:$rT), (ins GPRC:$rA, GPRC:$rB, VECREG:$rC),
2116 [/* no pattern */]>;
2118 multiclass ShuffleBytes
2120 def v16i8 : SHUFBVecInst<v16i8, v16i8>;
2121 def v16i8_m32 : SHUFBVecInst<v16i8, v4i32>;
2122 def v8i16 : SHUFBVecInst<v8i16, v16i8>;
2123 def v8i16_m32 : SHUFBVecInst<v8i16, v4i32>;
2124 def v4i32 : SHUFBVecInst<v4i32, v16i8>;
2125 def v4i32_m32 : SHUFBVecInst<v4i32, v4i32>;
2126 def v2i64 : SHUFBVecInst<v2i64, v16i8>;
2127 def v2i64_m32 : SHUFBVecInst<v2i64, v4i32>;
2129 def v4f32 : SHUFBVecInst<v4f32, v16i8>;
2130 def v4f32_m32 : SHUFBVecInst<v4f32, v4i32>;
2132 def v2f64 : SHUFBVecInst<v2f64, v16i8>;
2133 def v2f64_m32 : SHUFBVecInst<v2f64, v4i32>;
2135 def gprc : SHUFBGPRCInst;
2138 defm SHUFB : ShuffleBytes;
2140 //===----------------------------------------------------------------------===//
2141 // Shift and rotate group:
2142 //===----------------------------------------------------------------------===//
2144 class SHLHInst<dag OOL, dag IOL, list<dag> pattern>:
2145 RRForm<0b11111010000, OOL, IOL, "shlh\t$rT, $rA, $rB",
2146 RotateShift, pattern>;
2148 class SHLHVecInst<ValueType vectype>:
2149 SHLHInst<(outs VECREG:$rT), (ins VECREG:$rA, R16C:$rB),
2150 [(set (vectype VECREG:$rT),
2151 (SPUvec_shl (vectype VECREG:$rA), R16C:$rB))]>;
2153 // $rB gets promoted to 32-bit register type when confronted with
2154 // this llvm assembly code:
2156 // define i16 @shlh_i16_1(i16 %arg1, i16 %arg2) {
2157 // %A = shl i16 %arg1, %arg2
2161 multiclass ShiftLeftHalfword
2163 def v8i16: SHLHVecInst<v8i16>;
2164 def r16: SHLHInst<(outs R16C:$rT), (ins R16C:$rA, R16C:$rB),
2165 [(set R16C:$rT, (shl R16C:$rA, R16C:$rB))]>;
2166 def r16_r32: SHLHInst<(outs R16C:$rT), (ins R16C:$rA, R32C:$rB),
2167 [(set R16C:$rT, (shl R16C:$rA, R32C:$rB))]>;
2170 defm SHLH : ShiftLeftHalfword;
2172 //===----------------------------------------------------------------------===//
2174 class SHLHIInst<dag OOL, dag IOL, list<dag> pattern>:
2175 RI7Form<0b11111010000, OOL, IOL, "shlhi\t$rT, $rA, $val",
2176 RotateShift, pattern>;
2178 class SHLHIVecInst<ValueType vectype>:
2179 SHLHIInst<(outs VECREG:$rT), (ins VECREG:$rA, u7imm:$val),
2180 [(set (vectype VECREG:$rT),
2181 (SPUvec_shl (vectype VECREG:$rA), (i16 uimm7:$val)))]>;
2183 multiclass ShiftLeftHalfwordImm
2185 def v8i16: SHLHIVecInst<v8i16>;
2186 def r16: SHLHIInst<(outs R16C:$rT), (ins R16C:$rA, u7imm:$val),
2187 [(set R16C:$rT, (shl R16C:$rA, (i16 uimm7:$val)))]>;
2190 defm SHLHI : ShiftLeftHalfwordImm;
2192 def : Pat<(SPUvec_shl (v8i16 VECREG:$rA), (i32 uimm7:$val)),
2193 (SHLHIv8i16 VECREG:$rA, uimm7:$val)>;
2195 def : Pat<(shl R16C:$rA, (i32 uimm7:$val)),
2196 (SHLHIr16 R16C:$rA, uimm7:$val)>;
2198 //===----------------------------------------------------------------------===//
2200 class SHLInst<dag OOL, dag IOL, list<dag> pattern>:
2201 RRForm<0b11111010000, OOL, IOL, "shl\t$rT, $rA, $rB",
2202 RotateShift, pattern>;
2204 multiclass ShiftLeftWord
2207 SHLInst<(outs VECREG:$rT), (ins VECREG:$rA, R16C:$rB),
2208 [(set (v4i32 VECREG:$rT),
2209 (SPUvec_shl (v4i32 VECREG:$rA), R16C:$rB))]>;
2211 SHLInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
2212 [(set R32C:$rT, (shl R32C:$rA, R32C:$rB))]>;
2215 defm SHL: ShiftLeftWord;
2217 //===----------------------------------------------------------------------===//
2219 class SHLIInst<dag OOL, dag IOL, list<dag> pattern>:
2220 RI7Form<0b11111010000, OOL, IOL, "shli\t$rT, $rA, $val",
2221 RotateShift, pattern>;
2223 multiclass ShiftLeftWordImm
2226 SHLIInst<(outs VECREG:$rT), (ins VECREG:$rA, u7imm_i32:$val),
2227 [(set (v4i32 VECREG:$rT),
2228 (SPUvec_shl (v4i32 VECREG:$rA), (i32 uimm7:$val)))]>;
2231 SHLIInst<(outs R32C:$rT), (ins R32C:$rA, u7imm_i32:$val),
2232 [(set R32C:$rT, (shl R32C:$rA, (i32 uimm7:$val)))]>;
2235 defm SHLI : ShiftLeftWordImm;
2237 //===----------------------------------------------------------------------===//
2238 // SHLQBI vec form: Note that this will shift the entire vector (the 128-bit
2239 // register) to the left. Vector form is here to ensure type correctness.
2241 // The shift count is in the lowest 3 bits (29-31) of $rB, so only a bit shift
2242 // of 7 bits is actually possible.
2244 // Note also that SHLQBI/SHLQBII are used in conjunction with SHLQBY/SHLQBYI
2245 // to shift i64 and i128. SHLQBI is the residual left over after shifting by
2246 // bytes with SHLQBY.
2248 class SHLQBIInst<dag OOL, dag IOL, list<dag> pattern>:
2249 RRForm<0b11011011100, OOL, IOL, "shlqbi\t$rT, $rA, $rB",
2250 RotateShift, pattern>;
2252 class SHLQBIVecInst<ValueType vectype>:
2253 SHLQBIInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2254 [(set (vectype VECREG:$rT),
2255 (SPUshlquad_l_bits (vectype VECREG:$rA), R32C:$rB))]>;
2257 multiclass ShiftLeftQuadByBits
2259 def v16i8: SHLQBIVecInst<v16i8>;
2260 def v8i16: SHLQBIVecInst<v8i16>;
2261 def v4i32: SHLQBIVecInst<v4i32>;
2262 def v4f32: SHLQBIVecInst<v4f32>;
2263 def v2i64: SHLQBIVecInst<v2i64>;
2264 def v2f64: SHLQBIVecInst<v2f64>;
2267 defm SHLQBI : ShiftLeftQuadByBits;
2269 // See note above on SHLQBI. In this case, the predicate actually does then
2270 // enforcement, whereas with SHLQBI, we have to "take it on faith."
2271 class SHLQBIIInst<dag OOL, dag IOL, list<dag> pattern>:
2272 RI7Form<0b11011111100, OOL, IOL, "shlqbii\t$rT, $rA, $val",
2273 RotateShift, pattern>;
2275 class SHLQBIIVecInst<ValueType vectype>:
2276 SHLQBIIInst<(outs VECREG:$rT), (ins VECREG:$rA, u7imm_i32:$val),
2277 [(set (vectype VECREG:$rT),
2278 (SPUshlquad_l_bits (vectype VECREG:$rA), (i32 bitshift:$val)))]>;
2280 multiclass ShiftLeftQuadByBitsImm
2282 def v16i8 : SHLQBIIVecInst<v16i8>;
2283 def v8i16 : SHLQBIIVecInst<v8i16>;
2284 def v4i32 : SHLQBIIVecInst<v4i32>;
2285 def v4f32 : SHLQBIIVecInst<v4f32>;
2286 def v2i64 : SHLQBIIVecInst<v2i64>;
2287 def v2f64 : SHLQBIIVecInst<v2f64>;
2290 defm SHLQBII : ShiftLeftQuadByBitsImm;
2292 // SHLQBY, SHLQBYI vector forms: Shift the entire vector to the left by bytes,
2293 // not by bits. See notes above on SHLQBI.
2295 class SHLQBYInst<dag OOL, dag IOL, list<dag> pattern>:
2296 RI7Form<0b11111011100, OOL, IOL, "shlqby\t$rT, $rA, $rB",
2297 RotateShift, pattern>;
2299 class SHLQBYVecInst<ValueType vectype>:
2300 SHLQBYInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2301 [(set (vectype VECREG:$rT),
2302 (SPUshlquad_l_bytes (vectype VECREG:$rA), R32C:$rB))]>;
2304 multiclass ShiftLeftQuadBytes
2306 def v16i8: SHLQBYVecInst<v16i8>;
2307 def v8i16: SHLQBYVecInst<v8i16>;
2308 def v4i32: SHLQBYVecInst<v4i32>;
2309 def v4f32: SHLQBYVecInst<v4f32>;
2310 def v2i64: SHLQBYVecInst<v2i64>;
2311 def v2f64: SHLQBYVecInst<v2f64>;
2312 def r128: SHLQBYInst<(outs GPRC:$rT), (ins GPRC:$rA, R32C:$rB),
2313 [(set GPRC:$rT, (SPUshlquad_l_bytes GPRC:$rA, R32C:$rB))]>;
2316 defm SHLQBY: ShiftLeftQuadBytes;
2318 class SHLQBYIInst<dag OOL, dag IOL, list<dag> pattern>:
2319 RI7Form<0b11111111100, OOL, IOL, "shlqbyi\t$rT, $rA, $val",
2320 RotateShift, pattern>;
2322 class SHLQBYIVecInst<ValueType vectype>:
2323 SHLQBYIInst<(outs VECREG:$rT), (ins VECREG:$rA, u7imm_i32:$val),
2324 [(set (vectype VECREG:$rT),
2325 (SPUshlquad_l_bytes (vectype VECREG:$rA), (i32 uimm7:$val)))]>;
2327 multiclass ShiftLeftQuadBytesImm
2329 def v16i8: SHLQBYIVecInst<v16i8>;
2330 def v8i16: SHLQBYIVecInst<v8i16>;
2331 def v4i32: SHLQBYIVecInst<v4i32>;
2332 def v4f32: SHLQBYIVecInst<v4f32>;
2333 def v2i64: SHLQBYIVecInst<v2i64>;
2334 def v2f64: SHLQBYIVecInst<v2f64>;
2335 def r128: SHLQBYIInst<(outs GPRC:$rT), (ins GPRC:$rA, u7imm_i32:$val),
2337 (SPUshlquad_l_bytes GPRC:$rA, (i32 uimm7:$val)))]>;
2340 defm SHLQBYI : ShiftLeftQuadBytesImm;
2342 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2344 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2345 class ROTHInst<dag OOL, dag IOL, list<dag> pattern>:
2346 RRForm<0b00111010000, OOL, IOL, "roth\t$rT, $rA, $rB",
2347 RotateShift, pattern>;
2349 class ROTHVecInst<ValueType vectype>:
2350 ROTHInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
2351 [(set (vectype VECREG:$rT),
2352 (SPUvec_rotl VECREG:$rA, VECREG:$rB))]>;
2354 class ROTHRegInst<RegisterClass rclass>:
2355 ROTHInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
2356 [(set rclass:$rT, (rotl rclass:$rA, rclass:$rB))]>;
2358 multiclass RotateLeftHalfword
2360 def v8i16: ROTHVecInst<v8i16>;
2361 def r16: ROTHRegInst<R16C>;
2364 defm ROTH: RotateLeftHalfword;
2366 def ROTHr16_r32: ROTHInst<(outs R16C:$rT), (ins R16C:$rA, R32C:$rB),
2367 [(set R16C:$rT, (rotl R16C:$rA, R32C:$rB))]>;
2369 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2370 // Rotate halfword, immediate:
2371 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2372 class ROTHIInst<dag OOL, dag IOL, list<dag> pattern>:
2373 RI7Form<0b00111110000, OOL, IOL, "rothi\t$rT, $rA, $val",
2374 RotateShift, pattern>;
2376 class ROTHIVecInst<ValueType vectype>:
2377 ROTHIInst<(outs VECREG:$rT), (ins VECREG:$rA, u7imm:$val),
2378 [(set (vectype VECREG:$rT),
2379 (SPUvec_rotl VECREG:$rA, (i16 uimm7:$val)))]>;
2381 multiclass RotateLeftHalfwordImm
2383 def v8i16: ROTHIVecInst<v8i16>;
2384 def r16: ROTHIInst<(outs R16C:$rT), (ins R16C:$rA, u7imm:$val),
2385 [(set R16C:$rT, (rotl R16C:$rA, (i16 uimm7:$val)))]>;
2386 def r16_r32: ROTHIInst<(outs R16C:$rT), (ins R16C:$rA, u7imm_i32:$val),
2387 [(set R16C:$rT, (rotl R16C:$rA, (i32 uimm7:$val)))]>;
2390 defm ROTHI: RotateLeftHalfwordImm;
2392 def : Pat<(SPUvec_rotl VECREG:$rA, (i32 uimm7:$val)),
2393 (ROTHIv8i16 VECREG:$rA, imm:$val)>;
2395 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2397 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2399 class ROTInst<dag OOL, dag IOL, list<dag> pattern>:
2400 RRForm<0b00011010000, OOL, IOL, "rot\t$rT, $rA, $rB",
2401 RotateShift, pattern>;
2403 class ROTVecInst<ValueType vectype>:
2404 ROTInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2405 [(set (vectype VECREG:$rT),
2406 (SPUvec_rotl (vectype VECREG:$rA), R32C:$rB))]>;
2408 class ROTRegInst<RegisterClass rclass>:
2409 ROTInst<(outs rclass:$rT), (ins rclass:$rA, R32C:$rB),
2411 (rotl rclass:$rA, R32C:$rB))]>;
2413 multiclass RotateLeftWord
2415 def v4i32: ROTVecInst<v4i32>;
2416 def r32: ROTRegInst<R32C>;
2419 defm ROT: RotateLeftWord;
2421 // The rotate amount is in the same bits whether we've got an 8-bit, 16-bit or
2423 def ROTr32_r16_anyext:
2424 ROTInst<(outs R32C:$rT), (ins R32C:$rA, R16C:$rB),
2425 [(set R32C:$rT, (rotl R32C:$rA, (i32 (anyext R16C:$rB))))]>;
2427 def : Pat<(rotl R32C:$rA, (i32 (zext R16C:$rB))),
2428 (ROTr32_r16_anyext R32C:$rA, R16C:$rB)>;
2430 def : Pat<(rotl R32C:$rA, (i32 (sext R16C:$rB))),
2431 (ROTr32_r16_anyext R32C:$rA, R16C:$rB)>;
2433 def ROTr32_r8_anyext:
2434 ROTInst<(outs R32C:$rT), (ins R32C:$rA, R8C:$rB),
2435 [(set R32C:$rT, (rotl R32C:$rA, (i32 (anyext R8C:$rB))))]>;
2437 def : Pat<(rotl R32C:$rA, (i32 (zext R8C:$rB))),
2438 (ROTr32_r8_anyext R32C:$rA, R8C:$rB)>;
2440 def : Pat<(rotl R32C:$rA, (i32 (sext R8C:$rB))),
2441 (ROTr32_r8_anyext R32C:$rA, R8C:$rB)>;
2443 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2444 // Rotate word, immediate
2445 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2447 class ROTIInst<dag OOL, dag IOL, list<dag> pattern>:
2448 RI7Form<0b00011110000, OOL, IOL, "roti\t$rT, $rA, $val",
2449 RotateShift, pattern>;
2451 class ROTIVecInst<ValueType vectype, Operand optype, ValueType inttype, PatLeaf pred>:
2452 ROTIInst<(outs VECREG:$rT), (ins VECREG:$rA, optype:$val),
2453 [(set (vectype VECREG:$rT),
2454 (SPUvec_rotl (vectype VECREG:$rA), (inttype pred:$val)))]>;
2456 class ROTIRegInst<RegisterClass rclass, Operand optype, ValueType inttype, PatLeaf pred>:
2457 ROTIInst<(outs rclass:$rT), (ins rclass:$rA, optype:$val),
2458 [(set rclass:$rT, (rotl rclass:$rA, (inttype pred:$val)))]>;
2460 multiclass RotateLeftWordImm
2462 def v4i32: ROTIVecInst<v4i32, u7imm_i32, i32, uimm7>;
2463 def v4i32_i16: ROTIVecInst<v4i32, u7imm, i16, uimm7>;
2464 def v4i32_i8: ROTIVecInst<v4i32, u7imm_i8, i8, uimm7>;
2466 def r32: ROTIRegInst<R32C, u7imm_i32, i32, uimm7>;
2467 def r32_i16: ROTIRegInst<R32C, u7imm, i16, uimm7>;
2468 def r32_i8: ROTIRegInst<R32C, u7imm_i8, i8, uimm7>;
2471 defm ROTI : RotateLeftWordImm;
2473 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2474 // Rotate quad by byte (count)
2475 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2477 class ROTQBYInst<dag OOL, dag IOL, list<dag> pattern>:
2478 RRForm<0b00111011100, OOL, IOL, "rotqby\t$rT, $rA, $rB",
2479 RotateShift, pattern>;
2481 class ROTQBYVecInst<ValueType vectype>:
2482 ROTQBYInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2483 [(set (vectype VECREG:$rT),
2484 (SPUrotbytes_left (vectype VECREG:$rA), R32C:$rB))]>;
2486 multiclass RotateQuadLeftByBytes
2488 def v16i8: ROTQBYVecInst<v16i8>;
2489 def v8i16: ROTQBYVecInst<v8i16>;
2490 def v4i32: ROTQBYVecInst<v4i32>;
2491 def v4f32: ROTQBYVecInst<v4f32>;
2492 def v2i64: ROTQBYVecInst<v2i64>;
2493 def v2f64: ROTQBYVecInst<v2f64>;
2496 defm ROTQBY: RotateQuadLeftByBytes;
2498 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2499 // Rotate quad by byte (count), immediate
2500 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2502 class ROTQBYIInst<dag OOL, dag IOL, list<dag> pattern>:
2503 RI7Form<0b00111111100, OOL, IOL, "rotqbyi\t$rT, $rA, $val",
2504 RotateShift, pattern>;
2506 class ROTQBYIVecInst<ValueType vectype>:
2507 ROTQBYIInst<(outs VECREG:$rT), (ins VECREG:$rA, u7imm:$val),
2508 [(set (vectype VECREG:$rT),
2509 (SPUrotbytes_left (vectype VECREG:$rA), (i16 uimm7:$val)))]>;
2511 multiclass RotateQuadByBytesImm
2513 def v16i8: ROTQBYIVecInst<v16i8>;
2514 def v8i16: ROTQBYIVecInst<v8i16>;
2515 def v4i32: ROTQBYIVecInst<v4i32>;
2516 def v4f32: ROTQBYIVecInst<v4f32>;
2517 def v2i64: ROTQBYIVecInst<v2i64>;
2518 def vfi64: ROTQBYIVecInst<v2f64>;
2521 defm ROTQBYI: RotateQuadByBytesImm;
2523 // See ROTQBY note above.
2524 class ROTQBYBIInst<dag OOL, dag IOL, list<dag> pattern>:
2525 RI7Form<0b00110011100, OOL, IOL,
2526 "rotqbybi\t$rT, $rA, $shift",
2527 RotateShift, pattern>;
2529 class ROTQBYBIVecInst<ValueType vectype, RegisterClass rclass>:
2530 ROTQBYBIInst<(outs VECREG:$rT), (ins VECREG:$rA, rclass:$shift),
2531 [(set (vectype VECREG:$rT),
2532 (SPUrotbytes_left_bits (vectype VECREG:$rA), rclass:$shift))]>;
2534 multiclass RotateQuadByBytesByBitshift {
2535 def v16i8_r32: ROTQBYBIVecInst<v16i8, R32C>;
2536 def v8i16_r32: ROTQBYBIVecInst<v8i16, R32C>;
2537 def v4i32_r32: ROTQBYBIVecInst<v4i32, R32C>;
2538 def v2i64_r32: ROTQBYBIVecInst<v2i64, R32C>;
2541 defm ROTQBYBI : RotateQuadByBytesByBitshift;
2543 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2544 // See ROTQBY note above.
2546 // Assume that the user of this instruction knows to shift the rotate count
2548 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2550 class ROTQBIInst<dag OOL, dag IOL, list<dag> pattern>:
2551 RRForm<0b00011011100, OOL, IOL, "rotqbi\t$rT, $rA, $rB",
2552 RotateShift, pattern>;
2554 class ROTQBIVecInst<ValueType vectype>:
2555 ROTQBIInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2556 [/* no pattern yet */]>;
2558 class ROTQBIRegInst<RegisterClass rclass>:
2559 ROTQBIInst<(outs rclass:$rT), (ins rclass:$rA, R32C:$rB),
2560 [/* no pattern yet */]>;
2562 multiclass RotateQuadByBitCount
2564 def v16i8: ROTQBIVecInst<v16i8>;
2565 def v8i16: ROTQBIVecInst<v8i16>;
2566 def v4i32: ROTQBIVecInst<v4i32>;
2567 def v2i64: ROTQBIVecInst<v2i64>;
2569 def r128: ROTQBIRegInst<GPRC>;
2570 def r64: ROTQBIRegInst<R64C>;
2573 defm ROTQBI: RotateQuadByBitCount;
2575 class ROTQBIIInst<dag OOL, dag IOL, list<dag> pattern>:
2576 RI7Form<0b00011111100, OOL, IOL, "rotqbii\t$rT, $rA, $val",
2577 RotateShift, pattern>;
2579 class ROTQBIIVecInst<ValueType vectype, Operand optype, ValueType inttype,
2581 ROTQBIIInst<(outs VECREG:$rT), (ins VECREG:$rA, optype:$val),
2582 [/* no pattern yet */]>;
2584 class ROTQBIIRegInst<RegisterClass rclass, Operand optype, ValueType inttype,
2586 ROTQBIIInst<(outs rclass:$rT), (ins rclass:$rA, optype:$val),
2587 [/* no pattern yet */]>;
2589 multiclass RotateQuadByBitCountImm
2591 def v16i8: ROTQBIIVecInst<v16i8, u7imm_i32, i32, uimm7>;
2592 def v8i16: ROTQBIIVecInst<v8i16, u7imm_i32, i32, uimm7>;
2593 def v4i32: ROTQBIIVecInst<v4i32, u7imm_i32, i32, uimm7>;
2594 def v2i64: ROTQBIIVecInst<v2i64, u7imm_i32, i32, uimm7>;
2596 def r128: ROTQBIIRegInst<GPRC, u7imm_i32, i32, uimm7>;
2597 def r64: ROTQBIIRegInst<R64C, u7imm_i32, i32, uimm7>;
2600 defm ROTQBII : RotateQuadByBitCountImm;
2602 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2603 // ROTHM v8i16 form:
2604 // NOTE(1): No vector rotate is generated by the C/C++ frontend (today),
2605 // so this only matches a synthetically generated/lowered code
2607 // NOTE(2): $rB must be negated before the right rotate!
2608 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2610 class ROTHMInst<dag OOL, dag IOL, list<dag> pattern>:
2611 RRForm<0b10111010000, OOL, IOL, "rothm\t$rT, $rA, $rB",
2612 RotateShift, pattern>;
2615 ROTHMInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2616 [/* see patterns below - $rB must be negated */]>;
2618 def : Pat<(SPUvec_srl (v8i16 VECREG:$rA), R32C:$rB),
2619 (ROTHMv8i16 VECREG:$rA, (SFIr32 R32C:$rB, 0))>;
2621 def : Pat<(SPUvec_srl (v8i16 VECREG:$rA), R16C:$rB),
2622 (ROTHMv8i16 VECREG:$rA,
2623 (SFIr32 (XSHWr16 R16C:$rB), 0))>;
2625 def : Pat<(SPUvec_srl (v8i16 VECREG:$rA), R8C:$rB),
2626 (ROTHMv8i16 VECREG:$rA,
2627 (SFIr32 (XSHWr16 (XSBHr8 R8C:$rB) ), 0))>;
2629 // ROTHM r16 form: Rotate 16-bit quantity to right, zero fill at the left
2630 // Note: This instruction doesn't match a pattern because rB must be negated
2631 // for the instruction to work. Thus, the pattern below the instruction!
2634 ROTHMInst<(outs R16C:$rT), (ins R16C:$rA, R32C:$rB),
2635 [/* see patterns below - $rB must be negated! */]>;
2637 def : Pat<(srl R16C:$rA, R32C:$rB),
2638 (ROTHMr16 R16C:$rA, (SFIr32 R32C:$rB, 0))>;
2640 def : Pat<(srl R16C:$rA, R16C:$rB),
2642 (SFIr32 (XSHWr16 R16C:$rB), 0))>;
2644 def : Pat<(srl R16C:$rA, R8C:$rB),
2646 (SFIr32 (XSHWr16 (XSBHr8 R8C:$rB) ), 0))>;
2648 // ROTHMI v8i16 form: See the comment for ROTHM v8i16. The difference here is
2649 // that the immediate can be complemented, so that the user doesn't have to
2652 class ROTHMIInst<dag OOL, dag IOL, list<dag> pattern>:
2653 RI7Form<0b10111110000, OOL, IOL, "rothmi\t$rT, $rA, $val",
2654 RotateShift, pattern>;
2657 ROTHMIInst<(outs VECREG:$rT), (ins VECREG:$rA, rothNeg7imm:$val),
2658 [/* no pattern */]>;
2660 def : Pat<(SPUvec_srl (v8i16 VECREG:$rA), (i32 imm:$val)),
2661 (ROTHMIv8i16 VECREG:$rA, imm:$val)>;
2663 def: Pat<(SPUvec_srl (v8i16 VECREG:$rA), (i16 imm:$val)),
2664 (ROTHMIv8i16 VECREG:$rA, imm:$val)>;
2666 def: Pat<(SPUvec_srl (v8i16 VECREG:$rA), (i8 imm:$val)),
2667 (ROTHMIv8i16 VECREG:$rA, imm:$val)>;
2670 ROTHMIInst<(outs R16C:$rT), (ins R16C:$rA, rothNeg7imm:$val),
2671 [/* no pattern */]>;
2673 def: Pat<(srl R16C:$rA, (i32 uimm7:$val)),
2674 (ROTHMIr16 R16C:$rA, uimm7:$val)>;
2676 def: Pat<(srl R16C:$rA, (i16 uimm7:$val)),
2677 (ROTHMIr16 R16C:$rA, uimm7:$val)>;
2679 def: Pat<(srl R16C:$rA, (i8 uimm7:$val)),
2680 (ROTHMIr16 R16C:$rA, uimm7:$val)>;
2682 // ROTM v4i32 form: See the ROTHM v8i16 comments.
2683 class ROTMInst<dag OOL, dag IOL, list<dag> pattern>:
2684 RRForm<0b10011010000, OOL, IOL, "rotm\t$rT, $rA, $rB",
2685 RotateShift, pattern>;
2688 ROTMInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2689 [/* see patterns below - $rB must be negated */]>;
2691 def : Pat<(SPUvec_srl VECREG:$rA, R32C:$rB),
2692 (ROTMv4i32 VECREG:$rA, (SFIr32 R32C:$rB, 0))>;
2694 def : Pat<(SPUvec_srl VECREG:$rA, R16C:$rB),
2695 (ROTMv4i32 VECREG:$rA,
2696 (SFIr32 (XSHWr16 R16C:$rB), 0))>;
2698 def : Pat<(SPUvec_srl VECREG:$rA, R8C:$rB),
2699 (ROTMv4i32 VECREG:$rA,
2700 (SFIr32 (XSHWr16 (XSBHr8 R8C:$rB)), 0))>;
2703 ROTMInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
2704 [/* see patterns below - $rB must be negated */]>;
2706 def : Pat<(srl R32C:$rA, R32C:$rB),
2707 (ROTMr32 R32C:$rA, (SFIr32 R32C:$rB, 0))>;
2709 def : Pat<(srl R32C:$rA, R16C:$rB),
2711 (SFIr32 (XSHWr16 R16C:$rB), 0))>;
2713 def : Pat<(srl R32C:$rA, R8C:$rB),
2715 (SFIr32 (XSHWr16 (XSBHr8 R8C:$rB)), 0))>;
2717 // ROTMI v4i32 form: See the comment for ROTHM v8i16.
2719 RI7Form<0b10011110000, (outs VECREG:$rT), (ins VECREG:$rA, rotNeg7imm:$val),
2720 "rotmi\t$rT, $rA, $val", RotateShift,
2721 [(set (v4i32 VECREG:$rT),
2722 (SPUvec_srl VECREG:$rA, (i32 uimm7:$val)))]>;
2724 def : Pat<(SPUvec_srl VECREG:$rA, (i16 uimm7:$val)),
2725 (ROTMIv4i32 VECREG:$rA, uimm7:$val)>;
2727 def : Pat<(SPUvec_srl VECREG:$rA, (i8 uimm7:$val)),
2728 (ROTMIv4i32 VECREG:$rA, uimm7:$val)>;
2730 // ROTMI r32 form: know how to complement the immediate value.
2732 RI7Form<0b10011110000, (outs R32C:$rT), (ins R32C:$rA, rotNeg7imm:$val),
2733 "rotmi\t$rT, $rA, $val", RotateShift,
2734 [(set R32C:$rT, (srl R32C:$rA, (i32 uimm7:$val)))]>;
2736 def : Pat<(srl R32C:$rA, (i16 imm:$val)),
2737 (ROTMIr32 R32C:$rA, uimm7:$val)>;
2739 def : Pat<(srl R32C:$rA, (i8 imm:$val)),
2740 (ROTMIr32 R32C:$rA, uimm7:$val)>;
2742 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2743 // ROTQMBY: This is a vector form merely so that when used in an
2744 // instruction pattern, type checking will succeed. This instruction assumes
2745 // that the user knew to negate $rB.
2746 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2748 class ROTQMBYInst<dag OOL, dag IOL, list<dag> pattern>:
2749 RRForm<0b10111011100, OOL, IOL, "rotqmby\t$rT, $rA, $rB",
2750 RotateShift, pattern>;
2752 class ROTQMBYVecInst<ValueType vectype>:
2753 ROTQMBYInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2754 [/* no pattern, $rB must be negated */]>;
2756 class ROTQMBYRegInst<RegisterClass rclass>:
2757 ROTQMBYInst<(outs rclass:$rT), (ins rclass:$rA, R32C:$rB),
2758 [/* no pattern */]>;
2760 multiclass RotateQuadBytes
2762 def v16i8: ROTQMBYVecInst<v16i8>;
2763 def v8i16: ROTQMBYVecInst<v8i16>;
2764 def v4i32: ROTQMBYVecInst<v4i32>;
2765 def v2i64: ROTQMBYVecInst<v2i64>;
2767 def r128: ROTQMBYRegInst<GPRC>;
2768 def r64: ROTQMBYRegInst<R64C>;
2771 defm ROTQMBY : RotateQuadBytes;
2773 class ROTQMBYIInst<dag OOL, dag IOL, list<dag> pattern>:
2774 RI7Form<0b10111111100, OOL, IOL, "rotqmbyi\t$rT, $rA, $val",
2775 RotateShift, pattern>;
2777 class ROTQMBYIVecInst<ValueType vectype>:
2778 ROTQMBYIInst<(outs VECREG:$rT), (ins VECREG:$rA, rotNeg7imm:$val),
2779 [/* no pattern */]>;
2781 class ROTQMBYIRegInst<RegisterClass rclass, Operand optype, ValueType inttype,
2783 ROTQMBYIInst<(outs rclass:$rT), (ins rclass:$rA, optype:$val),
2784 [/* no pattern */]>;
2786 // 128-bit zero extension form:
2787 class ROTQMBYIZExtInst<RegisterClass rclass, Operand optype, PatLeaf pred>:
2788 ROTQMBYIInst<(outs GPRC:$rT), (ins rclass:$rA, optype:$val),
2789 [/* no pattern */]>;
2791 multiclass RotateQuadBytesImm
2793 def v16i8: ROTQMBYIVecInst<v16i8>;
2794 def v8i16: ROTQMBYIVecInst<v8i16>;
2795 def v4i32: ROTQMBYIVecInst<v4i32>;
2796 def v2i64: ROTQMBYIVecInst<v2i64>;
2798 def r128: ROTQMBYIRegInst<GPRC, rotNeg7imm, i32, uimm7>;
2799 def r64: ROTQMBYIRegInst<R64C, rotNeg7imm, i32, uimm7>;
2801 def r128_zext_r8: ROTQMBYIZExtInst<R8C, rotNeg7imm, uimm7>;
2802 def r128_zext_r16: ROTQMBYIZExtInst<R16C, rotNeg7imm, uimm7>;
2803 def r128_zext_r32: ROTQMBYIZExtInst<R32C, rotNeg7imm, uimm7>;
2804 def r128_zext_r64: ROTQMBYIZExtInst<R64C, rotNeg7imm, uimm7>;
2807 defm ROTQMBYI : RotateQuadBytesImm;
2809 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2810 // Rotate right and mask by bit count
2811 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2813 class ROTQMBYBIInst<dag OOL, dag IOL, list<dag> pattern>:
2814 RRForm<0b10110011100, OOL, IOL, "rotqmbybi\t$rT, $rA, $rB",
2815 RotateShift, pattern>;
2817 class ROTQMBYBIVecInst<ValueType vectype>:
2818 ROTQMBYBIInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2819 [/* no pattern, */]>;
2821 multiclass RotateMaskQuadByBitCount
2823 def v16i8: ROTQMBYBIVecInst<v16i8>;
2824 def v8i16: ROTQMBYBIVecInst<v8i16>;
2825 def v4i32: ROTQMBYBIVecInst<v4i32>;
2826 def v2i64: ROTQMBYBIVecInst<v2i64>;
2829 defm ROTQMBYBI: RotateMaskQuadByBitCount;
2831 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2832 // Rotate quad and mask by bits
2833 // Note that the rotate amount has to be negated
2834 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2836 class ROTQMBIInst<dag OOL, dag IOL, list<dag> pattern>:
2837 RRForm<0b10011011100, OOL, IOL, "rotqmbi\t$rT, $rA, $rB",
2838 RotateShift, pattern>;
2840 class ROTQMBIVecInst<ValueType vectype>:
2841 ROTQMBIInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2842 [/* no pattern */]>;
2844 class ROTQMBIRegInst<RegisterClass rclass>:
2845 ROTQMBIInst<(outs rclass:$rT), (ins rclass:$rA, R32C:$rB),
2846 [/* no pattern */]>;
2848 multiclass RotateMaskQuadByBits
2850 def v16i8: ROTQMBIVecInst<v16i8>;
2851 def v8i16: ROTQMBIVecInst<v8i16>;
2852 def v4i32: ROTQMBIVecInst<v4i32>;
2853 def v2i64: ROTQMBIVecInst<v2i64>;
2855 def r128: ROTQMBIRegInst<GPRC>;
2856 def r64: ROTQMBIRegInst<R64C>;
2859 defm ROTQMBI: RotateMaskQuadByBits;
2861 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2862 // Rotate quad and mask by bits, immediate
2863 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2865 class ROTQMBIIInst<dag OOL, dag IOL, list<dag> pattern>:
2866 RI7Form<0b10011111100, OOL, IOL, "rotqmbii\t$rT, $rA, $val",
2867 RotateShift, pattern>;
2869 class ROTQMBIIVecInst<ValueType vectype>:
2870 ROTQMBIIInst<(outs VECREG:$rT), (ins VECREG:$rA, rotNeg7imm:$val),
2871 [/* no pattern */]>;
2873 class ROTQMBIIRegInst<RegisterClass rclass>:
2874 ROTQMBIIInst<(outs rclass:$rT), (ins rclass:$rA, rotNeg7imm:$val),
2875 [/* no pattern */]>;
2877 multiclass RotateMaskQuadByBitsImm
2879 def v16i8: ROTQMBIIVecInst<v16i8>;
2880 def v8i16: ROTQMBIIVecInst<v8i16>;
2881 def v4i32: ROTQMBIIVecInst<v4i32>;
2882 def v2i64: ROTQMBIIVecInst<v2i64>;
2884 def r128: ROTQMBIIRegInst<GPRC>;
2885 def r64: ROTQMBIIRegInst<R64C>;
2888 defm ROTQMBII: RotateMaskQuadByBitsImm;
2890 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2891 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2894 RRForm<0b01111010000, (outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2895 "rotmah\t$rT, $rA, $rB", RotateShift,
2896 [/* see patterns below - $rB must be negated */]>;
2898 def : Pat<(SPUvec_sra VECREG:$rA, R32C:$rB),
2899 (ROTMAHv8i16 VECREG:$rA, (SFIr32 R32C:$rB, 0))>;
2901 def : Pat<(SPUvec_sra VECREG:$rA, R16C:$rB),
2902 (ROTMAHv8i16 VECREG:$rA,
2903 (SFIr32 (XSHWr16 R16C:$rB), 0))>;
2905 def : Pat<(SPUvec_sra VECREG:$rA, R8C:$rB),
2906 (ROTMAHv8i16 VECREG:$rA,
2907 (SFIr32 (XSHWr16 (XSBHr8 R8C:$rB)), 0))>;
2910 RRForm<0b01111010000, (outs R16C:$rT), (ins R16C:$rA, R32C:$rB),
2911 "rotmah\t$rT, $rA, $rB", RotateShift,
2912 [/* see patterns below - $rB must be negated */]>;
2914 def : Pat<(sra R16C:$rA, R32C:$rB),
2915 (ROTMAHr16 R16C:$rA, (SFIr32 R32C:$rB, 0))>;
2917 def : Pat<(sra R16C:$rA, R16C:$rB),
2918 (ROTMAHr16 R16C:$rA,
2919 (SFIr32 (XSHWr16 R16C:$rB), 0))>;
2921 def : Pat<(sra R16C:$rA, R8C:$rB),
2922 (ROTMAHr16 R16C:$rA,
2923 (SFIr32 (XSHWr16 (XSBHr8 R8C:$rB)), 0))>;
2926 RRForm<0b01111110000, (outs VECREG:$rT), (ins VECREG:$rA, rothNeg7imm:$val),
2927 "rotmahi\t$rT, $rA, $val", RotateShift,
2928 [(set (v8i16 VECREG:$rT),
2929 (SPUvec_sra (v8i16 VECREG:$rA), (i32 uimm7:$val)))]>;
2931 def : Pat<(SPUvec_sra (v8i16 VECREG:$rA), (i16 uimm7:$val)),
2932 (ROTMAHIv8i16 (v8i16 VECREG:$rA), (i32 uimm7:$val))>;
2934 def : Pat<(SPUvec_sra (v8i16 VECREG:$rA), (i8 uimm7:$val)),
2935 (ROTMAHIv8i16 (v8i16 VECREG:$rA), (i32 uimm7:$val))>;
2938 RRForm<0b01111110000, (outs R16C:$rT), (ins R16C:$rA, rothNeg7imm_i16:$val),
2939 "rotmahi\t$rT, $rA, $val", RotateShift,
2940 [(set R16C:$rT, (sra R16C:$rA, (i16 uimm7:$val)))]>;
2942 def : Pat<(sra R16C:$rA, (i32 imm:$val)),
2943 (ROTMAHIr16 R16C:$rA, uimm7:$val)>;
2945 def : Pat<(sra R16C:$rA, (i8 imm:$val)),
2946 (ROTMAHIr16 R16C:$rA, uimm7:$val)>;
2949 RRForm<0b01011010000, (outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2950 "rotma\t$rT, $rA, $rB", RotateShift,
2951 [/* see patterns below - $rB must be negated */]>;
2953 def : Pat<(SPUvec_sra VECREG:$rA, R32C:$rB),
2954 (ROTMAv4i32 (v4i32 VECREG:$rA), (SFIr32 R32C:$rB, 0))>;
2956 def : Pat<(SPUvec_sra VECREG:$rA, R16C:$rB),
2957 (ROTMAv4i32 (v4i32 VECREG:$rA),
2958 (SFIr32 (XSHWr16 R16C:$rB), 0))>;
2960 def : Pat<(SPUvec_sra VECREG:$rA, R8C:$rB),
2961 (ROTMAv4i32 (v4i32 VECREG:$rA),
2962 (SFIr32 (XSHWr16 (XSBHr8 R8C:$rB)), 0))>;
2965 RRForm<0b01011010000, (outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
2966 "rotma\t$rT, $rA, $rB", RotateShift,
2967 [/* see patterns below - $rB must be negated */]>;
2969 def : Pat<(sra R32C:$rA, R32C:$rB),
2970 (ROTMAr32 R32C:$rA, (SFIr32 R32C:$rB, 0))>;
2972 def : Pat<(sra R32C:$rA, R16C:$rB),
2974 (SFIr32 (XSHWr16 R16C:$rB), 0))>;
2976 def : Pat<(sra R32C:$rA, R8C:$rB),
2978 (SFIr32 (XSHWr16 (XSBHr8 R8C:$rB)), 0))>;
2980 class ROTMAIInst<dag OOL, dag IOL, list<dag> pattern>:
2981 RRForm<0b01011110000, OOL, IOL,
2982 "rotmai\t$rT, $rA, $val",
2983 RotateShift, pattern>;
2985 class ROTMAIVecInst<ValueType vectype, Operand intop, ValueType inttype>:
2986 ROTMAIInst<(outs VECREG:$rT), (ins VECREG:$rA, intop:$val),
2987 [(set (vectype VECREG:$rT),
2988 (SPUvec_sra VECREG:$rA, (inttype uimm7:$val)))]>;
2990 class ROTMAIRegInst<RegisterClass rclass, Operand intop, ValueType inttype>:
2991 ROTMAIInst<(outs rclass:$rT), (ins rclass:$rA, intop:$val),
2992 [(set rclass:$rT, (sra rclass:$rA, (inttype uimm7:$val)))]>;
2994 multiclass RotateMaskAlgebraicImm {
2995 def v2i64_i32 : ROTMAIVecInst<v2i64, rotNeg7imm, i32>;
2996 def v4i32_i32 : ROTMAIVecInst<v4i32, rotNeg7imm, i32>;
2997 def r64_i32 : ROTMAIRegInst<R64C, rotNeg7imm, i32>;
2998 def r32_i32 : ROTMAIRegInst<R32C, rotNeg7imm, i32>;
3001 defm ROTMAI : RotateMaskAlgebraicImm;
3003 //===----------------------------------------------------------------------===//
3004 // Branch and conditionals:
3005 //===----------------------------------------------------------------------===//
3007 let isTerminator = 1, isBarrier = 1 in {
3008 // Halt If Equal (r32 preferred slot only, no vector form)
3010 RRForm_3<0b00011011110, (outs), (ins R32C:$rA, R32C:$rB),
3011 "heq\t$rA, $rB", BranchResolv,
3012 [/* no pattern to match */]>;
3015 RI10Form_2<0b11111110, (outs), (ins R32C:$rA, s10imm:$val),
3016 "heqi\t$rA, $val", BranchResolv,
3017 [/* no pattern to match */]>;
3019 // HGT/HGTI: These instructions use signed arithmetic for the comparison,
3020 // contrasting with HLGT/HLGTI, which use unsigned comparison:
3022 RRForm_3<0b00011010010, (outs), (ins R32C:$rA, R32C:$rB),
3023 "hgt\t$rA, $rB", BranchResolv,
3024 [/* no pattern to match */]>;
3027 RI10Form_2<0b11110010, (outs), (ins R32C:$rA, s10imm:$val),
3028 "hgti\t$rA, $val", BranchResolv,
3029 [/* no pattern to match */]>;
3032 RRForm_3<0b00011011010, (outs), (ins R32C:$rA, R32C:$rB),
3033 "hlgt\t$rA, $rB", BranchResolv,
3034 [/* no pattern to match */]>;
3037 RI10Form_2<0b11111010, (outs), (ins R32C:$rA, s10imm:$val),
3038 "hlgti\t$rA, $val", BranchResolv,
3039 [/* no pattern to match */]>;
3042 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
3043 // Comparison operators for i8, i16 and i32:
3044 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
3046 class CEQBInst<dag OOL, dag IOL, list<dag> pattern> :
3047 RRForm<0b00001011110, OOL, IOL, "ceqb\t$rT, $rA, $rB",
3050 multiclass CmpEqualByte
3053 CEQBInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3054 [(set (v16i8 VECREG:$rT), (seteq (v8i16 VECREG:$rA),
3055 (v8i16 VECREG:$rB)))]>;
3058 CEQBInst<(outs R8C:$rT), (ins R8C:$rA, R8C:$rB),
3059 [(set R8C:$rT, (seteq R8C:$rA, R8C:$rB))]>;
3062 class CEQBIInst<dag OOL, dag IOL, list<dag> pattern> :
3063 RI10Form<0b01111110, OOL, IOL, "ceqbi\t$rT, $rA, $val",
3066 multiclass CmpEqualByteImm
3069 CEQBIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm_i8:$val),
3070 [(set (v16i8 VECREG:$rT), (seteq (v16i8 VECREG:$rA),
3071 v16i8SExt8Imm:$val))]>;
3073 CEQBIInst<(outs R8C:$rT), (ins R8C:$rA, s10imm_i8:$val),
3074 [(set R8C:$rT, (seteq R8C:$rA, immSExt8:$val))]>;
3077 class CEQHInst<dag OOL, dag IOL, list<dag> pattern> :
3078 RRForm<0b00010011110, OOL, IOL, "ceqh\t$rT, $rA, $rB",
3081 multiclass CmpEqualHalfword
3083 def v8i16 : CEQHInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3084 [(set (v8i16 VECREG:$rT), (seteq (v8i16 VECREG:$rA),
3085 (v8i16 VECREG:$rB)))]>;
3087 def r16 : CEQHInst<(outs R16C:$rT), (ins R16C:$rA, R16C:$rB),
3088 [(set R16C:$rT, (seteq R16C:$rA, R16C:$rB))]>;
3091 class CEQHIInst<dag OOL, dag IOL, list<dag> pattern> :
3092 RI10Form<0b10111110, OOL, IOL, "ceqhi\t$rT, $rA, $val",
3095 multiclass CmpEqualHalfwordImm
3097 def v8i16 : CEQHIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
3098 [(set (v8i16 VECREG:$rT),
3099 (seteq (v8i16 VECREG:$rA),
3100 (v8i16 v8i16SExt10Imm:$val)))]>;
3101 def r16 : CEQHIInst<(outs R16C:$rT), (ins R16C:$rA, s10imm:$val),
3102 [(set R16C:$rT, (seteq R16C:$rA, i16ImmSExt10:$val))]>;
3105 class CEQInst<dag OOL, dag IOL, list<dag> pattern> :
3106 RRForm<0b00000011110, OOL, IOL, "ceq\t$rT, $rA, $rB",
3109 multiclass CmpEqualWord
3111 def v4i32 : CEQInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3112 [(set (v4i32 VECREG:$rT),
3113 (seteq (v4i32 VECREG:$rA), (v4i32 VECREG:$rB)))]>;
3115 def r32 : CEQInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
3116 [(set R32C:$rT, (seteq R32C:$rA, R32C:$rB))]>;
3119 class CEQIInst<dag OOL, dag IOL, list<dag> pattern> :
3120 RI10Form<0b00111110, OOL, IOL, "ceqi\t$rT, $rA, $val",
3123 multiclass CmpEqualWordImm
3125 def v4i32 : CEQIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
3126 [(set (v4i32 VECREG:$rT),
3127 (seteq (v4i32 VECREG:$rA),
3128 (v4i32 v4i32SExt16Imm:$val)))]>;
3130 def r32: CEQIInst<(outs R32C:$rT), (ins R32C:$rA, s10imm_i32:$val),
3131 [(set R32C:$rT, (seteq R32C:$rA, i32ImmSExt10:$val))]>;
3134 class CGTBInst<dag OOL, dag IOL, list<dag> pattern> :
3135 RRForm<0b00001010010, OOL, IOL, "cgtb\t$rT, $rA, $rB",
3138 multiclass CmpGtrByte
3141 CGTBInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3142 [(set (v16i8 VECREG:$rT), (setgt (v8i16 VECREG:$rA),
3143 (v8i16 VECREG:$rB)))]>;
3146 CGTBInst<(outs R8C:$rT), (ins R8C:$rA, R8C:$rB),
3147 [(set R8C:$rT, (setgt R8C:$rA, R8C:$rB))]>;
3150 class CGTBIInst<dag OOL, dag IOL, list<dag> pattern> :
3151 RI10Form<0b01110010, OOL, IOL, "cgtbi\t$rT, $rA, $val",
3154 multiclass CmpGtrByteImm
3157 CGTBIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm_i8:$val),
3158 [(set (v16i8 VECREG:$rT), (setgt (v16i8 VECREG:$rA),
3159 v16i8SExt8Imm:$val))]>;
3161 CGTBIInst<(outs R8C:$rT), (ins R8C:$rA, s10imm_i8:$val),
3162 [(set R8C:$rT, (setgt R8C:$rA, immSExt8:$val))]>;
3165 class CGTHInst<dag OOL, dag IOL, list<dag> pattern> :
3166 RRForm<0b00010010010, OOL, IOL, "cgth\t$rT, $rA, $rB",
3169 multiclass CmpGtrHalfword
3171 def v8i16 : CGTHInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3172 [(set (v8i16 VECREG:$rT), (setgt (v8i16 VECREG:$rA),
3173 (v8i16 VECREG:$rB)))]>;
3175 def r16 : CGTHInst<(outs R16C:$rT), (ins R16C:$rA, R16C:$rB),
3176 [(set R16C:$rT, (setgt R16C:$rA, R16C:$rB))]>;
3179 class CGTHIInst<dag OOL, dag IOL, list<dag> pattern> :
3180 RI10Form<0b10110010, OOL, IOL, "cgthi\t$rT, $rA, $val",
3183 multiclass CmpGtrHalfwordImm
3185 def v8i16 : CGTHIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
3186 [(set (v8i16 VECREG:$rT),
3187 (setgt (v8i16 VECREG:$rA),
3188 (v8i16 v8i16SExt10Imm:$val)))]>;
3189 def r16 : CGTHIInst<(outs R16C:$rT), (ins R16C:$rA, s10imm:$val),
3190 [(set R16C:$rT, (setgt R16C:$rA, i16ImmSExt10:$val))]>;
3193 class CGTInst<dag OOL, dag IOL, list<dag> pattern> :
3194 RRForm<0b00000010010, OOL, IOL, "cgt\t$rT, $rA, $rB",
3197 multiclass CmpGtrWord
3199 def v4i32 : CGTInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3200 [(set (v4i32 VECREG:$rT),
3201 (setgt (v4i32 VECREG:$rA), (v4i32 VECREG:$rB)))]>;
3203 def r32 : CGTInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
3204 [(set R32C:$rT, (setgt R32C:$rA, R32C:$rB))]>;
3207 class CGTIInst<dag OOL, dag IOL, list<dag> pattern> :
3208 RI10Form<0b00110010, OOL, IOL, "cgti\t$rT, $rA, $val",
3211 multiclass CmpGtrWordImm
3213 def v4i32 : CGTIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
3214 [(set (v4i32 VECREG:$rT),
3215 (setgt (v4i32 VECREG:$rA),
3216 (v4i32 v4i32SExt16Imm:$val)))]>;
3218 def r32: CGTIInst<(outs R32C:$rT), (ins R32C:$rA, s10imm_i32:$val),
3219 [(set R32C:$rT, (setgt R32C:$rA, i32ImmSExt10:$val))]>;
3221 // CGTIv4f32, CGTIf32: These are used in the f32 fdiv instruction sequence:
3222 def v4f32: CGTIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
3223 [(set (v4i32 VECREG:$rT),
3224 (setgt (v4i32 (bitconvert (v4f32 VECREG:$rA))),
3225 (v4i32 v4i32SExt16Imm:$val)))]>;
3227 def f32: CGTIInst<(outs R32C:$rT), (ins R32FP:$rA, s10imm_i32:$val),
3228 [/* no pattern */]>;
3231 class CLGTBInst<dag OOL, dag IOL, list<dag> pattern> :
3232 RRForm<0b00001011010, OOL, IOL, "clgtb\t$rT, $rA, $rB",
3235 multiclass CmpLGtrByte
3238 CLGTBInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3239 [(set (v16i8 VECREG:$rT), (setugt (v8i16 VECREG:$rA),
3240 (v8i16 VECREG:$rB)))]>;
3243 CLGTBInst<(outs R8C:$rT), (ins R8C:$rA, R8C:$rB),
3244 [(set R8C:$rT, (setugt R8C:$rA, R8C:$rB))]>;
3247 class CLGTBIInst<dag OOL, dag IOL, list<dag> pattern> :
3248 RI10Form<0b01111010, OOL, IOL, "clgtbi\t$rT, $rA, $val",
3251 multiclass CmpLGtrByteImm
3254 CLGTBIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm_i8:$val),
3255 [(set (v16i8 VECREG:$rT), (setugt (v16i8 VECREG:$rA),
3256 v16i8SExt8Imm:$val))]>;
3258 CLGTBIInst<(outs R8C:$rT), (ins R8C:$rA, s10imm_i8:$val),
3259 [(set R8C:$rT, (setugt R8C:$rA, immSExt8:$val))]>;
3262 class CLGTHInst<dag OOL, dag IOL, list<dag> pattern> :
3263 RRForm<0b00010011010, OOL, IOL, "clgth\t$rT, $rA, $rB",
3266 multiclass CmpLGtrHalfword
3268 def v8i16 : CLGTHInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3269 [(set (v8i16 VECREG:$rT), (setugt (v8i16 VECREG:$rA),
3270 (v8i16 VECREG:$rB)))]>;
3272 def r16 : CLGTHInst<(outs R16C:$rT), (ins R16C:$rA, R16C:$rB),
3273 [(set R16C:$rT, (setugt R16C:$rA, R16C:$rB))]>;
3276 class CLGTHIInst<dag OOL, dag IOL, list<dag> pattern> :
3277 RI10Form<0b10111010, OOL, IOL, "clgthi\t$rT, $rA, $val",
3280 multiclass CmpLGtrHalfwordImm
3282 def v8i16 : CLGTHIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
3283 [(set (v8i16 VECREG:$rT),
3284 (setugt (v8i16 VECREG:$rA),
3285 (v8i16 v8i16SExt10Imm:$val)))]>;
3286 def r16 : CLGTHIInst<(outs R16C:$rT), (ins R16C:$rA, s10imm:$val),
3287 [(set R16C:$rT, (setugt R16C:$rA, i16ImmSExt10:$val))]>;
3290 class CLGTInst<dag OOL, dag IOL, list<dag> pattern> :
3291 RRForm<0b00000011010, OOL, IOL, "clgt\t$rT, $rA, $rB",
3294 multiclass CmpLGtrWord
3296 def v4i32 : CLGTInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3297 [(set (v4i32 VECREG:$rT),
3298 (setugt (v4i32 VECREG:$rA), (v4i32 VECREG:$rB)))]>;
3300 def r32 : CLGTInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
3301 [(set R32C:$rT, (setugt R32C:$rA, R32C:$rB))]>;
3304 class CLGTIInst<dag OOL, dag IOL, list<dag> pattern> :
3305 RI10Form<0b00111010, OOL, IOL, "clgti\t$rT, $rA, $val",
3308 multiclass CmpLGtrWordImm
3310 def v4i32 : CLGTIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
3311 [(set (v4i32 VECREG:$rT),
3312 (setugt (v4i32 VECREG:$rA),
3313 (v4i32 v4i32SExt16Imm:$val)))]>;
3315 def r32: CLGTIInst<(outs R32C:$rT), (ins R32C:$rA, s10imm_i32:$val),
3316 [(set R32C:$rT, (setugt R32C:$rA, i32ImmSExt10:$val))]>;
3319 defm CEQB : CmpEqualByte;
3320 defm CEQBI : CmpEqualByteImm;
3321 defm CEQH : CmpEqualHalfword;
3322 defm CEQHI : CmpEqualHalfwordImm;
3323 defm CEQ : CmpEqualWord;
3324 defm CEQI : CmpEqualWordImm;
3325 defm CGTB : CmpGtrByte;
3326 defm CGTBI : CmpGtrByteImm;
3327 defm CGTH : CmpGtrHalfword;
3328 defm CGTHI : CmpGtrHalfwordImm;
3329 defm CGT : CmpGtrWord;
3330 defm CGTI : CmpGtrWordImm;
3331 defm CLGTB : CmpLGtrByte;
3332 defm CLGTBI : CmpLGtrByteImm;
3333 defm CLGTH : CmpLGtrHalfword;
3334 defm CLGTHI : CmpLGtrHalfwordImm;
3335 defm CLGT : CmpLGtrWord;
3336 defm CLGTI : CmpLGtrWordImm;
3338 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
3339 // For SETCC primitives not supported above (setlt, setle, setge, etc.)
3340 // define a pattern to generate the right code, as a binary operator
3341 // (in a manner of speaking.)
3344 // 1. This only matches the setcc set of conditionals. Special pattern
3345 // matching is used for select conditionals.
3347 // 2. The "DAG" versions of these classes is almost exclusively used for
3348 // i64 comparisons. See the tblgen fundamentals documentation for what
3349 // ".ResultInstrs[0]" means; see TargetSelectionDAG.td and the Pattern
3350 // class for where ResultInstrs originates.
3351 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
3353 class SETCCNegCondReg<PatFrag cond, RegisterClass rclass, ValueType inttype,
3354 SPUInstr xorinst, SPUInstr cmpare>:
3355 Pat<(cond rclass:$rA, rclass:$rB),
3356 (xorinst (cmpare rclass:$rA, rclass:$rB), (inttype -1))>;
3358 class SETCCNegCondImm<PatFrag cond, RegisterClass rclass, ValueType inttype,
3359 PatLeaf immpred, SPUInstr xorinst, SPUInstr cmpare>:
3360 Pat<(cond rclass:$rA, (inttype immpred:$imm)),
3361 (xorinst (cmpare rclass:$rA, (inttype immpred:$imm)), (inttype -1))>;
3363 def : SETCCNegCondReg<setne, R8C, i8, XORBIr8, CEQBr8>;
3364 def : SETCCNegCondImm<setne, R8C, i8, immSExt8, XORBIr8, CEQBIr8>;
3366 def : SETCCNegCondReg<setne, R16C, i16, XORHIr16, CEQHr16>;
3367 def : SETCCNegCondImm<setne, R16C, i16, i16ImmSExt10, XORHIr16, CEQHIr16>;
3369 def : SETCCNegCondReg<setne, R32C, i32, XORIr32, CEQr32>;
3370 def : SETCCNegCondImm<setne, R32C, i32, i32ImmSExt10, XORIr32, CEQIr32>;
3372 class SETCCBinOpReg<PatFrag cond, RegisterClass rclass,
3373 SPUInstr binop, SPUInstr cmpOp1, SPUInstr cmpOp2>:
3374 Pat<(cond rclass:$rA, rclass:$rB),
3375 (binop (cmpOp1 rclass:$rA, rclass:$rB),
3376 (cmpOp2 rclass:$rA, rclass:$rB))>;
3378 class SETCCBinOpImm<PatFrag cond, RegisterClass rclass, PatLeaf immpred,
3380 SPUInstr binop, SPUInstr cmpOp1, SPUInstr cmpOp2>:
3381 Pat<(cond rclass:$rA, (immtype immpred:$imm)),
3382 (binop (cmpOp1 rclass:$rA, (immtype immpred:$imm)),
3383 (cmpOp2 rclass:$rA, (immtype immpred:$imm)))>;
3385 def : SETCCBinOpReg<setge, R8C, ORr8, CGTBr8, CEQBr8>;
3386 def : SETCCBinOpImm<setge, R8C, immSExt8, i8, ORr8, CGTBIr8, CEQBIr8>;
3387 def : SETCCBinOpReg<setlt, R8C, NORr8, CGTBr8, CEQBr8>;
3388 def : SETCCBinOpImm<setlt, R8C, immSExt8, i8, NORr8, CGTBIr8, CEQBIr8>;
3389 def : Pat<(setle R8C:$rA, R8C:$rB),
3390 (XORBIr8 (CGTBr8 R8C:$rA, R8C:$rB), 0xff)>;
3391 def : Pat<(setle R8C:$rA, immU8:$imm),
3392 (XORBIr8 (CGTBIr8 R8C:$rA, immU8:$imm), 0xff)>;
3394 def : SETCCBinOpReg<setge, R16C, ORr16, CGTHr16, CEQHr16>;
3395 def : SETCCBinOpImm<setge, R16C, i16ImmSExt10, i16,
3396 ORr16, CGTHIr16, CEQHIr16>;
3397 def : SETCCBinOpReg<setlt, R16C, NORr16, CGTHr16, CEQHr16>;
3398 def : SETCCBinOpImm<setlt, R16C, i16ImmSExt10, i16, NORr16, CGTHIr16, CEQHIr16>;
3399 def : Pat<(setle R16C:$rA, R16C:$rB),
3400 (XORHIr16 (CGTHr16 R16C:$rA, R16C:$rB), 0xffff)>;
3401 def : Pat<(setle R16C:$rA, i16ImmSExt10:$imm),
3402 (XORHIr16 (CGTHIr16 R16C:$rA, i16ImmSExt10:$imm), 0xffff)>;
3404 def : SETCCBinOpReg<setge, R32C, ORr32, CGTr32, CEQr32>;
3405 def : SETCCBinOpImm<setge, R32C, i32ImmSExt10, i32,
3406 ORr32, CGTIr32, CEQIr32>;
3407 def : SETCCBinOpReg<setlt, R32C, NORr32, CGTr32, CEQr32>;
3408 def : SETCCBinOpImm<setlt, R32C, i32ImmSExt10, i32, NORr32, CGTIr32, CEQIr32>;
3409 def : Pat<(setle R32C:$rA, R32C:$rB),
3410 (XORIr32 (CGTr32 R32C:$rA, R32C:$rB), 0xffffffff)>;
3411 def : Pat<(setle R32C:$rA, i32ImmSExt10:$imm),
3412 (XORIr32 (CGTIr32 R32C:$rA, i32ImmSExt10:$imm), 0xffffffff)>;
3414 def : SETCCBinOpReg<setuge, R8C, ORr8, CLGTBr8, CEQBr8>;
3415 def : SETCCBinOpImm<setuge, R8C, immSExt8, i8, ORr8, CLGTBIr8, CEQBIr8>;
3416 def : SETCCBinOpReg<setult, R8C, NORr8, CLGTBr8, CEQBr8>;
3417 def : SETCCBinOpImm<setult, R8C, immSExt8, i8, NORr8, CLGTBIr8, CEQBIr8>;
3418 def : Pat<(setule R8C:$rA, R8C:$rB),
3419 (XORBIr8 (CLGTBr8 R8C:$rA, R8C:$rB), 0xff)>;
3420 def : Pat<(setule R8C:$rA, immU8:$imm),
3421 (XORBIr8 (CLGTBIr8 R8C:$rA, immU8:$imm), 0xff)>;
3423 def : SETCCBinOpReg<setuge, R16C, ORr16, CLGTHr16, CEQHr16>;
3424 def : SETCCBinOpImm<setuge, R16C, i16ImmSExt10, i16,
3425 ORr16, CLGTHIr16, CEQHIr16>;
3426 def : SETCCBinOpReg<setult, R16C, NORr16, CLGTHr16, CEQHr16>;
3427 def : SETCCBinOpImm<setult, R16C, i16ImmSExt10, i16, NORr16,
3428 CLGTHIr16, CEQHIr16>;
3429 def : Pat<(setule R16C:$rA, R16C:$rB),
3430 (XORHIr16 (CLGTHr16 R16C:$rA, R16C:$rB), 0xffff)>;
3431 def : Pat<(setule R16C:$rA, i16ImmSExt10:$imm),
3432 (XORHIr16 (CLGTHIr16 R16C:$rA, i16ImmSExt10:$imm), 0xffff)>;
3434 def : SETCCBinOpReg<setuge, R32C, ORr32, CLGTr32, CEQr32>;
3435 def : SETCCBinOpImm<setuge, R32C, i32ImmSExt10, i32,
3436 ORr32, CLGTIr32, CEQIr32>;
3437 def : SETCCBinOpReg<setult, R32C, NORr32, CLGTr32, CEQr32>;
3438 def : SETCCBinOpImm<setult, R32C, i32ImmSExt10, i32, NORr32, CLGTIr32, CEQIr32>;
3439 def : Pat<(setule R32C:$rA, R32C:$rB),
3440 (XORIr32 (CLGTr32 R32C:$rA, R32C:$rB), 0xffffffff)>;
3441 def : Pat<(setule R32C:$rA, i32ImmSExt10:$imm),
3442 (XORIr32 (CLGTIr32 R32C:$rA, i32ImmSExt10:$imm), 0xffffffff)>;
3444 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
3445 // select conditional patterns:
3446 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
3448 class SELECTNegCondReg<PatFrag cond, RegisterClass rclass, ValueType inttype,
3449 SPUInstr selinstr, SPUInstr cmpare>:
3450 Pat<(select (inttype (cond rclass:$rA, rclass:$rB)),
3451 rclass:$rTrue, rclass:$rFalse),
3452 (selinstr rclass:$rTrue, rclass:$rFalse,
3453 (cmpare rclass:$rA, rclass:$rB))>;
3455 class SELECTNegCondImm<PatFrag cond, RegisterClass rclass, ValueType inttype,
3456 PatLeaf immpred, SPUInstr selinstr, SPUInstr cmpare>:
3457 Pat<(select (inttype (cond rclass:$rA, immpred:$imm)),
3458 rclass:$rTrue, rclass:$rFalse),
3459 (selinstr rclass:$rTrue, rclass:$rFalse,
3460 (cmpare rclass:$rA, immpred:$imm))>;
3462 def : SELECTNegCondReg<setne, R8C, i8, SELBr8, CEQBr8>;
3463 def : SELECTNegCondImm<setne, R8C, i8, immSExt8, SELBr8, CEQBIr8>;
3464 def : SELECTNegCondReg<setle, R8C, i8, SELBr8, CGTBr8>;
3465 def : SELECTNegCondImm<setle, R8C, i8, immSExt8, SELBr8, CGTBr8>;
3466 def : SELECTNegCondReg<setule, R8C, i8, SELBr8, CLGTBr8>;
3467 def : SELECTNegCondImm<setule, R8C, i8, immU8, SELBr8, CLGTBIr8>;
3469 def : SELECTNegCondReg<setne, R16C, i16, SELBr16, CEQHr16>;
3470 def : SELECTNegCondImm<setne, R16C, i16, i16ImmSExt10, SELBr16, CEQHIr16>;
3471 def : SELECTNegCondReg<setle, R16C, i16, SELBr16, CGTHr16>;
3472 def : SELECTNegCondImm<setle, R16C, i16, i16ImmSExt10, SELBr16, CGTHIr16>;
3473 def : SELECTNegCondReg<setule, R16C, i16, SELBr16, CLGTHr16>;
3474 def : SELECTNegCondImm<setule, R16C, i16, i16ImmSExt10, SELBr16, CLGTHIr16>;
3476 def : SELECTNegCondReg<setne, R32C, i32, SELBr32, CEQr32>;
3477 def : SELECTNegCondImm<setne, R32C, i32, i32ImmSExt10, SELBr32, CEQIr32>;
3478 def : SELECTNegCondReg<setle, R32C, i32, SELBr32, CGTr32>;
3479 def : SELECTNegCondImm<setle, R32C, i32, i32ImmSExt10, SELBr32, CGTIr32>;
3480 def : SELECTNegCondReg<setule, R32C, i32, SELBr32, CLGTr32>;
3481 def : SELECTNegCondImm<setule, R32C, i32, i32ImmSExt10, SELBr32, CLGTIr32>;
3483 class SELECTBinOpReg<PatFrag cond, RegisterClass rclass, ValueType inttype,
3484 SPUInstr selinstr, SPUInstr binop, SPUInstr cmpOp1,
3486 Pat<(select (inttype (cond rclass:$rA, rclass:$rB)),
3487 rclass:$rTrue, rclass:$rFalse),
3488 (selinstr rclass:$rFalse, rclass:$rTrue,
3489 (binop (cmpOp1 rclass:$rA, rclass:$rB),
3490 (cmpOp2 rclass:$rA, rclass:$rB)))>;
3492 class SELECTBinOpImm<PatFrag cond, RegisterClass rclass, PatLeaf immpred,
3494 SPUInstr selinstr, SPUInstr binop, SPUInstr cmpOp1,
3496 Pat<(select (inttype (cond rclass:$rA, (inttype immpred:$imm))),
3497 rclass:$rTrue, rclass:$rFalse),
3498 (selinstr rclass:$rFalse, rclass:$rTrue,
3499 (binop (cmpOp1 rclass:$rA, (inttype immpred:$imm)),
3500 (cmpOp2 rclass:$rA, (inttype immpred:$imm))))>;
3502 def : SELECTBinOpReg<setge, R8C, i8, SELBr8, ORr8, CGTBr8, CEQBr8>;
3503 def : SELECTBinOpImm<setge, R8C, immSExt8, i8,
3504 SELBr8, ORr8, CGTBIr8, CEQBIr8>;
3506 def : SELECTBinOpReg<setge, R16C, i16, SELBr16, ORr16, CGTHr16, CEQHr16>;
3507 def : SELECTBinOpImm<setge, R16C, i16ImmSExt10, i16,
3508 SELBr16, ORr16, CGTHIr16, CEQHIr16>;
3510 def : SELECTBinOpReg<setge, R32C, i32, SELBr32, ORr32, CGTr32, CEQr32>;
3511 def : SELECTBinOpImm<setge, R32C, i32ImmSExt10, i32,
3512 SELBr32, ORr32, CGTIr32, CEQIr32>;
3514 def : SELECTBinOpReg<setuge, R8C, i8, SELBr8, ORr8, CLGTBr8, CEQBr8>;
3515 def : SELECTBinOpImm<setuge, R8C, immSExt8, i8,
3516 SELBr8, ORr8, CLGTBIr8, CEQBIr8>;
3518 def : SELECTBinOpReg<setuge, R16C, i16, SELBr16, ORr16, CLGTHr16, CEQHr16>;
3519 def : SELECTBinOpImm<setuge, R16C, i16ImmUns10, i16,
3520 SELBr16, ORr16, CLGTHIr16, CEQHIr16>;
3522 def : SELECTBinOpReg<setuge, R32C, i32, SELBr32, ORr32, CLGTr32, CEQr32>;
3523 def : SELECTBinOpImm<setuge, R32C, i32ImmUns10, i32,
3524 SELBr32, ORr32, CLGTIr32, CEQIr32>;
3526 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
3529 // All calls clobber the non-callee-saved registers:
3530 Defs = [R0, R1, R2, R3, R4, R5, R6, R7, R8, R9,
3531 R10,R11,R12,R13,R14,R15,R16,R17,R18,R19,
3532 R20,R21,R22,R23,R24,R25,R26,R27,R28,R29,
3533 R30,R31,R32,R33,R34,R35,R36,R37,R38,R39,
3534 R40,R41,R42,R43,R44,R45,R46,R47,R48,R49,
3535 R50,R51,R52,R53,R54,R55,R56,R57,R58,R59,
3536 R60,R61,R62,R63,R64,R65,R66,R67,R68,R69,
3537 R70,R71,R72,R73,R74,R75,R76,R77,R78,R79],
3538 // All of these instructions use $lr (aka $0)
3540 // Branch relative and set link: Used if we actually know that the target
3541 // is within [-32768, 32767] bytes of the target
3543 BranchSetLink<0b011001100, (outs), (ins relcalltarget:$func, variable_ops),
3544 "brsl\t$$lr, $func",
3545 [(SPUcall (SPUpcrel tglobaladdr:$func, 0))]>;
3547 // Branch absolute and set link: Used if we actually know that the target
3548 // is an absolute address
3550 BranchSetLink<0b011001100, (outs), (ins calltarget:$func, variable_ops),
3551 "brasl\t$$lr, $func",
3552 [(SPUcall (SPUaform tglobaladdr:$func, 0))]>;
3554 // Branch indirect and set link if external data. These instructions are not
3555 // actually generated, matched by an intrinsic:
3556 def BISLED_00: BISLEDForm<0b11, "bisled\t$$lr, $func", [/* empty pattern */]>;
3557 def BISLED_E0: BISLEDForm<0b10, "bisled\t$$lr, $func", [/* empty pattern */]>;
3558 def BISLED_0D: BISLEDForm<0b01, "bisled\t$$lr, $func", [/* empty pattern */]>;
3559 def BISLED_ED: BISLEDForm<0b00, "bisled\t$$lr, $func", [/* empty pattern */]>;
3561 // Branch indirect and set link. This is the "X-form" address version of a
3564 BIForm<0b10010101100, "bisl\t$$lr, $func", [(SPUcall R32C:$func)]>;
3567 // Support calls to external symbols:
3568 def : Pat<(SPUcall (SPUpcrel texternalsym:$func, 0)),
3569 (BRSL texternalsym:$func)>;
3571 def : Pat<(SPUcall (SPUaform texternalsym:$func, 0)),
3572 (BRASL texternalsym:$func)>;
3574 // Unconditional branches:
3575 let isBranch = 1, isTerminator = 1, hasCtrlDep = 1, isBarrier = 1 in {
3577 UncondBranch<0b001001100, (outs), (ins brtarget:$dest),
3581 // Unconditional, absolute address branch
3583 UncondBranch<0b001100000, (outs), (ins brtarget:$dest),
3585 [/* no pattern */]>;
3589 BIForm<0b00010101100, "bi\t$func", [(brind R32C:$func)]>;
3591 // Conditional branches:
3592 class BRNZInst<dag IOL, list<dag> pattern>:
3593 RI16Form<0b010000100, (outs), IOL, "brnz\t$rCond,$dest",
3594 BranchResolv, pattern>;
3596 class BRNZRegInst<RegisterClass rclass>:
3597 BRNZInst<(ins rclass:$rCond, brtarget:$dest),
3598 [(brcond rclass:$rCond, bb:$dest)]>;
3600 class BRNZVecInst<ValueType vectype>:
3601 BRNZInst<(ins VECREG:$rCond, brtarget:$dest),
3602 [(brcond (vectype VECREG:$rCond), bb:$dest)]>;
3604 multiclass BranchNotZero {
3605 def v4i32 : BRNZVecInst<v4i32>;
3606 def r32 : BRNZRegInst<R32C>;
3609 defm BRNZ : BranchNotZero;
3611 class BRZInst<dag IOL, list<dag> pattern>:
3612 RI16Form<0b000000100, (outs), IOL, "brz\t$rT,$dest",
3613 BranchResolv, pattern>;
3615 class BRZRegInst<RegisterClass rclass>:
3616 BRZInst<(ins rclass:$rT, brtarget:$dest), [/* no pattern */]>;
3618 class BRZVecInst<ValueType vectype>:
3619 BRZInst<(ins VECREG:$rT, brtarget:$dest), [/* no pattern */]>;
3621 multiclass BranchZero {
3622 def v4i32: BRZVecInst<v4i32>;
3623 def r32: BRZRegInst<R32C>;
3626 defm BRZ: BranchZero;
3628 // Note: LLVM doesn't do branch conditional, indirect. Otherwise these would
3631 class BINZInst<dag IOL, list<dag> pattern>:
3632 BICondForm<0b10010100100, (outs), IOL, "binz\t$rA, $dest", pattern>;
3634 class BINZRegInst<RegisterClass rclass>:
3635 BINZInst<(ins rclass:$rA, brtarget:$dest),
3636 [(brcond rclass:$rA, R32C:$dest)]>;
3638 class BINZVecInst<ValueType vectype>:
3639 BINZInst<(ins VECREG:$rA, R32C:$dest),
3640 [(brcond (vectype VECREG:$rA), R32C:$dest)]>;
3642 multiclass BranchNotZeroIndirect {
3643 def v4i32: BINZVecInst<v4i32>;
3644 def r32: BINZRegInst<R32C>;
3647 defm BINZ: BranchNotZeroIndirect;
3649 class BIZInst<dag IOL, list<dag> pattern>:
3650 BICondForm<0b00010100100, (outs), IOL, "biz\t$rA, $func", pattern>;
3652 class BIZRegInst<RegisterClass rclass>:
3653 BIZInst<(ins rclass:$rA, R32C:$func), [/* no pattern */]>;
3655 class BIZVecInst<ValueType vectype>:
3656 BIZInst<(ins VECREG:$rA, R32C:$func), [/* no pattern */]>;
3658 multiclass BranchZeroIndirect {
3659 def v4i32: BIZVecInst<v4i32>;
3660 def r32: BIZRegInst<R32C>;
3663 defm BIZ: BranchZeroIndirect;
3666 class BRHNZInst<dag IOL, list<dag> pattern>:
3667 RI16Form<0b011000100, (outs), IOL, "brhnz\t$rCond,$dest", BranchResolv,
3670 class BRHNZRegInst<RegisterClass rclass>:
3671 BRHNZInst<(ins rclass:$rCond, brtarget:$dest),
3672 [(brcond rclass:$rCond, bb:$dest)]>;
3674 class BRHNZVecInst<ValueType vectype>:
3675 BRHNZInst<(ins VECREG:$rCond, brtarget:$dest), [/* no pattern */]>;
3677 multiclass BranchNotZeroHalfword {
3678 def v8i16: BRHNZVecInst<v8i16>;
3679 def r16: BRHNZRegInst<R16C>;
3682 defm BRHNZ: BranchNotZeroHalfword;
3684 class BRHZInst<dag IOL, list<dag> pattern>:
3685 RI16Form<0b001000100, (outs), IOL, "brhz\t$rT,$dest", BranchResolv,
3688 class BRHZRegInst<RegisterClass rclass>:
3689 BRHZInst<(ins rclass:$rT, brtarget:$dest), [/* no pattern */]>;
3691 class BRHZVecInst<ValueType vectype>:
3692 BRHZInst<(ins VECREG:$rT, brtarget:$dest), [/* no pattern */]>;
3694 multiclass BranchZeroHalfword {
3695 def v8i16: BRHZVecInst<v8i16>;
3696 def r16: BRHZRegInst<R16C>;
3699 defm BRHZ: BranchZeroHalfword;
3702 //===----------------------------------------------------------------------===//
3703 // setcc and brcond patterns:
3704 //===----------------------------------------------------------------------===//
3706 def : Pat<(brcond (i16 (seteq R16C:$rA, 0)), bb:$dest),
3707 (BRHZr16 R16C:$rA, bb:$dest)>;
3708 def : Pat<(brcond (i16 (setne R16C:$rA, 0)), bb:$dest),
3709 (BRHNZr16 R16C:$rA, bb:$dest)>;
3711 def : Pat<(brcond (i32 (seteq R32C:$rA, 0)), bb:$dest),
3712 (BRZr32 R32C:$rA, bb:$dest)>;
3713 def : Pat<(brcond (i32 (setne R32C:$rA, 0)), bb:$dest),
3714 (BRNZr32 R32C:$rA, bb:$dest)>;
3716 multiclass BranchCondEQ<PatFrag cond, SPUInstr brinst16, SPUInstr brinst32>
3718 def r16imm: Pat<(brcond (i16 (cond R16C:$rA, i16ImmSExt10:$val)), bb:$dest),
3719 (brinst16 (CEQHIr16 R16C:$rA, i16ImmSExt10:$val), bb:$dest)>;
3721 def r16 : Pat<(brcond (i16 (cond R16C:$rA, R16C:$rB)), bb:$dest),
3722 (brinst16 (CEQHr16 R16C:$rA, R16:$rB), bb:$dest)>;
3724 def r32imm : Pat<(brcond (i32 (cond R32C:$rA, i32ImmSExt10:$val)), bb:$dest),
3725 (brinst32 (CEQIr32 R32C:$rA, i32ImmSExt10:$val), bb:$dest)>;
3727 def r32 : Pat<(brcond (i32 (cond R32C:$rA, R32C:$rB)), bb:$dest),
3728 (brinst32 (CEQr32 R32C:$rA, R32C:$rB), bb:$dest)>;
3731 defm BRCONDeq : BranchCondEQ<seteq, BRHNZr16, BRNZr32>;
3732 defm BRCONDne : BranchCondEQ<setne, BRHZr16, BRZr32>;
3734 multiclass BranchCondLGT<PatFrag cond, SPUInstr brinst16, SPUInstr brinst32>
3736 def r16imm : Pat<(brcond (i16 (cond R16C:$rA, i16ImmSExt10:$val)), bb:$dest),
3737 (brinst16 (CLGTHIr16 R16C:$rA, i16ImmSExt10:$val), bb:$dest)>;
3739 def r16 : Pat<(brcond (i16 (cond R16C:$rA, R16C:$rB)), bb:$dest),
3740 (brinst16 (CLGTHr16 R16C:$rA, R16:$rB), bb:$dest)>;
3742 def r32imm : Pat<(brcond (i32 (cond R32C:$rA, i32ImmSExt10:$val)), bb:$dest),
3743 (brinst32 (CLGTIr32 R32C:$rA, i32ImmSExt10:$val), bb:$dest)>;
3745 def r32 : Pat<(brcond (i32 (cond R32C:$rA, R32C:$rB)), bb:$dest),
3746 (brinst32 (CLGTr32 R32C:$rA, R32C:$rB), bb:$dest)>;
3749 defm BRCONDugt : BranchCondLGT<setugt, BRHNZr16, BRNZr32>;
3750 defm BRCONDule : BranchCondLGT<setule, BRHZr16, BRZr32>;
3752 multiclass BranchCondLGTEQ<PatFrag cond, SPUInstr orinst16, SPUInstr brinst16,
3753 SPUInstr orinst32, SPUInstr brinst32>
3755 def r16imm: Pat<(brcond (i16 (cond R16C:$rA, i16ImmSExt10:$val)), bb:$dest),
3756 (brinst16 (orinst16 (CLGTHIr16 R16C:$rA, i16ImmSExt10:$val),
3757 (CEQHIr16 R16C:$rA, i16ImmSExt10:$val)),
3760 def r16: Pat<(brcond (i16 (cond R16C:$rA, R16C:$rB)), bb:$dest),
3761 (brinst16 (orinst16 (CLGTHr16 R16C:$rA, R16:$rB),
3762 (CEQHr16 R16C:$rA, R16:$rB)),
3765 def r32imm : Pat<(brcond (i32 (cond R32C:$rA, i32ImmSExt10:$val)), bb:$dest),
3766 (brinst32 (orinst32 (CLGTIr32 R32C:$rA, i32ImmSExt10:$val),
3767 (CEQIr32 R32C:$rA, i32ImmSExt10:$val)),
3770 def r32 : Pat<(brcond (i32 (cond R32C:$rA, R32C:$rB)), bb:$dest),
3771 (brinst32 (orinst32 (CLGTr32 R32C:$rA, R32C:$rB),
3772 (CEQr32 R32C:$rA, R32C:$rB)),
3776 defm BRCONDuge : BranchCondLGTEQ<setuge, ORr16, BRHNZr16, ORr32, BRNZr32>;
3777 defm BRCONDult : BranchCondLGTEQ<setult, ORr16, BRHZr16, ORr32, BRZr32>;
3779 multiclass BranchCondGT<PatFrag cond, SPUInstr brinst16, SPUInstr brinst32>
3781 def r16imm : Pat<(brcond (i16 (cond R16C:$rA, i16ImmSExt10:$val)), bb:$dest),
3782 (brinst16 (CGTHIr16 R16C:$rA, i16ImmSExt10:$val), bb:$dest)>;
3784 def r16 : Pat<(brcond (i16 (cond R16C:$rA, R16C:$rB)), bb:$dest),
3785 (brinst16 (CGTHr16 R16C:$rA, R16:$rB), bb:$dest)>;
3787 def r32imm : Pat<(brcond (i32 (cond R32C:$rA, i32ImmSExt10:$val)), bb:$dest),
3788 (brinst32 (CGTIr32 R32C:$rA, i32ImmSExt10:$val), bb:$dest)>;
3790 def r32 : Pat<(brcond (i32 (cond R32C:$rA, R32C:$rB)), bb:$dest),
3791 (brinst32 (CGTr32 R32C:$rA, R32C:$rB), bb:$dest)>;
3794 defm BRCONDgt : BranchCondGT<setgt, BRHNZr16, BRNZr32>;
3795 defm BRCONDle : BranchCondGT<setle, BRHZr16, BRZr32>;
3797 multiclass BranchCondGTEQ<PatFrag cond, SPUInstr orinst16, SPUInstr brinst16,
3798 SPUInstr orinst32, SPUInstr brinst32>
3800 def r16imm: Pat<(brcond (i16 (cond R16C:$rA, i16ImmSExt10:$val)), bb:$dest),
3801 (brinst16 (orinst16 (CGTHIr16 R16C:$rA, i16ImmSExt10:$val),
3802 (CEQHIr16 R16C:$rA, i16ImmSExt10:$val)),
3805 def r16: Pat<(brcond (i16 (cond R16C:$rA, R16C:$rB)), bb:$dest),
3806 (brinst16 (orinst16 (CGTHr16 R16C:$rA, R16:$rB),
3807 (CEQHr16 R16C:$rA, R16:$rB)),
3810 def r32imm : Pat<(brcond (i32 (cond R32C:$rA, i32ImmSExt10:$val)), bb:$dest),
3811 (brinst32 (orinst32 (CGTIr32 R32C:$rA, i32ImmSExt10:$val),
3812 (CEQIr32 R32C:$rA, i32ImmSExt10:$val)),
3815 def r32 : Pat<(brcond (i32 (cond R32C:$rA, R32C:$rB)), bb:$dest),
3816 (brinst32 (orinst32 (CGTr32 R32C:$rA, R32C:$rB),
3817 (CEQr32 R32C:$rA, R32C:$rB)),
3821 defm BRCONDge : BranchCondGTEQ<setge, ORr16, BRHNZr16, ORr32, BRNZr32>;
3822 defm BRCONDlt : BranchCondGTEQ<setlt, ORr16, BRHZr16, ORr32, BRZr32>;
3824 let isTerminator = 1, isBarrier = 1 in {
3825 let isReturn = 1 in {
3827 RETForm<"bi\t$$lr", [(retflag)]>;
3831 //===----------------------------------------------------------------------===//
3832 // Single precision floating point instructions
3833 //===----------------------------------------------------------------------===//
3835 class FAInst<dag OOL, dag IOL, list<dag> pattern>:
3836 RRForm<0b01011000100, OOL, IOL, "fa\t$rT, $rA, $rB",
3839 class FAVecInst<ValueType vectype>:
3840 FAInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3841 [(set (vectype VECREG:$rT),
3842 (fadd (vectype VECREG:$rA), (vectype VECREG:$rB)))]>;
3846 def v4f32: FAVecInst<v4f32>;
3847 def f32: FAInst<(outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB),
3848 [(set R32FP:$rT, (fadd R32FP:$rA, R32FP:$rB))]>;
3853 class FSInst<dag OOL, dag IOL, list<dag> pattern>:
3854 RRForm<0b01011000100, OOL, IOL, "fs\t$rT, $rA, $rB",
3857 class FSVecInst<ValueType vectype>:
3858 FSInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3859 [(set (vectype VECREG:$rT),
3860 (fsub (vectype VECREG:$rA), (vectype VECREG:$rB)))]>;
3864 def v4f32: FSVecInst<v4f32>;
3865 def f32: FSInst<(outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB),
3866 [(set R32FP:$rT, (fsub R32FP:$rA, R32FP:$rB))]>;
3871 // Floating point reciprocal estimate
3873 class FRESTInst<dag OOL, dag IOL>:
3874 RRForm_1<0b00110111000, OOL, IOL,
3875 "frest\t$rT, $rA", SPrecFP,
3876 [/* no pattern */]>;
3879 FRESTInst<(outs VECREG:$rT), (ins VECREG:$rA)>;
3882 FRESTInst<(outs R32FP:$rT), (ins R32FP:$rA)>;
3884 // Floating point interpolate (used in conjunction with reciprocal estimate)
3886 RRForm<0b00101011110, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3887 "fi\t$rT, $rA, $rB", SPrecFP,
3888 [/* no pattern */]>;
3891 RRForm<0b00101011110, (outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB),
3892 "fi\t$rT, $rA, $rB", SPrecFP,
3893 [/* no pattern */]>;
3895 //--------------------------------------------------------------------------
3896 // Basic single precision floating point comparisons:
3898 // Note: There is no support on SPU for single precision NaN. Consequently,
3899 // ordered and unordered comparisons are the same.
3900 //--------------------------------------------------------------------------
3903 RRForm<0b01000011110, (outs R32C:$rT), (ins R32FP:$rA, R32FP:$rB),
3904 "fceq\t$rT, $rA, $rB", SPrecFP,
3905 [(set R32C:$rT, (setueq R32FP:$rA, R32FP:$rB))]>;
3907 def : Pat<(setoeq R32FP:$rA, R32FP:$rB),
3908 (FCEQf32 R32FP:$rA, R32FP:$rB)>;
3911 RRForm<0b01010011110, (outs R32C:$rT), (ins R32FP:$rA, R32FP:$rB),
3912 "fcmeq\t$rT, $rA, $rB", SPrecFP,
3913 [(set R32C:$rT, (setueq (fabs R32FP:$rA), (fabs R32FP:$rB)))]>;
3915 def : Pat<(setoeq (fabs R32FP:$rA), (fabs R32FP:$rB)),
3916 (FCMEQf32 R32FP:$rA, R32FP:$rB)>;
3919 RRForm<0b01000011010, (outs R32C:$rT), (ins R32FP:$rA, R32FP:$rB),
3920 "fcgt\t$rT, $rA, $rB", SPrecFP,
3921 [(set R32C:$rT, (setugt R32FP:$rA, R32FP:$rB))]>;
3923 def : Pat<(setugt R32FP:$rA, R32FP:$rB),
3924 (FCGTf32 R32FP:$rA, R32FP:$rB)>;
3927 RRForm<0b01010011010, (outs R32C:$rT), (ins R32FP:$rA, R32FP:$rB),
3928 "fcmgt\t$rT, $rA, $rB", SPrecFP,
3929 [(set R32C:$rT, (setugt (fabs R32FP:$rA), (fabs R32FP:$rB)))]>;
3931 def : Pat<(setugt (fabs R32FP:$rA), (fabs R32FP:$rB)),
3932 (FCMGTf32 R32FP:$rA, R32FP:$rB)>;
3934 //--------------------------------------------------------------------------
3935 // Single precision floating point comparisons and SETCC equivalents:
3936 //--------------------------------------------------------------------------
3938 def : SETCCNegCondReg<setune, R32FP, i32, XORIr32, FCEQf32>;
3939 def : SETCCNegCondReg<setone, R32FP, i32, XORIr32, FCEQf32>;
3941 def : SETCCBinOpReg<setuge, R32FP, ORr32, FCGTf32, FCEQf32>;
3942 def : SETCCBinOpReg<setoge, R32FP, ORr32, FCGTf32, FCEQf32>;
3944 def : SETCCBinOpReg<setult, R32FP, NORr32, FCGTf32, FCEQf32>;
3945 def : SETCCBinOpReg<setolt, R32FP, NORr32, FCGTf32, FCEQf32>;
3947 def : Pat<(setule R32FP:$rA, R32FP:$rB),
3948 (XORIr32 (FCGTf32 R32FP:$rA, R32FP:$rB), 0xffffffff)>;
3949 def : Pat<(setole R32FP:$rA, R32FP:$rB),
3950 (XORIr32 (FCGTf32 R32FP:$rA, R32FP:$rB), 0xffffffff)>;
3952 // FP Status and Control Register Write
3953 // Why isn't rT a don't care in the ISA?
3954 // Should we create a special RRForm_3 for this guy and zero out the rT?
3956 RRForm_1<0b01011101110, (outs R32FP:$rT), (ins R32FP:$rA),
3957 "fscrwr\t$rA", SPrecFP,
3958 [/* This instruction requires an intrinsic. Note: rT is unused. */]>;
3960 // FP Status and Control Register Read
3962 RRForm_2<0b01011101110, (outs R32FP:$rT), (ins),
3963 "fscrrd\t$rT", SPrecFP,
3964 [/* This instruction requires an intrinsic */]>;
3966 // llvm instruction space
3967 // How do these map onto cell instructions?
3969 // frest rC rB # c = 1/b (both lines)
3971 // fm rD rA rC # d = a * 1/b
3972 // fnms rB rD rB rA # b = - (d * b - a) --should == 0 in a perfect world
3973 // fma rB rB rC rD # b = b * c + d
3974 // = -(d *b -a) * c + d
3975 // = a * c - c ( a *b *c - a)
3980 // These llvm instructions will actually map to library calls.
3981 // All that's needed, then, is to check that the appropriate library is
3982 // imported and do a brsl to the proper function name.
3983 // frem # fmod(x, y): x - (x/y) * y
3984 // (Note: fmod(double, double), fmodf(float,float)
3988 // Unimplemented SPU instruction space
3989 // floating reciprocal absolute square root estimate (frsqest)
3991 // The following are probably just intrinsics
3992 // status and control register write
3993 // status and control register read
3995 //--------------------------------------
3996 // Floating point multiply instructions
3997 //--------------------------------------
4000 RRForm<0b00100011010, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
4001 "fm\t$rT, $rA, $rB", SPrecFP,
4002 [(set (v4f32 VECREG:$rT), (fmul (v4f32 VECREG:$rA),
4003 (v4f32 VECREG:$rB)))]>;
4006 RRForm<0b01100011010, (outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB),
4007 "fm\t$rT, $rA, $rB", SPrecFP,
4008 [(set R32FP:$rT, (fmul R32FP:$rA, R32FP:$rB))]>;
4010 // Floating point multiply and add
4011 // e.g. d = c + (a * b)
4013 RRRForm<0b0111, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
4014 "fma\t$rT, $rA, $rB, $rC", SPrecFP,
4015 [(set (v4f32 VECREG:$rT),
4016 (fadd (v4f32 VECREG:$rC),
4017 (fmul (v4f32 VECREG:$rA), (v4f32 VECREG:$rB))))]>;
4020 RRRForm<0b0111, (outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB, R32FP:$rC),
4021 "fma\t$rT, $rA, $rB, $rC", SPrecFP,
4022 [(set R32FP:$rT, (fadd R32FP:$rC, (fmul R32FP:$rA, R32FP:$rB)))]>;
4024 // FP multiply and subtract
4025 // Subtracts value in rC from product
4028 RRRForm<0b0111, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
4029 "fms\t$rT, $rA, $rB, $rC", SPrecFP,
4030 [(set (v4f32 VECREG:$rT),
4031 (fsub (fmul (v4f32 VECREG:$rA), (v4f32 VECREG:$rB)),
4032 (v4f32 VECREG:$rC)))]>;
4035 RRRForm<0b0111, (outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB, R32FP:$rC),
4036 "fms\t$rT, $rA, $rB, $rC", SPrecFP,
4038 (fsub (fmul R32FP:$rA, R32FP:$rB), R32FP:$rC))]>;
4040 // Floating Negative Mulitply and Subtract
4041 // Subtracts product from value in rC
4042 // res = fneg(fms a b c)
4045 // NOTE: subtraction order
4049 RRRForm<0b1101, (outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB, R32FP:$rC),
4050 "fnms\t$rT, $rA, $rB, $rC", SPrecFP,
4051 [(set R32FP:$rT, (fsub R32FP:$rC, (fmul R32FP:$rA, R32FP:$rB)))]>;
4054 RRRForm<0b1101, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
4055 "fnms\t$rT, $rA, $rB, $rC", SPrecFP,
4056 [(set (v4f32 VECREG:$rT),
4057 (fsub (v4f32 VECREG:$rC),
4058 (fmul (v4f32 VECREG:$rA),
4059 (v4f32 VECREG:$rB))))]>;
4061 //--------------------------------------
4062 // Floating Point Conversions
4063 // Signed conversions:
4065 CVTIntFPForm<0b0101101110, (outs VECREG:$rT), (ins VECREG:$rA),
4066 "csflt\t$rT, $rA, 0", SPrecFP,
4067 [(set (v4f32 VECREG:$rT), (sint_to_fp (v4i32 VECREG:$rA)))]>;
4069 // Convert signed integer to floating point
4071 CVTIntFPForm<0b0101101110, (outs R32FP:$rT), (ins R32C:$rA),
4072 "csflt\t$rT, $rA, 0", SPrecFP,
4073 [(set R32FP:$rT, (sint_to_fp R32C:$rA))]>;
4075 // Convert unsigned into to float
4077 CVTIntFPForm<0b1101101110, (outs VECREG:$rT), (ins VECREG:$rA),
4078 "cuflt\t$rT, $rA, 0", SPrecFP,
4079 [(set (v4f32 VECREG:$rT), (uint_to_fp (v4i32 VECREG:$rA)))]>;
4082 CVTIntFPForm<0b1101101110, (outs R32FP:$rT), (ins R32C:$rA),
4083 "cuflt\t$rT, $rA, 0", SPrecFP,
4084 [(set R32FP:$rT, (uint_to_fp R32C:$rA))]>;
4086 // Convert float to unsigned int
4087 // Assume that scale = 0
4090 CVTIntFPForm<0b1101101110, (outs VECREG:$rT), (ins VECREG:$rA),
4091 "cfltu\t$rT, $rA, 0", SPrecFP,
4092 [(set (v4i32 VECREG:$rT), (fp_to_uint (v4f32 VECREG:$rA)))]>;
4095 CVTIntFPForm<0b1101101110, (outs R32C:$rT), (ins R32FP:$rA),
4096 "cfltu\t$rT, $rA, 0", SPrecFP,
4097 [(set R32C:$rT, (fp_to_uint R32FP:$rA))]>;
4099 // Convert float to signed int
4100 // Assume that scale = 0
4103 CVTIntFPForm<0b1101101110, (outs VECREG:$rT), (ins VECREG:$rA),
4104 "cflts\t$rT, $rA, 0", SPrecFP,
4105 [(set (v4i32 VECREG:$rT), (fp_to_sint (v4f32 VECREG:$rA)))]>;
4108 CVTIntFPForm<0b1101101110, (outs R32C:$rT), (ins R32FP:$rA),
4109 "cflts\t$rT, $rA, 0", SPrecFP,
4110 [(set R32C:$rT, (fp_to_sint R32FP:$rA))]>;
4112 //===----------------------------------------------------------------------==//
4113 // Single<->Double precision conversions
4114 //===----------------------------------------------------------------------==//
4116 // NOTE: We use "vec" name suffix here to avoid confusion (e.g. input is a
4117 // v4f32, output is v2f64--which goes in the name?)
4119 // Floating point extend single to double
4120 // NOTE: Not sure if passing in v4f32 to FESDvec is correct since it
4121 // operates on two double-word slots (i.e. 1st and 3rd fp numbers
4124 RRForm_1<0b00011101110, (outs VECREG:$rT), (ins VECREG:$rA),
4125 "fesd\t$rT, $rA", SPrecFP,
4126 [(set (v2f64 VECREG:$rT), (fextend (v4f32 VECREG:$rA)))]>;
4129 RRForm_1<0b00011101110, (outs R64FP:$rT), (ins R32FP:$rA),
4130 "fesd\t$rT, $rA", SPrecFP,
4131 [(set R64FP:$rT, (fextend R32FP:$rA))]>;
4133 // Floating point round double to single
4135 // RRForm_1<0b10011101110, (outs VECREG:$rT), (ins VECREG:$rA),
4136 // "frds\t$rT, $rA,", SPrecFP,
4137 // [(set (v4f32 R32FP:$rT), (fround (v2f64 R64FP:$rA)))]>;
4140 RRForm_1<0b10011101110, (outs R32FP:$rT), (ins R64FP:$rA),
4141 "frds\t$rT, $rA", SPrecFP,
4142 [(set R32FP:$rT, (fround R64FP:$rA))]>;
4144 //ToDo include anyextend?
4146 //===----------------------------------------------------------------------==//
4147 // Double precision floating point instructions
4148 //===----------------------------------------------------------------------==//
4150 RRForm<0b00110011010, (outs R64FP:$rT), (ins R64FP:$rA, R64FP:$rB),
4151 "dfa\t$rT, $rA, $rB", DPrecFP,
4152 [(set R64FP:$rT, (fadd R64FP:$rA, R64FP:$rB))]>;
4155 RRForm<0b00110011010, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
4156 "dfa\t$rT, $rA, $rB", DPrecFP,
4157 [(set (v2f64 VECREG:$rT), (fadd (v2f64 VECREG:$rA), (v2f64 VECREG:$rB)))]>;
4160 RRForm<0b10100011010, (outs R64FP:$rT), (ins R64FP:$rA, R64FP:$rB),
4161 "dfs\t$rT, $rA, $rB", DPrecFP,
4162 [(set R64FP:$rT, (fsub R64FP:$rA, R64FP:$rB))]>;
4165 RRForm<0b10100011010, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
4166 "dfs\t$rT, $rA, $rB", DPrecFP,
4167 [(set (v2f64 VECREG:$rT),
4168 (fsub (v2f64 VECREG:$rA), (v2f64 VECREG:$rB)))]>;
4171 RRForm<0b01100011010, (outs R64FP:$rT), (ins R64FP:$rA, R64FP:$rB),
4172 "dfm\t$rT, $rA, $rB", DPrecFP,
4173 [(set R64FP:$rT, (fmul R64FP:$rA, R64FP:$rB))]>;
4176 RRForm<0b00100011010, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
4177 "dfm\t$rT, $rA, $rB", DPrecFP,
4178 [(set (v2f64 VECREG:$rT),
4179 (fmul (v2f64 VECREG:$rA), (v2f64 VECREG:$rB)))]>;
4182 RRForm<0b00111010110, (outs R64FP:$rT),
4183 (ins R64FP:$rA, R64FP:$rB, R64FP:$rC),
4184 "dfma\t$rT, $rA, $rB", DPrecFP,
4185 [(set R64FP:$rT, (fadd R64FP:$rC, (fmul R64FP:$rA, R64FP:$rB)))]>,
4186 RegConstraint<"$rC = $rT">,
4190 RRForm<0b00111010110, (outs VECREG:$rT),
4191 (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
4192 "dfma\t$rT, $rA, $rB", DPrecFP,
4193 [(set (v2f64 VECREG:$rT),
4194 (fadd (v2f64 VECREG:$rC),
4195 (fmul (v2f64 VECREG:$rA), (v2f64 VECREG:$rB))))]>,
4196 RegConstraint<"$rC = $rT">,
4200 RRForm<0b10111010110, (outs R64FP:$rT),
4201 (ins R64FP:$rA, R64FP:$rB, R64FP:$rC),
4202 "dfms\t$rT, $rA, $rB", DPrecFP,
4203 [(set R64FP:$rT, (fsub (fmul R64FP:$rA, R64FP:$rB), R64FP:$rC))]>,
4204 RegConstraint<"$rC = $rT">,
4208 RRForm<0b10111010110, (outs VECREG:$rT),
4209 (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
4210 "dfms\t$rT, $rA, $rB", DPrecFP,
4211 [(set (v2f64 VECREG:$rT),
4212 (fsub (fmul (v2f64 VECREG:$rA), (v2f64 VECREG:$rB)),
4213 (v2f64 VECREG:$rC)))]>;
4215 // FNMS: - (a * b - c)
4216 // - (a * b) + c => c - (a * b)
4218 RRForm<0b01111010110, (outs R64FP:$rT),
4219 (ins R64FP:$rA, R64FP:$rB, R64FP:$rC),
4220 "dfnms\t$rT, $rA, $rB", DPrecFP,
4221 [(set R64FP:$rT, (fsub R64FP:$rC, (fmul R64FP:$rA, R64FP:$rB)))]>,
4222 RegConstraint<"$rC = $rT">,
4225 def : Pat<(fneg (fsub (fmul R64FP:$rA, R64FP:$rB), R64FP:$rC)),
4226 (FNMSf64 R64FP:$rA, R64FP:$rB, R64FP:$rC)>;
4229 RRForm<0b01111010110, (outs VECREG:$rT),
4230 (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
4231 "dfnms\t$rT, $rA, $rB", DPrecFP,
4232 [(set (v2f64 VECREG:$rT),
4233 (fsub (v2f64 VECREG:$rC),
4234 (fmul (v2f64 VECREG:$rA),
4235 (v2f64 VECREG:$rB))))]>,
4236 RegConstraint<"$rC = $rT">,
4239 def : Pat<(fneg (fsub (fmul (v2f64 VECREG:$rA), (v2f64 VECREG:$rB)),
4240 (v2f64 VECREG:$rC))),
4241 (FNMSv2f64 VECREG:$rA, VECREG:$rB, VECREG:$rC)>;
4246 RRForm<0b11111010110, (outs R64FP:$rT),
4247 (ins R64FP:$rA, R64FP:$rB, R64FP:$rC),
4248 "dfnma\t$rT, $rA, $rB", DPrecFP,
4249 [(set R64FP:$rT, (fneg (fadd R64FP:$rC, (fmul R64FP:$rA, R64FP:$rB))))]>,
4250 RegConstraint<"$rC = $rT">,
4254 RRForm<0b11111010110, (outs VECREG:$rT),
4255 (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
4256 "dfnma\t$rT, $rA, $rB", DPrecFP,
4257 [(set (v2f64 VECREG:$rT),
4258 (fneg (fadd (v2f64 VECREG:$rC),
4259 (fmul (v2f64 VECREG:$rA),
4260 (v2f64 VECREG:$rB)))))]>,
4261 RegConstraint<"$rC = $rT">,
4264 //===----------------------------------------------------------------------==//
4265 // Floating point negation and absolute value
4266 //===----------------------------------------------------------------------==//
4268 def : Pat<(fneg (v4f32 VECREG:$rA)),
4269 (XORfnegvec (v4f32 VECREG:$rA),
4270 (v4f32 (ILHUv4i32 0x8000)))>;
4272 def : Pat<(fneg R32FP:$rA),
4273 (XORfneg32 R32FP:$rA, (ILHUr32 0x8000))>;
4275 def : Pat<(fneg (v2f64 VECREG:$rA)),
4276 (XORfnegvec (v2f64 VECREG:$rA),
4277 (v2f64 (ANDBIv16i8 (FSMBIv16i8 0x8080), 0x80)))>;
4279 def : Pat<(fneg R64FP:$rA),
4280 (XORfneg64 R64FP:$rA,
4281 (ANDBIv16i8 (FSMBIv16i8 0x8080), 0x80))>;
4283 // Floating point absolute value
4285 def : Pat<(fabs R32FP:$rA),
4286 (ANDfabs32 R32FP:$rA, (IOHLr32 (ILHUr32 0x7fff), 0xffff))>;
4288 def : Pat<(fabs (v4f32 VECREG:$rA)),
4289 (ANDfabsvec (v4f32 VECREG:$rA),
4290 (v4f32 (ANDBIv16i8 (FSMBIv16i8 0xffff), 0x7f)))>;
4292 def : Pat<(fabs R64FP:$rA),
4293 (ANDfabs64 R64FP:$rA, (ANDBIv16i8 (FSMBIv16i8 0xffff), 0x7f))>;
4295 def : Pat<(fabs (v2f64 VECREG:$rA)),
4296 (ANDfabsvec (v2f64 VECREG:$rA),
4297 (v2f64 (ANDBIv16i8 (FSMBIv16i8 0xffff), 0x7f)))>;
4299 //===----------------------------------------------------------------------===//
4300 // Hint for branch instructions:
4301 //===----------------------------------------------------------------------===//
4303 /* def HBR : SPUInstr<(outs), (ins), "hbr\t" */
4305 //===----------------------------------------------------------------------===//
4306 // Execution, Load NOP (execute NOPs belong in even pipeline, load NOPs belong
4307 // in the odd pipeline)
4308 //===----------------------------------------------------------------------===//
4310 def ENOP : SPUInstr<(outs), (ins), "enop", ExecNOP> {
4313 let Inst{0-10} = 0b10000000010;
4314 let Inst{11-17} = 0;
4315 let Inst{18-24} = 0;
4316 let Inst{25-31} = 0;
4319 def LNOP : SPUInstr<(outs), (ins), "lnop", LoadNOP> {
4322 let Inst{0-10} = 0b10000000000;
4323 let Inst{11-17} = 0;
4324 let Inst{18-24} = 0;
4325 let Inst{25-31} = 0;
4328 //===----------------------------------------------------------------------===//
4329 // Bit conversions (type conversions between vector/packed types)
4330 // NOTE: Promotions are handled using the XS* instructions. Truncation
4332 //===----------------------------------------------------------------------===//
4333 def : Pat<(v16i8 (bitconvert (v8i16 VECREG:$src))), (v16i8 VECREG:$src)>;
4334 def : Pat<(v16i8 (bitconvert (v4i32 VECREG:$src))), (v16i8 VECREG:$src)>;
4335 def : Pat<(v16i8 (bitconvert (v2i64 VECREG:$src))), (v16i8 VECREG:$src)>;
4336 def : Pat<(v16i8 (bitconvert (v4f32 VECREG:$src))), (v16i8 VECREG:$src)>;
4337 def : Pat<(v16i8 (bitconvert (v2f64 VECREG:$src))), (v16i8 VECREG:$src)>;
4339 def : Pat<(v8i16 (bitconvert (v16i8 VECREG:$src))), (v8i16 VECREG:$src)>;
4340 def : Pat<(v8i16 (bitconvert (v4i32 VECREG:$src))), (v8i16 VECREG:$src)>;
4341 def : Pat<(v8i16 (bitconvert (v2i64 VECREG:$src))), (v8i16 VECREG:$src)>;
4342 def : Pat<(v8i16 (bitconvert (v4f32 VECREG:$src))), (v8i16 VECREG:$src)>;
4343 def : Pat<(v8i16 (bitconvert (v2f64 VECREG:$src))), (v8i16 VECREG:$src)>;
4345 def : Pat<(v4i32 (bitconvert (v16i8 VECREG:$src))), (v4i32 VECREG:$src)>;
4346 def : Pat<(v4i32 (bitconvert (v8i16 VECREG:$src))), (v4i32 VECREG:$src)>;
4347 def : Pat<(v4i32 (bitconvert (v2i64 VECREG:$src))), (v4i32 VECREG:$src)>;
4348 def : Pat<(v4i32 (bitconvert (v4f32 VECREG:$src))), (v4i32 VECREG:$src)>;
4349 def : Pat<(v4i32 (bitconvert (v2f64 VECREG:$src))), (v4i32 VECREG:$src)>;
4351 def : Pat<(v2i64 (bitconvert (v16i8 VECREG:$src))), (v2i64 VECREG:$src)>;
4352 def : Pat<(v2i64 (bitconvert (v8i16 VECREG:$src))), (v2i64 VECREG:$src)>;
4353 def : Pat<(v2i64 (bitconvert (v4i32 VECREG:$src))), (v2i64 VECREG:$src)>;
4354 def : Pat<(v2i64 (bitconvert (v4f32 VECREG:$src))), (v2i64 VECREG:$src)>;
4355 def : Pat<(v2i64 (bitconvert (v2f64 VECREG:$src))), (v2i64 VECREG:$src)>;
4357 def : Pat<(v4f32 (bitconvert (v16i8 VECREG:$src))), (v4f32 VECREG:$src)>;
4358 def : Pat<(v4f32 (bitconvert (v8i16 VECREG:$src))), (v4f32 VECREG:$src)>;
4359 def : Pat<(v4f32 (bitconvert (v2i64 VECREG:$src))), (v4f32 VECREG:$src)>;
4360 def : Pat<(v4f32 (bitconvert (v4i32 VECREG:$src))), (v4f32 VECREG:$src)>;
4361 def : Pat<(v4f32 (bitconvert (v2f64 VECREG:$src))), (v4f32 VECREG:$src)>;
4363 def : Pat<(v2f64 (bitconvert (v16i8 VECREG:$src))), (v2f64 VECREG:$src)>;
4364 def : Pat<(v2f64 (bitconvert (v8i16 VECREG:$src))), (v2f64 VECREG:$src)>;
4365 def : Pat<(v2f64 (bitconvert (v4i32 VECREG:$src))), (v2f64 VECREG:$src)>;
4366 def : Pat<(v2f64 (bitconvert (v2i64 VECREG:$src))), (v2f64 VECREG:$src)>;
4367 def : Pat<(v2f64 (bitconvert (v2f64 VECREG:$src))), (v2f64 VECREG:$src)>;
4369 def : Pat<(f32 (bitconvert (i32 R32C:$src))), (f32 R32FP:$src)>;
4370 def : Pat<(f64 (bitconvert (i64 R64C:$src))), (f64 R64FP:$src)>;
4372 //===----------------------------------------------------------------------===//
4373 // Instruction patterns:
4374 //===----------------------------------------------------------------------===//
4376 // General 32-bit constants:
4377 def : Pat<(i32 imm:$imm),
4378 (IOHLr32 (ILHUr32 (HI16 imm:$imm)), (LO16 imm:$imm))>;
4380 // Single precision float constants:
4381 def : Pat<(f32 fpimm:$imm),
4382 (IOHLf32 (ILHUf32 (HI16_f32 fpimm:$imm)), (LO16_f32 fpimm:$imm))>;
4384 // General constant 32-bit vectors
4385 def : Pat<(v4i32 v4i32Imm:$imm),
4386 (IOHLv4i32 (v4i32 (ILHUv4i32 (HI16_vec v4i32Imm:$imm))),
4387 (LO16_vec v4i32Imm:$imm))>;
4390 def : Pat<(i8 imm:$imm),
4393 //===----------------------------------------------------------------------===//
4394 // Call instruction patterns:
4395 //===----------------------------------------------------------------------===//
4400 //===----------------------------------------------------------------------===//
4401 // Zero/Any/Sign extensions
4402 //===----------------------------------------------------------------------===//
4404 // sext 8->32: Sign extend bytes to words
4405 def : Pat<(sext_inreg R32C:$rSrc, i8),
4406 (XSHWr32 (XSBHr32 R32C:$rSrc))>;
4408 def : Pat<(i32 (sext R8C:$rSrc)),
4409 (XSHWr16 (XSBHr8 R8C:$rSrc))>;
4411 // sext 8->64: Sign extend bytes to double word
4412 def : Pat<(sext_inreg R64C:$rSrc, i8),
4413 (XSWDr64_inreg (XSHWr64 (XSBHr64 R64C:$rSrc)))>;
4415 def : Pat<(i64 (sext R8C:$rSrc)),
4416 (XSWDr64 (XSHWr16 (XSBHr8 R8C:$rSrc)))>;
4418 // zext 8->16: Zero extend bytes to halfwords
4419 def : Pat<(i16 (zext R8C:$rSrc)),
4420 (ANDHIi8i16 R8C:$rSrc, 0xff)>;
4422 // zext 8->32: Zero extend bytes to words
4423 def : Pat<(i32 (zext R8C:$rSrc)),
4424 (ANDIi8i32 R8C:$rSrc, 0xff)>;
4426 // zext 8->64: Zero extend bytes to double words
4427 def : Pat<(i64 (zext R8C:$rSrc)),
4428 (ORi64_v2i64 (SELBv4i32 (ROTQMBYv4i32
4429 (ORv4i32_i32 (ANDIi8i32 R8C:$rSrc, 0xff)),
4432 (FSMBIv4i32 0x0f0f)))>;
4434 // anyext 8->16: Extend 8->16 bits, irrespective of sign, preserves high bits
4435 def : Pat<(i16 (anyext R8C:$rSrc)),
4436 (ORHIi8i16 R8C:$rSrc, 0)>;
4438 // anyext 8->32: Extend 8->32 bits, irrespective of sign, preserves high bits
4439 def : Pat<(i32 (anyext R8C:$rSrc)),
4440 (ORIi8i32 R8C:$rSrc, 0)>;
4442 // sext 16->64: Sign extend halfword to double word
4443 def : Pat<(sext_inreg R64C:$rSrc, i16),
4444 (XSWDr64_inreg (XSHWr64 R64C:$rSrc))>;
4446 def : Pat<(sext R16C:$rSrc),
4447 (XSWDr64 (XSHWr16 R16C:$rSrc))>;
4449 // zext 16->32: Zero extend halfwords to words
4450 def : Pat<(i32 (zext R16C:$rSrc)),
4451 (ANDi16i32 R16C:$rSrc, (ILAr32 0xffff))>;
4453 def : Pat<(i32 (zext (and R16C:$rSrc, 0xf))),
4454 (ANDIi16i32 R16C:$rSrc, 0xf)>;
4456 def : Pat<(i32 (zext (and R16C:$rSrc, 0xff))),
4457 (ANDIi16i32 R16C:$rSrc, 0xff)>;
4459 def : Pat<(i32 (zext (and R16C:$rSrc, 0xfff))),
4460 (ANDIi16i32 R16C:$rSrc, 0xfff)>;
4462 // anyext 16->32: Extend 16->32 bits, irrespective of sign
4463 def : Pat<(i32 (anyext R16C:$rSrc)),
4464 (ORIi16i32 R16C:$rSrc, 0)>;
4466 //===----------------------------------------------------------------------===//
4468 // These truncates are for the SPU's supported types (i8, i16, i32). i64 and
4469 // above are custom lowered.
4470 //===----------------------------------------------------------------------===//
4472 def : Pat<(i8 (trunc GPRC:$src)),
4474 (SHUFBgprc GPRC:$src, GPRC:$src,
4475 (IOHLv4i32 (ILHUv4i32 0x0f0f), 0x0f0f)))>;
4477 def : Pat<(i8 (trunc R64C:$src)),
4480 (ORv2i64_i64 R64C:$src),
4481 (ORv2i64_i64 R64C:$src),
4482 (IOHLv4i32 (ILHUv4i32 0x0707), 0x0707)))>;
4484 def : Pat<(i8 (trunc R32C:$src)),
4487 (ORv4i32_i32 R32C:$src),
4488 (ORv4i32_i32 R32C:$src),
4489 (IOHLv4i32 (ILHUv4i32 0x0303), 0x0303)))>;
4491 def : Pat<(i8 (trunc R16C:$src)),
4494 (ORv8i16_i16 R16C:$src),
4495 (ORv8i16_i16 R16C:$src),
4496 (IOHLv4i32 (ILHUv4i32 0x0303), 0x0303)))>;
4498 def : Pat<(i16 (trunc GPRC:$src)),
4500 (SHUFBgprc GPRC:$src, GPRC:$src,
4501 (IOHLv4i32 (ILHUv4i32 0x0e0f), 0x0e0f)))>;
4503 def : Pat<(i16 (trunc R64C:$src)),
4506 (ORv2i64_i64 R64C:$src),
4507 (ORv2i64_i64 R64C:$src),
4508 (IOHLv4i32 (ILHUv4i32 0x0607), 0x0607)))>;
4510 def : Pat<(i16 (trunc R32C:$src)),
4513 (ORv4i32_i32 R32C:$src),
4514 (ORv4i32_i32 R32C:$src),
4515 (IOHLv4i32 (ILHUv4i32 0x0203), 0x0203)))>;
4517 def : Pat<(i32 (trunc GPRC:$src)),
4519 (SHUFBgprc GPRC:$src, GPRC:$src,
4520 (IOHLv4i32 (ILHUv4i32 0x0c0d), 0x0e0f)))>;
4522 def : Pat<(i32 (trunc R64C:$src)),
4525 (ORv2i64_i64 R64C:$src),
4526 (ORv2i64_i64 R64C:$src),
4527 (IOHLv4i32 (ILHUv4i32 0x0405), 0x0607)))>;
4529 //===----------------------------------------------------------------------===//
4530 // Address generation: SPU, like PPC, has to split addresses into high and
4531 // low parts in order to load them into a register.
4532 //===----------------------------------------------------------------------===//
4534 def : Pat<(SPUaform tglobaladdr:$in, 0), (ILAlsa tglobaladdr:$in)>;
4535 def : Pat<(SPUaform texternalsym:$in, 0), (ILAlsa texternalsym:$in)>;
4536 def : Pat<(SPUaform tjumptable:$in, 0), (ILAlsa tjumptable:$in)>;
4537 def : Pat<(SPUaform tconstpool:$in, 0), (ILAlsa tconstpool:$in)>;
4539 def : Pat<(SPUindirect (SPUhi tglobaladdr:$in, 0),
4540 (SPUlo tglobaladdr:$in, 0)),
4541 (IOHLlo (ILHUhi tglobaladdr:$in), tglobaladdr:$in)>;
4543 def : Pat<(SPUindirect (SPUhi texternalsym:$in, 0),
4544 (SPUlo texternalsym:$in, 0)),
4545 (IOHLlo (ILHUhi texternalsym:$in), texternalsym:$in)>;
4547 def : Pat<(SPUindirect (SPUhi tjumptable:$in, 0),
4548 (SPUlo tjumptable:$in, 0)),
4549 (IOHLlo (ILHUhi tjumptable:$in), tjumptable:$in)>;
4551 def : Pat<(SPUindirect (SPUhi tconstpool:$in, 0),
4552 (SPUlo tconstpool:$in, 0)),
4553 (IOHLlo (ILHUhi tconstpool:$in), tconstpool:$in)>;
4555 def : Pat<(add (SPUhi tglobaladdr:$in, 0), (SPUlo tglobaladdr:$in, 0)),
4556 (IOHLlo (ILHUhi tglobaladdr:$in), tglobaladdr:$in)>;
4558 def : Pat<(add (SPUhi texternalsym:$in, 0), (SPUlo texternalsym:$in, 0)),
4559 (IOHLlo (ILHUhi texternalsym:$in), texternalsym:$in)>;
4561 def : Pat<(add (SPUhi tjumptable:$in, 0), (SPUlo tjumptable:$in, 0)),
4562 (IOHLlo (ILHUhi tjumptable:$in), tjumptable:$in)>;
4564 def : Pat<(add (SPUhi tconstpool:$in, 0), (SPUlo tconstpool:$in, 0)),
4565 (IOHLlo (ILHUhi tconstpool:$in), tconstpool:$in)>;
4568 include "CellSDKIntrinsics.td"
4569 // Various math operator instruction sequences
4570 include "SPUMathInstr.td"
4571 // 64-bit "instructions"/support
4572 include "SPU64InstrInfo.td"
4573 // 128-bit "instructions"/support
4574 include "SPU128InstrInfo.td"