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:$rCarry),
709 RegConstraint<"$rCarry = $rT">,
712 class ADDXRegInst<RegisterClass rclass>:
713 ADDXInst<(outs rclass:$rT),
714 (ins rclass:$rA, rclass:$rB, rclass:$rCarry),
716 RegConstraint<"$rCarry = $rT">,
719 multiclass AddExtended {
720 def v2i64 : ADDXVecInst<v2i64>;
721 def v4i32 : ADDXVecInst<v4i32>;
722 def r64 : ADDXRegInst<R64C>;
723 def r32 : ADDXRegInst<R32C>;
726 defm ADDX : AddExtended;
728 // CG: Generate carry for add
729 class CGInst<dag OOL, dag IOL, list<dag> pattern>:
730 RRForm<0b01000011000, OOL, IOL,
734 class CGVecInst<ValueType vectype>:
735 CGInst<(outs VECREG:$rT),
736 (ins VECREG:$rA, VECREG:$rB),
739 class CGRegInst<RegisterClass rclass>:
740 CGInst<(outs rclass:$rT),
741 (ins rclass:$rA, rclass:$rB),
744 multiclass CarryGenerate {
745 def v2i64 : CGVecInst<v2i64>;
746 def v4i32 : CGVecInst<v4i32>;
747 def r64 : CGRegInst<R64C>;
748 def r32 : CGRegInst<R32C>;
751 defm CG : CarryGenerate;
753 // SFX: Subract from, extended. This is used in conjunction with BG to subtract
754 // with carry (borrow, in this case)
755 class SFXInst<dag OOL, dag IOL, list<dag> pattern>:
756 RRForm<0b10000010110, OOL, IOL,
757 "sfx\t$rT, $rA, $rB",
760 class SFXVecInst<ValueType vectype>:
761 SFXInst<(outs VECREG:$rT),
762 (ins VECREG:$rA, VECREG:$rB, VECREG:$rCarry),
764 RegConstraint<"$rCarry = $rT">,
767 class SFXRegInst<RegisterClass rclass>:
768 SFXInst<(outs rclass:$rT),
769 (ins rclass:$rA, rclass:$rB, rclass:$rCarry),
771 RegConstraint<"$rCarry = $rT">,
774 multiclass SubtractExtended {
775 def v2i64 : SFXVecInst<v2i64>;
776 def v4i32 : SFXVecInst<v4i32>;
777 def r64 : SFXRegInst<R64C>;
778 def r32 : SFXRegInst<R32C>;
781 defm SFX : SubtractExtended;
783 // BG: only available in vector form, doesn't match a pattern.
784 class BGInst<dag OOL, dag IOL, list<dag> pattern>:
785 RRForm<0b01000010000, OOL, IOL,
789 class BGVecInst<ValueType vectype>:
790 BGInst<(outs VECREG:$rT),
791 (ins VECREG:$rA, VECREG:$rB),
794 class BGRegInst<RegisterClass rclass>:
795 BGInst<(outs rclass:$rT),
796 (ins rclass:$rA, rclass:$rB),
799 multiclass BorrowGenerate {
800 def v4i32 : BGVecInst<v4i32>;
801 def v2i64 : BGVecInst<v2i64>;
802 def r64 : BGRegInst<R64C>;
803 def r32 : BGRegInst<R32C>;
806 defm BG : BorrowGenerate;
808 // BGX: Borrow generate, extended.
810 RRForm<0b11000010110, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB,
812 "bgx\t$rT, $rA, $rB", IntegerOp,
814 RegConstraint<"$rCarry = $rT">,
817 // Halfword multiply variants:
818 // N.B: These can be used to build up larger quantities (16x16 -> 32)
821 RRForm<0b00100011110, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
822 "mpy\t$rT, $rA, $rB", IntegerMulDiv,
826 RRForm<0b00100011110, (outs R16C:$rT), (ins R16C:$rA, R16C:$rB),
827 "mpy\t$rT, $rA, $rB", IntegerMulDiv,
828 [(set R16C:$rT, (mul R16C:$rA, R16C:$rB))]>;
830 // Unsigned 16-bit multiply:
832 class MPYUInst<dag OOL, dag IOL, list<dag> pattern>:
833 RRForm<0b00110011110, OOL, IOL,
834 "mpyu\t$rT, $rA, $rB", IntegerMulDiv,
838 MPYUInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
842 MPYUInst<(outs R32C:$rT), (ins R16C:$rA, R16C:$rB),
843 [(set R32C:$rT, (mul (zext R16C:$rA), (zext R16C:$rB)))]>;
846 MPYUInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
849 // mpyi: multiply 16 x s10imm -> 32 result.
851 class MPYIInst<dag OOL, dag IOL, list<dag> pattern>:
852 RI10Form<0b00101110, OOL, IOL,
853 "mpyi\t$rT, $rA, $val", IntegerMulDiv,
857 MPYIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
858 [(set (v8i16 VECREG:$rT),
859 (mul (v8i16 VECREG:$rA), v8i16SExt10Imm:$val))]>;
862 MPYIInst<(outs R16C:$rT), (ins R16C:$rA, s10imm:$val),
863 [(set R16C:$rT, (mul R16C:$rA, i16ImmSExt10:$val))]>;
865 // mpyui: same issues as other multiplies, plus, this doesn't match a
866 // pattern... but may be used during target DAG selection or lowering
868 class MPYUIInst<dag OOL, dag IOL, list<dag> pattern>:
869 RI10Form<0b10101110, OOL, IOL,
870 "mpyui\t$rT, $rA, $val", IntegerMulDiv,
874 MPYUIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
878 MPYUIInst<(outs R16C:$rT), (ins R16C:$rA, s10imm:$val),
881 // mpya: 16 x 16 + 16 -> 32 bit result
882 class MPYAInst<dag OOL, dag IOL, list<dag> pattern>:
883 RRRForm<0b0011, OOL, IOL,
884 "mpya\t$rT, $rA, $rB, $rC", IntegerMulDiv,
888 MPYAInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
889 [(set (v4i32 VECREG:$rT),
890 (add (v4i32 (bitconvert (mul (v8i16 VECREG:$rA),
891 (v8i16 VECREG:$rB)))),
892 (v4i32 VECREG:$rC)))]>;
895 MPYAInst<(outs R32C:$rT), (ins R16C:$rA, R16C:$rB, R32C:$rC),
896 [(set R32C:$rT, (add (sext (mul R16C:$rA, R16C:$rB)),
900 MPYAInst<(outs R32C:$rT), (ins R16C:$rA, R16C:$rB, R32C:$rC),
901 [(set R32C:$rT, (add (mul (sext R16C:$rA), (sext R16C:$rB)),
904 def MPYAr32_sextinreg:
905 MPYAInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB, R32C:$rC),
906 [(set R32C:$rT, (add (mul (sext_inreg R32C:$rA, i16),
907 (sext_inreg R32C:$rB, i16)),
910 // mpyh: multiply high, used to synthesize 32-bit multiplies
911 class MPYHInst<dag OOL, dag IOL, list<dag> pattern>:
912 RRForm<0b10100011110, OOL, IOL,
913 "mpyh\t$rT, $rA, $rB", IntegerMulDiv,
917 MPYHInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
921 MPYHInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
924 // mpys: multiply high and shift right (returns the top half of
925 // a 16-bit multiply, sign extended to 32 bits.)
927 class MPYSInst<dag OOL, dag IOL>:
928 RRForm<0b11100011110, OOL, IOL,
929 "mpys\t$rT, $rA, $rB", IntegerMulDiv,
933 MPYSInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB)>;
936 MPYSInst<(outs R32C:$rT), (ins R16C:$rA, R16C:$rB)>;
938 // mpyhh: multiply high-high (returns the 32-bit result from multiplying
939 // the top 16 bits of the $rA, $rB)
941 class MPYHHInst<dag OOL, dag IOL>:
942 RRForm<0b01100011110, OOL, IOL,
943 "mpyhh\t$rT, $rA, $rB", IntegerMulDiv,
947 MPYHHInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB)>;
950 MPYHHInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB)>;
952 // mpyhha: Multiply high-high, add to $rT:
954 class MPYHHAInst<dag OOL, dag IOL>:
955 RRForm<0b01100010110, OOL, IOL,
956 "mpyhha\t$rT, $rA, $rB", IntegerMulDiv,
960 MPYHHAInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB)>;
963 MPYHHAInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB)>;
965 // mpyhhu: Multiply high-high, unsigned, e.g.:
967 // +-------+-------+ +-------+-------+ +---------+
968 // | a0 . a1 | x | b0 . b1 | = | a0 x b0 |
969 // +-------+-------+ +-------+-------+ +---------+
971 // where a0, b0 are the upper 16 bits of the 32-bit word
973 class MPYHHUInst<dag OOL, dag IOL>:
974 RRForm<0b01110011110, OOL, IOL,
975 "mpyhhu\t$rT, $rA, $rB", IntegerMulDiv,
979 MPYHHUInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB)>;
982 MPYHHUInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB)>;
984 // mpyhhau: Multiply high-high, unsigned
986 class MPYHHAUInst<dag OOL, dag IOL>:
987 RRForm<0b01110010110, OOL, IOL,
988 "mpyhhau\t$rT, $rA, $rB", IntegerMulDiv,
992 MPYHHAUInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB)>;
995 MPYHHAUInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB)>;
997 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
998 // clz: Count leading zeroes
999 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
1000 class CLZInst<dag OOL, dag IOL, list<dag> pattern>:
1001 RRForm_1<0b10100101010, OOL, IOL, "clz\t$rT, $rA",
1002 IntegerOp, pattern>;
1004 class CLZRegInst<RegisterClass rclass>:
1005 CLZInst<(outs rclass:$rT), (ins rclass:$rA),
1006 [(set rclass:$rT, (ctlz rclass:$rA))]>;
1008 class CLZVecInst<ValueType vectype>:
1009 CLZInst<(outs VECREG:$rT), (ins VECREG:$rA),
1010 [(set (vectype VECREG:$rT), (ctlz (vectype VECREG:$rA)))]>;
1012 multiclass CountLeadingZeroes {
1013 def v4i32 : CLZVecInst<v4i32>;
1014 def r32 : CLZRegInst<R32C>;
1017 defm CLZ : CountLeadingZeroes;
1019 // cntb: Count ones in bytes (aka "population count")
1021 // NOTE: This instruction is really a vector instruction, but the custom
1022 // lowering code uses it in unorthodox ways to support CTPOP for other
1026 RRForm_1<0b00101101010, (outs VECREG:$rT), (ins VECREG:$rA),
1027 "cntb\t$rT, $rA", IntegerOp,
1028 [(set (v16i8 VECREG:$rT), (SPUcntb (v16i8 VECREG:$rA)))]>;
1031 RRForm_1<0b00101101010, (outs VECREG:$rT), (ins VECREG:$rA),
1032 "cntb\t$rT, $rA", IntegerOp,
1033 [(set (v8i16 VECREG:$rT), (SPUcntb (v8i16 VECREG:$rA)))]>;
1036 RRForm_1<0b00101101010, (outs VECREG:$rT), (ins VECREG:$rA),
1037 "cntb\t$rT, $rA", IntegerOp,
1038 [(set (v4i32 VECREG:$rT), (SPUcntb (v4i32 VECREG:$rA)))]>;
1040 // gbb: Gather the low order bits from each byte in $rA into a single 16-bit
1041 // quantity stored into $rT's slot 0, upper 16 bits are zeroed, as are
1044 // Note: This instruction "pairs" with the fsmb instruction for all of the
1045 // various types defined here.
1047 // Note 2: The "VecInst" and "RegInst" forms refer to the result being either
1048 // a vector or register.
1050 class GBBInst<dag OOL, dag IOL, list<dag> pattern>:
1051 RRForm_1<0b01001101100, OOL, IOL, "gbb\t$rT, $rA", GatherOp, pattern>;
1053 class GBBRegInst<RegisterClass rclass, ValueType vectype>:
1054 GBBInst<(outs rclass:$rT), (ins VECREG:$rA),
1055 [/* no pattern */]>;
1057 class GBBVecInst<ValueType vectype>:
1058 GBBInst<(outs VECREG:$rT), (ins VECREG:$rA),
1059 [/* no pattern */]>;
1061 multiclass GatherBitsFromBytes {
1062 def v16i8_r32: GBBRegInst<R32C, v16i8>;
1063 def v16i8_r16: GBBRegInst<R16C, v16i8>;
1064 def v16i8: GBBVecInst<v16i8>;
1067 defm GBB: GatherBitsFromBytes;
1069 // gbh: Gather all low order bits from each halfword in $rA into a single
1070 // 8-bit quantity stored in $rT's slot 0, with the upper bits of $rT set to 0
1071 // and slots 1-3 also set to 0.
1073 // See notes for GBBInst, above.
1075 class GBHInst<dag OOL, dag IOL, list<dag> pattern>:
1076 RRForm_1<0b10001101100, OOL, IOL, "gbh\t$rT, $rA", GatherOp,
1079 class GBHRegInst<RegisterClass rclass, ValueType vectype>:
1080 GBHInst<(outs rclass:$rT), (ins VECREG:$rA),
1081 [/* no pattern */]>;
1083 class GBHVecInst<ValueType vectype>:
1084 GBHInst<(outs VECREG:$rT), (ins VECREG:$rA),
1085 [/* no pattern */]>;
1087 multiclass GatherBitsHalfword {
1088 def v8i16_r32: GBHRegInst<R32C, v8i16>;
1089 def v8i16_r16: GBHRegInst<R16C, v8i16>;
1090 def v8i16: GBHVecInst<v8i16>;
1093 defm GBH: GatherBitsHalfword;
1095 // gb: Gather all low order bits from each word in $rA into a single
1096 // 4-bit quantity stored in $rT's slot 0, upper bits in $rT set to 0,
1097 // as well as slots 1-3.
1099 // See notes for gbb, above.
1101 class GBInst<dag OOL, dag IOL, list<dag> pattern>:
1102 RRForm_1<0b00001101100, OOL, IOL, "gb\t$rT, $rA", GatherOp,
1105 class GBRegInst<RegisterClass rclass, ValueType vectype>:
1106 GBInst<(outs rclass:$rT), (ins VECREG:$rA),
1107 [/* no pattern */]>;
1109 class GBVecInst<ValueType vectype>:
1110 GBInst<(outs VECREG:$rT), (ins VECREG:$rA),
1111 [/* no pattern */]>;
1113 multiclass GatherBitsWord {
1114 def v4i32_r32: GBRegInst<R32C, v4i32>;
1115 def v4i32_r16: GBRegInst<R16C, v4i32>;
1116 def v4i32: GBVecInst<v4i32>;
1119 defm GB: GatherBitsWord;
1121 // avgb: average bytes
1123 RRForm<0b11001011000, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1124 "avgb\t$rT, $rA, $rB", ByteOp,
1127 // absdb: absolute difference of bytes
1129 RRForm<0b11001010000, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1130 "absdb\t$rT, $rA, $rB", ByteOp,
1133 // sumb: sum bytes into halfwords
1135 RRForm<0b11001010010, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1136 "sumb\t$rT, $rA, $rB", ByteOp,
1139 // Sign extension operations:
1140 class XSBHInst<dag OOL, dag IOL, list<dag> pattern>:
1141 RRForm_1<0b01101101010, OOL, IOL,
1142 "xsbh\t$rDst, $rSrc",
1143 IntegerOp, pattern>;
1145 class XSBHVecInst<ValueType vectype>:
1146 XSBHInst<(outs VECREG:$rDst), (ins VECREG:$rSrc),
1147 [(set (v8i16 VECREG:$rDst), (sext (vectype VECREG:$rSrc)))]>;
1149 class XSBHInRegInst<RegisterClass rclass, list<dag> pattern>:
1150 XSBHInst<(outs rclass:$rDst), (ins rclass:$rSrc),
1153 multiclass ExtendByteHalfword {
1154 def v16i8: XSBHVecInst<v8i16>;
1155 def r8: XSBHInst<(outs R16C:$rDst), (ins R8C:$rSrc),
1156 [(set R16C:$rDst, (sext R8C:$rSrc))]>;
1157 def r16: XSBHInRegInst<R16C,
1158 [(set R16C:$rDst, (sext_inreg R16C:$rSrc, i8))]>;
1160 // 32-bit form for XSBH: used to sign extend 8-bit quantities to 16-bit
1161 // quantities to 32-bit quantities via a 32-bit register (see the sext 8->32
1162 // pattern below). Intentionally doesn't match a pattern because we want the
1163 // sext 8->32 pattern to do the work for us, namely because we need the extra
1165 def r32: XSBHInRegInst<R32C, [/* no pattern */]>;
1167 // Same as the 32-bit version, but for i64
1168 def r64: XSBHInRegInst<R64C, [/* no pattern */]>;
1171 defm XSBH : ExtendByteHalfword;
1173 // Sign extend halfwords to words:
1175 class XSHWInst<dag OOL, dag IOL, list<dag> pattern>:
1176 RRForm_1<0b01101101010, OOL, IOL, "xshw\t$rDest, $rSrc",
1177 IntegerOp, pattern>;
1179 class XSHWVecInst<ValueType in_vectype, ValueType out_vectype>:
1180 XSHWInst<(outs VECREG:$rDest), (ins VECREG:$rSrc),
1181 [(set (out_vectype VECREG:$rDest),
1182 (sext (in_vectype VECREG:$rSrc)))]>;
1184 class XSHWInRegInst<RegisterClass rclass, list<dag> pattern>:
1185 XSHWInst<(outs rclass:$rDest), (ins rclass:$rSrc),
1188 class XSHWRegInst<RegisterClass rclass>:
1189 XSHWInst<(outs rclass:$rDest), (ins R16C:$rSrc),
1190 [(set rclass:$rDest, (sext R16C:$rSrc))]>;
1192 multiclass ExtendHalfwordWord {
1193 def v4i32: XSHWVecInst<v4i32, v8i16>;
1195 def r16: XSHWRegInst<R32C>;
1197 def r32: XSHWInRegInst<R32C,
1198 [(set R32C:$rDest, (sext_inreg R32C:$rSrc, i16))]>;
1199 def r64: XSHWInRegInst<R64C, [/* no pattern */]>;
1202 defm XSHW : ExtendHalfwordWord;
1204 // Sign-extend words to doublewords (32->64 bits)
1206 class XSWDInst<dag OOL, dag IOL, list<dag> pattern>:
1207 RRForm_1<0b01100101010, OOL, IOL, "xswd\t$rDst, $rSrc",
1208 IntegerOp, pattern>;
1210 class XSWDVecInst<ValueType in_vectype, ValueType out_vectype>:
1211 XSWDInst<(outs VECREG:$rDst), (ins VECREG:$rSrc),
1212 [(set (out_vectype VECREG:$rDst),
1213 (sext (out_vectype VECREG:$rSrc)))]>;
1215 class XSWDRegInst<RegisterClass in_rclass, RegisterClass out_rclass>:
1216 XSWDInst<(outs out_rclass:$rDst), (ins in_rclass:$rSrc),
1217 [(set out_rclass:$rDst, (sext in_rclass:$rSrc))]>;
1219 multiclass ExtendWordToDoubleWord {
1220 def v2i64: XSWDVecInst<v4i32, v2i64>;
1221 def r64: XSWDRegInst<R32C, R64C>;
1223 def r64_inreg: XSWDInst<(outs R64C:$rDst), (ins R64C:$rSrc),
1224 [(set R64C:$rDst, (sext_inreg R64C:$rSrc, i32))]>;
1227 defm XSWD : ExtendWordToDoubleWord;
1231 class ANDInst<dag OOL, dag IOL, list<dag> pattern> :
1232 RRForm<0b10000011000, OOL, IOL, "and\t$rT, $rA, $rB",
1233 IntegerOp, pattern>;
1235 class ANDVecInst<ValueType vectype>:
1236 ANDInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1237 [(set (vectype VECREG:$rT), (and (vectype VECREG:$rA),
1238 (vectype VECREG:$rB)))]>;
1240 class ANDRegInst<RegisterClass rclass>:
1241 ANDInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
1242 [(set rclass:$rT, (and rclass:$rA, rclass:$rB))]>;
1244 multiclass BitwiseAnd
1246 def v16i8: ANDVecInst<v16i8>;
1247 def v8i16: ANDVecInst<v8i16>;
1248 def v4i32: ANDVecInst<v4i32>;
1249 def v2i64: ANDVecInst<v2i64>;
1251 def r128: ANDRegInst<GPRC>;
1252 def r64: ANDRegInst<R64C>;
1253 def r32: ANDRegInst<R32C>;
1254 def r16: ANDRegInst<R16C>;
1255 def r8: ANDRegInst<R8C>;
1257 //===---------------------------------------------
1258 // Special instructions to perform the fabs instruction
1259 def fabs32: ANDInst<(outs R32FP:$rT), (ins R32FP:$rA, R32C:$rB),
1260 [/* Intentionally does not match a pattern */]>;
1262 def fabs64: ANDInst<(outs R64FP:$rT), (ins R64FP:$rA, VECREG:$rB),
1263 [/* Intentionally does not match a pattern */]>;
1265 // Could use v4i32, but won't for clarity
1266 def fabsvec: ANDInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1267 [/* Intentionally does not match a pattern */]>;
1269 //===---------------------------------------------
1271 // Hacked form of AND to zero-extend 16-bit quantities to 32-bit
1272 // quantities -- see 16->32 zext pattern.
1274 // This pattern is somewhat artificial, since it might match some
1275 // compiler generated pattern but it is unlikely to do so.
1277 def i16i32: ANDInst<(outs R32C:$rT), (ins R16C:$rA, R32C:$rB),
1278 [(set R32C:$rT, (and (zext R16C:$rA), R32C:$rB))]>;
1281 defm AND : BitwiseAnd;
1283 // N.B.: vnot_conv is one of those special target selection pattern fragments,
1284 // in which we expect there to be a bit_convert on the constant. Bear in mind
1285 // that llvm translates "not <reg>" to "xor <reg>, -1" (or in this case, a
1286 // constant -1 vector.)
1288 class ANDCInst<dag OOL, dag IOL, list<dag> pattern>:
1289 RRForm<0b10000011010, OOL, IOL, "andc\t$rT, $rA, $rB",
1290 IntegerOp, pattern>;
1292 class ANDCVecInst<ValueType vectype>:
1293 ANDCInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1294 [(set (vectype VECREG:$rT), (and (vectype VECREG:$rA),
1295 (vnot (vectype VECREG:$rB))))]>;
1297 class ANDCRegInst<RegisterClass rclass>:
1298 ANDCInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
1299 [(set rclass:$rT, (and rclass:$rA, (not rclass:$rB)))]>;
1301 multiclass AndComplement
1303 def v16i8: ANDCVecInst<v16i8>;
1304 def v8i16: ANDCVecInst<v8i16>;
1305 def v4i32: ANDCVecInst<v4i32>;
1306 def v2i64: ANDCVecInst<v2i64>;
1308 def r128: ANDCRegInst<GPRC>;
1309 def r64: ANDCRegInst<R64C>;
1310 def r32: ANDCRegInst<R32C>;
1311 def r16: ANDCRegInst<R16C>;
1312 def r8: ANDCRegInst<R8C>;
1315 defm ANDC : AndComplement;
1317 class ANDBIInst<dag OOL, dag IOL, list<dag> pattern>:
1318 RI10Form<0b01101000, OOL, IOL, "andbi\t$rT, $rA, $val",
1321 multiclass AndByteImm
1323 def v16i8: ANDBIInst<(outs VECREG:$rT), (ins VECREG:$rA, u10imm:$val),
1324 [(set (v16i8 VECREG:$rT),
1325 (and (v16i8 VECREG:$rA),
1326 (v16i8 v16i8U8Imm:$val)))]>;
1328 def r8: ANDBIInst<(outs R8C:$rT), (ins R8C:$rA, u10imm_i8:$val),
1329 [(set R8C:$rT, (and R8C:$rA, immU8:$val))]>;
1332 defm ANDBI : AndByteImm;
1334 class ANDHIInst<dag OOL, dag IOL, list<dag> pattern> :
1335 RI10Form<0b10101000, OOL, IOL, "andhi\t$rT, $rA, $val",
1338 multiclass AndHalfwordImm
1340 def v8i16: ANDHIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
1341 [(set (v8i16 VECREG:$rT),
1342 (and (v8i16 VECREG:$rA), v8i16SExt10Imm:$val))]>;
1344 def r16: ANDHIInst<(outs R16C:$rT), (ins R16C:$rA, u10imm:$val),
1345 [(set R16C:$rT, (and R16C:$rA, i16ImmUns10:$val))]>;
1347 // Zero-extend i8 to i16:
1348 def i8i16: ANDHIInst<(outs R16C:$rT), (ins R8C:$rA, u10imm:$val),
1349 [(set R16C:$rT, (and (zext R8C:$rA), i16ImmUns10:$val))]>;
1352 defm ANDHI : AndHalfwordImm;
1354 class ANDIInst<dag OOL, dag IOL, list<dag> pattern> :
1355 RI10Form<0b00101000, OOL, IOL, "andi\t$rT, $rA, $val",
1356 IntegerOp, pattern>;
1358 multiclass AndWordImm
1360 def v4i32: ANDIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
1361 [(set (v4i32 VECREG:$rT),
1362 (and (v4i32 VECREG:$rA), v4i32SExt10Imm:$val))]>;
1364 def r32: ANDIInst<(outs R32C:$rT), (ins R32C:$rA, s10imm_i32:$val),
1365 [(set R32C:$rT, (and R32C:$rA, i32ImmSExt10:$val))]>;
1367 // Hacked form of ANDI to zero-extend i8 quantities to i32. See the zext 8->32
1369 def i8i32: ANDIInst<(outs R32C:$rT), (ins R8C:$rA, s10imm_i32:$val),
1371 (and (zext R8C:$rA), i32ImmSExt10:$val))]>;
1373 // Hacked form of ANDI to zero-extend i16 quantities to i32. See the
1374 // zext 16->32 pattern below.
1376 // Note that this pattern is somewhat artificial, since it might match
1377 // something the compiler generates but is unlikely to occur in practice.
1378 def i16i32: ANDIInst<(outs R32C:$rT), (ins R16C:$rA, s10imm_i32:$val),
1380 (and (zext R16C:$rA), i32ImmSExt10:$val))]>;
1383 defm ANDI : AndWordImm;
1385 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
1386 // Bitwise OR group:
1387 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
1389 // Bitwise "or" (N.B.: These are also register-register copy instructions...)
1390 class ORInst<dag OOL, dag IOL, list<dag> pattern>:
1391 RRForm<0b10000010000, OOL, IOL, "or\t$rT, $rA, $rB",
1392 IntegerOp, pattern>;
1394 class ORVecInst<ValueType vectype>:
1395 ORInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1396 [(set (vectype VECREG:$rT), (or (vectype VECREG:$rA),
1397 (vectype VECREG:$rB)))]>;
1399 class ORRegInst<RegisterClass rclass>:
1400 ORInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
1401 [(set rclass:$rT, (or rclass:$rA, rclass:$rB))]>;
1403 // ORCvtForm: OR conversion form
1405 // This is used to "convert" the preferred slot to its vector equivalent, as
1406 // well as convert a vector back to its preferred slot.
1408 // These are effectively no-ops, but need to exist for proper type conversion
1409 // and type coercion.
1411 class ORCvtForm<dag OOL, dag IOL, list<dag> pattern = [/* no pattern */]>
1412 : SPUInstr<OOL, IOL, "or\t$rT, $rA, $rA", IntegerOp> {
1416 let Pattern = pattern;
1418 let Inst{0-10} = 0b10000010000;
1419 let Inst{11-17} = RA;
1420 let Inst{18-24} = RA;
1421 let Inst{25-31} = RT;
1424 class ORPromoteScalar<RegisterClass rclass>:
1425 ORCvtForm<(outs VECREG:$rT), (ins rclass:$rA)>;
1427 class ORExtractElt<RegisterClass rclass>:
1428 ORCvtForm<(outs rclass:$rT), (ins VECREG:$rA)>;
1430 /* class ORCvtRegGPRC<RegisterClass rclass>:
1431 ORCvtForm<(outs GPRC:$rT), (ins rclass:$rA)>; */
1433 /* class ORCvtVecGPRC:
1434 ORCvtForm<(outs GPRC:$rT), (ins VECREG:$rA)>; */
1436 /* class ORCvtGPRCReg<RegisterClass rclass>:
1437 ORCvtForm<(outs rclass:$rT), (ins GPRC:$rA)>; */
1439 class ORCvtFormR32Reg<RegisterClass rclass, list<dag> pattern = [ ]>:
1440 ORCvtForm<(outs rclass:$rT), (ins R32C:$rA), pattern>;
1442 class ORCvtFormRegR32<RegisterClass rclass, list<dag> pattern = [ ]>:
1443 ORCvtForm<(outs R32C:$rT), (ins rclass:$rA), pattern>;
1445 class ORCvtFormR64Reg<RegisterClass rclass, list<dag> pattern = [ ]>:
1446 ORCvtForm<(outs rclass:$rT), (ins R64C:$rA), pattern>;
1448 class ORCvtFormRegR64<RegisterClass rclass, list<dag> pattern = [ ]>:
1449 ORCvtForm<(outs R64C:$rT), (ins rclass:$rA), pattern>;
1451 /* class ORCvtGPRCVec:
1452 ORCvtForm<(outs VECREG:$rT), (ins GPRC:$rA)>; */
1454 multiclass BitwiseOr
1456 def v16i8: ORVecInst<v16i8>;
1457 def v8i16: ORVecInst<v8i16>;
1458 def v4i32: ORVecInst<v4i32>;
1459 def v2i64: ORVecInst<v2i64>;
1461 def v4f32: ORInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1462 [(set (v4f32 VECREG:$rT),
1463 (v4f32 (bitconvert (or (v4i32 VECREG:$rA),
1464 (v4i32 VECREG:$rB)))))]>;
1466 def v2f64: ORInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1467 [(set (v2f64 VECREG:$rT),
1468 (v2f64 (bitconvert (or (v2i64 VECREG:$rA),
1469 (v2i64 VECREG:$rB)))))]>;
1471 def r128: ORRegInst<GPRC>;
1472 def r64: ORRegInst<R64C>;
1473 def r32: ORRegInst<R32C>;
1474 def r16: ORRegInst<R16C>;
1475 def r8: ORRegInst<R8C>;
1477 // OR instructions used to copy f32 and f64 registers.
1478 def f32: ORInst<(outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB),
1479 [/* no pattern */]>;
1481 def f64: ORInst<(outs R64FP:$rT), (ins R64FP:$rA, R64FP:$rB),
1482 [/* no pattern */]>;
1484 // scalar->vector promotion, prefslot2vec:
1485 def v16i8_i8: ORPromoteScalar<R8C>;
1486 def v8i16_i16: ORPromoteScalar<R16C>;
1487 def v4i32_i32: ORPromoteScalar<R32C>;
1488 def v2i64_i64: ORPromoteScalar<R64C>;
1489 def v4f32_f32: ORPromoteScalar<R32FP>;
1490 def v2f64_f64: ORPromoteScalar<R64FP>;
1492 // vector->scalar demotion, vec2prefslot:
1493 def i8_v16i8: ORExtractElt<R8C>;
1494 def i16_v8i16: ORExtractElt<R16C>;
1495 def i32_v4i32: ORExtractElt<R32C>;
1496 def i64_v2i64: ORExtractElt<R64C>;
1497 def f32_v4f32: ORExtractElt<R32FP>;
1498 def f64_v2f64: ORExtractElt<R64FP>;
1501 // Conversion from GPRC to register
1502 def i128_r64: ORCvtRegGPRC<R64C>;
1503 def i128_f64: ORCvtRegGPRC<R64FP>;
1504 def i128_r32: ORCvtRegGPRC<R32C>;
1505 def i128_f32: ORCvtRegGPRC<R32FP>;
1506 def i128_r16: ORCvtRegGPRC<R16C>;
1507 def i128_r8: ORCvtRegGPRC<R8C>;
1509 // Conversion from GPRC to vector
1510 def i128_vec: ORCvtVecGPRC;
1512 // Conversion from register to GPRC
1513 def r64_i128: ORCvtGPRCReg<R64C>;
1514 def f64_i128: ORCvtGPRCReg<R64FP>;
1515 def r32_i128: ORCvtGPRCReg<R32C>;
1516 def f32_i128: ORCvtGPRCReg<R32FP>;
1517 def r16_i128: ORCvtGPRCReg<R16C>;
1518 def r8_i128: ORCvtGPRCReg<R8C>;
1520 // Conversion from vector to GPRC
1521 def vec_i128: ORCvtGPRCVec;
1524 // Conversion from register to R32C:
1525 def r16_r32: ORCvtFormRegR32<R16C>;
1526 def r8_r32: ORCvtFormRegR32<R8C>;
1528 // Conversion from R32C to register
1529 def r32_r16: ORCvtFormR32Reg<R16C>;
1530 def r32_r8: ORCvtFormR32Reg<R8C>;
1533 // Conversion to register from R64C:
1534 def r32_r64: ORCvtFormR64Reg<R32C>;
1535 // def r16_r64: ORCvtFormR64Reg<R16C>;
1536 // def r8_r64: ORCvtFormR64Reg<R8C>;
1538 // Conversion to R64C from register
1539 def r64_r32: ORCvtFormRegR64<R32C>;
1540 // def r64_r16: ORCvtFormRegR64<R16C>;
1541 // def r64_r8: ORCvtFormRegR64<R8C>;
1543 // bitconvert patterns:
1544 def r32_f32: ORCvtFormR32Reg<R32FP,
1545 [(set R32FP:$rT, (bitconvert R32C:$rA))]>;
1546 def f32_r32: ORCvtFormRegR32<R32FP,
1547 [(set R32C:$rT, (bitconvert R32FP:$rA))]>;
1549 def r64_f64: ORCvtFormR64Reg<R64FP,
1550 [(set R64FP:$rT, (bitconvert R64C:$rA))]>;
1551 def f64_r64: ORCvtFormRegR64<R64FP,
1552 [(set R64C:$rT, (bitconvert R64FP:$rA))]>;
1555 defm OR : BitwiseOr;
1557 // scalar->vector promotion patterns (preferred slot to vector):
1558 def : Pat<(v16i8 (SPUprefslot2vec R8C:$rA)),
1559 (ORv16i8_i8 R8C:$rA)>;
1561 def : Pat<(v8i16 (SPUprefslot2vec R16C:$rA)),
1562 (ORv8i16_i16 R16C:$rA)>;
1564 def : Pat<(v4i32 (SPUprefslot2vec R32C:$rA)),
1565 (ORv4i32_i32 R32C:$rA)>;
1567 def : Pat<(v2i64 (SPUprefslot2vec R64C:$rA)),
1568 (ORv2i64_i64 R64C:$rA)>;
1570 def : Pat<(v4f32 (SPUprefslot2vec R32FP:$rA)),
1571 (ORv4f32_f32 R32FP:$rA)>;
1573 def : Pat<(v2f64 (SPUprefslot2vec R64FP:$rA)),
1574 (ORv2f64_f64 R64FP:$rA)>;
1576 // ORi*_v*: Used to extract vector element 0 (the preferred slot), otherwise
1577 // known as converting the vector back to its preferred slot
1579 def : Pat<(SPUvec2prefslot (v16i8 VECREG:$rA)),
1580 (ORi8_v16i8 VECREG:$rA)>;
1582 def : Pat<(SPUvec2prefslot (v8i16 VECREG:$rA)),
1583 (ORi16_v8i16 VECREG:$rA)>;
1585 def : Pat<(SPUvec2prefslot (v4i32 VECREG:$rA)),
1586 (ORi32_v4i32 VECREG:$rA)>;
1588 def : Pat<(SPUvec2prefslot (v2i64 VECREG:$rA)),
1589 (ORi64_v2i64 VECREG:$rA)>;
1591 def : Pat<(SPUvec2prefslot (v4f32 VECREG:$rA)),
1592 (ORf32_v4f32 VECREG:$rA)>;
1594 def : Pat<(SPUvec2prefslot (v2f64 VECREG:$rA)),
1595 (ORf64_v2f64 VECREG:$rA)>;
1597 // Load Register: This is an assembler alias for a bitwise OR of a register
1598 // against itself. It's here because it brings some clarity to assembly
1601 let hasCtrlDep = 1 in {
1602 class LRInst<dag OOL, dag IOL>
1603 : SPUInstr<OOL, IOL, "lr\t$rT, $rA", IntegerOp> {
1607 let Pattern = [/*no pattern*/];
1609 let Inst{0-10} = 0b10000010000; /* It's an OR operation */
1610 let Inst{11-17} = RA;
1611 let Inst{18-24} = RA;
1612 let Inst{25-31} = RT;
1615 class LRVecInst<ValueType vectype>:
1616 LRInst<(outs VECREG:$rT), (ins VECREG:$rA)>;
1618 class LRRegInst<RegisterClass rclass>:
1619 LRInst<(outs rclass:$rT), (ins rclass:$rA)>;
1621 multiclass LoadRegister {
1622 def v2i64: LRVecInst<v2i64>;
1623 def v2f64: LRVecInst<v2f64>;
1624 def v4i32: LRVecInst<v4i32>;
1625 def v4f32: LRVecInst<v4f32>;
1626 def v8i16: LRVecInst<v8i16>;
1627 def v16i8: LRVecInst<v16i8>;
1629 def r128: LRRegInst<GPRC>;
1630 def r64: LRRegInst<R64C>;
1631 def f64: LRRegInst<R64FP>;
1632 def r32: LRRegInst<R32C>;
1633 def f32: LRRegInst<R32FP>;
1634 def r16: LRRegInst<R16C>;
1635 def r8: LRRegInst<R8C>;
1638 defm LR: LoadRegister;
1641 // ORC: Bitwise "or" with complement (c = a | ~b)
1643 class ORCInst<dag OOL, dag IOL, list<dag> pattern>:
1644 RRForm<0b10010010000, OOL, IOL, "orc\t$rT, $rA, $rB",
1645 IntegerOp, pattern>;
1647 class ORCVecInst<ValueType vectype>:
1648 ORCInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1649 [(set (vectype VECREG:$rT), (or (vectype VECREG:$rA),
1650 (vnot (vectype VECREG:$rB))))]>;
1652 class ORCRegInst<RegisterClass rclass>:
1653 ORCInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
1654 [(set rclass:$rT, (or rclass:$rA, (not rclass:$rB)))]>;
1656 multiclass BitwiseOrComplement
1658 def v16i8: ORCVecInst<v16i8>;
1659 def v8i16: ORCVecInst<v8i16>;
1660 def v4i32: ORCVecInst<v4i32>;
1661 def v2i64: ORCVecInst<v2i64>;
1663 def r64: ORCRegInst<R64C>;
1664 def r32: ORCRegInst<R32C>;
1665 def r16: ORCRegInst<R16C>;
1666 def r8: ORCRegInst<R8C>;
1669 defm ORC : BitwiseOrComplement;
1671 // OR byte immediate
1672 class ORBIInst<dag OOL, dag IOL, list<dag> pattern>:
1673 RI10Form<0b01100000, OOL, IOL, "orbi\t$rT, $rA, $val",
1674 IntegerOp, pattern>;
1676 class ORBIVecInst<ValueType vectype, PatLeaf immpred>:
1677 ORBIInst<(outs VECREG:$rT), (ins VECREG:$rA, u10imm:$val),
1678 [(set (v16i8 VECREG:$rT), (or (vectype VECREG:$rA),
1679 (vectype immpred:$val)))]>;
1681 multiclass BitwiseOrByteImm
1683 def v16i8: ORBIVecInst<v16i8, v16i8U8Imm>;
1685 def r8: ORBIInst<(outs R8C:$rT), (ins R8C:$rA, u10imm_i8:$val),
1686 [(set R8C:$rT, (or R8C:$rA, immU8:$val))]>;
1689 defm ORBI : BitwiseOrByteImm;
1691 // OR halfword immediate
1692 class ORHIInst<dag OOL, dag IOL, list<dag> pattern>:
1693 RI10Form<0b10100000, OOL, IOL, "orhi\t$rT, $rA, $val",
1694 IntegerOp, pattern>;
1696 class ORHIVecInst<ValueType vectype, PatLeaf immpred>:
1697 ORHIInst<(outs VECREG:$rT), (ins VECREG:$rA, u10imm:$val),
1698 [(set (vectype VECREG:$rT), (or (vectype VECREG:$rA),
1701 multiclass BitwiseOrHalfwordImm
1703 def v8i16: ORHIVecInst<v8i16, v8i16Uns10Imm>;
1705 def r16: ORHIInst<(outs R16C:$rT), (ins R16C:$rA, u10imm:$val),
1706 [(set R16C:$rT, (or R16C:$rA, i16ImmUns10:$val))]>;
1708 // Specialized ORHI form used to promote 8-bit registers to 16-bit
1709 def i8i16: ORHIInst<(outs R16C:$rT), (ins R8C:$rA, s10imm:$val),
1710 [(set R16C:$rT, (or (anyext R8C:$rA),
1711 i16ImmSExt10:$val))]>;
1714 defm ORHI : BitwiseOrHalfwordImm;
1716 class ORIInst<dag OOL, dag IOL, list<dag> pattern>:
1717 RI10Form<0b00100000, OOL, IOL, "ori\t$rT, $rA, $val",
1718 IntegerOp, pattern>;
1720 class ORIVecInst<ValueType vectype, PatLeaf immpred>:
1721 ORIInst<(outs VECREG:$rT), (ins VECREG:$rA, u10imm:$val),
1722 [(set (vectype VECREG:$rT), (or (vectype VECREG:$rA),
1725 // Bitwise "or" with immediate
1726 multiclass BitwiseOrImm
1728 def v4i32: ORIVecInst<v4i32, v4i32Uns10Imm>;
1730 def r32: ORIInst<(outs R32C:$rT), (ins R32C:$rA, u10imm_i32:$val),
1731 [(set R32C:$rT, (or R32C:$rA, i32ImmUns10:$val))]>;
1733 // i16i32: hacked version of the ori instruction to extend 16-bit quantities
1734 // to 32-bit quantities. used exclusively to match "anyext" conversions (vide
1735 // infra "anyext 16->32" pattern.)
1736 def i16i32: ORIInst<(outs R32C:$rT), (ins R16C:$rA, s10imm_i32:$val),
1737 [(set R32C:$rT, (or (anyext R16C:$rA),
1738 i32ImmSExt10:$val))]>;
1740 // i8i32: Hacked version of the ORI instruction to extend 16-bit quantities
1741 // to 32-bit quantities. Used exclusively to match "anyext" conversions (vide
1742 // infra "anyext 16->32" pattern.)
1743 def i8i32: ORIInst<(outs R32C:$rT), (ins R8C:$rA, s10imm_i32:$val),
1744 [(set R32C:$rT, (or (anyext R8C:$rA),
1745 i32ImmSExt10:$val))]>;
1748 defm ORI : BitwiseOrImm;
1750 // ORX: "or" across the vector: or's $rA's word slots leaving the result in
1751 // $rT[0], slots 1-3 are zeroed.
1753 // FIXME: Needs to match an intrinsic pattern.
1755 RRForm<0b10010010000, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1756 "orx\t$rT, $rA, $rB", IntegerOp,
1761 class XORInst<dag OOL, dag IOL, list<dag> pattern> :
1762 RRForm<0b10010010000, OOL, IOL, "xor\t$rT, $rA, $rB",
1763 IntegerOp, pattern>;
1765 class XORVecInst<ValueType vectype>:
1766 XORInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1767 [(set (vectype VECREG:$rT), (xor (vectype VECREG:$rA),
1768 (vectype VECREG:$rB)))]>;
1770 class XORRegInst<RegisterClass rclass>:
1771 XORInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
1772 [(set rclass:$rT, (xor rclass:$rA, rclass:$rB))]>;
1774 multiclass BitwiseExclusiveOr
1776 def v16i8: XORVecInst<v16i8>;
1777 def v8i16: XORVecInst<v8i16>;
1778 def v4i32: XORVecInst<v4i32>;
1779 def v2i64: XORVecInst<v2i64>;
1781 def r128: XORRegInst<GPRC>;
1782 def r64: XORRegInst<R64C>;
1783 def r32: XORRegInst<R32C>;
1784 def r16: XORRegInst<R16C>;
1785 def r8: XORRegInst<R8C>;
1787 // Special forms for floating point instructions.
1788 // fneg and fabs require bitwise logical ops to manipulate the sign bit.
1790 def fneg32: XORInst<(outs R32FP:$rT), (ins R32FP:$rA, R32C:$rB),
1791 [/* no pattern */]>;
1793 def fneg64: XORInst<(outs R64FP:$rT), (ins R64FP:$rA, VECREG:$rB),
1794 [/* no pattern */]>;
1796 def fnegvec: XORInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1797 [/* no pattern, see fneg{32,64} */]>;
1800 defm XOR : BitwiseExclusiveOr;
1802 //==----------------------------------------------------------
1804 class XORBIInst<dag OOL, dag IOL, list<dag> pattern>:
1805 RI10Form<0b01100000, OOL, IOL, "xorbi\t$rT, $rA, $val",
1806 IntegerOp, pattern>;
1808 multiclass XorByteImm
1811 XORBIInst<(outs VECREG:$rT), (ins VECREG:$rA, u10imm:$val),
1812 [(set (v16i8 VECREG:$rT), (xor (v16i8 VECREG:$rA), v16i8U8Imm:$val))]>;
1815 XORBIInst<(outs R8C:$rT), (ins R8C:$rA, u10imm_i8:$val),
1816 [(set R8C:$rT, (xor R8C:$rA, immU8:$val))]>;
1819 defm XORBI : XorByteImm;
1822 RI10Form<0b10100000, (outs VECREG:$rT), (ins VECREG:$rA, u10imm:$val),
1823 "xorhi\t$rT, $rA, $val", IntegerOp,
1824 [(set (v8i16 VECREG:$rT), (xor (v8i16 VECREG:$rA),
1825 v8i16SExt10Imm:$val))]>;
1828 RI10Form<0b10100000, (outs R16C:$rT), (ins R16C:$rA, s10imm:$val),
1829 "xorhi\t$rT, $rA, $val", IntegerOp,
1830 [(set R16C:$rT, (xor R16C:$rA, i16ImmSExt10:$val))]>;
1833 RI10Form<0b00100000, (outs VECREG:$rT), (ins VECREG:$rA, s10imm_i32:$val),
1834 "xori\t$rT, $rA, $val", IntegerOp,
1835 [(set (v4i32 VECREG:$rT), (xor (v4i32 VECREG:$rA),
1836 v4i32SExt10Imm:$val))]>;
1839 RI10Form<0b00100000, (outs R32C:$rT), (ins R32C:$rA, s10imm_i32:$val),
1840 "xori\t$rT, $rA, $val", IntegerOp,
1841 [(set R32C:$rT, (xor R32C:$rA, i32ImmSExt10:$val))]>;
1845 RRForm<0b10010010000, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1846 "nand\t$rT, $rA, $rB", IntegerOp,
1847 [(set (v16i8 VECREG:$rT), (vnot (and (v16i8 VECREG:$rA),
1848 (v16i8 VECREG:$rB))))]>;
1851 RRForm<0b10010010000, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1852 "nand\t$rT, $rA, $rB", IntegerOp,
1853 [(set (v8i16 VECREG:$rT), (vnot (and (v8i16 VECREG:$rA),
1854 (v8i16 VECREG:$rB))))]>;
1857 RRForm<0b10010010000, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1858 "nand\t$rT, $rA, $rB", IntegerOp,
1859 [(set (v4i32 VECREG:$rT), (vnot (and (v4i32 VECREG:$rA),
1860 (v4i32 VECREG:$rB))))]>;
1863 RRForm<0b10010010000, (outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
1864 "nand\t$rT, $rA, $rB", IntegerOp,
1865 [(set R32C:$rT, (not (and R32C:$rA, R32C:$rB)))]>;
1868 RRForm<0b10010010000, (outs R16C:$rT), (ins R16C:$rA, R16C:$rB),
1869 "nand\t$rT, $rA, $rB", IntegerOp,
1870 [(set R16C:$rT, (not (and R16C:$rA, R16C:$rB)))]>;
1873 RRForm<0b10010010000, (outs R8C:$rT), (ins R8C:$rA, R8C:$rB),
1874 "nand\t$rT, $rA, $rB", IntegerOp,
1875 [(set R8C:$rT, (not (and R8C:$rA, R8C:$rB)))]>;
1879 RRForm<0b10010010000, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1880 "nor\t$rT, $rA, $rB", IntegerOp,
1881 [(set (v16i8 VECREG:$rT), (vnot (or (v16i8 VECREG:$rA),
1882 (v16i8 VECREG:$rB))))]>;
1885 RRForm<0b10010010000, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1886 "nor\t$rT, $rA, $rB", IntegerOp,
1887 [(set (v8i16 VECREG:$rT), (vnot (or (v8i16 VECREG:$rA),
1888 (v8i16 VECREG:$rB))))]>;
1891 RRForm<0b10010010000, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
1892 "nor\t$rT, $rA, $rB", IntegerOp,
1893 [(set (v4i32 VECREG:$rT), (vnot (or (v4i32 VECREG:$rA),
1894 (v4i32 VECREG:$rB))))]>;
1897 RRForm<0b10010010000, (outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
1898 "nor\t$rT, $rA, $rB", IntegerOp,
1899 [(set R32C:$rT, (not (or R32C:$rA, R32C:$rB)))]>;
1902 RRForm<0b10010010000, (outs R16C:$rT), (ins R16C:$rA, R16C:$rB),
1903 "nor\t$rT, $rA, $rB", IntegerOp,
1904 [(set R16C:$rT, (not (or R16C:$rA, R16C:$rB)))]>;
1907 RRForm<0b10010010000, (outs R8C:$rT), (ins R8C:$rA, R8C:$rB),
1908 "nor\t$rT, $rA, $rB", IntegerOp,
1909 [(set R8C:$rT, (not (or R8C:$rA, R8C:$rB)))]>;
1912 class SELBInst<dag OOL, dag IOL, list<dag> pattern>:
1913 RRRForm<0b1000, OOL, IOL, "selb\t$rT, $rA, $rB, $rC",
1914 IntegerOp, pattern>;
1916 class SELBVecInst<ValueType vectype, PatFrag vnot_frag = vnot>:
1917 SELBInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
1918 [(set (vectype VECREG:$rT),
1919 (or (and (vectype VECREG:$rC), (vectype VECREG:$rB)),
1920 (and (vnot_frag (vectype VECREG:$rC)),
1921 (vectype VECREG:$rA))))]>;
1923 class SELBVecVCondInst<ValueType vectype>:
1924 SELBInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
1925 [(set (vectype VECREG:$rT),
1926 (select (vectype VECREG:$rC),
1927 (vectype VECREG:$rB),
1928 (vectype VECREG:$rA)))]>;
1930 class SELBVecCondInst<ValueType vectype>:
1931 SELBInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, R32C:$rC),
1932 [(set (vectype VECREG:$rT),
1934 (vectype VECREG:$rB),
1935 (vectype VECREG:$rA)))]>;
1937 class SELBRegInst<RegisterClass rclass>:
1938 SELBInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB, rclass:$rC),
1940 (or (and rclass:$rB, rclass:$rC),
1941 (and rclass:$rA, (not rclass:$rC))))]>;
1943 class SELBRegCondInst<RegisterClass rcond, RegisterClass rclass>:
1944 SELBInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB, rcond:$rC),
1946 (select rcond:$rC, rclass:$rB, rclass:$rA))]>;
1948 multiclass SelectBits
1950 def v16i8: SELBVecInst<v16i8>;
1951 def v8i16: SELBVecInst<v8i16>;
1952 def v4i32: SELBVecInst<v4i32>;
1953 def v2i64: SELBVecInst<v2i64, vnot_conv>;
1955 def r128: SELBRegInst<GPRC>;
1956 def r64: SELBRegInst<R64C>;
1957 def r32: SELBRegInst<R32C>;
1958 def r16: SELBRegInst<R16C>;
1959 def r8: SELBRegInst<R8C>;
1961 def v16i8_cond: SELBVecCondInst<v16i8>;
1962 def v8i16_cond: SELBVecCondInst<v8i16>;
1963 def v4i32_cond: SELBVecCondInst<v4i32>;
1964 def v2i64_cond: SELBVecCondInst<v2i64>;
1966 def v16i8_vcond: SELBVecCondInst<v16i8>;
1967 def v8i16_vcond: SELBVecCondInst<v8i16>;
1968 def v4i32_vcond: SELBVecCondInst<v4i32>;
1969 def v2i64_vcond: SELBVecCondInst<v2i64>;
1972 SELBInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
1973 [(set (v4f32 VECREG:$rT),
1974 (select (v4i32 VECREG:$rC),
1976 (v4f32 VECREG:$rA)))]>;
1978 // SELBr64_cond is defined in SPU64InstrInfo.td
1979 def r32_cond: SELBRegCondInst<R32C, R32C>;
1980 def f32_cond: SELBRegCondInst<R32C, R32FP>;
1981 def r16_cond: SELBRegCondInst<R16C, R16C>;
1982 def r8_cond: SELBRegCondInst<R8C, R8C>;
1985 defm SELB : SelectBits;
1987 class SPUselbPatVec<ValueType vectype, SPUInstr inst>:
1988 Pat<(SPUselb (vectype VECREG:$rA), (vectype VECREG:$rB), (vectype VECREG:$rC)),
1989 (inst VECREG:$rA, VECREG:$rB, VECREG:$rC)>;
1991 def : SPUselbPatVec<v16i8, SELBv16i8>;
1992 def : SPUselbPatVec<v8i16, SELBv8i16>;
1993 def : SPUselbPatVec<v4i32, SELBv4i32>;
1994 def : SPUselbPatVec<v2i64, SELBv2i64>;
1996 class SPUselbPatReg<RegisterClass rclass, SPUInstr inst>:
1997 Pat<(SPUselb rclass:$rA, rclass:$rB, rclass:$rC),
1998 (inst rclass:$rA, rclass:$rB, rclass:$rC)>;
2000 def : SPUselbPatReg<R8C, SELBr8>;
2001 def : SPUselbPatReg<R16C, SELBr16>;
2002 def : SPUselbPatReg<R32C, SELBr32>;
2003 def : SPUselbPatReg<R64C, SELBr64>;
2005 // EQV: Equivalence (1 for each same bit, otherwise 0)
2007 // Note: There are a lot of ways to match this bit operator and these patterns
2008 // attempt to be as exhaustive as possible.
2010 class EQVInst<dag OOL, dag IOL, list<dag> pattern>:
2011 RRForm<0b10010010000, OOL, IOL, "eqv\t$rT, $rA, $rB",
2012 IntegerOp, pattern>;
2014 class EQVVecInst<ValueType vectype>:
2015 EQVInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
2016 [(set (vectype VECREG:$rT),
2017 (or (and (vectype VECREG:$rA), (vectype VECREG:$rB)),
2018 (and (vnot (vectype VECREG:$rA)),
2019 (vnot (vectype VECREG:$rB)))))]>;
2021 class EQVRegInst<RegisterClass rclass>:
2022 EQVInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
2023 [(set rclass:$rT, (or (and rclass:$rA, rclass:$rB),
2024 (and (not rclass:$rA), (not rclass:$rB))))]>;
2026 class EQVVecPattern1<ValueType vectype>:
2027 EQVInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
2028 [(set (vectype VECREG:$rT),
2029 (xor (vectype VECREG:$rA), (vnot (vectype VECREG:$rB))))]>;
2031 class EQVRegPattern1<RegisterClass rclass>:
2032 EQVInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
2033 [(set rclass:$rT, (xor rclass:$rA, (not rclass:$rB)))]>;
2035 class EQVVecPattern2<ValueType vectype>:
2036 EQVInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
2037 [(set (vectype VECREG:$rT),
2038 (or (and (vectype VECREG:$rA), (vectype VECREG:$rB)),
2039 (vnot (or (vectype VECREG:$rA), (vectype VECREG:$rB)))))]>;
2041 class EQVRegPattern2<RegisterClass rclass>:
2042 EQVInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
2044 (or (and rclass:$rA, rclass:$rB),
2045 (not (or rclass:$rA, rclass:$rB))))]>;
2047 class EQVVecPattern3<ValueType vectype>:
2048 EQVInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
2049 [(set (vectype VECREG:$rT),
2050 (not (xor (vectype VECREG:$rA), (vectype VECREG:$rB))))]>;
2052 class EQVRegPattern3<RegisterClass rclass>:
2053 EQVInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
2054 [(set rclass:$rT, (not (xor rclass:$rA, rclass:$rB)))]>;
2056 multiclass BitEquivalence
2058 def v16i8: EQVVecInst<v16i8>;
2059 def v8i16: EQVVecInst<v8i16>;
2060 def v4i32: EQVVecInst<v4i32>;
2061 def v2i64: EQVVecInst<v2i64>;
2063 def v16i8_1: EQVVecPattern1<v16i8>;
2064 def v8i16_1: EQVVecPattern1<v8i16>;
2065 def v4i32_1: EQVVecPattern1<v4i32>;
2066 def v2i64_1: EQVVecPattern1<v2i64>;
2068 def v16i8_2: EQVVecPattern2<v16i8>;
2069 def v8i16_2: EQVVecPattern2<v8i16>;
2070 def v4i32_2: EQVVecPattern2<v4i32>;
2071 def v2i64_2: EQVVecPattern2<v2i64>;
2073 def v16i8_3: EQVVecPattern3<v16i8>;
2074 def v8i16_3: EQVVecPattern3<v8i16>;
2075 def v4i32_3: EQVVecPattern3<v4i32>;
2076 def v2i64_3: EQVVecPattern3<v2i64>;
2078 def r128: EQVRegInst<GPRC>;
2079 def r64: EQVRegInst<R64C>;
2080 def r32: EQVRegInst<R32C>;
2081 def r16: EQVRegInst<R16C>;
2082 def r8: EQVRegInst<R8C>;
2084 def r128_1: EQVRegPattern1<GPRC>;
2085 def r64_1: EQVRegPattern1<R64C>;
2086 def r32_1: EQVRegPattern1<R32C>;
2087 def r16_1: EQVRegPattern1<R16C>;
2088 def r8_1: EQVRegPattern1<R8C>;
2090 def r128_2: EQVRegPattern2<GPRC>;
2091 def r64_2: EQVRegPattern2<R64C>;
2092 def r32_2: EQVRegPattern2<R32C>;
2093 def r16_2: EQVRegPattern2<R16C>;
2094 def r8_2: EQVRegPattern2<R8C>;
2096 def r128_3: EQVRegPattern3<GPRC>;
2097 def r64_3: EQVRegPattern3<R64C>;
2098 def r32_3: EQVRegPattern3<R32C>;
2099 def r16_3: EQVRegPattern3<R16C>;
2100 def r8_3: EQVRegPattern3<R8C>;
2103 defm EQV: BitEquivalence;
2105 //===----------------------------------------------------------------------===//
2106 // Vector shuffle...
2107 //===----------------------------------------------------------------------===//
2108 // SPUshuffle is generated in LowerVECTOR_SHUFFLE and gets replaced with SHUFB.
2109 // See the SPUshuffle SDNode operand above, which sets up the DAG pattern
2110 // matcher to emit something when the LowerVECTOR_SHUFFLE generates a node with
2111 // the SPUISD::SHUFB opcode.
2112 //===----------------------------------------------------------------------===//
2114 class SHUFBInst<dag OOL, dag IOL, list<dag> pattern>:
2115 RRRForm<0b1000, OOL, IOL, "shufb\t$rT, $rA, $rB, $rC",
2116 IntegerOp, pattern>;
2118 class SHUFBVecInst<ValueType resultvec, ValueType maskvec>:
2119 SHUFBInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
2120 [(set (resultvec VECREG:$rT),
2121 (SPUshuffle (resultvec VECREG:$rA),
2122 (resultvec VECREG:$rB),
2123 (maskvec VECREG:$rC)))]>;
2125 class SHUFBGPRCInst:
2126 SHUFBInst<(outs VECREG:$rT), (ins GPRC:$rA, GPRC:$rB, VECREG:$rC),
2127 [/* no pattern */]>;
2129 multiclass ShuffleBytes
2131 def v16i8 : SHUFBVecInst<v16i8, v16i8>;
2132 def v16i8_m32 : SHUFBVecInst<v16i8, v4i32>;
2133 def v8i16 : SHUFBVecInst<v8i16, v16i8>;
2134 def v8i16_m32 : SHUFBVecInst<v8i16, v4i32>;
2135 def v4i32 : SHUFBVecInst<v4i32, v16i8>;
2136 def v4i32_m32 : SHUFBVecInst<v4i32, v4i32>;
2137 def v2i64 : SHUFBVecInst<v2i64, v16i8>;
2138 def v2i64_m32 : SHUFBVecInst<v2i64, v4i32>;
2140 def v4f32 : SHUFBVecInst<v4f32, v16i8>;
2141 def v4f32_m32 : SHUFBVecInst<v4f32, v4i32>;
2143 def v2f64 : SHUFBVecInst<v2f64, v16i8>;
2144 def v2f64_m32 : SHUFBVecInst<v2f64, v4i32>;
2146 def gprc : SHUFBGPRCInst;
2149 defm SHUFB : ShuffleBytes;
2151 //===----------------------------------------------------------------------===//
2152 // Shift and rotate group:
2153 //===----------------------------------------------------------------------===//
2155 class SHLHInst<dag OOL, dag IOL, list<dag> pattern>:
2156 RRForm<0b11111010000, OOL, IOL, "shlh\t$rT, $rA, $rB",
2157 RotateShift, pattern>;
2159 class SHLHVecInst<ValueType vectype>:
2160 SHLHInst<(outs VECREG:$rT), (ins VECREG:$rA, R16C:$rB),
2161 [(set (vectype VECREG:$rT),
2162 (SPUvec_shl (vectype VECREG:$rA), R16C:$rB))]>;
2164 multiclass ShiftLeftHalfword
2166 def v8i16: SHLHVecInst<v8i16>;
2167 def r16: SHLHInst<(outs R16C:$rT), (ins R16C:$rA, R16C:$rB),
2168 [(set R16C:$rT, (shl R16C:$rA, R16C:$rB))]>;
2169 def r16_r32: SHLHInst<(outs R16C:$rT), (ins R16C:$rA, R32C:$rB),
2170 [(set R16C:$rT, (shl R16C:$rA, R32C:$rB))]>;
2173 defm SHLH : ShiftLeftHalfword;
2175 //===----------------------------------------------------------------------===//
2177 class SHLHIInst<dag OOL, dag IOL, list<dag> pattern>:
2178 RI7Form<0b11111010000, OOL, IOL, "shlhi\t$rT, $rA, $val",
2179 RotateShift, pattern>;
2181 class SHLHIVecInst<ValueType vectype>:
2182 SHLHIInst<(outs VECREG:$rT), (ins VECREG:$rA, u7imm:$val),
2183 [(set (vectype VECREG:$rT),
2184 (SPUvec_shl (vectype VECREG:$rA), (i16 uimm7:$val)))]>;
2186 multiclass ShiftLeftHalfwordImm
2188 def v8i16: SHLHIVecInst<v8i16>;
2189 def r16: SHLHIInst<(outs R16C:$rT), (ins R16C:$rA, u7imm:$val),
2190 [(set R16C:$rT, (shl R16C:$rA, (i16 uimm7:$val)))]>;
2193 defm SHLHI : ShiftLeftHalfwordImm;
2195 def : Pat<(SPUvec_shl (v8i16 VECREG:$rA), (i32 uimm7:$val)),
2196 (SHLHIv8i16 VECREG:$rA, uimm7:$val)>;
2198 def : Pat<(shl R16C:$rA, (i32 uimm7:$val)),
2199 (SHLHIr16 R16C:$rA, uimm7:$val)>;
2201 //===----------------------------------------------------------------------===//
2203 class SHLInst<dag OOL, dag IOL, list<dag> pattern>:
2204 RRForm<0b11111010000, OOL, IOL, "shl\t$rT, $rA, $rB",
2205 RotateShift, pattern>;
2207 multiclass ShiftLeftWord
2210 SHLInst<(outs VECREG:$rT), (ins VECREG:$rA, R16C:$rB),
2211 [(set (v4i32 VECREG:$rT),
2212 (SPUvec_shl (v4i32 VECREG:$rA), R16C:$rB))]>;
2214 SHLInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
2215 [(set R32C:$rT, (shl R32C:$rA, R32C:$rB))]>;
2218 defm SHL: ShiftLeftWord;
2220 //===----------------------------------------------------------------------===//
2222 class SHLIInst<dag OOL, dag IOL, list<dag> pattern>:
2223 RI7Form<0b11111010000, OOL, IOL, "shli\t$rT, $rA, $val",
2224 RotateShift, pattern>;
2226 multiclass ShiftLeftWordImm
2229 SHLIInst<(outs VECREG:$rT), (ins VECREG:$rA, u7imm_i32:$val),
2230 [(set (v4i32 VECREG:$rT),
2231 (SPUvec_shl (v4i32 VECREG:$rA), (i32 uimm7:$val)))]>;
2234 SHLIInst<(outs R32C:$rT), (ins R32C:$rA, u7imm_i32:$val),
2235 [(set R32C:$rT, (shl R32C:$rA, (i32 uimm7:$val)))]>;
2238 defm SHLI : ShiftLeftWordImm;
2240 //===----------------------------------------------------------------------===//
2241 // SHLQBI vec form: Note that this will shift the entire vector (the 128-bit
2242 // register) to the left. Vector form is here to ensure type correctness.
2244 // The shift count is in the lowest 3 bits (29-31) of $rB, so only a bit shift
2245 // of 7 bits is actually possible.
2247 // Note also that SHLQBI/SHLQBII are used in conjunction with SHLQBY/SHLQBYI
2248 // to shift i64 and i128. SHLQBI is the residual left over after shifting by
2249 // bytes with SHLQBY.
2251 class SHLQBIInst<dag OOL, dag IOL, list<dag> pattern>:
2252 RRForm<0b11011011100, OOL, IOL, "shlqbi\t$rT, $rA, $rB",
2253 RotateShift, pattern>;
2255 class SHLQBIVecInst<ValueType vectype>:
2256 SHLQBIInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2257 [(set (vectype VECREG:$rT),
2258 (SPUshlquad_l_bits (vectype VECREG:$rA), R32C:$rB))]>;
2260 class SHLQBIRegInst<RegisterClass rclass>:
2261 SHLQBIInst<(outs rclass:$rT), (ins rclass:$rA, R32C:$rB),
2262 [/* no pattern */]>;
2264 multiclass ShiftLeftQuadByBits
2266 def v16i8: SHLQBIVecInst<v16i8>;
2267 def v8i16: SHLQBIVecInst<v8i16>;
2268 def v4i32: SHLQBIVecInst<v4i32>;
2269 def v4f32: SHLQBIVecInst<v4f32>;
2270 def v2i64: SHLQBIVecInst<v2i64>;
2271 def v2f64: SHLQBIVecInst<v2f64>;
2273 def r128: SHLQBIRegInst<GPRC>;
2276 defm SHLQBI : ShiftLeftQuadByBits;
2278 // See note above on SHLQBI. In this case, the predicate actually does then
2279 // enforcement, whereas with SHLQBI, we have to "take it on faith."
2280 class SHLQBIIInst<dag OOL, dag IOL, list<dag> pattern>:
2281 RI7Form<0b11011111100, OOL, IOL, "shlqbii\t$rT, $rA, $val",
2282 RotateShift, pattern>;
2284 class SHLQBIIVecInst<ValueType vectype>:
2285 SHLQBIIInst<(outs VECREG:$rT), (ins VECREG:$rA, u7imm_i32:$val),
2286 [(set (vectype VECREG:$rT),
2287 (SPUshlquad_l_bits (vectype VECREG:$rA), (i32 bitshift:$val)))]>;
2289 multiclass ShiftLeftQuadByBitsImm
2291 def v16i8 : SHLQBIIVecInst<v16i8>;
2292 def v8i16 : SHLQBIIVecInst<v8i16>;
2293 def v4i32 : SHLQBIIVecInst<v4i32>;
2294 def v4f32 : SHLQBIIVecInst<v4f32>;
2295 def v2i64 : SHLQBIIVecInst<v2i64>;
2296 def v2f64 : SHLQBIIVecInst<v2f64>;
2299 defm SHLQBII : ShiftLeftQuadByBitsImm;
2301 // SHLQBY, SHLQBYI vector forms: Shift the entire vector to the left by bytes,
2302 // not by bits. See notes above on SHLQBI.
2304 class SHLQBYInst<dag OOL, dag IOL, list<dag> pattern>:
2305 RI7Form<0b11111011100, OOL, IOL, "shlqby\t$rT, $rA, $rB",
2306 RotateShift, pattern>;
2308 class SHLQBYVecInst<ValueType vectype>:
2309 SHLQBYInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2310 [(set (vectype VECREG:$rT),
2311 (SPUshlquad_l_bytes (vectype VECREG:$rA), R32C:$rB))]>;
2313 multiclass ShiftLeftQuadBytes
2315 def v16i8: SHLQBYVecInst<v16i8>;
2316 def v8i16: SHLQBYVecInst<v8i16>;
2317 def v4i32: SHLQBYVecInst<v4i32>;
2318 def v4f32: SHLQBYVecInst<v4f32>;
2319 def v2i64: SHLQBYVecInst<v2i64>;
2320 def v2f64: SHLQBYVecInst<v2f64>;
2321 def r128: SHLQBYInst<(outs GPRC:$rT), (ins GPRC:$rA, R32C:$rB),
2322 [(set GPRC:$rT, (SPUshlquad_l_bytes GPRC:$rA, R32C:$rB))]>;
2325 defm SHLQBY: ShiftLeftQuadBytes;
2327 class SHLQBYIInst<dag OOL, dag IOL, list<dag> pattern>:
2328 RI7Form<0b11111111100, OOL, IOL, "shlqbyi\t$rT, $rA, $val",
2329 RotateShift, pattern>;
2331 class SHLQBYIVecInst<ValueType vectype>:
2332 SHLQBYIInst<(outs VECREG:$rT), (ins VECREG:$rA, u7imm_i32:$val),
2333 [(set (vectype VECREG:$rT),
2334 (SPUshlquad_l_bytes (vectype VECREG:$rA), (i32 uimm7:$val)))]>;
2336 multiclass ShiftLeftQuadBytesImm
2338 def v16i8: SHLQBYIVecInst<v16i8>;
2339 def v8i16: SHLQBYIVecInst<v8i16>;
2340 def v4i32: SHLQBYIVecInst<v4i32>;
2341 def v4f32: SHLQBYIVecInst<v4f32>;
2342 def v2i64: SHLQBYIVecInst<v2i64>;
2343 def v2f64: SHLQBYIVecInst<v2f64>;
2344 def r128: SHLQBYIInst<(outs GPRC:$rT), (ins GPRC:$rA, u7imm_i32:$val),
2346 (SPUshlquad_l_bytes GPRC:$rA, (i32 uimm7:$val)))]>;
2349 defm SHLQBYI : ShiftLeftQuadBytesImm;
2351 class SHLQBYBIInst<dag OOL, dag IOL, list<dag> pattern>:
2352 RRForm<0b00111001111, OOL, IOL, "shlqbybi\t$rT, $rA, $rB",
2353 RotateShift, pattern>;
2355 class SHLQBYBIVecInst<ValueType vectype>:
2356 SHLQBYBIInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2357 [/* no pattern */]>;
2359 class SHLQBYBIRegInst<RegisterClass rclass>:
2360 SHLQBYBIInst<(outs rclass:$rT), (ins rclass:$rA, R32C:$rB),
2361 [/* no pattern */]>;
2363 multiclass ShiftLeftQuadBytesBitCount
2365 def v16i8: SHLQBYBIVecInst<v16i8>;
2366 def v8i16: SHLQBYBIVecInst<v8i16>;
2367 def v4i32: SHLQBYBIVecInst<v4i32>;
2368 def v4f32: SHLQBYBIVecInst<v4f32>;
2369 def v2i64: SHLQBYBIVecInst<v2i64>;
2370 def v2f64: SHLQBYBIVecInst<v2f64>;
2372 def r128: SHLQBYBIRegInst<GPRC>;
2375 defm SHLQBYBI : ShiftLeftQuadBytesBitCount;
2377 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2379 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2380 class ROTHInst<dag OOL, dag IOL, list<dag> pattern>:
2381 RRForm<0b00111010000, OOL, IOL, "roth\t$rT, $rA, $rB",
2382 RotateShift, pattern>;
2384 class ROTHVecInst<ValueType vectype>:
2385 ROTHInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
2386 [(set (vectype VECREG:$rT),
2387 (SPUvec_rotl VECREG:$rA, VECREG:$rB))]>;
2389 class ROTHRegInst<RegisterClass rclass>:
2390 ROTHInst<(outs rclass:$rT), (ins rclass:$rA, rclass:$rB),
2391 [(set rclass:$rT, (rotl rclass:$rA, rclass:$rB))]>;
2393 multiclass RotateLeftHalfword
2395 def v8i16: ROTHVecInst<v8i16>;
2396 def r16: ROTHRegInst<R16C>;
2399 defm ROTH: RotateLeftHalfword;
2401 def ROTHr16_r32: ROTHInst<(outs R16C:$rT), (ins R16C:$rA, R32C:$rB),
2402 [(set R16C:$rT, (rotl R16C:$rA, R32C:$rB))]>;
2404 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2405 // Rotate halfword, immediate:
2406 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2407 class ROTHIInst<dag OOL, dag IOL, list<dag> pattern>:
2408 RI7Form<0b00111110000, OOL, IOL, "rothi\t$rT, $rA, $val",
2409 RotateShift, pattern>;
2411 class ROTHIVecInst<ValueType vectype>:
2412 ROTHIInst<(outs VECREG:$rT), (ins VECREG:$rA, u7imm:$val),
2413 [(set (vectype VECREG:$rT),
2414 (SPUvec_rotl VECREG:$rA, (i16 uimm7:$val)))]>;
2416 multiclass RotateLeftHalfwordImm
2418 def v8i16: ROTHIVecInst<v8i16>;
2419 def r16: ROTHIInst<(outs R16C:$rT), (ins R16C:$rA, u7imm:$val),
2420 [(set R16C:$rT, (rotl R16C:$rA, (i16 uimm7:$val)))]>;
2421 def r16_r32: ROTHIInst<(outs R16C:$rT), (ins R16C:$rA, u7imm_i32:$val),
2422 [(set R16C:$rT, (rotl R16C:$rA, (i32 uimm7:$val)))]>;
2425 defm ROTHI: RotateLeftHalfwordImm;
2427 def : Pat<(SPUvec_rotl VECREG:$rA, (i32 uimm7:$val)),
2428 (ROTHIv8i16 VECREG:$rA, imm:$val)>;
2430 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2432 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2434 class ROTInst<dag OOL, dag IOL, list<dag> pattern>:
2435 RRForm<0b00011010000, OOL, IOL, "rot\t$rT, $rA, $rB",
2436 RotateShift, pattern>;
2438 class ROTVecInst<ValueType vectype>:
2439 ROTInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2440 [(set (vectype VECREG:$rT),
2441 (SPUvec_rotl (vectype VECREG:$rA), R32C:$rB))]>;
2443 class ROTRegInst<RegisterClass rclass>:
2444 ROTInst<(outs rclass:$rT), (ins rclass:$rA, R32C:$rB),
2446 (rotl rclass:$rA, R32C:$rB))]>;
2448 multiclass RotateLeftWord
2450 def v4i32: ROTVecInst<v4i32>;
2451 def r32: ROTRegInst<R32C>;
2454 defm ROT: RotateLeftWord;
2456 // The rotate amount is in the same bits whether we've got an 8-bit, 16-bit or
2458 def ROTr32_r16_anyext:
2459 ROTInst<(outs R32C:$rT), (ins R32C:$rA, R16C:$rB),
2460 [(set R32C:$rT, (rotl R32C:$rA, (i32 (anyext R16C:$rB))))]>;
2462 def : Pat<(rotl R32C:$rA, (i32 (zext R16C:$rB))),
2463 (ROTr32_r16_anyext R32C:$rA, R16C:$rB)>;
2465 def : Pat<(rotl R32C:$rA, (i32 (sext R16C:$rB))),
2466 (ROTr32_r16_anyext R32C:$rA, R16C:$rB)>;
2468 def ROTr32_r8_anyext:
2469 ROTInst<(outs R32C:$rT), (ins R32C:$rA, R8C:$rB),
2470 [(set R32C:$rT, (rotl R32C:$rA, (i32 (anyext R8C:$rB))))]>;
2472 def : Pat<(rotl R32C:$rA, (i32 (zext R8C:$rB))),
2473 (ROTr32_r8_anyext R32C:$rA, R8C:$rB)>;
2475 def : Pat<(rotl R32C:$rA, (i32 (sext R8C:$rB))),
2476 (ROTr32_r8_anyext R32C:$rA, R8C:$rB)>;
2478 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2479 // Rotate word, immediate
2480 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2482 class ROTIInst<dag OOL, dag IOL, list<dag> pattern>:
2483 RI7Form<0b00011110000, OOL, IOL, "roti\t$rT, $rA, $val",
2484 RotateShift, pattern>;
2486 class ROTIVecInst<ValueType vectype, Operand optype, ValueType inttype, PatLeaf pred>:
2487 ROTIInst<(outs VECREG:$rT), (ins VECREG:$rA, optype:$val),
2488 [(set (vectype VECREG:$rT),
2489 (SPUvec_rotl (vectype VECREG:$rA), (inttype pred:$val)))]>;
2491 class ROTIRegInst<RegisterClass rclass, Operand optype, ValueType inttype, PatLeaf pred>:
2492 ROTIInst<(outs rclass:$rT), (ins rclass:$rA, optype:$val),
2493 [(set rclass:$rT, (rotl rclass:$rA, (inttype pred:$val)))]>;
2495 multiclass RotateLeftWordImm
2497 def v4i32: ROTIVecInst<v4i32, u7imm_i32, i32, uimm7>;
2498 def v4i32_i16: ROTIVecInst<v4i32, u7imm, i16, uimm7>;
2499 def v4i32_i8: ROTIVecInst<v4i32, u7imm_i8, i8, uimm7>;
2501 def r32: ROTIRegInst<R32C, u7imm_i32, i32, uimm7>;
2502 def r32_i16: ROTIRegInst<R32C, u7imm, i16, uimm7>;
2503 def r32_i8: ROTIRegInst<R32C, u7imm_i8, i8, uimm7>;
2506 defm ROTI : RotateLeftWordImm;
2508 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2509 // Rotate quad by byte (count)
2510 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2512 class ROTQBYInst<dag OOL, dag IOL, list<dag> pattern>:
2513 RRForm<0b00111011100, OOL, IOL, "rotqby\t$rT, $rA, $rB",
2514 RotateShift, pattern>;
2516 class ROTQBYVecInst<ValueType vectype>:
2517 ROTQBYInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2518 [(set (vectype VECREG:$rT),
2519 (SPUrotbytes_left (vectype VECREG:$rA), R32C:$rB))]>;
2521 multiclass RotateQuadLeftByBytes
2523 def v16i8: ROTQBYVecInst<v16i8>;
2524 def v8i16: ROTQBYVecInst<v8i16>;
2525 def v4i32: ROTQBYVecInst<v4i32>;
2526 def v4f32: ROTQBYVecInst<v4f32>;
2527 def v2i64: ROTQBYVecInst<v2i64>;
2528 def v2f64: ROTQBYVecInst<v2f64>;
2531 defm ROTQBY: RotateQuadLeftByBytes;
2533 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2534 // Rotate quad by byte (count), immediate
2535 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2537 class ROTQBYIInst<dag OOL, dag IOL, list<dag> pattern>:
2538 RI7Form<0b00111111100, OOL, IOL, "rotqbyi\t$rT, $rA, $val",
2539 RotateShift, pattern>;
2541 class ROTQBYIVecInst<ValueType vectype>:
2542 ROTQBYIInst<(outs VECREG:$rT), (ins VECREG:$rA, u7imm:$val),
2543 [(set (vectype VECREG:$rT),
2544 (SPUrotbytes_left (vectype VECREG:$rA), (i16 uimm7:$val)))]>;
2546 multiclass RotateQuadByBytesImm
2548 def v16i8: ROTQBYIVecInst<v16i8>;
2549 def v8i16: ROTQBYIVecInst<v8i16>;
2550 def v4i32: ROTQBYIVecInst<v4i32>;
2551 def v4f32: ROTQBYIVecInst<v4f32>;
2552 def v2i64: ROTQBYIVecInst<v2i64>;
2553 def vfi64: ROTQBYIVecInst<v2f64>;
2556 defm ROTQBYI: RotateQuadByBytesImm;
2558 // See ROTQBY note above.
2559 class ROTQBYBIInst<dag OOL, dag IOL, list<dag> pattern>:
2560 RI7Form<0b00110011100, OOL, IOL,
2561 "rotqbybi\t$rT, $rA, $shift",
2562 RotateShift, pattern>;
2564 class ROTQBYBIVecInst<ValueType vectype, RegisterClass rclass>:
2565 ROTQBYBIInst<(outs VECREG:$rT), (ins VECREG:$rA, rclass:$shift),
2566 [(set (vectype VECREG:$rT),
2567 (SPUrotbytes_left_bits (vectype VECREG:$rA), rclass:$shift))]>;
2569 multiclass RotateQuadByBytesByBitshift {
2570 def v16i8_r32: ROTQBYBIVecInst<v16i8, R32C>;
2571 def v8i16_r32: ROTQBYBIVecInst<v8i16, R32C>;
2572 def v4i32_r32: ROTQBYBIVecInst<v4i32, R32C>;
2573 def v2i64_r32: ROTQBYBIVecInst<v2i64, R32C>;
2576 defm ROTQBYBI : RotateQuadByBytesByBitshift;
2578 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2579 // See ROTQBY note above.
2581 // Assume that the user of this instruction knows to shift the rotate count
2583 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2585 class ROTQBIInst<dag OOL, dag IOL, list<dag> pattern>:
2586 RRForm<0b00011011100, OOL, IOL, "rotqbi\t$rT, $rA, $rB",
2587 RotateShift, pattern>;
2589 class ROTQBIVecInst<ValueType vectype>:
2590 ROTQBIInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2591 [/* no pattern yet */]>;
2593 class ROTQBIRegInst<RegisterClass rclass>:
2594 ROTQBIInst<(outs rclass:$rT), (ins rclass:$rA, R32C:$rB),
2595 [/* no pattern yet */]>;
2597 multiclass RotateQuadByBitCount
2599 def v16i8: ROTQBIVecInst<v16i8>;
2600 def v8i16: ROTQBIVecInst<v8i16>;
2601 def v4i32: ROTQBIVecInst<v4i32>;
2602 def v2i64: ROTQBIVecInst<v2i64>;
2604 def r128: ROTQBIRegInst<GPRC>;
2605 def r64: ROTQBIRegInst<R64C>;
2608 defm ROTQBI: RotateQuadByBitCount;
2610 class ROTQBIIInst<dag OOL, dag IOL, list<dag> pattern>:
2611 RI7Form<0b00011111100, OOL, IOL, "rotqbii\t$rT, $rA, $val",
2612 RotateShift, pattern>;
2614 class ROTQBIIVecInst<ValueType vectype, Operand optype, ValueType inttype,
2616 ROTQBIIInst<(outs VECREG:$rT), (ins VECREG:$rA, optype:$val),
2617 [/* no pattern yet */]>;
2619 class ROTQBIIRegInst<RegisterClass rclass, Operand optype, ValueType inttype,
2621 ROTQBIIInst<(outs rclass:$rT), (ins rclass:$rA, optype:$val),
2622 [/* no pattern yet */]>;
2624 multiclass RotateQuadByBitCountImm
2626 def v16i8: ROTQBIIVecInst<v16i8, u7imm_i32, i32, uimm7>;
2627 def v8i16: ROTQBIIVecInst<v8i16, u7imm_i32, i32, uimm7>;
2628 def v4i32: ROTQBIIVecInst<v4i32, u7imm_i32, i32, uimm7>;
2629 def v2i64: ROTQBIIVecInst<v2i64, u7imm_i32, i32, uimm7>;
2631 def r128: ROTQBIIRegInst<GPRC, u7imm_i32, i32, uimm7>;
2632 def r64: ROTQBIIRegInst<R64C, u7imm_i32, i32, uimm7>;
2635 defm ROTQBII : RotateQuadByBitCountImm;
2637 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2638 // ROTHM v8i16 form:
2639 // NOTE(1): No vector rotate is generated by the C/C++ frontend (today),
2640 // so this only matches a synthetically generated/lowered code
2642 // NOTE(2): $rB must be negated before the right rotate!
2643 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2645 class ROTHMInst<dag OOL, dag IOL, list<dag> pattern>:
2646 RRForm<0b10111010000, OOL, IOL, "rothm\t$rT, $rA, $rB",
2647 RotateShift, pattern>;
2650 ROTHMInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2651 [/* see patterns below - $rB must be negated */]>;
2653 def : Pat<(SPUvec_srl (v8i16 VECREG:$rA), R32C:$rB),
2654 (ROTHMv8i16 VECREG:$rA, (SFIr32 R32C:$rB, 0))>;
2656 def : Pat<(SPUvec_srl (v8i16 VECREG:$rA), R16C:$rB),
2657 (ROTHMv8i16 VECREG:$rA,
2658 (SFIr32 (XSHWr16 R16C:$rB), 0))>;
2660 def : Pat<(SPUvec_srl (v8i16 VECREG:$rA), R8C:$rB),
2661 (ROTHMv8i16 VECREG:$rA,
2662 (SFIr32 (XSHWr16 (XSBHr8 R8C:$rB) ), 0))>;
2664 // ROTHM r16 form: Rotate 16-bit quantity to right, zero fill at the left
2665 // Note: This instruction doesn't match a pattern because rB must be negated
2666 // for the instruction to work. Thus, the pattern below the instruction!
2669 ROTHMInst<(outs R16C:$rT), (ins R16C:$rA, R32C:$rB),
2670 [/* see patterns below - $rB must be negated! */]>;
2672 def : Pat<(srl R16C:$rA, R32C:$rB),
2673 (ROTHMr16 R16C:$rA, (SFIr32 R32C:$rB, 0))>;
2675 def : Pat<(srl R16C:$rA, R16C:$rB),
2677 (SFIr32 (XSHWr16 R16C:$rB), 0))>;
2679 def : Pat<(srl R16C:$rA, R8C:$rB),
2681 (SFIr32 (XSHWr16 (XSBHr8 R8C:$rB) ), 0))>;
2683 // ROTHMI v8i16 form: See the comment for ROTHM v8i16. The difference here is
2684 // that the immediate can be complemented, so that the user doesn't have to
2687 class ROTHMIInst<dag OOL, dag IOL, list<dag> pattern>:
2688 RI7Form<0b10111110000, OOL, IOL, "rothmi\t$rT, $rA, $val",
2689 RotateShift, pattern>;
2692 ROTHMIInst<(outs VECREG:$rT), (ins VECREG:$rA, rothNeg7imm:$val),
2693 [/* no pattern */]>;
2695 def : Pat<(SPUvec_srl (v8i16 VECREG:$rA), (i32 imm:$val)),
2696 (ROTHMIv8i16 VECREG:$rA, imm:$val)>;
2698 def: Pat<(SPUvec_srl (v8i16 VECREG:$rA), (i16 imm:$val)),
2699 (ROTHMIv8i16 VECREG:$rA, imm:$val)>;
2701 def: Pat<(SPUvec_srl (v8i16 VECREG:$rA), (i8 imm:$val)),
2702 (ROTHMIv8i16 VECREG:$rA, imm:$val)>;
2705 ROTHMIInst<(outs R16C:$rT), (ins R16C:$rA, rothNeg7imm:$val),
2706 [/* no pattern */]>;
2708 def: Pat<(srl R16C:$rA, (i32 uimm7:$val)),
2709 (ROTHMIr16 R16C:$rA, uimm7:$val)>;
2711 def: Pat<(srl R16C:$rA, (i16 uimm7:$val)),
2712 (ROTHMIr16 R16C:$rA, uimm7:$val)>;
2714 def: Pat<(srl R16C:$rA, (i8 uimm7:$val)),
2715 (ROTHMIr16 R16C:$rA, uimm7:$val)>;
2717 // ROTM v4i32 form: See the ROTHM v8i16 comments.
2718 class ROTMInst<dag OOL, dag IOL, list<dag> pattern>:
2719 RRForm<0b10011010000, OOL, IOL, "rotm\t$rT, $rA, $rB",
2720 RotateShift, pattern>;
2723 ROTMInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2724 [/* see patterns below - $rB must be negated */]>;
2726 def : Pat<(SPUvec_srl VECREG:$rA, R32C:$rB),
2727 (ROTMv4i32 VECREG:$rA, (SFIr32 R32C:$rB, 0))>;
2729 def : Pat<(SPUvec_srl VECREG:$rA, R16C:$rB),
2730 (ROTMv4i32 VECREG:$rA,
2731 (SFIr32 (XSHWr16 R16C:$rB), 0))>;
2733 def : Pat<(SPUvec_srl VECREG:$rA, R8C:$rB),
2734 (ROTMv4i32 VECREG:$rA,
2735 (SFIr32 (XSHWr16 (XSBHr8 R8C:$rB)), 0))>;
2738 ROTMInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
2739 [/* see patterns below - $rB must be negated */]>;
2741 def : Pat<(srl R32C:$rA, R32C:$rB),
2742 (ROTMr32 R32C:$rA, (SFIr32 R32C:$rB, 0))>;
2744 def : Pat<(srl R32C:$rA, R16C:$rB),
2746 (SFIr32 (XSHWr16 R16C:$rB), 0))>;
2748 def : Pat<(srl R32C:$rA, R8C:$rB),
2750 (SFIr32 (XSHWr16 (XSBHr8 R8C:$rB)), 0))>;
2752 // ROTMI v4i32 form: See the comment for ROTHM v8i16.
2754 RI7Form<0b10011110000, (outs VECREG:$rT), (ins VECREG:$rA, rotNeg7imm:$val),
2755 "rotmi\t$rT, $rA, $val", RotateShift,
2756 [(set (v4i32 VECREG:$rT),
2757 (SPUvec_srl VECREG:$rA, (i32 uimm7:$val)))]>;
2759 def : Pat<(SPUvec_srl VECREG:$rA, (i16 uimm7:$val)),
2760 (ROTMIv4i32 VECREG:$rA, uimm7:$val)>;
2762 def : Pat<(SPUvec_srl VECREG:$rA, (i8 uimm7:$val)),
2763 (ROTMIv4i32 VECREG:$rA, uimm7:$val)>;
2765 // ROTMI r32 form: know how to complement the immediate value.
2767 RI7Form<0b10011110000, (outs R32C:$rT), (ins R32C:$rA, rotNeg7imm:$val),
2768 "rotmi\t$rT, $rA, $val", RotateShift,
2769 [(set R32C:$rT, (srl R32C:$rA, (i32 uimm7:$val)))]>;
2771 def : Pat<(srl R32C:$rA, (i16 imm:$val)),
2772 (ROTMIr32 R32C:$rA, uimm7:$val)>;
2774 def : Pat<(srl R32C:$rA, (i8 imm:$val)),
2775 (ROTMIr32 R32C:$rA, uimm7:$val)>;
2777 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2778 // ROTQMBY: This is a vector form merely so that when used in an
2779 // instruction pattern, type checking will succeed. This instruction assumes
2780 // that the user knew to negate $rB.
2781 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2783 class ROTQMBYInst<dag OOL, dag IOL, list<dag> pattern>:
2784 RRForm<0b10111011100, OOL, IOL, "rotqmby\t$rT, $rA, $rB",
2785 RotateShift, pattern>;
2787 class ROTQMBYVecInst<ValueType vectype>:
2788 ROTQMBYInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2789 [/* no pattern, $rB must be negated */]>;
2791 class ROTQMBYRegInst<RegisterClass rclass>:
2792 ROTQMBYInst<(outs rclass:$rT), (ins rclass:$rA, R32C:$rB),
2793 [/* no pattern */]>;
2795 multiclass RotateQuadBytes
2797 def v16i8: ROTQMBYVecInst<v16i8>;
2798 def v8i16: ROTQMBYVecInst<v8i16>;
2799 def v4i32: ROTQMBYVecInst<v4i32>;
2800 def v2i64: ROTQMBYVecInst<v2i64>;
2802 def r128: ROTQMBYRegInst<GPRC>;
2803 def r64: ROTQMBYRegInst<R64C>;
2806 defm ROTQMBY : RotateQuadBytes;
2808 class ROTQMBYIInst<dag OOL, dag IOL, list<dag> pattern>:
2809 RI7Form<0b10111111100, OOL, IOL, "rotqmbyi\t$rT, $rA, $val",
2810 RotateShift, pattern>;
2812 class ROTQMBYIVecInst<ValueType vectype>:
2813 ROTQMBYIInst<(outs VECREG:$rT), (ins VECREG:$rA, rotNeg7imm:$val),
2814 [/* no pattern */]>;
2816 class ROTQMBYIRegInst<RegisterClass rclass, Operand optype, ValueType inttype,
2818 ROTQMBYIInst<(outs rclass:$rT), (ins rclass:$rA, optype:$val),
2819 [/* no pattern */]>;
2821 // 128-bit zero extension form:
2822 class ROTQMBYIZExtInst<RegisterClass rclass, Operand optype, PatLeaf pred>:
2823 ROTQMBYIInst<(outs GPRC:$rT), (ins rclass:$rA, optype:$val),
2824 [/* no pattern */]>;
2826 multiclass RotateQuadBytesImm
2828 def v16i8: ROTQMBYIVecInst<v16i8>;
2829 def v8i16: ROTQMBYIVecInst<v8i16>;
2830 def v4i32: ROTQMBYIVecInst<v4i32>;
2831 def v2i64: ROTQMBYIVecInst<v2i64>;
2833 def r128: ROTQMBYIRegInst<GPRC, rotNeg7imm, i32, uimm7>;
2834 def r64: ROTQMBYIRegInst<R64C, rotNeg7imm, i32, uimm7>;
2836 def r128_zext_r8: ROTQMBYIZExtInst<R8C, rotNeg7imm, uimm7>;
2837 def r128_zext_r16: ROTQMBYIZExtInst<R16C, rotNeg7imm, uimm7>;
2838 def r128_zext_r32: ROTQMBYIZExtInst<R32C, rotNeg7imm, uimm7>;
2839 def r128_zext_r64: ROTQMBYIZExtInst<R64C, rotNeg7imm, uimm7>;
2842 defm ROTQMBYI : RotateQuadBytesImm;
2844 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2845 // Rotate right and mask by bit count
2846 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2848 class ROTQMBYBIInst<dag OOL, dag IOL, list<dag> pattern>:
2849 RRForm<0b10110011100, OOL, IOL, "rotqmbybi\t$rT, $rA, $rB",
2850 RotateShift, pattern>;
2852 class ROTQMBYBIVecInst<ValueType vectype>:
2853 ROTQMBYBIInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2854 [/* no pattern, */]>;
2856 multiclass RotateMaskQuadByBitCount
2858 def v16i8: ROTQMBYBIVecInst<v16i8>;
2859 def v8i16: ROTQMBYBIVecInst<v8i16>;
2860 def v4i32: ROTQMBYBIVecInst<v4i32>;
2861 def v2i64: ROTQMBYBIVecInst<v2i64>;
2864 defm ROTQMBYBI: RotateMaskQuadByBitCount;
2866 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2867 // Rotate quad and mask by bits
2868 // Note that the rotate amount has to be negated
2869 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2871 class ROTQMBIInst<dag OOL, dag IOL, list<dag> pattern>:
2872 RRForm<0b10011011100, OOL, IOL, "rotqmbi\t$rT, $rA, $rB",
2873 RotateShift, pattern>;
2875 class ROTQMBIVecInst<ValueType vectype>:
2876 ROTQMBIInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2877 [/* no pattern */]>;
2879 class ROTQMBIRegInst<RegisterClass rclass>:
2880 ROTQMBIInst<(outs rclass:$rT), (ins rclass:$rA, R32C:$rB),
2881 [/* no pattern */]>;
2883 multiclass RotateMaskQuadByBits
2885 def v16i8: ROTQMBIVecInst<v16i8>;
2886 def v8i16: ROTQMBIVecInst<v8i16>;
2887 def v4i32: ROTQMBIVecInst<v4i32>;
2888 def v2i64: ROTQMBIVecInst<v2i64>;
2890 def r128: ROTQMBIRegInst<GPRC>;
2891 def r64: ROTQMBIRegInst<R64C>;
2894 defm ROTQMBI: RotateMaskQuadByBits;
2896 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2897 // Rotate quad and mask by bits, immediate
2898 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2900 class ROTQMBIIInst<dag OOL, dag IOL, list<dag> pattern>:
2901 RI7Form<0b10011111100, OOL, IOL, "rotqmbii\t$rT, $rA, $val",
2902 RotateShift, pattern>;
2904 class ROTQMBIIVecInst<ValueType vectype>:
2905 ROTQMBIIInst<(outs VECREG:$rT), (ins VECREG:$rA, rotNeg7imm:$val),
2906 [/* no pattern */]>;
2908 class ROTQMBIIRegInst<RegisterClass rclass>:
2909 ROTQMBIIInst<(outs rclass:$rT), (ins rclass:$rA, rotNeg7imm:$val),
2910 [/* no pattern */]>;
2912 multiclass RotateMaskQuadByBitsImm
2914 def v16i8: ROTQMBIIVecInst<v16i8>;
2915 def v8i16: ROTQMBIIVecInst<v8i16>;
2916 def v4i32: ROTQMBIIVecInst<v4i32>;
2917 def v2i64: ROTQMBIIVecInst<v2i64>;
2919 def r128: ROTQMBIIRegInst<GPRC>;
2920 def r64: ROTQMBIIRegInst<R64C>;
2923 defm ROTQMBII: RotateMaskQuadByBitsImm;
2925 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2926 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
2929 RRForm<0b01111010000, (outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2930 "rotmah\t$rT, $rA, $rB", RotateShift,
2931 [/* see patterns below - $rB must be negated */]>;
2933 def : Pat<(SPUvec_sra VECREG:$rA, R32C:$rB),
2934 (ROTMAHv8i16 VECREG:$rA, (SFIr32 R32C:$rB, 0))>;
2936 def : Pat<(SPUvec_sra VECREG:$rA, R16C:$rB),
2937 (ROTMAHv8i16 VECREG:$rA,
2938 (SFIr32 (XSHWr16 R16C:$rB), 0))>;
2940 def : Pat<(SPUvec_sra VECREG:$rA, R8C:$rB),
2941 (ROTMAHv8i16 VECREG:$rA,
2942 (SFIr32 (XSHWr16 (XSBHr8 R8C:$rB)), 0))>;
2945 RRForm<0b01111010000, (outs R16C:$rT), (ins R16C:$rA, R32C:$rB),
2946 "rotmah\t$rT, $rA, $rB", RotateShift,
2947 [/* see patterns below - $rB must be negated */]>;
2949 def : Pat<(sra R16C:$rA, R32C:$rB),
2950 (ROTMAHr16 R16C:$rA, (SFIr32 R32C:$rB, 0))>;
2952 def : Pat<(sra R16C:$rA, R16C:$rB),
2953 (ROTMAHr16 R16C:$rA,
2954 (SFIr32 (XSHWr16 R16C:$rB), 0))>;
2956 def : Pat<(sra R16C:$rA, R8C:$rB),
2957 (ROTMAHr16 R16C:$rA,
2958 (SFIr32 (XSHWr16 (XSBHr8 R8C:$rB)), 0))>;
2961 RRForm<0b01111110000, (outs VECREG:$rT), (ins VECREG:$rA, rothNeg7imm:$val),
2962 "rotmahi\t$rT, $rA, $val", RotateShift,
2963 [(set (v8i16 VECREG:$rT),
2964 (SPUvec_sra (v8i16 VECREG:$rA), (i32 uimm7:$val)))]>;
2966 def : Pat<(SPUvec_sra (v8i16 VECREG:$rA), (i16 uimm7:$val)),
2967 (ROTMAHIv8i16 (v8i16 VECREG:$rA), (i32 uimm7:$val))>;
2969 def : Pat<(SPUvec_sra (v8i16 VECREG:$rA), (i8 uimm7:$val)),
2970 (ROTMAHIv8i16 (v8i16 VECREG:$rA), (i32 uimm7:$val))>;
2973 RRForm<0b01111110000, (outs R16C:$rT), (ins R16C:$rA, rothNeg7imm_i16:$val),
2974 "rotmahi\t$rT, $rA, $val", RotateShift,
2975 [(set R16C:$rT, (sra R16C:$rA, (i16 uimm7:$val)))]>;
2977 def : Pat<(sra R16C:$rA, (i32 imm:$val)),
2978 (ROTMAHIr16 R16C:$rA, uimm7:$val)>;
2980 def : Pat<(sra R16C:$rA, (i8 imm:$val)),
2981 (ROTMAHIr16 R16C:$rA, uimm7:$val)>;
2984 RRForm<0b01011010000, (outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
2985 "rotma\t$rT, $rA, $rB", RotateShift,
2986 [/* see patterns below - $rB must be negated */]>;
2988 def : Pat<(SPUvec_sra VECREG:$rA, R32C:$rB),
2989 (ROTMAv4i32 (v4i32 VECREG:$rA), (SFIr32 R32C:$rB, 0))>;
2991 def : Pat<(SPUvec_sra VECREG:$rA, R16C:$rB),
2992 (ROTMAv4i32 (v4i32 VECREG:$rA),
2993 (SFIr32 (XSHWr16 R16C:$rB), 0))>;
2995 def : Pat<(SPUvec_sra VECREG:$rA, R8C:$rB),
2996 (ROTMAv4i32 (v4i32 VECREG:$rA),
2997 (SFIr32 (XSHWr16 (XSBHr8 R8C:$rB)), 0))>;
3000 RRForm<0b01011010000, (outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
3001 "rotma\t$rT, $rA, $rB", RotateShift,
3002 [/* see patterns below - $rB must be negated */]>;
3004 def : Pat<(sra R32C:$rA, R32C:$rB),
3005 (ROTMAr32 R32C:$rA, (SFIr32 R32C:$rB, 0))>;
3007 def : Pat<(sra R32C:$rA, R16C:$rB),
3009 (SFIr32 (XSHWr16 R16C:$rB), 0))>;
3011 def : Pat<(sra R32C:$rA, R8C:$rB),
3013 (SFIr32 (XSHWr16 (XSBHr8 R8C:$rB)), 0))>;
3015 class ROTMAIInst<dag OOL, dag IOL, list<dag> pattern>:
3016 RRForm<0b01011110000, OOL, IOL,
3017 "rotmai\t$rT, $rA, $val",
3018 RotateShift, pattern>;
3020 class ROTMAIVecInst<ValueType vectype, Operand intop, ValueType inttype>:
3021 ROTMAIInst<(outs VECREG:$rT), (ins VECREG:$rA, intop:$val),
3022 [(set (vectype VECREG:$rT),
3023 (SPUvec_sra VECREG:$rA, (inttype uimm7:$val)))]>;
3025 class ROTMAIRegInst<RegisterClass rclass, Operand intop, ValueType inttype>:
3026 ROTMAIInst<(outs rclass:$rT), (ins rclass:$rA, intop:$val),
3027 [(set rclass:$rT, (sra rclass:$rA, (inttype uimm7:$val)))]>;
3029 multiclass RotateMaskAlgebraicImm {
3030 def v2i64_i32 : ROTMAIVecInst<v2i64, rotNeg7imm, i32>;
3031 def v4i32_i32 : ROTMAIVecInst<v4i32, rotNeg7imm, i32>;
3032 def r64_i32 : ROTMAIRegInst<R64C, rotNeg7imm, i32>;
3033 def r32_i32 : ROTMAIRegInst<R32C, rotNeg7imm, i32>;
3036 defm ROTMAI : RotateMaskAlgebraicImm;
3038 //===----------------------------------------------------------------------===//
3039 // Branch and conditionals:
3040 //===----------------------------------------------------------------------===//
3042 let isTerminator = 1, isBarrier = 1 in {
3043 // Halt If Equal (r32 preferred slot only, no vector form)
3045 RRForm_3<0b00011011110, (outs), (ins R32C:$rA, R32C:$rB),
3046 "heq\t$rA, $rB", BranchResolv,
3047 [/* no pattern to match */]>;
3050 RI10Form_2<0b11111110, (outs), (ins R32C:$rA, s10imm:$val),
3051 "heqi\t$rA, $val", BranchResolv,
3052 [/* no pattern to match */]>;
3054 // HGT/HGTI: These instructions use signed arithmetic for the comparison,
3055 // contrasting with HLGT/HLGTI, which use unsigned comparison:
3057 RRForm_3<0b00011010010, (outs), (ins R32C:$rA, R32C:$rB),
3058 "hgt\t$rA, $rB", BranchResolv,
3059 [/* no pattern to match */]>;
3062 RI10Form_2<0b11110010, (outs), (ins R32C:$rA, s10imm:$val),
3063 "hgti\t$rA, $val", BranchResolv,
3064 [/* no pattern to match */]>;
3067 RRForm_3<0b00011011010, (outs), (ins R32C:$rA, R32C:$rB),
3068 "hlgt\t$rA, $rB", BranchResolv,
3069 [/* no pattern to match */]>;
3072 RI10Form_2<0b11111010, (outs), (ins R32C:$rA, s10imm:$val),
3073 "hlgti\t$rA, $val", BranchResolv,
3074 [/* no pattern to match */]>;
3077 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
3078 // Comparison operators for i8, i16 and i32:
3079 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
3081 class CEQBInst<dag OOL, dag IOL, list<dag> pattern> :
3082 RRForm<0b00001011110, OOL, IOL, "ceqb\t$rT, $rA, $rB",
3085 multiclass CmpEqualByte
3088 CEQBInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3089 [(set (v16i8 VECREG:$rT), (seteq (v8i16 VECREG:$rA),
3090 (v8i16 VECREG:$rB)))]>;
3093 CEQBInst<(outs R8C:$rT), (ins R8C:$rA, R8C:$rB),
3094 [(set R8C:$rT, (seteq R8C:$rA, R8C:$rB))]>;
3097 class CEQBIInst<dag OOL, dag IOL, list<dag> pattern> :
3098 RI10Form<0b01111110, OOL, IOL, "ceqbi\t$rT, $rA, $val",
3101 multiclass CmpEqualByteImm
3104 CEQBIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm_i8:$val),
3105 [(set (v16i8 VECREG:$rT), (seteq (v16i8 VECREG:$rA),
3106 v16i8SExt8Imm:$val))]>;
3108 CEQBIInst<(outs R8C:$rT), (ins R8C:$rA, s10imm_i8:$val),
3109 [(set R8C:$rT, (seteq R8C:$rA, immSExt8:$val))]>;
3112 class CEQHInst<dag OOL, dag IOL, list<dag> pattern> :
3113 RRForm<0b00010011110, OOL, IOL, "ceqh\t$rT, $rA, $rB",
3116 multiclass CmpEqualHalfword
3118 def v8i16 : CEQHInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3119 [(set (v8i16 VECREG:$rT), (seteq (v8i16 VECREG:$rA),
3120 (v8i16 VECREG:$rB)))]>;
3122 def r16 : CEQHInst<(outs R16C:$rT), (ins R16C:$rA, R16C:$rB),
3123 [(set R16C:$rT, (seteq R16C:$rA, R16C:$rB))]>;
3126 class CEQHIInst<dag OOL, dag IOL, list<dag> pattern> :
3127 RI10Form<0b10111110, OOL, IOL, "ceqhi\t$rT, $rA, $val",
3130 multiclass CmpEqualHalfwordImm
3132 def v8i16 : CEQHIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
3133 [(set (v8i16 VECREG:$rT),
3134 (seteq (v8i16 VECREG:$rA),
3135 (v8i16 v8i16SExt10Imm:$val)))]>;
3136 def r16 : CEQHIInst<(outs R16C:$rT), (ins R16C:$rA, s10imm:$val),
3137 [(set R16C:$rT, (seteq R16C:$rA, i16ImmSExt10:$val))]>;
3140 class CEQInst<dag OOL, dag IOL, list<dag> pattern> :
3141 RRForm<0b00000011110, OOL, IOL, "ceq\t$rT, $rA, $rB",
3144 multiclass CmpEqualWord
3146 def v4i32 : CEQInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3147 [(set (v4i32 VECREG:$rT),
3148 (seteq (v4i32 VECREG:$rA), (v4i32 VECREG:$rB)))]>;
3150 def r32 : CEQInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
3151 [(set R32C:$rT, (seteq R32C:$rA, R32C:$rB))]>;
3154 class CEQIInst<dag OOL, dag IOL, list<dag> pattern> :
3155 RI10Form<0b00111110, OOL, IOL, "ceqi\t$rT, $rA, $val",
3158 multiclass CmpEqualWordImm
3160 def v4i32 : CEQIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
3161 [(set (v4i32 VECREG:$rT),
3162 (seteq (v4i32 VECREG:$rA),
3163 (v4i32 v4i32SExt16Imm:$val)))]>;
3165 def r32: CEQIInst<(outs R32C:$rT), (ins R32C:$rA, s10imm_i32:$val),
3166 [(set R32C:$rT, (seteq R32C:$rA, i32ImmSExt10:$val))]>;
3169 class CGTBInst<dag OOL, dag IOL, list<dag> pattern> :
3170 RRForm<0b00001010010, OOL, IOL, "cgtb\t$rT, $rA, $rB",
3173 multiclass CmpGtrByte
3176 CGTBInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3177 [(set (v16i8 VECREG:$rT), (setgt (v8i16 VECREG:$rA),
3178 (v8i16 VECREG:$rB)))]>;
3181 CGTBInst<(outs R8C:$rT), (ins R8C:$rA, R8C:$rB),
3182 [(set R8C:$rT, (setgt R8C:$rA, R8C:$rB))]>;
3185 class CGTBIInst<dag OOL, dag IOL, list<dag> pattern> :
3186 RI10Form<0b01110010, OOL, IOL, "cgtbi\t$rT, $rA, $val",
3189 multiclass CmpGtrByteImm
3192 CGTBIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm_i8:$val),
3193 [(set (v16i8 VECREG:$rT), (setgt (v16i8 VECREG:$rA),
3194 v16i8SExt8Imm:$val))]>;
3196 CGTBIInst<(outs R8C:$rT), (ins R8C:$rA, s10imm_i8:$val),
3197 [(set R8C:$rT, (setgt R8C:$rA, immSExt8:$val))]>;
3200 class CGTHInst<dag OOL, dag IOL, list<dag> pattern> :
3201 RRForm<0b00010010010, OOL, IOL, "cgth\t$rT, $rA, $rB",
3204 multiclass CmpGtrHalfword
3206 def v8i16 : CGTHInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3207 [(set (v8i16 VECREG:$rT), (setgt (v8i16 VECREG:$rA),
3208 (v8i16 VECREG:$rB)))]>;
3210 def r16 : CGTHInst<(outs R16C:$rT), (ins R16C:$rA, R16C:$rB),
3211 [(set R16C:$rT, (setgt R16C:$rA, R16C:$rB))]>;
3214 class CGTHIInst<dag OOL, dag IOL, list<dag> pattern> :
3215 RI10Form<0b10110010, OOL, IOL, "cgthi\t$rT, $rA, $val",
3218 multiclass CmpGtrHalfwordImm
3220 def v8i16 : CGTHIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
3221 [(set (v8i16 VECREG:$rT),
3222 (setgt (v8i16 VECREG:$rA),
3223 (v8i16 v8i16SExt10Imm:$val)))]>;
3224 def r16 : CGTHIInst<(outs R16C:$rT), (ins R16C:$rA, s10imm:$val),
3225 [(set R16C:$rT, (setgt R16C:$rA, i16ImmSExt10:$val))]>;
3228 class CGTInst<dag OOL, dag IOL, list<dag> pattern> :
3229 RRForm<0b00000010010, OOL, IOL, "cgt\t$rT, $rA, $rB",
3232 multiclass CmpGtrWord
3234 def v4i32 : CGTInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3235 [(set (v4i32 VECREG:$rT),
3236 (setgt (v4i32 VECREG:$rA), (v4i32 VECREG:$rB)))]>;
3238 def r32 : CGTInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
3239 [(set R32C:$rT, (setgt R32C:$rA, R32C:$rB))]>;
3242 class CGTIInst<dag OOL, dag IOL, list<dag> pattern> :
3243 RI10Form<0b00110010, OOL, IOL, "cgti\t$rT, $rA, $val",
3246 multiclass CmpGtrWordImm
3248 def v4i32 : CGTIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
3249 [(set (v4i32 VECREG:$rT),
3250 (setgt (v4i32 VECREG:$rA),
3251 (v4i32 v4i32SExt16Imm:$val)))]>;
3253 def r32: CGTIInst<(outs R32C:$rT), (ins R32C:$rA, s10imm_i32:$val),
3254 [(set R32C:$rT, (setgt R32C:$rA, i32ImmSExt10:$val))]>;
3256 // CGTIv4f32, CGTIf32: These are used in the f32 fdiv instruction sequence:
3257 def v4f32: CGTIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
3258 [(set (v4i32 VECREG:$rT),
3259 (setgt (v4i32 (bitconvert (v4f32 VECREG:$rA))),
3260 (v4i32 v4i32SExt16Imm:$val)))]>;
3262 def f32: CGTIInst<(outs R32C:$rT), (ins R32FP:$rA, s10imm_i32:$val),
3263 [/* no pattern */]>;
3266 class CLGTBInst<dag OOL, dag IOL, list<dag> pattern> :
3267 RRForm<0b00001011010, OOL, IOL, "clgtb\t$rT, $rA, $rB",
3270 multiclass CmpLGtrByte
3273 CLGTBInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3274 [(set (v16i8 VECREG:$rT), (setugt (v8i16 VECREG:$rA),
3275 (v8i16 VECREG:$rB)))]>;
3278 CLGTBInst<(outs R8C:$rT), (ins R8C:$rA, R8C:$rB),
3279 [(set R8C:$rT, (setugt R8C:$rA, R8C:$rB))]>;
3282 class CLGTBIInst<dag OOL, dag IOL, list<dag> pattern> :
3283 RI10Form<0b01111010, OOL, IOL, "clgtbi\t$rT, $rA, $val",
3286 multiclass CmpLGtrByteImm
3289 CLGTBIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm_i8:$val),
3290 [(set (v16i8 VECREG:$rT), (setugt (v16i8 VECREG:$rA),
3291 v16i8SExt8Imm:$val))]>;
3293 CLGTBIInst<(outs R8C:$rT), (ins R8C:$rA, s10imm_i8:$val),
3294 [(set R8C:$rT, (setugt R8C:$rA, immSExt8:$val))]>;
3297 class CLGTHInst<dag OOL, dag IOL, list<dag> pattern> :
3298 RRForm<0b00010011010, OOL, IOL, "clgth\t$rT, $rA, $rB",
3301 multiclass CmpLGtrHalfword
3303 def v8i16 : CLGTHInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3304 [(set (v8i16 VECREG:$rT), (setugt (v8i16 VECREG:$rA),
3305 (v8i16 VECREG:$rB)))]>;
3307 def r16 : CLGTHInst<(outs R16C:$rT), (ins R16C:$rA, R16C:$rB),
3308 [(set R16C:$rT, (setugt R16C:$rA, R16C:$rB))]>;
3311 class CLGTHIInst<dag OOL, dag IOL, list<dag> pattern> :
3312 RI10Form<0b10111010, OOL, IOL, "clgthi\t$rT, $rA, $val",
3315 multiclass CmpLGtrHalfwordImm
3317 def v8i16 : CLGTHIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
3318 [(set (v8i16 VECREG:$rT),
3319 (setugt (v8i16 VECREG:$rA),
3320 (v8i16 v8i16SExt10Imm:$val)))]>;
3321 def r16 : CLGTHIInst<(outs R16C:$rT), (ins R16C:$rA, s10imm:$val),
3322 [(set R16C:$rT, (setugt R16C:$rA, i16ImmSExt10:$val))]>;
3325 class CLGTInst<dag OOL, dag IOL, list<dag> pattern> :
3326 RRForm<0b00000011010, OOL, IOL, "clgt\t$rT, $rA, $rB",
3329 multiclass CmpLGtrWord
3331 def v4i32 : CLGTInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3332 [(set (v4i32 VECREG:$rT),
3333 (setugt (v4i32 VECREG:$rA), (v4i32 VECREG:$rB)))]>;
3335 def r32 : CLGTInst<(outs R32C:$rT), (ins R32C:$rA, R32C:$rB),
3336 [(set R32C:$rT, (setugt R32C:$rA, R32C:$rB))]>;
3339 class CLGTIInst<dag OOL, dag IOL, list<dag> pattern> :
3340 RI10Form<0b00111010, OOL, IOL, "clgti\t$rT, $rA, $val",
3343 multiclass CmpLGtrWordImm
3345 def v4i32 : CLGTIInst<(outs VECREG:$rT), (ins VECREG:$rA, s10imm:$val),
3346 [(set (v4i32 VECREG:$rT),
3347 (setugt (v4i32 VECREG:$rA),
3348 (v4i32 v4i32SExt16Imm:$val)))]>;
3350 def r32: CLGTIInst<(outs R32C:$rT), (ins R32C:$rA, s10imm_i32:$val),
3351 [(set R32C:$rT, (setugt R32C:$rA, i32ImmSExt10:$val))]>;
3354 defm CEQB : CmpEqualByte;
3355 defm CEQBI : CmpEqualByteImm;
3356 defm CEQH : CmpEqualHalfword;
3357 defm CEQHI : CmpEqualHalfwordImm;
3358 defm CEQ : CmpEqualWord;
3359 defm CEQI : CmpEqualWordImm;
3360 defm CGTB : CmpGtrByte;
3361 defm CGTBI : CmpGtrByteImm;
3362 defm CGTH : CmpGtrHalfword;
3363 defm CGTHI : CmpGtrHalfwordImm;
3364 defm CGT : CmpGtrWord;
3365 defm CGTI : CmpGtrWordImm;
3366 defm CLGTB : CmpLGtrByte;
3367 defm CLGTBI : CmpLGtrByteImm;
3368 defm CLGTH : CmpLGtrHalfword;
3369 defm CLGTHI : CmpLGtrHalfwordImm;
3370 defm CLGT : CmpLGtrWord;
3371 defm CLGTI : CmpLGtrWordImm;
3373 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
3374 // For SETCC primitives not supported above (setlt, setle, setge, etc.)
3375 // define a pattern to generate the right code, as a binary operator
3376 // (in a manner of speaking.)
3379 // 1. This only matches the setcc set of conditionals. Special pattern
3380 // matching is used for select conditionals.
3382 // 2. The "DAG" versions of these classes is almost exclusively used for
3383 // i64 comparisons. See the tblgen fundamentals documentation for what
3384 // ".ResultInstrs[0]" means; see TargetSelectionDAG.td and the Pattern
3385 // class for where ResultInstrs originates.
3386 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
3388 class SETCCNegCondReg<PatFrag cond, RegisterClass rclass, ValueType inttype,
3389 SPUInstr xorinst, SPUInstr cmpare>:
3390 Pat<(cond rclass:$rA, rclass:$rB),
3391 (xorinst (cmpare rclass:$rA, rclass:$rB), (inttype -1))>;
3393 class SETCCNegCondImm<PatFrag cond, RegisterClass rclass, ValueType inttype,
3394 PatLeaf immpred, SPUInstr xorinst, SPUInstr cmpare>:
3395 Pat<(cond rclass:$rA, (inttype immpred:$imm)),
3396 (xorinst (cmpare rclass:$rA, (inttype immpred:$imm)), (inttype -1))>;
3398 def : SETCCNegCondReg<setne, R8C, i8, XORBIr8, CEQBr8>;
3399 def : SETCCNegCondImm<setne, R8C, i8, immSExt8, XORBIr8, CEQBIr8>;
3401 def : SETCCNegCondReg<setne, R16C, i16, XORHIr16, CEQHr16>;
3402 def : SETCCNegCondImm<setne, R16C, i16, i16ImmSExt10, XORHIr16, CEQHIr16>;
3404 def : SETCCNegCondReg<setne, R32C, i32, XORIr32, CEQr32>;
3405 def : SETCCNegCondImm<setne, R32C, i32, i32ImmSExt10, XORIr32, CEQIr32>;
3407 class SETCCBinOpReg<PatFrag cond, RegisterClass rclass,
3408 SPUInstr binop, SPUInstr cmpOp1, SPUInstr cmpOp2>:
3409 Pat<(cond rclass:$rA, rclass:$rB),
3410 (binop (cmpOp1 rclass:$rA, rclass:$rB),
3411 (cmpOp2 rclass:$rA, rclass:$rB))>;
3413 class SETCCBinOpImm<PatFrag cond, RegisterClass rclass, PatLeaf immpred,
3415 SPUInstr binop, SPUInstr cmpOp1, SPUInstr cmpOp2>:
3416 Pat<(cond rclass:$rA, (immtype immpred:$imm)),
3417 (binop (cmpOp1 rclass:$rA, (immtype immpred:$imm)),
3418 (cmpOp2 rclass:$rA, (immtype immpred:$imm)))>;
3420 def : SETCCBinOpReg<setge, R8C, ORr8, CGTBr8, CEQBr8>;
3421 def : SETCCBinOpImm<setge, R8C, immSExt8, i8, ORr8, CGTBIr8, CEQBIr8>;
3422 def : SETCCBinOpReg<setlt, R8C, NORr8, CGTBr8, CEQBr8>;
3423 def : SETCCBinOpImm<setlt, R8C, immSExt8, i8, NORr8, CGTBIr8, CEQBIr8>;
3424 def : Pat<(setle R8C:$rA, R8C:$rB),
3425 (XORBIr8 (CGTBr8 R8C:$rA, R8C:$rB), 0xff)>;
3426 def : Pat<(setle R8C:$rA, immU8:$imm),
3427 (XORBIr8 (CGTBIr8 R8C:$rA, immU8:$imm), 0xff)>;
3429 def : SETCCBinOpReg<setge, R16C, ORr16, CGTHr16, CEQHr16>;
3430 def : SETCCBinOpImm<setge, R16C, i16ImmSExt10, i16,
3431 ORr16, CGTHIr16, CEQHIr16>;
3432 def : SETCCBinOpReg<setlt, R16C, NORr16, CGTHr16, CEQHr16>;
3433 def : SETCCBinOpImm<setlt, R16C, i16ImmSExt10, i16, NORr16, CGTHIr16, CEQHIr16>;
3434 def : Pat<(setle R16C:$rA, R16C:$rB),
3435 (XORHIr16 (CGTHr16 R16C:$rA, R16C:$rB), 0xffff)>;
3436 def : Pat<(setle R16C:$rA, i16ImmSExt10:$imm),
3437 (XORHIr16 (CGTHIr16 R16C:$rA, i16ImmSExt10:$imm), 0xffff)>;
3439 def : SETCCBinOpReg<setge, R32C, ORr32, CGTr32, CEQr32>;
3440 def : SETCCBinOpImm<setge, R32C, i32ImmSExt10, i32,
3441 ORr32, CGTIr32, CEQIr32>;
3442 def : SETCCBinOpReg<setlt, R32C, NORr32, CGTr32, CEQr32>;
3443 def : SETCCBinOpImm<setlt, R32C, i32ImmSExt10, i32, NORr32, CGTIr32, CEQIr32>;
3444 def : Pat<(setle R32C:$rA, R32C:$rB),
3445 (XORIr32 (CGTr32 R32C:$rA, R32C:$rB), 0xffffffff)>;
3446 def : Pat<(setle R32C:$rA, i32ImmSExt10:$imm),
3447 (XORIr32 (CGTIr32 R32C:$rA, i32ImmSExt10:$imm), 0xffffffff)>;
3449 def : SETCCBinOpReg<setuge, R8C, ORr8, CLGTBr8, CEQBr8>;
3450 def : SETCCBinOpImm<setuge, R8C, immSExt8, i8, ORr8, CLGTBIr8, CEQBIr8>;
3451 def : SETCCBinOpReg<setult, R8C, NORr8, CLGTBr8, CEQBr8>;
3452 def : SETCCBinOpImm<setult, R8C, immSExt8, i8, NORr8, CLGTBIr8, CEQBIr8>;
3453 def : Pat<(setule R8C:$rA, R8C:$rB),
3454 (XORBIr8 (CLGTBr8 R8C:$rA, R8C:$rB), 0xff)>;
3455 def : Pat<(setule R8C:$rA, immU8:$imm),
3456 (XORBIr8 (CLGTBIr8 R8C:$rA, immU8:$imm), 0xff)>;
3458 def : SETCCBinOpReg<setuge, R16C, ORr16, CLGTHr16, CEQHr16>;
3459 def : SETCCBinOpImm<setuge, R16C, i16ImmSExt10, i16,
3460 ORr16, CLGTHIr16, CEQHIr16>;
3461 def : SETCCBinOpReg<setult, R16C, NORr16, CLGTHr16, CEQHr16>;
3462 def : SETCCBinOpImm<setult, R16C, i16ImmSExt10, i16, NORr16,
3463 CLGTHIr16, CEQHIr16>;
3464 def : Pat<(setule R16C:$rA, R16C:$rB),
3465 (XORHIr16 (CLGTHr16 R16C:$rA, R16C:$rB), 0xffff)>;
3466 def : Pat<(setule R16C:$rA, i16ImmSExt10:$imm),
3467 (XORHIr16 (CLGTHIr16 R16C:$rA, i16ImmSExt10:$imm), 0xffff)>;
3469 def : SETCCBinOpReg<setuge, R32C, ORr32, CLGTr32, CEQr32>;
3470 def : SETCCBinOpImm<setuge, R32C, i32ImmSExt10, i32,
3471 ORr32, CLGTIr32, CEQIr32>;
3472 def : SETCCBinOpReg<setult, R32C, NORr32, CLGTr32, CEQr32>;
3473 def : SETCCBinOpImm<setult, R32C, i32ImmSExt10, i32, NORr32, CLGTIr32, CEQIr32>;
3474 def : Pat<(setule R32C:$rA, R32C:$rB),
3475 (XORIr32 (CLGTr32 R32C:$rA, R32C:$rB), 0xffffffff)>;
3476 def : Pat<(setule R32C:$rA, i32ImmSExt10:$imm),
3477 (XORIr32 (CLGTIr32 R32C:$rA, i32ImmSExt10:$imm), 0xffffffff)>;
3479 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
3480 // select conditional patterns:
3481 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
3483 class SELECTNegCondReg<PatFrag cond, RegisterClass rclass, ValueType inttype,
3484 SPUInstr selinstr, SPUInstr cmpare>:
3485 Pat<(select (inttype (cond rclass:$rA, rclass:$rB)),
3486 rclass:$rTrue, rclass:$rFalse),
3487 (selinstr rclass:$rTrue, rclass:$rFalse,
3488 (cmpare rclass:$rA, rclass:$rB))>;
3490 class SELECTNegCondImm<PatFrag cond, RegisterClass rclass, ValueType inttype,
3491 PatLeaf immpred, SPUInstr selinstr, SPUInstr cmpare>:
3492 Pat<(select (inttype (cond rclass:$rA, immpred:$imm)),
3493 rclass:$rTrue, rclass:$rFalse),
3494 (selinstr rclass:$rTrue, rclass:$rFalse,
3495 (cmpare rclass:$rA, immpred:$imm))>;
3497 def : SELECTNegCondReg<setne, R8C, i8, SELBr8, CEQBr8>;
3498 def : SELECTNegCondImm<setne, R8C, i8, immSExt8, SELBr8, CEQBIr8>;
3499 def : SELECTNegCondReg<setle, R8C, i8, SELBr8, CGTBr8>;
3500 def : SELECTNegCondImm<setle, R8C, i8, immSExt8, SELBr8, CGTBr8>;
3501 def : SELECTNegCondReg<setule, R8C, i8, SELBr8, CLGTBr8>;
3502 def : SELECTNegCondImm<setule, R8C, i8, immU8, SELBr8, CLGTBIr8>;
3504 def : SELECTNegCondReg<setne, R16C, i16, SELBr16, CEQHr16>;
3505 def : SELECTNegCondImm<setne, R16C, i16, i16ImmSExt10, SELBr16, CEQHIr16>;
3506 def : SELECTNegCondReg<setle, R16C, i16, SELBr16, CGTHr16>;
3507 def : SELECTNegCondImm<setle, R16C, i16, i16ImmSExt10, SELBr16, CGTHIr16>;
3508 def : SELECTNegCondReg<setule, R16C, i16, SELBr16, CLGTHr16>;
3509 def : SELECTNegCondImm<setule, R16C, i16, i16ImmSExt10, SELBr16, CLGTHIr16>;
3511 def : SELECTNegCondReg<setne, R32C, i32, SELBr32, CEQr32>;
3512 def : SELECTNegCondImm<setne, R32C, i32, i32ImmSExt10, SELBr32, CEQIr32>;
3513 def : SELECTNegCondReg<setle, R32C, i32, SELBr32, CGTr32>;
3514 def : SELECTNegCondImm<setle, R32C, i32, i32ImmSExt10, SELBr32, CGTIr32>;
3515 def : SELECTNegCondReg<setule, R32C, i32, SELBr32, CLGTr32>;
3516 def : SELECTNegCondImm<setule, R32C, i32, i32ImmSExt10, SELBr32, CLGTIr32>;
3518 class SELECTBinOpReg<PatFrag cond, RegisterClass rclass, ValueType inttype,
3519 SPUInstr selinstr, SPUInstr binop, SPUInstr cmpOp1,
3521 Pat<(select (inttype (cond rclass:$rA, rclass:$rB)),
3522 rclass:$rTrue, rclass:$rFalse),
3523 (selinstr rclass:$rFalse, rclass:$rTrue,
3524 (binop (cmpOp1 rclass:$rA, rclass:$rB),
3525 (cmpOp2 rclass:$rA, rclass:$rB)))>;
3527 class SELECTBinOpImm<PatFrag cond, RegisterClass rclass, PatLeaf immpred,
3529 SPUInstr selinstr, SPUInstr binop, SPUInstr cmpOp1,
3531 Pat<(select (inttype (cond rclass:$rA, (inttype immpred:$imm))),
3532 rclass:$rTrue, rclass:$rFalse),
3533 (selinstr rclass:$rFalse, rclass:$rTrue,
3534 (binop (cmpOp1 rclass:$rA, (inttype immpred:$imm)),
3535 (cmpOp2 rclass:$rA, (inttype immpred:$imm))))>;
3537 def : SELECTBinOpReg<setge, R8C, i8, SELBr8, ORr8, CGTBr8, CEQBr8>;
3538 def : SELECTBinOpImm<setge, R8C, immSExt8, i8,
3539 SELBr8, ORr8, CGTBIr8, CEQBIr8>;
3541 def : SELECTBinOpReg<setge, R16C, i16, SELBr16, ORr16, CGTHr16, CEQHr16>;
3542 def : SELECTBinOpImm<setge, R16C, i16ImmSExt10, i16,
3543 SELBr16, ORr16, CGTHIr16, CEQHIr16>;
3545 def : SELECTBinOpReg<setge, R32C, i32, SELBr32, ORr32, CGTr32, CEQr32>;
3546 def : SELECTBinOpImm<setge, R32C, i32ImmSExt10, i32,
3547 SELBr32, ORr32, CGTIr32, CEQIr32>;
3549 def : SELECTBinOpReg<setuge, R8C, i8, SELBr8, ORr8, CLGTBr8, CEQBr8>;
3550 def : SELECTBinOpImm<setuge, R8C, immSExt8, i8,
3551 SELBr8, ORr8, CLGTBIr8, CEQBIr8>;
3553 def : SELECTBinOpReg<setuge, R16C, i16, SELBr16, ORr16, CLGTHr16, CEQHr16>;
3554 def : SELECTBinOpImm<setuge, R16C, i16ImmUns10, i16,
3555 SELBr16, ORr16, CLGTHIr16, CEQHIr16>;
3557 def : SELECTBinOpReg<setuge, R32C, i32, SELBr32, ORr32, CLGTr32, CEQr32>;
3558 def : SELECTBinOpImm<setuge, R32C, i32ImmUns10, i32,
3559 SELBr32, ORr32, CLGTIr32, CEQIr32>;
3561 //-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
3564 // All calls clobber the non-callee-saved registers:
3565 Defs = [R0, R1, R2, R3, R4, R5, R6, R7, R8, R9,
3566 R10,R11,R12,R13,R14,R15,R16,R17,R18,R19,
3567 R20,R21,R22,R23,R24,R25,R26,R27,R28,R29,
3568 R30,R31,R32,R33,R34,R35,R36,R37,R38,R39,
3569 R40,R41,R42,R43,R44,R45,R46,R47,R48,R49,
3570 R50,R51,R52,R53,R54,R55,R56,R57,R58,R59,
3571 R60,R61,R62,R63,R64,R65,R66,R67,R68,R69,
3572 R70,R71,R72,R73,R74,R75,R76,R77,R78,R79],
3573 // All of these instructions use $lr (aka $0)
3575 // Branch relative and set link: Used if we actually know that the target
3576 // is within [-32768, 32767] bytes of the target
3578 BranchSetLink<0b011001100, (outs), (ins relcalltarget:$func, variable_ops),
3579 "brsl\t$$lr, $func",
3580 [(SPUcall (SPUpcrel tglobaladdr:$func, 0))]>;
3582 // Branch absolute and set link: Used if we actually know that the target
3583 // is an absolute address
3585 BranchSetLink<0b011001100, (outs), (ins calltarget:$func, variable_ops),
3586 "brasl\t$$lr, $func",
3587 [(SPUcall (SPUaform tglobaladdr:$func, 0))]>;
3589 // Branch indirect and set link if external data. These instructions are not
3590 // actually generated, matched by an intrinsic:
3591 def BISLED_00: BISLEDForm<0b11, "bisled\t$$lr, $func", [/* empty pattern */]>;
3592 def BISLED_E0: BISLEDForm<0b10, "bisled\t$$lr, $func", [/* empty pattern */]>;
3593 def BISLED_0D: BISLEDForm<0b01, "bisled\t$$lr, $func", [/* empty pattern */]>;
3594 def BISLED_ED: BISLEDForm<0b00, "bisled\t$$lr, $func", [/* empty pattern */]>;
3596 // Branch indirect and set link. This is the "X-form" address version of a
3599 BIForm<0b10010101100, "bisl\t$$lr, $func", [(SPUcall R32C:$func)]>;
3602 // Support calls to external symbols:
3603 def : Pat<(SPUcall (SPUpcrel texternalsym:$func, 0)),
3604 (BRSL texternalsym:$func)>;
3606 def : Pat<(SPUcall (SPUaform texternalsym:$func, 0)),
3607 (BRASL texternalsym:$func)>;
3609 // Unconditional branches:
3610 let isBranch = 1, isTerminator = 1, hasCtrlDep = 1, isBarrier = 1 in {
3612 UncondBranch<0b001001100, (outs), (ins brtarget:$dest),
3616 // Unconditional, absolute address branch
3618 UncondBranch<0b001100000, (outs), (ins brtarget:$dest),
3620 [/* no pattern */]>;
3624 BIForm<0b00010101100, "bi\t$func", [(brind R32C:$func)]>;
3626 // Conditional branches:
3627 class BRNZInst<dag IOL, list<dag> pattern>:
3628 RI16Form<0b010000100, (outs), IOL, "brnz\t$rCond,$dest",
3629 BranchResolv, pattern>;
3631 class BRNZRegInst<RegisterClass rclass>:
3632 BRNZInst<(ins rclass:$rCond, brtarget:$dest),
3633 [(brcond rclass:$rCond, bb:$dest)]>;
3635 class BRNZVecInst<ValueType vectype>:
3636 BRNZInst<(ins VECREG:$rCond, brtarget:$dest),
3637 [(brcond (vectype VECREG:$rCond), bb:$dest)]>;
3639 multiclass BranchNotZero {
3640 def v4i32 : BRNZVecInst<v4i32>;
3641 def r32 : BRNZRegInst<R32C>;
3644 defm BRNZ : BranchNotZero;
3646 class BRZInst<dag IOL, list<dag> pattern>:
3647 RI16Form<0b000000100, (outs), IOL, "brz\t$rT,$dest",
3648 BranchResolv, pattern>;
3650 class BRZRegInst<RegisterClass rclass>:
3651 BRZInst<(ins rclass:$rT, brtarget:$dest), [/* no pattern */]>;
3653 class BRZVecInst<ValueType vectype>:
3654 BRZInst<(ins VECREG:$rT, brtarget:$dest), [/* no pattern */]>;
3656 multiclass BranchZero {
3657 def v4i32: BRZVecInst<v4i32>;
3658 def r32: BRZRegInst<R32C>;
3661 defm BRZ: BranchZero;
3663 // Note: LLVM doesn't do branch conditional, indirect. Otherwise these would
3666 class BINZInst<dag IOL, list<dag> pattern>:
3667 BICondForm<0b10010100100, (outs), IOL, "binz\t$rA, $dest", pattern>;
3669 class BINZRegInst<RegisterClass rclass>:
3670 BINZInst<(ins rclass:$rA, brtarget:$dest),
3671 [(brcond rclass:$rA, R32C:$dest)]>;
3673 class BINZVecInst<ValueType vectype>:
3674 BINZInst<(ins VECREG:$rA, R32C:$dest),
3675 [(brcond (vectype VECREG:$rA), R32C:$dest)]>;
3677 multiclass BranchNotZeroIndirect {
3678 def v4i32: BINZVecInst<v4i32>;
3679 def r32: BINZRegInst<R32C>;
3682 defm BINZ: BranchNotZeroIndirect;
3684 class BIZInst<dag IOL, list<dag> pattern>:
3685 BICondForm<0b00010100100, (outs), IOL, "biz\t$rA, $func", pattern>;
3687 class BIZRegInst<RegisterClass rclass>:
3688 BIZInst<(ins rclass:$rA, R32C:$func), [/* no pattern */]>;
3690 class BIZVecInst<ValueType vectype>:
3691 BIZInst<(ins VECREG:$rA, R32C:$func), [/* no pattern */]>;
3693 multiclass BranchZeroIndirect {
3694 def v4i32: BIZVecInst<v4i32>;
3695 def r32: BIZRegInst<R32C>;
3698 defm BIZ: BranchZeroIndirect;
3701 class BRHNZInst<dag IOL, list<dag> pattern>:
3702 RI16Form<0b011000100, (outs), IOL, "brhnz\t$rCond,$dest", BranchResolv,
3705 class BRHNZRegInst<RegisterClass rclass>:
3706 BRHNZInst<(ins rclass:$rCond, brtarget:$dest),
3707 [(brcond rclass:$rCond, bb:$dest)]>;
3709 class BRHNZVecInst<ValueType vectype>:
3710 BRHNZInst<(ins VECREG:$rCond, brtarget:$dest), [/* no pattern */]>;
3712 multiclass BranchNotZeroHalfword {
3713 def v8i16: BRHNZVecInst<v8i16>;
3714 def r16: BRHNZRegInst<R16C>;
3717 defm BRHNZ: BranchNotZeroHalfword;
3719 class BRHZInst<dag IOL, list<dag> pattern>:
3720 RI16Form<0b001000100, (outs), IOL, "brhz\t$rT,$dest", BranchResolv,
3723 class BRHZRegInst<RegisterClass rclass>:
3724 BRHZInst<(ins rclass:$rT, brtarget:$dest), [/* no pattern */]>;
3726 class BRHZVecInst<ValueType vectype>:
3727 BRHZInst<(ins VECREG:$rT, brtarget:$dest), [/* no pattern */]>;
3729 multiclass BranchZeroHalfword {
3730 def v8i16: BRHZVecInst<v8i16>;
3731 def r16: BRHZRegInst<R16C>;
3734 defm BRHZ: BranchZeroHalfword;
3737 //===----------------------------------------------------------------------===//
3738 // setcc and brcond patterns:
3739 //===----------------------------------------------------------------------===//
3741 def : Pat<(brcond (i16 (seteq R16C:$rA, 0)), bb:$dest),
3742 (BRHZr16 R16C:$rA, bb:$dest)>;
3743 def : Pat<(brcond (i16 (setne R16C:$rA, 0)), bb:$dest),
3744 (BRHNZr16 R16C:$rA, bb:$dest)>;
3746 def : Pat<(brcond (i32 (seteq R32C:$rA, 0)), bb:$dest),
3747 (BRZr32 R32C:$rA, bb:$dest)>;
3748 def : Pat<(brcond (i32 (setne R32C:$rA, 0)), bb:$dest),
3749 (BRNZr32 R32C:$rA, bb:$dest)>;
3751 multiclass BranchCondEQ<PatFrag cond, SPUInstr brinst16, SPUInstr brinst32>
3753 def r16imm: Pat<(brcond (i16 (cond R16C:$rA, i16ImmSExt10:$val)), bb:$dest),
3754 (brinst16 (CEQHIr16 R16C:$rA, i16ImmSExt10:$val), bb:$dest)>;
3756 def r16 : Pat<(brcond (i16 (cond R16C:$rA, R16C:$rB)), bb:$dest),
3757 (brinst16 (CEQHr16 R16C:$rA, R16:$rB), bb:$dest)>;
3759 def r32imm : Pat<(brcond (i32 (cond R32C:$rA, i32ImmSExt10:$val)), bb:$dest),
3760 (brinst32 (CEQIr32 R32C:$rA, i32ImmSExt10:$val), bb:$dest)>;
3762 def r32 : Pat<(brcond (i32 (cond R32C:$rA, R32C:$rB)), bb:$dest),
3763 (brinst32 (CEQr32 R32C:$rA, R32C:$rB), bb:$dest)>;
3766 defm BRCONDeq : BranchCondEQ<seteq, BRHNZr16, BRNZr32>;
3767 defm BRCONDne : BranchCondEQ<setne, BRHZr16, BRZr32>;
3769 multiclass BranchCondLGT<PatFrag cond, SPUInstr brinst16, SPUInstr brinst32>
3771 def r16imm : Pat<(brcond (i16 (cond R16C:$rA, i16ImmSExt10:$val)), bb:$dest),
3772 (brinst16 (CLGTHIr16 R16C:$rA, i16ImmSExt10:$val), bb:$dest)>;
3774 def r16 : Pat<(brcond (i16 (cond R16C:$rA, R16C:$rB)), bb:$dest),
3775 (brinst16 (CLGTHr16 R16C:$rA, R16:$rB), bb:$dest)>;
3777 def r32imm : Pat<(brcond (i32 (cond R32C:$rA, i32ImmSExt10:$val)), bb:$dest),
3778 (brinst32 (CLGTIr32 R32C:$rA, i32ImmSExt10:$val), bb:$dest)>;
3780 def r32 : Pat<(brcond (i32 (cond R32C:$rA, R32C:$rB)), bb:$dest),
3781 (brinst32 (CLGTr32 R32C:$rA, R32C:$rB), bb:$dest)>;
3784 defm BRCONDugt : BranchCondLGT<setugt, BRHNZr16, BRNZr32>;
3785 defm BRCONDule : BranchCondLGT<setule, BRHZr16, BRZr32>;
3787 multiclass BranchCondLGTEQ<PatFrag cond, SPUInstr orinst16, SPUInstr brinst16,
3788 SPUInstr orinst32, SPUInstr brinst32>
3790 def r16imm: Pat<(brcond (i16 (cond R16C:$rA, i16ImmSExt10:$val)), bb:$dest),
3791 (brinst16 (orinst16 (CLGTHIr16 R16C:$rA, i16ImmSExt10:$val),
3792 (CEQHIr16 R16C:$rA, i16ImmSExt10:$val)),
3795 def r16: Pat<(brcond (i16 (cond R16C:$rA, R16C:$rB)), bb:$dest),
3796 (brinst16 (orinst16 (CLGTHr16 R16C:$rA, R16:$rB),
3797 (CEQHr16 R16C:$rA, R16:$rB)),
3800 def r32imm : Pat<(brcond (i32 (cond R32C:$rA, i32ImmSExt10:$val)), bb:$dest),
3801 (brinst32 (orinst32 (CLGTIr32 R32C:$rA, i32ImmSExt10:$val),
3802 (CEQIr32 R32C:$rA, i32ImmSExt10:$val)),
3805 def r32 : Pat<(brcond (i32 (cond R32C:$rA, R32C:$rB)), bb:$dest),
3806 (brinst32 (orinst32 (CLGTr32 R32C:$rA, R32C:$rB),
3807 (CEQr32 R32C:$rA, R32C:$rB)),
3811 defm BRCONDuge : BranchCondLGTEQ<setuge, ORr16, BRHNZr16, ORr32, BRNZr32>;
3812 defm BRCONDult : BranchCondLGTEQ<setult, ORr16, BRHZr16, ORr32, BRZr32>;
3814 multiclass BranchCondGT<PatFrag cond, SPUInstr brinst16, SPUInstr brinst32>
3816 def r16imm : Pat<(brcond (i16 (cond R16C:$rA, i16ImmSExt10:$val)), bb:$dest),
3817 (brinst16 (CGTHIr16 R16C:$rA, i16ImmSExt10:$val), bb:$dest)>;
3819 def r16 : Pat<(brcond (i16 (cond R16C:$rA, R16C:$rB)), bb:$dest),
3820 (brinst16 (CGTHr16 R16C:$rA, R16:$rB), bb:$dest)>;
3822 def r32imm : Pat<(brcond (i32 (cond R32C:$rA, i32ImmSExt10:$val)), bb:$dest),
3823 (brinst32 (CGTIr32 R32C:$rA, i32ImmSExt10:$val), bb:$dest)>;
3825 def r32 : Pat<(brcond (i32 (cond R32C:$rA, R32C:$rB)), bb:$dest),
3826 (brinst32 (CGTr32 R32C:$rA, R32C:$rB), bb:$dest)>;
3829 defm BRCONDgt : BranchCondGT<setgt, BRHNZr16, BRNZr32>;
3830 defm BRCONDle : BranchCondGT<setle, BRHZr16, BRZr32>;
3832 multiclass BranchCondGTEQ<PatFrag cond, SPUInstr orinst16, SPUInstr brinst16,
3833 SPUInstr orinst32, SPUInstr brinst32>
3835 def r16imm: Pat<(brcond (i16 (cond R16C:$rA, i16ImmSExt10:$val)), bb:$dest),
3836 (brinst16 (orinst16 (CGTHIr16 R16C:$rA, i16ImmSExt10:$val),
3837 (CEQHIr16 R16C:$rA, i16ImmSExt10:$val)),
3840 def r16: Pat<(brcond (i16 (cond R16C:$rA, R16C:$rB)), bb:$dest),
3841 (brinst16 (orinst16 (CGTHr16 R16C:$rA, R16:$rB),
3842 (CEQHr16 R16C:$rA, R16:$rB)),
3845 def r32imm : Pat<(brcond (i32 (cond R32C:$rA, i32ImmSExt10:$val)), bb:$dest),
3846 (brinst32 (orinst32 (CGTIr32 R32C:$rA, i32ImmSExt10:$val),
3847 (CEQIr32 R32C:$rA, i32ImmSExt10:$val)),
3850 def r32 : Pat<(brcond (i32 (cond R32C:$rA, R32C:$rB)), bb:$dest),
3851 (brinst32 (orinst32 (CGTr32 R32C:$rA, R32C:$rB),
3852 (CEQr32 R32C:$rA, R32C:$rB)),
3856 defm BRCONDge : BranchCondGTEQ<setge, ORr16, BRHNZr16, ORr32, BRNZr32>;
3857 defm BRCONDlt : BranchCondGTEQ<setlt, ORr16, BRHZr16, ORr32, BRZr32>;
3859 let isTerminator = 1, isBarrier = 1 in {
3860 let isReturn = 1 in {
3862 RETForm<"bi\t$$lr", [(retflag)]>;
3866 //===----------------------------------------------------------------------===//
3867 // Single precision floating point instructions
3868 //===----------------------------------------------------------------------===//
3870 class FAInst<dag OOL, dag IOL, list<dag> pattern>:
3871 RRForm<0b01011000100, OOL, IOL, "fa\t$rT, $rA, $rB",
3874 class FAVecInst<ValueType vectype>:
3875 FAInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3876 [(set (vectype VECREG:$rT),
3877 (fadd (vectype VECREG:$rA), (vectype VECREG:$rB)))]>;
3881 def v4f32: FAVecInst<v4f32>;
3882 def f32: FAInst<(outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB),
3883 [(set R32FP:$rT, (fadd R32FP:$rA, R32FP:$rB))]>;
3888 class FSInst<dag OOL, dag IOL, list<dag> pattern>:
3889 RRForm<0b01011000100, OOL, IOL, "fs\t$rT, $rA, $rB",
3892 class FSVecInst<ValueType vectype>:
3893 FSInst<(outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3894 [(set (vectype VECREG:$rT),
3895 (fsub (vectype VECREG:$rA), (vectype VECREG:$rB)))]>;
3899 def v4f32: FSVecInst<v4f32>;
3900 def f32: FSInst<(outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB),
3901 [(set R32FP:$rT, (fsub R32FP:$rA, R32FP:$rB))]>;
3906 // Floating point reciprocal estimate
3908 class FRESTInst<dag OOL, dag IOL>:
3909 RRForm_1<0b00110111000, OOL, IOL,
3910 "frest\t$rT, $rA", SPrecFP,
3911 [/* no pattern */]>;
3914 FRESTInst<(outs VECREG:$rT), (ins VECREG:$rA)>;
3917 FRESTInst<(outs R32FP:$rT), (ins R32FP:$rA)>;
3919 // Floating point interpolate (used in conjunction with reciprocal estimate)
3921 RRForm<0b00101011110, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
3922 "fi\t$rT, $rA, $rB", SPrecFP,
3923 [/* no pattern */]>;
3926 RRForm<0b00101011110, (outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB),
3927 "fi\t$rT, $rA, $rB", SPrecFP,
3928 [/* no pattern */]>;
3930 //--------------------------------------------------------------------------
3931 // Basic single precision floating point comparisons:
3933 // Note: There is no support on SPU for single precision NaN. Consequently,
3934 // ordered and unordered comparisons are the same.
3935 //--------------------------------------------------------------------------
3938 RRForm<0b01000011110, (outs R32C:$rT), (ins R32FP:$rA, R32FP:$rB),
3939 "fceq\t$rT, $rA, $rB", SPrecFP,
3940 [(set R32C:$rT, (setueq R32FP:$rA, R32FP:$rB))]>;
3942 def : Pat<(setoeq R32FP:$rA, R32FP:$rB),
3943 (FCEQf32 R32FP:$rA, R32FP:$rB)>;
3946 RRForm<0b01010011110, (outs R32C:$rT), (ins R32FP:$rA, R32FP:$rB),
3947 "fcmeq\t$rT, $rA, $rB", SPrecFP,
3948 [(set R32C:$rT, (setueq (fabs R32FP:$rA), (fabs R32FP:$rB)))]>;
3950 def : Pat<(setoeq (fabs R32FP:$rA), (fabs R32FP:$rB)),
3951 (FCMEQf32 R32FP:$rA, R32FP:$rB)>;
3954 RRForm<0b01000011010, (outs R32C:$rT), (ins R32FP:$rA, R32FP:$rB),
3955 "fcgt\t$rT, $rA, $rB", SPrecFP,
3956 [(set R32C:$rT, (setugt R32FP:$rA, R32FP:$rB))]>;
3958 def : Pat<(setugt R32FP:$rA, R32FP:$rB),
3959 (FCGTf32 R32FP:$rA, R32FP:$rB)>;
3962 RRForm<0b01010011010, (outs R32C:$rT), (ins R32FP:$rA, R32FP:$rB),
3963 "fcmgt\t$rT, $rA, $rB", SPrecFP,
3964 [(set R32C:$rT, (setugt (fabs R32FP:$rA), (fabs R32FP:$rB)))]>;
3966 def : Pat<(setugt (fabs R32FP:$rA), (fabs R32FP:$rB)),
3967 (FCMGTf32 R32FP:$rA, R32FP:$rB)>;
3969 //--------------------------------------------------------------------------
3970 // Single precision floating point comparisons and SETCC equivalents:
3971 //--------------------------------------------------------------------------
3973 def : SETCCNegCondReg<setune, R32FP, i32, XORIr32, FCEQf32>;
3974 def : SETCCNegCondReg<setone, R32FP, i32, XORIr32, FCEQf32>;
3976 def : SETCCBinOpReg<setuge, R32FP, ORr32, FCGTf32, FCEQf32>;
3977 def : SETCCBinOpReg<setoge, R32FP, ORr32, FCGTf32, FCEQf32>;
3979 def : SETCCBinOpReg<setult, R32FP, NORr32, FCGTf32, FCEQf32>;
3980 def : SETCCBinOpReg<setolt, R32FP, NORr32, FCGTf32, FCEQf32>;
3982 def : Pat<(setule R32FP:$rA, R32FP:$rB),
3983 (XORIr32 (FCGTf32 R32FP:$rA, R32FP:$rB), 0xffffffff)>;
3984 def : Pat<(setole R32FP:$rA, R32FP:$rB),
3985 (XORIr32 (FCGTf32 R32FP:$rA, R32FP:$rB), 0xffffffff)>;
3987 // FP Status and Control Register Write
3988 // Why isn't rT a don't care in the ISA?
3989 // Should we create a special RRForm_3 for this guy and zero out the rT?
3991 RRForm_1<0b01011101110, (outs R32FP:$rT), (ins R32FP:$rA),
3992 "fscrwr\t$rA", SPrecFP,
3993 [/* This instruction requires an intrinsic. Note: rT is unused. */]>;
3995 // FP Status and Control Register Read
3997 RRForm_2<0b01011101110, (outs R32FP:$rT), (ins),
3998 "fscrrd\t$rT", SPrecFP,
3999 [/* This instruction requires an intrinsic */]>;
4001 // llvm instruction space
4002 // How do these map onto cell instructions?
4004 // frest rC rB # c = 1/b (both lines)
4006 // fm rD rA rC # d = a * 1/b
4007 // fnms rB rD rB rA # b = - (d * b - a) --should == 0 in a perfect world
4008 // fma rB rB rC rD # b = b * c + d
4009 // = -(d *b -a) * c + d
4010 // = a * c - c ( a *b *c - a)
4015 // These llvm instructions will actually map to library calls.
4016 // All that's needed, then, is to check that the appropriate library is
4017 // imported and do a brsl to the proper function name.
4018 // frem # fmod(x, y): x - (x/y) * y
4019 // (Note: fmod(double, double), fmodf(float,float)
4023 // Unimplemented SPU instruction space
4024 // floating reciprocal absolute square root estimate (frsqest)
4026 // The following are probably just intrinsics
4027 // status and control register write
4028 // status and control register read
4030 //--------------------------------------
4031 // Floating point multiply instructions
4032 //--------------------------------------
4035 RRForm<0b00100011010, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
4036 "fm\t$rT, $rA, $rB", SPrecFP,
4037 [(set (v4f32 VECREG:$rT), (fmul (v4f32 VECREG:$rA),
4038 (v4f32 VECREG:$rB)))]>;
4041 RRForm<0b01100011010, (outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB),
4042 "fm\t$rT, $rA, $rB", SPrecFP,
4043 [(set R32FP:$rT, (fmul R32FP:$rA, R32FP:$rB))]>;
4045 // Floating point multiply and add
4046 // e.g. d = c + (a * b)
4048 RRRForm<0b0111, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
4049 "fma\t$rT, $rA, $rB, $rC", SPrecFP,
4050 [(set (v4f32 VECREG:$rT),
4051 (fadd (v4f32 VECREG:$rC),
4052 (fmul (v4f32 VECREG:$rA), (v4f32 VECREG:$rB))))]>;
4055 RRRForm<0b0111, (outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB, R32FP:$rC),
4056 "fma\t$rT, $rA, $rB, $rC", SPrecFP,
4057 [(set R32FP:$rT, (fadd R32FP:$rC, (fmul R32FP:$rA, R32FP:$rB)))]>;
4059 // FP multiply and subtract
4060 // Subtracts value in rC from product
4063 RRRForm<0b0111, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
4064 "fms\t$rT, $rA, $rB, $rC", SPrecFP,
4065 [(set (v4f32 VECREG:$rT),
4066 (fsub (fmul (v4f32 VECREG:$rA), (v4f32 VECREG:$rB)),
4067 (v4f32 VECREG:$rC)))]>;
4070 RRRForm<0b0111, (outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB, R32FP:$rC),
4071 "fms\t$rT, $rA, $rB, $rC", SPrecFP,
4073 (fsub (fmul R32FP:$rA, R32FP:$rB), R32FP:$rC))]>;
4075 // Floating Negative Mulitply and Subtract
4076 // Subtracts product from value in rC
4077 // res = fneg(fms a b c)
4080 // NOTE: subtraction order
4084 RRRForm<0b1101, (outs R32FP:$rT), (ins R32FP:$rA, R32FP:$rB, R32FP:$rC),
4085 "fnms\t$rT, $rA, $rB, $rC", SPrecFP,
4086 [(set R32FP:$rT, (fsub R32FP:$rC, (fmul R32FP:$rA, R32FP:$rB)))]>;
4089 RRRForm<0b1101, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
4090 "fnms\t$rT, $rA, $rB, $rC", SPrecFP,
4091 [(set (v4f32 VECREG:$rT),
4092 (fsub (v4f32 VECREG:$rC),
4093 (fmul (v4f32 VECREG:$rA),
4094 (v4f32 VECREG:$rB))))]>;
4096 //--------------------------------------
4097 // Floating Point Conversions
4098 // Signed conversions:
4100 CVTIntFPForm<0b0101101110, (outs VECREG:$rT), (ins VECREG:$rA),
4101 "csflt\t$rT, $rA, 0", SPrecFP,
4102 [(set (v4f32 VECREG:$rT), (sint_to_fp (v4i32 VECREG:$rA)))]>;
4104 // Convert signed integer to floating point
4106 CVTIntFPForm<0b0101101110, (outs R32FP:$rT), (ins R32C:$rA),
4107 "csflt\t$rT, $rA, 0", SPrecFP,
4108 [(set R32FP:$rT, (sint_to_fp R32C:$rA))]>;
4110 // Convert unsigned into to float
4112 CVTIntFPForm<0b1101101110, (outs VECREG:$rT), (ins VECREG:$rA),
4113 "cuflt\t$rT, $rA, 0", SPrecFP,
4114 [(set (v4f32 VECREG:$rT), (uint_to_fp (v4i32 VECREG:$rA)))]>;
4117 CVTIntFPForm<0b1101101110, (outs R32FP:$rT), (ins R32C:$rA),
4118 "cuflt\t$rT, $rA, 0", SPrecFP,
4119 [(set R32FP:$rT, (uint_to_fp R32C:$rA))]>;
4121 // Convert float to unsigned int
4122 // Assume that scale = 0
4125 CVTIntFPForm<0b1101101110, (outs VECREG:$rT), (ins VECREG:$rA),
4126 "cfltu\t$rT, $rA, 0", SPrecFP,
4127 [(set (v4i32 VECREG:$rT), (fp_to_uint (v4f32 VECREG:$rA)))]>;
4130 CVTIntFPForm<0b1101101110, (outs R32C:$rT), (ins R32FP:$rA),
4131 "cfltu\t$rT, $rA, 0", SPrecFP,
4132 [(set R32C:$rT, (fp_to_uint R32FP:$rA))]>;
4134 // Convert float to signed int
4135 // Assume that scale = 0
4138 CVTIntFPForm<0b1101101110, (outs VECREG:$rT), (ins VECREG:$rA),
4139 "cflts\t$rT, $rA, 0", SPrecFP,
4140 [(set (v4i32 VECREG:$rT), (fp_to_sint (v4f32 VECREG:$rA)))]>;
4143 CVTIntFPForm<0b1101101110, (outs R32C:$rT), (ins R32FP:$rA),
4144 "cflts\t$rT, $rA, 0", SPrecFP,
4145 [(set R32C:$rT, (fp_to_sint R32FP:$rA))]>;
4147 //===----------------------------------------------------------------------==//
4148 // Single<->Double precision conversions
4149 //===----------------------------------------------------------------------==//
4151 // NOTE: We use "vec" name suffix here to avoid confusion (e.g. input is a
4152 // v4f32, output is v2f64--which goes in the name?)
4154 // Floating point extend single to double
4155 // NOTE: Not sure if passing in v4f32 to FESDvec is correct since it
4156 // operates on two double-word slots (i.e. 1st and 3rd fp numbers
4159 RRForm_1<0b00011101110, (outs VECREG:$rT), (ins VECREG:$rA),
4160 "fesd\t$rT, $rA", SPrecFP,
4161 [(set (v2f64 VECREG:$rT), (fextend (v4f32 VECREG:$rA)))]>;
4164 RRForm_1<0b00011101110, (outs R64FP:$rT), (ins R32FP:$rA),
4165 "fesd\t$rT, $rA", SPrecFP,
4166 [(set R64FP:$rT, (fextend R32FP:$rA))]>;
4168 // Floating point round double to single
4170 // RRForm_1<0b10011101110, (outs VECREG:$rT), (ins VECREG:$rA),
4171 // "frds\t$rT, $rA,", SPrecFP,
4172 // [(set (v4f32 R32FP:$rT), (fround (v2f64 R64FP:$rA)))]>;
4175 RRForm_1<0b10011101110, (outs R32FP:$rT), (ins R64FP:$rA),
4176 "frds\t$rT, $rA", SPrecFP,
4177 [(set R32FP:$rT, (fround R64FP:$rA))]>;
4179 //ToDo include anyextend?
4181 //===----------------------------------------------------------------------==//
4182 // Double precision floating point instructions
4183 //===----------------------------------------------------------------------==//
4185 RRForm<0b00110011010, (outs R64FP:$rT), (ins R64FP:$rA, R64FP:$rB),
4186 "dfa\t$rT, $rA, $rB", DPrecFP,
4187 [(set R64FP:$rT, (fadd R64FP:$rA, R64FP:$rB))]>;
4190 RRForm<0b00110011010, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
4191 "dfa\t$rT, $rA, $rB", DPrecFP,
4192 [(set (v2f64 VECREG:$rT), (fadd (v2f64 VECREG:$rA), (v2f64 VECREG:$rB)))]>;
4195 RRForm<0b10100011010, (outs R64FP:$rT), (ins R64FP:$rA, R64FP:$rB),
4196 "dfs\t$rT, $rA, $rB", DPrecFP,
4197 [(set R64FP:$rT, (fsub R64FP:$rA, R64FP:$rB))]>;
4200 RRForm<0b10100011010, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
4201 "dfs\t$rT, $rA, $rB", DPrecFP,
4202 [(set (v2f64 VECREG:$rT),
4203 (fsub (v2f64 VECREG:$rA), (v2f64 VECREG:$rB)))]>;
4206 RRForm<0b01100011010, (outs R64FP:$rT), (ins R64FP:$rA, R64FP:$rB),
4207 "dfm\t$rT, $rA, $rB", DPrecFP,
4208 [(set R64FP:$rT, (fmul R64FP:$rA, R64FP:$rB))]>;
4211 RRForm<0b00100011010, (outs VECREG:$rT), (ins VECREG:$rA, VECREG:$rB),
4212 "dfm\t$rT, $rA, $rB", DPrecFP,
4213 [(set (v2f64 VECREG:$rT),
4214 (fmul (v2f64 VECREG:$rA), (v2f64 VECREG:$rB)))]>;
4217 RRForm<0b00111010110, (outs R64FP:$rT),
4218 (ins R64FP:$rA, R64FP:$rB, R64FP:$rC),
4219 "dfma\t$rT, $rA, $rB", DPrecFP,
4220 [(set R64FP:$rT, (fadd R64FP:$rC, (fmul R64FP:$rA, R64FP:$rB)))]>,
4221 RegConstraint<"$rC = $rT">,
4225 RRForm<0b00111010110, (outs VECREG:$rT),
4226 (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
4227 "dfma\t$rT, $rA, $rB", DPrecFP,
4228 [(set (v2f64 VECREG:$rT),
4229 (fadd (v2f64 VECREG:$rC),
4230 (fmul (v2f64 VECREG:$rA), (v2f64 VECREG:$rB))))]>,
4231 RegConstraint<"$rC = $rT">,
4235 RRForm<0b10111010110, (outs R64FP:$rT),
4236 (ins R64FP:$rA, R64FP:$rB, R64FP:$rC),
4237 "dfms\t$rT, $rA, $rB", DPrecFP,
4238 [(set R64FP:$rT, (fsub (fmul R64FP:$rA, R64FP:$rB), R64FP:$rC))]>,
4239 RegConstraint<"$rC = $rT">,
4243 RRForm<0b10111010110, (outs VECREG:$rT),
4244 (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
4245 "dfms\t$rT, $rA, $rB", DPrecFP,
4246 [(set (v2f64 VECREG:$rT),
4247 (fsub (fmul (v2f64 VECREG:$rA), (v2f64 VECREG:$rB)),
4248 (v2f64 VECREG:$rC)))]>;
4250 // FNMS: - (a * b - c)
4251 // - (a * b) + c => c - (a * b)
4253 RRForm<0b01111010110, (outs R64FP:$rT),
4254 (ins R64FP:$rA, R64FP:$rB, R64FP:$rC),
4255 "dfnms\t$rT, $rA, $rB", DPrecFP,
4256 [(set R64FP:$rT, (fsub R64FP:$rC, (fmul R64FP:$rA, R64FP:$rB)))]>,
4257 RegConstraint<"$rC = $rT">,
4260 def : Pat<(fneg (fsub (fmul R64FP:$rA, R64FP:$rB), R64FP:$rC)),
4261 (FNMSf64 R64FP:$rA, R64FP:$rB, R64FP:$rC)>;
4264 RRForm<0b01111010110, (outs VECREG:$rT),
4265 (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
4266 "dfnms\t$rT, $rA, $rB", DPrecFP,
4267 [(set (v2f64 VECREG:$rT),
4268 (fsub (v2f64 VECREG:$rC),
4269 (fmul (v2f64 VECREG:$rA),
4270 (v2f64 VECREG:$rB))))]>,
4271 RegConstraint<"$rC = $rT">,
4274 def : Pat<(fneg (fsub (fmul (v2f64 VECREG:$rA), (v2f64 VECREG:$rB)),
4275 (v2f64 VECREG:$rC))),
4276 (FNMSv2f64 VECREG:$rA, VECREG:$rB, VECREG:$rC)>;
4281 RRForm<0b11111010110, (outs R64FP:$rT),
4282 (ins R64FP:$rA, R64FP:$rB, R64FP:$rC),
4283 "dfnma\t$rT, $rA, $rB", DPrecFP,
4284 [(set R64FP:$rT, (fneg (fadd R64FP:$rC, (fmul R64FP:$rA, R64FP:$rB))))]>,
4285 RegConstraint<"$rC = $rT">,
4289 RRForm<0b11111010110, (outs VECREG:$rT),
4290 (ins VECREG:$rA, VECREG:$rB, VECREG:$rC),
4291 "dfnma\t$rT, $rA, $rB", DPrecFP,
4292 [(set (v2f64 VECREG:$rT),
4293 (fneg (fadd (v2f64 VECREG:$rC),
4294 (fmul (v2f64 VECREG:$rA),
4295 (v2f64 VECREG:$rB)))))]>,
4296 RegConstraint<"$rC = $rT">,
4299 //===----------------------------------------------------------------------==//
4300 // Floating point negation and absolute value
4301 //===----------------------------------------------------------------------==//
4303 def : Pat<(fneg (v4f32 VECREG:$rA)),
4304 (XORfnegvec (v4f32 VECREG:$rA),
4305 (v4f32 (ILHUv4i32 0x8000)))>;
4307 def : Pat<(fneg R32FP:$rA),
4308 (XORfneg32 R32FP:$rA, (ILHUr32 0x8000))>;
4310 def : Pat<(fneg (v2f64 VECREG:$rA)),
4311 (XORfnegvec (v2f64 VECREG:$rA),
4312 (v2f64 (ANDBIv16i8 (FSMBIv16i8 0x8080), 0x80)))>;
4314 def : Pat<(fneg R64FP:$rA),
4315 (XORfneg64 R64FP:$rA,
4316 (ANDBIv16i8 (FSMBIv16i8 0x8080), 0x80))>;
4318 // Floating point absolute value
4320 def : Pat<(fabs R32FP:$rA),
4321 (ANDfabs32 R32FP:$rA, (IOHLr32 (ILHUr32 0x7fff), 0xffff))>;
4323 def : Pat<(fabs (v4f32 VECREG:$rA)),
4324 (ANDfabsvec (v4f32 VECREG:$rA),
4325 (v4f32 (ANDBIv16i8 (FSMBIv16i8 0xffff), 0x7f)))>;
4327 def : Pat<(fabs R64FP:$rA),
4328 (ANDfabs64 R64FP:$rA, (ANDBIv16i8 (FSMBIv16i8 0xffff), 0x7f))>;
4330 def : Pat<(fabs (v2f64 VECREG:$rA)),
4331 (ANDfabsvec (v2f64 VECREG:$rA),
4332 (v2f64 (ANDBIv16i8 (FSMBIv16i8 0xffff), 0x7f)))>;
4334 //===----------------------------------------------------------------------===//
4335 // Hint for branch instructions:
4336 //===----------------------------------------------------------------------===//
4338 /* def HBR : SPUInstr<(outs), (ins), "hbr\t" */
4340 //===----------------------------------------------------------------------===//
4341 // Execution, Load NOP (execute NOPs belong in even pipeline, load NOPs belong
4342 // in the odd pipeline)
4343 //===----------------------------------------------------------------------===//
4345 def ENOP : SPUInstr<(outs), (ins), "enop", ExecNOP> {
4348 let Inst{0-10} = 0b10000000010;
4349 let Inst{11-17} = 0;
4350 let Inst{18-24} = 0;
4351 let Inst{25-31} = 0;
4354 def LNOP : SPUInstr<(outs), (ins), "lnop", LoadNOP> {
4357 let Inst{0-10} = 0b10000000000;
4358 let Inst{11-17} = 0;
4359 let Inst{18-24} = 0;
4360 let Inst{25-31} = 0;
4363 //===----------------------------------------------------------------------===//
4364 // Bit conversions (type conversions between vector/packed types)
4365 // NOTE: Promotions are handled using the XS* instructions. Truncation
4367 //===----------------------------------------------------------------------===//
4368 def : Pat<(v16i8 (bitconvert (v8i16 VECREG:$src))), (v16i8 VECREG:$src)>;
4369 def : Pat<(v16i8 (bitconvert (v4i32 VECREG:$src))), (v16i8 VECREG:$src)>;
4370 def : Pat<(v16i8 (bitconvert (v2i64 VECREG:$src))), (v16i8 VECREG:$src)>;
4371 def : Pat<(v16i8 (bitconvert (v4f32 VECREG:$src))), (v16i8 VECREG:$src)>;
4372 def : Pat<(v16i8 (bitconvert (v2f64 VECREG:$src))), (v16i8 VECREG:$src)>;
4374 def : Pat<(v8i16 (bitconvert (v16i8 VECREG:$src))), (v8i16 VECREG:$src)>;
4375 def : Pat<(v8i16 (bitconvert (v4i32 VECREG:$src))), (v8i16 VECREG:$src)>;
4376 def : Pat<(v8i16 (bitconvert (v2i64 VECREG:$src))), (v8i16 VECREG:$src)>;
4377 def : Pat<(v8i16 (bitconvert (v4f32 VECREG:$src))), (v8i16 VECREG:$src)>;
4378 def : Pat<(v8i16 (bitconvert (v2f64 VECREG:$src))), (v8i16 VECREG:$src)>;
4380 def : Pat<(v4i32 (bitconvert (v16i8 VECREG:$src))), (v4i32 VECREG:$src)>;
4381 def : Pat<(v4i32 (bitconvert (v8i16 VECREG:$src))), (v4i32 VECREG:$src)>;
4382 def : Pat<(v4i32 (bitconvert (v2i64 VECREG:$src))), (v4i32 VECREG:$src)>;
4383 def : Pat<(v4i32 (bitconvert (v4f32 VECREG:$src))), (v4i32 VECREG:$src)>;
4384 def : Pat<(v4i32 (bitconvert (v2f64 VECREG:$src))), (v4i32 VECREG:$src)>;
4386 def : Pat<(v2i64 (bitconvert (v16i8 VECREG:$src))), (v2i64 VECREG:$src)>;
4387 def : Pat<(v2i64 (bitconvert (v8i16 VECREG:$src))), (v2i64 VECREG:$src)>;
4388 def : Pat<(v2i64 (bitconvert (v4i32 VECREG:$src))), (v2i64 VECREG:$src)>;
4389 def : Pat<(v2i64 (bitconvert (v4f32 VECREG:$src))), (v2i64 VECREG:$src)>;
4390 def : Pat<(v2i64 (bitconvert (v2f64 VECREG:$src))), (v2i64 VECREG:$src)>;
4392 def : Pat<(v4f32 (bitconvert (v16i8 VECREG:$src))), (v4f32 VECREG:$src)>;
4393 def : Pat<(v4f32 (bitconvert (v8i16 VECREG:$src))), (v4f32 VECREG:$src)>;
4394 def : Pat<(v4f32 (bitconvert (v2i64 VECREG:$src))), (v4f32 VECREG:$src)>;
4395 def : Pat<(v4f32 (bitconvert (v4i32 VECREG:$src))), (v4f32 VECREG:$src)>;
4396 def : Pat<(v4f32 (bitconvert (v2f64 VECREG:$src))), (v4f32 VECREG:$src)>;
4398 def : Pat<(v2f64 (bitconvert (v16i8 VECREG:$src))), (v2f64 VECREG:$src)>;
4399 def : Pat<(v2f64 (bitconvert (v8i16 VECREG:$src))), (v2f64 VECREG:$src)>;
4400 def : Pat<(v2f64 (bitconvert (v4i32 VECREG:$src))), (v2f64 VECREG:$src)>;
4401 def : Pat<(v2f64 (bitconvert (v2i64 VECREG:$src))), (v2f64 VECREG:$src)>;
4402 def : Pat<(v2f64 (bitconvert (v2f64 VECREG:$src))), (v2f64 VECREG:$src)>;
4404 def : Pat<(f32 (bitconvert (i32 R32C:$src))), (f32 R32FP:$src)>;
4405 def : Pat<(f64 (bitconvert (i64 R64C:$src))), (f64 R64FP:$src)>;
4407 //===----------------------------------------------------------------------===//
4408 // Instruction patterns:
4409 //===----------------------------------------------------------------------===//
4411 // General 32-bit constants:
4412 def : Pat<(i32 imm:$imm),
4413 (IOHLr32 (ILHUr32 (HI16 imm:$imm)), (LO16 imm:$imm))>;
4415 // Single precision float constants:
4416 def : Pat<(f32 fpimm:$imm),
4417 (IOHLf32 (ILHUf32 (HI16_f32 fpimm:$imm)), (LO16_f32 fpimm:$imm))>;
4419 // General constant 32-bit vectors
4420 def : Pat<(v4i32 v4i32Imm:$imm),
4421 (IOHLv4i32 (v4i32 (ILHUv4i32 (HI16_vec v4i32Imm:$imm))),
4422 (LO16_vec v4i32Imm:$imm))>;
4425 def : Pat<(i8 imm:$imm),
4428 //===----------------------------------------------------------------------===//
4429 // Call instruction patterns:
4430 //===----------------------------------------------------------------------===//
4435 //===----------------------------------------------------------------------===//
4436 // Zero/Any/Sign extensions
4437 //===----------------------------------------------------------------------===//
4439 // sext 8->32: Sign extend bytes to words
4440 def : Pat<(sext_inreg R32C:$rSrc, i8),
4441 (XSHWr32 (XSBHr32 R32C:$rSrc))>;
4443 def : Pat<(i32 (sext R8C:$rSrc)),
4444 (XSHWr16 (XSBHr8 R8C:$rSrc))>;
4446 // sext 8->64: Sign extend bytes to double word
4447 def : Pat<(sext_inreg R64C:$rSrc, i8),
4448 (XSWDr64_inreg (XSHWr64 (XSBHr64 R64C:$rSrc)))>;
4450 def : Pat<(i64 (sext R8C:$rSrc)),
4451 (XSWDr64 (XSHWr16 (XSBHr8 R8C:$rSrc)))>;
4453 // zext 8->16: Zero extend bytes to halfwords
4454 def : Pat<(i16 (zext R8C:$rSrc)),
4455 (ANDHIi8i16 R8C:$rSrc, 0xff)>;
4457 // zext 8->32: Zero extend bytes to words
4458 def : Pat<(i32 (zext R8C:$rSrc)),
4459 (ANDIi8i32 R8C:$rSrc, 0xff)>;
4461 // zext 8->64: Zero extend bytes to double words
4462 def : Pat<(i64 (zext R8C:$rSrc)),
4463 (ORi64_v2i64 (SELBv4i32 (ROTQMBYv4i32
4464 (ORv4i32_i32 (ANDIi8i32 R8C:$rSrc, 0xff)),
4467 (FSMBIv4i32 0x0f0f)))>;
4469 // anyext 8->16: Extend 8->16 bits, irrespective of sign, preserves high bits
4470 def : Pat<(i16 (anyext R8C:$rSrc)),
4471 (ORHIi8i16 R8C:$rSrc, 0)>;
4473 // anyext 8->32: Extend 8->32 bits, irrespective of sign, preserves high bits
4474 def : Pat<(i32 (anyext R8C:$rSrc)),
4475 (ORIi8i32 R8C:$rSrc, 0)>;
4477 // sext 16->64: Sign extend halfword to double word
4478 def : Pat<(sext_inreg R64C:$rSrc, i16),
4479 (XSWDr64_inreg (XSHWr64 R64C:$rSrc))>;
4481 def : Pat<(sext R16C:$rSrc),
4482 (XSWDr64 (XSHWr16 R16C:$rSrc))>;
4484 // zext 16->32: Zero extend halfwords to words
4485 def : Pat<(i32 (zext R16C:$rSrc)),
4486 (ANDi16i32 R16C:$rSrc, (ILAr32 0xffff))>;
4488 def : Pat<(i32 (zext (and R16C:$rSrc, 0xf))),
4489 (ANDIi16i32 R16C:$rSrc, 0xf)>;
4491 def : Pat<(i32 (zext (and R16C:$rSrc, 0xff))),
4492 (ANDIi16i32 R16C:$rSrc, 0xff)>;
4494 def : Pat<(i32 (zext (and R16C:$rSrc, 0xfff))),
4495 (ANDIi16i32 R16C:$rSrc, 0xfff)>;
4497 // anyext 16->32: Extend 16->32 bits, irrespective of sign
4498 def : Pat<(i32 (anyext R16C:$rSrc)),
4499 (ORIi16i32 R16C:$rSrc, 0)>;
4501 //===----------------------------------------------------------------------===//
4503 // These truncates are for the SPU's supported types (i8, i16, i32). i64 and
4504 // above are custom lowered.
4505 //===----------------------------------------------------------------------===//
4507 def : Pat<(i8 (trunc GPRC:$src)),
4509 (SHUFBgprc GPRC:$src, GPRC:$src,
4510 (IOHLv4i32 (ILHUv4i32 0x0f0f), 0x0f0f)))>;
4512 def : Pat<(i8 (trunc R64C:$src)),
4515 (ORv2i64_i64 R64C:$src),
4516 (ORv2i64_i64 R64C:$src),
4517 (IOHLv4i32 (ILHUv4i32 0x0707), 0x0707)))>;
4519 def : Pat<(i8 (trunc R32C:$src)),
4522 (ORv4i32_i32 R32C:$src),
4523 (ORv4i32_i32 R32C:$src),
4524 (IOHLv4i32 (ILHUv4i32 0x0303), 0x0303)))>;
4526 def : Pat<(i8 (trunc R16C:$src)),
4529 (ORv8i16_i16 R16C:$src),
4530 (ORv8i16_i16 R16C:$src),
4531 (IOHLv4i32 (ILHUv4i32 0x0303), 0x0303)))>;
4533 def : Pat<(i16 (trunc GPRC:$src)),
4535 (SHUFBgprc GPRC:$src, GPRC:$src,
4536 (IOHLv4i32 (ILHUv4i32 0x0e0f), 0x0e0f)))>;
4538 def : Pat<(i16 (trunc R64C:$src)),
4541 (ORv2i64_i64 R64C:$src),
4542 (ORv2i64_i64 R64C:$src),
4543 (IOHLv4i32 (ILHUv4i32 0x0607), 0x0607)))>;
4545 def : Pat<(i16 (trunc R32C:$src)),
4548 (ORv4i32_i32 R32C:$src),
4549 (ORv4i32_i32 R32C:$src),
4550 (IOHLv4i32 (ILHUv4i32 0x0203), 0x0203)))>;
4552 def : Pat<(i32 (trunc GPRC:$src)),
4554 (SHUFBgprc GPRC:$src, GPRC:$src,
4555 (IOHLv4i32 (ILHUv4i32 0x0c0d), 0x0e0f)))>;
4557 def : Pat<(i32 (trunc R64C:$src)),
4560 (ORv2i64_i64 R64C:$src),
4561 (ORv2i64_i64 R64C:$src),
4562 (IOHLv4i32 (ILHUv4i32 0x0405), 0x0607)))>;
4564 //===----------------------------------------------------------------------===//
4565 // Address generation: SPU, like PPC, has to split addresses into high and
4566 // low parts in order to load them into a register.
4567 //===----------------------------------------------------------------------===//
4569 def : Pat<(SPUaform tglobaladdr:$in, 0), (ILAlsa tglobaladdr:$in)>;
4570 def : Pat<(SPUaform texternalsym:$in, 0), (ILAlsa texternalsym:$in)>;
4571 def : Pat<(SPUaform tjumptable:$in, 0), (ILAlsa tjumptable:$in)>;
4572 def : Pat<(SPUaform tconstpool:$in, 0), (ILAlsa tconstpool:$in)>;
4574 def : Pat<(SPUindirect (SPUhi tglobaladdr:$in, 0),
4575 (SPUlo tglobaladdr:$in, 0)),
4576 (IOHLlo (ILHUhi tglobaladdr:$in), tglobaladdr:$in)>;
4578 def : Pat<(SPUindirect (SPUhi texternalsym:$in, 0),
4579 (SPUlo texternalsym:$in, 0)),
4580 (IOHLlo (ILHUhi texternalsym:$in), texternalsym:$in)>;
4582 def : Pat<(SPUindirect (SPUhi tjumptable:$in, 0),
4583 (SPUlo tjumptable:$in, 0)),
4584 (IOHLlo (ILHUhi tjumptable:$in), tjumptable:$in)>;
4586 def : Pat<(SPUindirect (SPUhi tconstpool:$in, 0),
4587 (SPUlo tconstpool:$in, 0)),
4588 (IOHLlo (ILHUhi tconstpool:$in), tconstpool:$in)>;
4590 def : Pat<(add (SPUhi tglobaladdr:$in, 0), (SPUlo tglobaladdr:$in, 0)),
4591 (IOHLlo (ILHUhi tglobaladdr:$in), tglobaladdr:$in)>;
4593 def : Pat<(add (SPUhi texternalsym:$in, 0), (SPUlo texternalsym:$in, 0)),
4594 (IOHLlo (ILHUhi texternalsym:$in), texternalsym:$in)>;
4596 def : Pat<(add (SPUhi tjumptable:$in, 0), (SPUlo tjumptable:$in, 0)),
4597 (IOHLlo (ILHUhi tjumptable:$in), tjumptable:$in)>;
4599 def : Pat<(add (SPUhi tconstpool:$in, 0), (SPUlo tconstpool:$in, 0)),
4600 (IOHLlo (ILHUhi tconstpool:$in), tconstpool:$in)>;
4603 include "CellSDKIntrinsics.td"
4604 // Various math operator instruction sequences
4605 include "SPUMathInstr.td"
4606 // 64-bit "instructions"/support
4607 include "SPU64InstrInfo.td"
4608 // 128-bit "instructions"/support
4609 include "SPU128InstrInfo.td"