1 | //===-- X86ShuffleDecode.cpp - X86 shuffle decode logic -------------------===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | // |
9 | // Define several functions to decode x86 specific shuffle semantics into a |
10 | // generic vector mask. |
11 | // |
12 | //===----------------------------------------------------------------------===// |
13 | |
14 | #include "X86ShuffleDecode.h" |
15 | #include "llvm/ADT/APInt.h" |
16 | #include "llvm/ADT/ArrayRef.h" |
17 | #include "llvm/ADT/SmallVector.h" |
18 | #include "llvm/Support/MathExtras.h" |
19 | |
20 | //===----------------------------------------------------------------------===// |
21 | // Vector Mask Decoding |
22 | //===----------------------------------------------------------------------===// |
23 | |
24 | namespace llvm { |
25 | |
26 | void DecodeINSERTPSMask(unsigned Imm, SmallVectorImpl<int> &ShuffleMask) { |
27 | // Defaults the copying the dest value. |
28 | ShuffleMask.push_back(Elt: 0); |
29 | ShuffleMask.push_back(Elt: 1); |
30 | ShuffleMask.push_back(Elt: 2); |
31 | ShuffleMask.push_back(Elt: 3); |
32 | |
33 | // Decode the immediate. |
34 | unsigned ZMask = Imm & 15; |
35 | unsigned CountD = (Imm >> 4) & 3; |
36 | unsigned CountS = (Imm >> 6) & 3; |
37 | |
38 | // CountS selects which input element to use. |
39 | unsigned InVal = 4 + CountS; |
40 | // CountD specifies which element of destination to update. |
41 | ShuffleMask[CountD] = InVal; |
42 | // ZMask zaps values, potentially overriding the CountD elt. |
43 | if (ZMask & 1) ShuffleMask[0] = SM_SentinelZero; |
44 | if (ZMask & 2) ShuffleMask[1] = SM_SentinelZero; |
45 | if (ZMask & 4) ShuffleMask[2] = SM_SentinelZero; |
46 | if (ZMask & 8) ShuffleMask[3] = SM_SentinelZero; |
47 | } |
48 | |
49 | void DecodeInsertElementMask(unsigned NumElts, unsigned Idx, unsigned Len, |
50 | SmallVectorImpl<int> &ShuffleMask) { |
51 | assert((Idx + Len) <= NumElts && "Insertion out of range" ); |
52 | |
53 | for (unsigned i = 0; i != NumElts; ++i) |
54 | ShuffleMask.push_back(Elt: i); |
55 | for (unsigned i = 0; i != Len; ++i) |
56 | ShuffleMask[Idx + i] = NumElts + i; |
57 | } |
58 | |
59 | // <3,1> or <6,7,2,3> |
60 | void DecodeMOVHLPSMask(unsigned NElts, SmallVectorImpl<int> &ShuffleMask) { |
61 | for (unsigned i = NElts / 2; i != NElts; ++i) |
62 | ShuffleMask.push_back(Elt: NElts + i); |
63 | |
64 | for (unsigned i = NElts / 2; i != NElts; ++i) |
65 | ShuffleMask.push_back(Elt: i); |
66 | } |
67 | |
68 | // <0,2> or <0,1,4,5> |
69 | void DecodeMOVLHPSMask(unsigned NElts, SmallVectorImpl<int> &ShuffleMask) { |
70 | for (unsigned i = 0; i != NElts / 2; ++i) |
71 | ShuffleMask.push_back(Elt: i); |
72 | |
73 | for (unsigned i = 0; i != NElts / 2; ++i) |
74 | ShuffleMask.push_back(Elt: NElts + i); |
75 | } |
76 | |
77 | void DecodeMOVSLDUPMask(unsigned NumElts, SmallVectorImpl<int> &ShuffleMask) { |
78 | for (int i = 0, e = NumElts / 2; i < e; ++i) { |
79 | ShuffleMask.push_back(Elt: 2 * i); |
80 | ShuffleMask.push_back(Elt: 2 * i); |
81 | } |
82 | } |
83 | |
84 | void DecodeMOVSHDUPMask(unsigned NumElts, SmallVectorImpl<int> &ShuffleMask) { |
85 | for (int i = 0, e = NumElts / 2; i < e; ++i) { |
86 | ShuffleMask.push_back(Elt: 2 * i + 1); |
87 | ShuffleMask.push_back(Elt: 2 * i + 1); |
88 | } |
89 | } |
90 | |
91 | void DecodeMOVDDUPMask(unsigned NumElts, SmallVectorImpl<int> &ShuffleMask) { |
92 | const unsigned NumLaneElts = 2; |
93 | |
94 | for (unsigned l = 0; l < NumElts; l += NumLaneElts) |
95 | for (unsigned i = 0; i < NumLaneElts; ++i) |
96 | ShuffleMask.push_back(Elt: l); |
97 | } |
98 | |
99 | void DecodePSLLDQMask(unsigned NumElts, unsigned Imm, |
100 | SmallVectorImpl<int> &ShuffleMask) { |
101 | const unsigned NumLaneElts = 16; |
102 | |
103 | for (unsigned l = 0; l < NumElts; l += NumLaneElts) |
104 | for (unsigned i = 0; i < NumLaneElts; ++i) { |
105 | int M = SM_SentinelZero; |
106 | if (i >= Imm) M = i - Imm + l; |
107 | ShuffleMask.push_back(Elt: M); |
108 | } |
109 | } |
110 | |
111 | void DecodePSRLDQMask(unsigned NumElts, unsigned Imm, |
112 | SmallVectorImpl<int> &ShuffleMask) { |
113 | const unsigned NumLaneElts = 16; |
114 | |
115 | for (unsigned l = 0; l < NumElts; l += NumLaneElts) |
116 | for (unsigned i = 0; i < NumLaneElts; ++i) { |
117 | unsigned Base = i + Imm; |
118 | int M = Base + l; |
119 | if (Base >= NumLaneElts) M = SM_SentinelZero; |
120 | ShuffleMask.push_back(Elt: M); |
121 | } |
122 | } |
123 | |
124 | void DecodePALIGNRMask(unsigned NumElts, unsigned Imm, |
125 | SmallVectorImpl<int> &ShuffleMask) { |
126 | const unsigned NumLaneElts = 16; |
127 | |
128 | for (unsigned l = 0; l != NumElts; l += NumLaneElts) { |
129 | for (unsigned i = 0; i != NumLaneElts; ++i) { |
130 | unsigned Base = i + Imm; |
131 | // if i+imm is out of this lane then we actually need the other source |
132 | if (Base >= NumLaneElts) Base += NumElts - NumLaneElts; |
133 | ShuffleMask.push_back(Elt: Base + l); |
134 | } |
135 | } |
136 | } |
137 | |
138 | void DecodeVALIGNMask(unsigned NumElts, unsigned Imm, |
139 | SmallVectorImpl<int> &ShuffleMask) { |
140 | // Not all bits of the immediate are used so mask it. |
141 | assert(isPowerOf2_32(NumElts) && "NumElts should be power of 2" ); |
142 | Imm = Imm & (NumElts - 1); |
143 | for (unsigned i = 0; i != NumElts; ++i) |
144 | ShuffleMask.push_back(Elt: i + Imm); |
145 | } |
146 | |
147 | void DecodePSHUFMask(unsigned NumElts, unsigned ScalarBits, unsigned Imm, |
148 | SmallVectorImpl<int> &ShuffleMask) { |
149 | unsigned Size = NumElts * ScalarBits; |
150 | unsigned NumLanes = Size / 128; |
151 | if (NumLanes == 0) NumLanes = 1; // Handle MMX |
152 | unsigned NumLaneElts = NumElts / NumLanes; |
153 | |
154 | uint32_t SplatImm = (Imm & 0xff) * 0x01010101; |
155 | for (unsigned l = 0; l != NumElts; l += NumLaneElts) { |
156 | for (unsigned i = 0; i != NumLaneElts; ++i) { |
157 | ShuffleMask.push_back(Elt: SplatImm % NumLaneElts + l); |
158 | SplatImm /= NumLaneElts; |
159 | } |
160 | } |
161 | } |
162 | |
163 | void DecodePSHUFHWMask(unsigned NumElts, unsigned Imm, |
164 | SmallVectorImpl<int> &ShuffleMask) { |
165 | for (unsigned l = 0; l != NumElts; l += 8) { |
166 | unsigned NewImm = Imm; |
167 | for (unsigned i = 0, e = 4; i != e; ++i) { |
168 | ShuffleMask.push_back(Elt: l + i); |
169 | } |
170 | for (unsigned i = 4, e = 8; i != e; ++i) { |
171 | ShuffleMask.push_back(Elt: l + 4 + (NewImm & 3)); |
172 | NewImm >>= 2; |
173 | } |
174 | } |
175 | } |
176 | |
177 | void DecodePSHUFLWMask(unsigned NumElts, unsigned Imm, |
178 | SmallVectorImpl<int> &ShuffleMask) { |
179 | for (unsigned l = 0; l != NumElts; l += 8) { |
180 | unsigned NewImm = Imm; |
181 | for (unsigned i = 0, e = 4; i != e; ++i) { |
182 | ShuffleMask.push_back(Elt: l + (NewImm & 3)); |
183 | NewImm >>= 2; |
184 | } |
185 | for (unsigned i = 4, e = 8; i != e; ++i) { |
186 | ShuffleMask.push_back(Elt: l + i); |
187 | } |
188 | } |
189 | } |
190 | |
191 | void DecodePSWAPMask(unsigned NumElts, SmallVectorImpl<int> &ShuffleMask) { |
192 | unsigned NumHalfElts = NumElts / 2; |
193 | |
194 | for (unsigned l = 0; l != NumHalfElts; ++l) |
195 | ShuffleMask.push_back(Elt: l + NumHalfElts); |
196 | for (unsigned h = 0; h != NumHalfElts; ++h) |
197 | ShuffleMask.push_back(Elt: h); |
198 | } |
199 | |
200 | void DecodeSHUFPMask(unsigned NumElts, unsigned ScalarBits, |
201 | unsigned Imm, SmallVectorImpl<int> &ShuffleMask) { |
202 | unsigned NumLaneElts = 128 / ScalarBits; |
203 | |
204 | unsigned NewImm = Imm; |
205 | for (unsigned l = 0; l != NumElts; l += NumLaneElts) { |
206 | // each half of a lane comes from different source |
207 | for (unsigned s = 0; s != NumElts * 2; s += NumElts) { |
208 | for (unsigned i = 0; i != NumLaneElts / 2; ++i) { |
209 | ShuffleMask.push_back(Elt: NewImm % NumLaneElts + s + l); |
210 | NewImm /= NumLaneElts; |
211 | } |
212 | } |
213 | if (NumLaneElts == 4) NewImm = Imm; // reload imm |
214 | } |
215 | } |
216 | |
217 | void DecodeUNPCKHMask(unsigned NumElts, unsigned ScalarBits, |
218 | SmallVectorImpl<int> &ShuffleMask) { |
219 | // Handle 128 and 256-bit vector lengths. AVX defines UNPCK* to operate |
220 | // independently on 128-bit lanes. |
221 | unsigned NumLanes = (NumElts * ScalarBits) / 128; |
222 | if (NumLanes == 0) NumLanes = 1; // Handle MMX |
223 | unsigned NumLaneElts = NumElts / NumLanes; |
224 | |
225 | for (unsigned l = 0; l != NumElts; l += NumLaneElts) { |
226 | for (unsigned i = l + NumLaneElts / 2, e = l + NumLaneElts; i != e; ++i) { |
227 | ShuffleMask.push_back(Elt: i); // Reads from dest/src1 |
228 | ShuffleMask.push_back(Elt: i + NumElts); // Reads from src/src2 |
229 | } |
230 | } |
231 | } |
232 | |
233 | void DecodeUNPCKLMask(unsigned NumElts, unsigned ScalarBits, |
234 | SmallVectorImpl<int> &ShuffleMask) { |
235 | // Handle 128 and 256-bit vector lengths. AVX defines UNPCK* to operate |
236 | // independently on 128-bit lanes. |
237 | unsigned NumLanes = (NumElts * ScalarBits) / 128; |
238 | if (NumLanes == 0 ) NumLanes = 1; // Handle MMX |
239 | unsigned NumLaneElts = NumElts / NumLanes; |
240 | |
241 | for (unsigned l = 0; l != NumElts; l += NumLaneElts) { |
242 | for (unsigned i = l, e = l + NumLaneElts / 2; i != e; ++i) { |
243 | ShuffleMask.push_back(Elt: i); // Reads from dest/src1 |
244 | ShuffleMask.push_back(Elt: i + NumElts); // Reads from src/src2 |
245 | } |
246 | } |
247 | } |
248 | |
249 | void DecodeVectorBroadcast(unsigned NumElts, |
250 | SmallVectorImpl<int> &ShuffleMask) { |
251 | ShuffleMask.append(NumInputs: NumElts, Elt: 0); |
252 | } |
253 | |
254 | void DecodeSubVectorBroadcast(unsigned DstNumElts, unsigned SrcNumElts, |
255 | SmallVectorImpl<int> &ShuffleMask) { |
256 | unsigned Scale = DstNumElts / SrcNumElts; |
257 | |
258 | for (unsigned i = 0; i != Scale; ++i) |
259 | for (unsigned j = 0; j != SrcNumElts; ++j) |
260 | ShuffleMask.push_back(Elt: j); |
261 | } |
262 | |
263 | void decodeVSHUF64x2FamilyMask(unsigned NumElts, unsigned ScalarSize, |
264 | unsigned Imm, |
265 | SmallVectorImpl<int> &ShuffleMask) { |
266 | unsigned NumElementsInLane = 128 / ScalarSize; |
267 | unsigned NumLanes = NumElts / NumElementsInLane; |
268 | |
269 | for (unsigned l = 0; l != NumElts; l += NumElementsInLane) { |
270 | unsigned Index = (Imm % NumLanes) * NumElementsInLane; |
271 | Imm /= NumLanes; // Discard the bits we just used. |
272 | // We actually need the other source. |
273 | if (l >= (NumElts / 2)) |
274 | Index += NumElts; |
275 | for (unsigned i = 0; i != NumElementsInLane; ++i) |
276 | ShuffleMask.push_back(Elt: Index + i); |
277 | } |
278 | } |
279 | |
280 | void DecodeVPERM2X128Mask(unsigned NumElts, unsigned Imm, |
281 | SmallVectorImpl<int> &ShuffleMask) { |
282 | unsigned HalfSize = NumElts / 2; |
283 | |
284 | for (unsigned l = 0; l != 2; ++l) { |
285 | unsigned HalfMask = Imm >> (l * 4); |
286 | unsigned HalfBegin = (HalfMask & 0x3) * HalfSize; |
287 | for (unsigned i = HalfBegin, e = HalfBegin + HalfSize; i != e; ++i) |
288 | ShuffleMask.push_back(Elt: (HalfMask & 8) ? SM_SentinelZero : (int)i); |
289 | } |
290 | } |
291 | |
292 | void DecodePSHUFBMask(ArrayRef<uint64_t> RawMask, const APInt &UndefElts, |
293 | SmallVectorImpl<int> &ShuffleMask) { |
294 | for (int i = 0, e = RawMask.size(); i < e; ++i) { |
295 | uint64_t M = RawMask[i]; |
296 | if (UndefElts[i]) { |
297 | ShuffleMask.push_back(Elt: SM_SentinelUndef); |
298 | continue; |
299 | } |
300 | // For 256/512-bit vectors the base of the shuffle is the 128-bit |
301 | // subvector we're inside. |
302 | int Base = (i / 16) * 16; |
303 | // If the high bit (7) of the byte is set, the element is zeroed. |
304 | if (M & (1 << 7)) |
305 | ShuffleMask.push_back(Elt: SM_SentinelZero); |
306 | else { |
307 | // Only the least significant 4 bits of the byte are used. |
308 | int Index = Base + (M & 0xf); |
309 | ShuffleMask.push_back(Elt: Index); |
310 | } |
311 | } |
312 | } |
313 | |
314 | void DecodeBLENDMask(unsigned NumElts, unsigned Imm, |
315 | SmallVectorImpl<int> &ShuffleMask) { |
316 | for (unsigned i = 0; i < NumElts; ++i) { |
317 | // If there are more than 8 elements in the vector, then any immediate blend |
318 | // mask wraps around. |
319 | unsigned Bit = i % 8; |
320 | ShuffleMask.push_back(Elt: ((Imm >> Bit) & 1) ? NumElts + i : i); |
321 | } |
322 | } |
323 | |
324 | void DecodeVPPERMMask(ArrayRef<uint64_t> RawMask, const APInt &UndefElts, |
325 | SmallVectorImpl<int> &ShuffleMask) { |
326 | assert(RawMask.size() == 16 && "Illegal VPPERM shuffle mask size" ); |
327 | |
328 | // VPPERM Operation |
329 | // Bits[4:0] - Byte Index (0 - 31) |
330 | // Bits[7:5] - Permute Operation |
331 | // |
332 | // Permute Operation: |
333 | // 0 - Source byte (no logical operation). |
334 | // 1 - Invert source byte. |
335 | // 2 - Bit reverse of source byte. |
336 | // 3 - Bit reverse of inverted source byte. |
337 | // 4 - 00h (zero - fill). |
338 | // 5 - FFh (ones - fill). |
339 | // 6 - Most significant bit of source byte replicated in all bit positions. |
340 | // 7 - Invert most significant bit of source byte and replicate in all bit positions. |
341 | for (int i = 0, e = RawMask.size(); i < e; ++i) { |
342 | if (UndefElts[i]) { |
343 | ShuffleMask.push_back(Elt: SM_SentinelUndef); |
344 | continue; |
345 | } |
346 | |
347 | uint64_t M = RawMask[i]; |
348 | uint64_t PermuteOp = (M >> 5) & 0x7; |
349 | if (PermuteOp == 4) { |
350 | ShuffleMask.push_back(Elt: SM_SentinelZero); |
351 | continue; |
352 | } |
353 | if (PermuteOp != 0) { |
354 | ShuffleMask.clear(); |
355 | return; |
356 | } |
357 | |
358 | uint64_t Index = M & 0x1F; |
359 | ShuffleMask.push_back(Elt: (int)Index); |
360 | } |
361 | } |
362 | |
363 | void DecodeVPERMMask(unsigned NumElts, unsigned Imm, |
364 | SmallVectorImpl<int> &ShuffleMask) { |
365 | for (unsigned l = 0; l != NumElts; l += 4) |
366 | for (unsigned i = 0; i != 4; ++i) |
367 | ShuffleMask.push_back(Elt: l + ((Imm >> (2 * i)) & 3)); |
368 | } |
369 | |
370 | void DecodeZeroExtendMask(unsigned SrcScalarBits, unsigned DstScalarBits, |
371 | unsigned NumDstElts, bool IsAnyExtend, |
372 | SmallVectorImpl<int> &ShuffleMask) { |
373 | unsigned Scale = DstScalarBits / SrcScalarBits; |
374 | assert(SrcScalarBits < DstScalarBits && |
375 | "Expected zero extension mask to increase scalar size" ); |
376 | |
377 | int Sentinel = IsAnyExtend ? SM_SentinelUndef : SM_SentinelZero; |
378 | for (unsigned i = 0; i != NumDstElts; i++) { |
379 | ShuffleMask.push_back(Elt: i); |
380 | ShuffleMask.append(NumInputs: Scale - 1, Elt: Sentinel); |
381 | } |
382 | } |
383 | |
384 | void DecodeZeroMoveLowMask(unsigned NumElts, |
385 | SmallVectorImpl<int> &ShuffleMask) { |
386 | ShuffleMask.push_back(Elt: 0); |
387 | ShuffleMask.append(NumInputs: NumElts - 1, Elt: SM_SentinelZero); |
388 | } |
389 | |
390 | void DecodeScalarMoveMask(unsigned NumElts, bool IsLoad, |
391 | SmallVectorImpl<int> &ShuffleMask) { |
392 | // First element comes from the first element of second source. |
393 | // Remaining elements: Load zero extends / Move copies from first source. |
394 | ShuffleMask.push_back(Elt: NumElts); |
395 | for (unsigned i = 1; i < NumElts; i++) |
396 | ShuffleMask.push_back(Elt: IsLoad ? static_cast<int>(SM_SentinelZero) : i); |
397 | } |
398 | |
399 | void DecodeEXTRQIMask(unsigned NumElts, unsigned EltSize, int Len, int Idx, |
400 | SmallVectorImpl<int> &ShuffleMask) { |
401 | unsigned HalfElts = NumElts / 2; |
402 | |
403 | // Only the bottom 6 bits are valid for each immediate. |
404 | Len &= 0x3F; |
405 | Idx &= 0x3F; |
406 | |
407 | // We can only decode this bit extraction instruction as a shuffle if both the |
408 | // length and index work with whole elements. |
409 | if (0 != (Len % EltSize) || 0 != (Idx % EltSize)) |
410 | return; |
411 | |
412 | // A length of zero is equivalent to a bit length of 64. |
413 | if (Len == 0) |
414 | Len = 64; |
415 | |
416 | // If the length + index exceeds the bottom 64 bits the result is undefined. |
417 | if ((Len + Idx) > 64) { |
418 | ShuffleMask.append(NumInputs: NumElts, Elt: SM_SentinelUndef); |
419 | return; |
420 | } |
421 | |
422 | // Convert index and index to work with elements. |
423 | Len /= EltSize; |
424 | Idx /= EltSize; |
425 | |
426 | // EXTRQ: Extract Len elements starting from Idx. Zero pad the remaining |
427 | // elements of the lower 64-bits. The upper 64-bits are undefined. |
428 | for (int i = 0; i != Len; ++i) |
429 | ShuffleMask.push_back(Elt: i + Idx); |
430 | for (int i = Len; i != (int)HalfElts; ++i) |
431 | ShuffleMask.push_back(Elt: SM_SentinelZero); |
432 | for (int i = HalfElts; i != (int)NumElts; ++i) |
433 | ShuffleMask.push_back(Elt: SM_SentinelUndef); |
434 | } |
435 | |
436 | void DecodeINSERTQIMask(unsigned NumElts, unsigned EltSize, int Len, int Idx, |
437 | SmallVectorImpl<int> &ShuffleMask) { |
438 | unsigned HalfElts = NumElts / 2; |
439 | |
440 | // Only the bottom 6 bits are valid for each immediate. |
441 | Len &= 0x3F; |
442 | Idx &= 0x3F; |
443 | |
444 | // We can only decode this bit insertion instruction as a shuffle if both the |
445 | // length and index work with whole elements. |
446 | if (0 != (Len % EltSize) || 0 != (Idx % EltSize)) |
447 | return; |
448 | |
449 | // A length of zero is equivalent to a bit length of 64. |
450 | if (Len == 0) |
451 | Len = 64; |
452 | |
453 | // If the length + index exceeds the bottom 64 bits the result is undefined. |
454 | if ((Len + Idx) > 64) { |
455 | ShuffleMask.append(NumInputs: NumElts, Elt: SM_SentinelUndef); |
456 | return; |
457 | } |
458 | |
459 | // Convert index and index to work with elements. |
460 | Len /= EltSize; |
461 | Idx /= EltSize; |
462 | |
463 | // INSERTQ: Extract lowest Len elements from lower half of second source and |
464 | // insert over first source starting at Idx element. The upper 64-bits are |
465 | // undefined. |
466 | for (int i = 0; i != Idx; ++i) |
467 | ShuffleMask.push_back(Elt: i); |
468 | for (int i = 0; i != Len; ++i) |
469 | ShuffleMask.push_back(Elt: i + NumElts); |
470 | for (int i = Idx + Len; i != (int)HalfElts; ++i) |
471 | ShuffleMask.push_back(Elt: i); |
472 | for (int i = HalfElts; i != (int)NumElts; ++i) |
473 | ShuffleMask.push_back(Elt: SM_SentinelUndef); |
474 | } |
475 | |
476 | void DecodeVPERMILPMask(unsigned NumElts, unsigned ScalarBits, |
477 | ArrayRef<uint64_t> RawMask, const APInt &UndefElts, |
478 | SmallVectorImpl<int> &ShuffleMask) { |
479 | unsigned VecSize = NumElts * ScalarBits; |
480 | unsigned NumLanes = VecSize / 128; |
481 | unsigned NumEltsPerLane = NumElts / NumLanes; |
482 | assert((VecSize == 128 || VecSize == 256 || VecSize == 512) && |
483 | "Unexpected vector size" ); |
484 | assert((ScalarBits == 32 || ScalarBits == 64) && "Unexpected element size" ); |
485 | |
486 | for (unsigned i = 0, e = RawMask.size(); i < e; ++i) { |
487 | if (UndefElts[i]) { |
488 | ShuffleMask.push_back(Elt: SM_SentinelUndef); |
489 | continue; |
490 | } |
491 | uint64_t M = RawMask[i]; |
492 | M = (ScalarBits == 64 ? ((M >> 1) & 0x1) : (M & 0x3)); |
493 | unsigned LaneOffset = i & ~(NumEltsPerLane - 1); |
494 | ShuffleMask.push_back(Elt: (int)(LaneOffset + M)); |
495 | } |
496 | } |
497 | |
498 | void DecodeVPERMIL2PMask(unsigned NumElts, unsigned ScalarBits, unsigned M2Z, |
499 | ArrayRef<uint64_t> RawMask, const APInt &UndefElts, |
500 | SmallVectorImpl<int> &ShuffleMask) { |
501 | unsigned VecSize = NumElts * ScalarBits; |
502 | unsigned NumLanes = VecSize / 128; |
503 | unsigned NumEltsPerLane = NumElts / NumLanes; |
504 | assert((VecSize == 128 || VecSize == 256) && "Unexpected vector size" ); |
505 | assert((ScalarBits == 32 || ScalarBits == 64) && "Unexpected element size" ); |
506 | assert((NumElts == RawMask.size()) && "Unexpected mask size" ); |
507 | |
508 | for (unsigned i = 0, e = RawMask.size(); i < e; ++i) { |
509 | if (UndefElts[i]) { |
510 | ShuffleMask.push_back(Elt: SM_SentinelUndef); |
511 | continue; |
512 | } |
513 | |
514 | // VPERMIL2 Operation. |
515 | // Bits[3] - Match Bit. |
516 | // Bits[2:1] - (Per Lane) PD Shuffle Mask. |
517 | // Bits[2:0] - (Per Lane) PS Shuffle Mask. |
518 | uint64_t Selector = RawMask[i]; |
519 | unsigned MatchBit = (Selector >> 3) & 0x1; |
520 | |
521 | // M2Z[0:1] MatchBit |
522 | // 0Xb X Source selected by Selector index. |
523 | // 10b 0 Source selected by Selector index. |
524 | // 10b 1 Zero. |
525 | // 11b 0 Zero. |
526 | // 11b 1 Source selected by Selector index. |
527 | if ((M2Z & 0x2) != 0 && MatchBit != (M2Z & 0x1)) { |
528 | ShuffleMask.push_back(Elt: SM_SentinelZero); |
529 | continue; |
530 | } |
531 | |
532 | int Index = i & ~(NumEltsPerLane - 1); |
533 | if (ScalarBits == 64) |
534 | Index += (Selector >> 1) & 0x1; |
535 | else |
536 | Index += Selector & 0x3; |
537 | |
538 | int Src = (Selector >> 2) & 0x1; |
539 | Index += Src * NumElts; |
540 | ShuffleMask.push_back(Elt: Index); |
541 | } |
542 | } |
543 | |
544 | void DecodeVPERMVMask(ArrayRef<uint64_t> RawMask, const APInt &UndefElts, |
545 | SmallVectorImpl<int> &ShuffleMask) { |
546 | uint64_t EltMaskSize = RawMask.size() - 1; |
547 | for (int i = 0, e = RawMask.size(); i != e; ++i) { |
548 | if (UndefElts[i]) { |
549 | ShuffleMask.push_back(Elt: SM_SentinelUndef); |
550 | continue; |
551 | } |
552 | uint64_t M = RawMask[i]; |
553 | M &= EltMaskSize; |
554 | ShuffleMask.push_back(Elt: (int)M); |
555 | } |
556 | } |
557 | |
558 | void DecodeVPERMV3Mask(ArrayRef<uint64_t> RawMask, const APInt &UndefElts, |
559 | SmallVectorImpl<int> &ShuffleMask) { |
560 | uint64_t EltMaskSize = (RawMask.size() * 2) - 1; |
561 | for (int i = 0, e = RawMask.size(); i != e; ++i) { |
562 | if (UndefElts[i]) { |
563 | ShuffleMask.push_back(Elt: SM_SentinelUndef); |
564 | continue; |
565 | } |
566 | uint64_t M = RawMask[i]; |
567 | M &= EltMaskSize; |
568 | ShuffleMask.push_back(Elt: (int)M); |
569 | } |
570 | } |
571 | |
572 | } // namespace llvm |
573 | |