xref: /freebsd/contrib/llvm-project/llvm/lib/Target/X86/X86InstrVecCompiler.td (revision 1db9f3b21e39176dd5b67cf8ac378633b172463e)
1//===- X86InstrVecCompiler.td - Vector Compiler Patterns ---*- tablegen -*-===//
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// This file describes the various vector pseudo instructions used by the
10// compiler, as well as Pat patterns used during instruction selection.
11//
12//===----------------------------------------------------------------------===//
13
14//===----------------------------------------------------------------------===//
15//  Non-instruction patterns
16//===----------------------------------------------------------------------===//
17
18let Predicates = [NoAVX512] in {
19  // A vector extract of the first f32/f64 position is a subregister copy
20  def : Pat<(f16 (extractelt (v8f16 VR128:$src), (iPTR 0))),
21            (COPY_TO_REGCLASS (v8f16 VR128:$src), FR16)>;
22  def : Pat<(f32 (extractelt (v4f32 VR128:$src), (iPTR 0))),
23            (COPY_TO_REGCLASS (v4f32 VR128:$src), FR32)>;
24  def : Pat<(f64 (extractelt (v2f64 VR128:$src), (iPTR 0))),
25            (COPY_TO_REGCLASS (v2f64 VR128:$src), FR64)>;
26}
27
28let Predicates = [HasAVX512] in {
29  // A vector extract of the first f32/f64 position is a subregister copy
30  def : Pat<(f16 (extractelt (v8f16 VR128X:$src), (iPTR 0))),
31            (COPY_TO_REGCLASS (v8f16 VR128X:$src), FR16X)>;
32  def : Pat<(f32 (extractelt (v4f32 VR128X:$src), (iPTR 0))),
33            (COPY_TO_REGCLASS (v4f32 VR128X:$src), FR32X)>;
34  def : Pat<(f64 (extractelt (v2f64 VR128X:$src), (iPTR 0))),
35            (COPY_TO_REGCLASS (v2f64 VR128X:$src), FR64X)>;
36}
37
38let Predicates = [NoVLX] in {
39  def : Pat<(v8f16 (scalar_to_vector FR16:$src)),
40            (COPY_TO_REGCLASS FR16:$src, VR128)>;
41  // Implicitly promote a 32-bit scalar to a vector.
42  def : Pat<(v4f32 (scalar_to_vector FR32:$src)),
43            (COPY_TO_REGCLASS FR32:$src, VR128)>;
44  // Implicitly promote a 64-bit scalar to a vector.
45  def : Pat<(v2f64 (scalar_to_vector FR64:$src)),
46            (COPY_TO_REGCLASS FR64:$src, VR128)>;
47}
48
49let Predicates = [HasVLX] in {
50  def : Pat<(v8f16 (scalar_to_vector FR16X:$src)),
51            (COPY_TO_REGCLASS FR16X:$src, VR128X)>;
52  // Implicitly promote a 32-bit scalar to a vector.
53  def : Pat<(v4f32 (scalar_to_vector FR32X:$src)),
54            (COPY_TO_REGCLASS FR32X:$src, VR128X)>;
55  // Implicitly promote a 64-bit scalar to a vector.
56  def : Pat<(v2f64 (scalar_to_vector FR64X:$src)),
57            (COPY_TO_REGCLASS FR64X:$src, VR128X)>;
58}
59
60//===----------------------------------------------------------------------===//
61// Subvector tricks
62//===----------------------------------------------------------------------===//
63
64// Patterns for insert_subvector/extract_subvector to/from index=0
65multiclass subvector_subreg_lowering<RegisterClass subRC, ValueType subVT,
66                                     RegisterClass RC, ValueType VT,
67                                     SubRegIndex subIdx> {
68  def : Pat<(subVT (extract_subvector (VT RC:$src), (iPTR 0))),
69            (subVT (EXTRACT_SUBREG RC:$src, subIdx))>;
70
71  def : Pat<(VT (insert_subvector undef_or_freeze_undef, subRC:$src, (iPTR 0))),
72            (VT (INSERT_SUBREG (IMPLICIT_DEF), subRC:$src, subIdx))>;
73}
74
75// A 128-bit subvector extract from the first 256-bit vector position is a
76// subregister copy that needs no instruction. Likewise, a 128-bit subvector
77// insert to the first 256-bit vector position is a subregister copy that needs
78// no instruction.
79defm : subvector_subreg_lowering<VR128, v4i32, VR256, v8i32,  sub_xmm>;
80defm : subvector_subreg_lowering<VR128, v4f32, VR256, v8f32,  sub_xmm>;
81defm : subvector_subreg_lowering<VR128, v2i64, VR256, v4i64,  sub_xmm>;
82defm : subvector_subreg_lowering<VR128, v2f64, VR256, v4f64,  sub_xmm>;
83defm : subvector_subreg_lowering<VR128, v8i16, VR256, v16i16, sub_xmm>;
84defm : subvector_subreg_lowering<VR128, v16i8, VR256, v32i8,  sub_xmm>;
85defm : subvector_subreg_lowering<VR128, v8f16, VR256, v16f16, sub_xmm>;
86
87// A 128-bit subvector extract from the first 512-bit vector position is a
88// subregister copy that needs no instruction. Likewise, a 128-bit subvector
89// insert to the first 512-bit vector position is a subregister copy that needs
90// no instruction.
91defm : subvector_subreg_lowering<VR128, v4i32, VR512, v16i32, sub_xmm>;
92defm : subvector_subreg_lowering<VR128, v4f32, VR512, v16f32, sub_xmm>;
93defm : subvector_subreg_lowering<VR128, v2i64, VR512, v8i64,  sub_xmm>;
94defm : subvector_subreg_lowering<VR128, v2f64, VR512, v8f64,  sub_xmm>;
95defm : subvector_subreg_lowering<VR128, v8i16, VR512, v32i16, sub_xmm>;
96defm : subvector_subreg_lowering<VR128, v16i8, VR512, v64i8,  sub_xmm>;
97defm : subvector_subreg_lowering<VR128, v8f16, VR512, v32f16, sub_xmm>;
98
99// A 128-bit subvector extract from the first 512-bit vector position is a
100// subregister copy that needs no instruction. Likewise, a 128-bit subvector
101// insert to the first 512-bit vector position is a subregister copy that needs
102// no instruction.
103defm : subvector_subreg_lowering<VR256, v8i32,  VR512, v16i32, sub_ymm>;
104defm : subvector_subreg_lowering<VR256, v8f32,  VR512, v16f32, sub_ymm>;
105defm : subvector_subreg_lowering<VR256, v4i64,  VR512, v8i64,  sub_ymm>;
106defm : subvector_subreg_lowering<VR256, v4f64,  VR512, v8f64,  sub_ymm>;
107defm : subvector_subreg_lowering<VR256, v16i16, VR512, v32i16, sub_ymm>;
108defm : subvector_subreg_lowering<VR256, v32i8,  VR512, v64i8,  sub_ymm>;
109defm : subvector_subreg_lowering<VR256, v16f16, VR512, v32f16, sub_ymm>;
110
111
112// If we're inserting into an all zeros vector, just use a plain move which
113// will zero the upper bits. A post-isel hook will take care of removing
114// any moves that we can prove are unnecessary.
115multiclass subvec_zero_lowering<string MoveStr,
116                                RegisterClass RC, ValueType DstTy,
117                                ValueType SrcTy, SubRegIndex SubIdx> {
118  def : Pat<(DstTy (insert_subvector immAllZerosV,
119                                     (SrcTy RC:$src), (iPTR 0))),
120            (SUBREG_TO_REG (i64 0),
121             (SrcTy (!cast<Instruction>("VMOV"#MoveStr#"rr") RC:$src)), SubIdx)>;
122}
123
124let Predicates = [HasAVX, NoVLX] in {
125  defm : subvec_zero_lowering<"APD", VR128, v4f64, v2f64, sub_xmm>;
126  defm : subvec_zero_lowering<"APS", VR128, v8f32, v4f32, sub_xmm>;
127  defm : subvec_zero_lowering<"DQA", VR128, v4i64, v2i64, sub_xmm>;
128  defm : subvec_zero_lowering<"DQA", VR128, v8i32, v4i32, sub_xmm>;
129  defm : subvec_zero_lowering<"DQA", VR128, v16i16, v8i16, sub_xmm>;
130  defm : subvec_zero_lowering<"DQA", VR128, v32i8, v16i8, sub_xmm>;
131}
132
133let Predicates = [HasAVXNECONVERT, NoVLX] in
134  defm : subvec_zero_lowering<"DQA", VR128, v16bf16, v8bf16, sub_xmm>;
135
136let Predicates = [HasVLX] in {
137  defm : subvec_zero_lowering<"APDZ128", VR128X, v4f64, v2f64, sub_xmm>;
138  defm : subvec_zero_lowering<"APSZ128", VR128X, v8f32, v4f32, sub_xmm>;
139  defm : subvec_zero_lowering<"DQA64Z128", VR128X, v4i64, v2i64, sub_xmm>;
140  defm : subvec_zero_lowering<"DQA64Z128", VR128X, v8i32, v4i32, sub_xmm>;
141  defm : subvec_zero_lowering<"DQA64Z128", VR128X, v16i16, v8i16, sub_xmm>;
142  defm : subvec_zero_lowering<"DQA64Z128", VR128X, v32i8, v16i8, sub_xmm>;
143
144  defm : subvec_zero_lowering<"APDZ128", VR128X, v8f64, v2f64, sub_xmm>;
145  defm : subvec_zero_lowering<"APSZ128", VR128X, v16f32, v4f32, sub_xmm>;
146  defm : subvec_zero_lowering<"DQA64Z128", VR128X, v8i64, v2i64, sub_xmm>;
147  defm : subvec_zero_lowering<"DQA64Z128", VR128X, v16i32, v4i32, sub_xmm>;
148  defm : subvec_zero_lowering<"DQA64Z128", VR128X, v32i16, v8i16, sub_xmm>;
149  defm : subvec_zero_lowering<"DQA64Z128", VR128X, v64i8, v16i8, sub_xmm>;
150
151  defm : subvec_zero_lowering<"APDZ256", VR256X, v8f64, v4f64, sub_ymm>;
152  defm : subvec_zero_lowering<"APSZ256", VR256X, v16f32, v8f32, sub_ymm>;
153  defm : subvec_zero_lowering<"DQA64Z256", VR256X, v8i64, v4i64, sub_ymm>;
154  defm : subvec_zero_lowering<"DQA64Z256", VR256X, v16i32, v8i32, sub_ymm>;
155  defm : subvec_zero_lowering<"DQA64Z256", VR256X, v32i16, v16i16, sub_ymm>;
156  defm : subvec_zero_lowering<"DQA64Z256", VR256X, v64i8, v32i8, sub_ymm>;
157}
158
159let Predicates = [HasAVX512, NoVLX] in {
160  defm : subvec_zero_lowering<"APD", VR128, v8f64, v2f64, sub_xmm>;
161  defm : subvec_zero_lowering<"APS", VR128, v16f32, v4f32, sub_xmm>;
162  defm : subvec_zero_lowering<"DQA", VR128, v8i64, v2i64, sub_xmm>;
163  defm : subvec_zero_lowering<"DQA", VR128, v16i32, v4i32, sub_xmm>;
164  defm : subvec_zero_lowering<"DQA", VR128, v32i16, v8i16, sub_xmm>;
165  defm : subvec_zero_lowering<"DQA", VR128, v64i8, v16i8, sub_xmm>;
166
167  defm : subvec_zero_lowering<"APDY", VR256, v8f64, v4f64, sub_ymm>;
168  defm : subvec_zero_lowering<"APSY", VR256, v16f32, v8f32, sub_ymm>;
169  defm : subvec_zero_lowering<"DQAY", VR256, v8i64, v4i64, sub_ymm>;
170  defm : subvec_zero_lowering<"DQAY", VR256, v16i32, v8i32, sub_ymm>;
171  defm : subvec_zero_lowering<"DQAY", VR256, v32i16, v16i16, sub_ymm>;
172  defm : subvec_zero_lowering<"DQAY", VR256, v64i8, v32i8, sub_ymm>;
173}
174
175let Predicates = [HasFP16, HasVLX] in {
176  defm : subvec_zero_lowering<"APSZ128", VR128X, v16f16, v8f16, sub_xmm>;
177  defm : subvec_zero_lowering<"APSZ128", VR128X, v32f16, v8f16, sub_xmm>;
178  defm : subvec_zero_lowering<"APSZ256", VR256X, v32f16, v16f16, sub_ymm>;
179}
180
181let Predicates = [HasBF16, HasVLX] in {
182  defm : subvec_zero_lowering<"APSZ128", VR128X, v16bf16, v8bf16, sub_xmm>;
183  defm : subvec_zero_lowering<"APSZ128", VR128X, v32bf16, v8bf16, sub_xmm>;
184  defm : subvec_zero_lowering<"APSZ256", VR256X, v32bf16, v16bf16, sub_ymm>;
185}
186
187class maskzeroupper<ValueType vt, RegisterClass RC> :
188  PatLeaf<(vt RC:$src), [{
189    return isMaskZeroExtended(N);
190  }]>;
191
192def maskzeroupperv1i1  : maskzeroupper<v1i1,  VK1>;
193def maskzeroupperv2i1  : maskzeroupper<v2i1,  VK2>;
194def maskzeroupperv4i1  : maskzeroupper<v4i1,  VK4>;
195def maskzeroupperv8i1  : maskzeroupper<v8i1,  VK8>;
196def maskzeroupperv16i1 : maskzeroupper<v16i1, VK16>;
197def maskzeroupperv32i1 : maskzeroupper<v32i1, VK32>;
198
199// The patterns determine if we can depend on the upper bits of a mask register
200// being zeroed by the previous operation so that we can skip explicit
201// zeroing.
202let Predicates = [HasBWI] in {
203  def : Pat<(v32i1 (insert_subvector (v32i1 immAllZerosV),
204                                     maskzeroupperv1i1:$src, (iPTR 0))),
205            (COPY_TO_REGCLASS VK1:$src, VK32)>;
206  def : Pat<(v32i1 (insert_subvector (v32i1 immAllZerosV),
207                                     maskzeroupperv8i1:$src, (iPTR 0))),
208            (COPY_TO_REGCLASS VK8:$src, VK32)>;
209  def : Pat<(v32i1 (insert_subvector (v32i1 immAllZerosV),
210                                     maskzeroupperv16i1:$src, (iPTR 0))),
211            (COPY_TO_REGCLASS VK16:$src, VK32)>;
212
213  def : Pat<(v64i1 (insert_subvector (v64i1 immAllZerosV),
214                                     maskzeroupperv1i1:$src, (iPTR 0))),
215            (COPY_TO_REGCLASS VK1:$src, VK64)>;
216  def : Pat<(v64i1 (insert_subvector (v64i1 immAllZerosV),
217                                     maskzeroupperv8i1:$src, (iPTR 0))),
218            (COPY_TO_REGCLASS VK8:$src, VK64)>;
219  def : Pat<(v64i1 (insert_subvector (v64i1 immAllZerosV),
220                                     maskzeroupperv16i1:$src, (iPTR 0))),
221            (COPY_TO_REGCLASS VK16:$src, VK64)>;
222  def : Pat<(v64i1 (insert_subvector (v64i1 immAllZerosV),
223                                     maskzeroupperv32i1:$src, (iPTR 0))),
224            (COPY_TO_REGCLASS VK32:$src, VK64)>;
225}
226
227let Predicates = [HasAVX512] in {
228  def : Pat<(v16i1 (insert_subvector (v16i1 immAllZerosV),
229                                     maskzeroupperv1i1:$src, (iPTR 0))),
230            (COPY_TO_REGCLASS VK1:$src, VK16)>;
231  def : Pat<(v16i1 (insert_subvector (v16i1 immAllZerosV),
232                                     maskzeroupperv8i1:$src, (iPTR 0))),
233            (COPY_TO_REGCLASS VK8:$src, VK16)>;
234}
235
236let Predicates = [HasDQI] in {
237  def : Pat<(v8i1 (insert_subvector (v8i1 immAllZerosV),
238                                    maskzeroupperv1i1:$src, (iPTR 0))),
239            (COPY_TO_REGCLASS VK1:$src, VK8)>;
240}
241
242let Predicates = [HasVLX, HasDQI] in {
243  def : Pat<(v8i1 (insert_subvector (v8i1 immAllZerosV),
244                                    maskzeroupperv2i1:$src, (iPTR 0))),
245            (COPY_TO_REGCLASS VK2:$src, VK8)>;
246  def : Pat<(v8i1 (insert_subvector (v8i1 immAllZerosV),
247                                    maskzeroupperv4i1:$src, (iPTR 0))),
248            (COPY_TO_REGCLASS VK4:$src, VK8)>;
249}
250
251let Predicates = [HasVLX] in {
252  def : Pat<(v16i1 (insert_subvector (v16i1 immAllZerosV),
253                                     maskzeroupperv2i1:$src, (iPTR 0))),
254            (COPY_TO_REGCLASS VK2:$src, VK16)>;
255  def : Pat<(v16i1 (insert_subvector (v16i1 immAllZerosV),
256                                     maskzeroupperv4i1:$src, (iPTR 0))),
257            (COPY_TO_REGCLASS VK4:$src, VK16)>;
258}
259
260let Predicates = [HasBWI, HasVLX] in {
261  def : Pat<(v32i1 (insert_subvector (v32i1 immAllZerosV),
262                                     maskzeroupperv2i1:$src, (iPTR 0))),
263            (COPY_TO_REGCLASS VK2:$src, VK32)>;
264  def : Pat<(v32i1 (insert_subvector (v32i1 immAllZerosV),
265                                     maskzeroupperv4i1:$src, (iPTR 0))),
266            (COPY_TO_REGCLASS VK4:$src, VK32)>;
267  def : Pat<(v64i1 (insert_subvector (v64i1 immAllZerosV),
268                                     maskzeroupperv2i1:$src, (iPTR 0))),
269            (COPY_TO_REGCLASS VK2:$src, VK64)>;
270  def : Pat<(v64i1 (insert_subvector (v64i1 immAllZerosV),
271                                     maskzeroupperv4i1:$src, (iPTR 0))),
272            (COPY_TO_REGCLASS VK4:$src, VK64)>;
273}
274
275// If the bits are not zero we have to fall back to explicitly zeroing by
276// using shifts.
277let Predicates = [HasAVX512] in {
278  def : Pat<(v16i1 (insert_subvector (v16i1 immAllZerosV),
279                                     (v1i1 VK1:$mask), (iPTR 0))),
280            (KSHIFTRWri (KSHIFTLWri (COPY_TO_REGCLASS VK1:$mask, VK16),
281                                    (i8 15)), (i8 15))>;
282
283  def : Pat<(v16i1 (insert_subvector (v16i1 immAllZerosV),
284                                     (v2i1 VK2:$mask), (iPTR 0))),
285            (KSHIFTRWri (KSHIFTLWri (COPY_TO_REGCLASS VK2:$mask, VK16),
286                                    (i8 14)), (i8 14))>;
287
288  def : Pat<(v16i1 (insert_subvector (v16i1 immAllZerosV),
289                                     (v4i1 VK4:$mask), (iPTR 0))),
290            (KSHIFTRWri (KSHIFTLWri (COPY_TO_REGCLASS VK4:$mask, VK16),
291                                    (i8 12)), (i8 12))>;
292}
293
294let Predicates = [HasAVX512, NoDQI] in {
295  def : Pat<(v16i1 (insert_subvector (v16i1 immAllZerosV),
296                                     (v8i1 VK8:$mask), (iPTR 0))),
297            (KSHIFTRWri (KSHIFTLWri (COPY_TO_REGCLASS VK8:$mask, VK16),
298                                    (i8 8)), (i8 8))>;
299}
300
301let Predicates = [HasDQI] in {
302  def : Pat<(v16i1 (insert_subvector (v16i1 immAllZerosV),
303                                     (v8i1 VK8:$mask), (iPTR 0))),
304            (COPY_TO_REGCLASS (KMOVBkk VK8:$mask), VK16)>;
305
306  def : Pat<(v8i1 (insert_subvector (v8i1 immAllZerosV),
307                                    (v1i1 VK1:$mask), (iPTR 0))),
308            (KSHIFTRBri (KSHIFTLBri (COPY_TO_REGCLASS VK1:$mask, VK8),
309                                    (i8 7)), (i8 7))>;
310  def : Pat<(v8i1 (insert_subvector (v8i1 immAllZerosV),
311                                    (v2i1 VK2:$mask), (iPTR 0))),
312            (KSHIFTRBri (KSHIFTLBri (COPY_TO_REGCLASS VK2:$mask, VK8),
313                                    (i8 6)), (i8 6))>;
314  def : Pat<(v8i1 (insert_subvector (v8i1 immAllZerosV),
315                                    (v4i1 VK4:$mask), (iPTR 0))),
316            (KSHIFTRBri (KSHIFTLBri (COPY_TO_REGCLASS VK4:$mask, VK8),
317                                    (i8 4)), (i8 4))>;
318}
319
320let Predicates = [HasBWI] in {
321  def : Pat<(v32i1 (insert_subvector (v32i1 immAllZerosV),
322                                     (v16i1 VK16:$mask), (iPTR 0))),
323            (COPY_TO_REGCLASS (KMOVWkk VK16:$mask), VK32)>;
324
325  def : Pat<(v64i1 (insert_subvector (v64i1 immAllZerosV),
326                                     (v16i1 VK16:$mask), (iPTR 0))),
327            (COPY_TO_REGCLASS (KMOVWkk VK16:$mask), VK64)>;
328  def : Pat<(v64i1 (insert_subvector (v64i1 immAllZerosV),
329                                     (v32i1 VK32:$mask), (iPTR 0))),
330            (COPY_TO_REGCLASS (KMOVDkk VK32:$mask), VK64)>;
331}
332
333let Predicates = [HasBWI, NoDQI] in {
334  def : Pat<(v32i1 (insert_subvector (v32i1 immAllZerosV),
335                                     (v8i1 VK8:$mask), (iPTR 0))),
336            (KSHIFTRDri (KSHIFTLDri (COPY_TO_REGCLASS VK8:$mask, VK32),
337                                    (i8 24)), (i8 24))>;
338
339  def : Pat<(v64i1 (insert_subvector (v64i1 immAllZerosV),
340                                     (v8i1 VK8:$mask), (iPTR 0))),
341            (KSHIFTRQri (KSHIFTLQri (COPY_TO_REGCLASS VK8:$mask, VK64),
342                                    (i8 56)), (i8 56))>;
343}
344
345let Predicates = [HasBWI, HasDQI] in {
346  def : Pat<(v32i1 (insert_subvector (v32i1 immAllZerosV),
347                                     (v8i1 VK8:$mask), (iPTR 0))),
348            (COPY_TO_REGCLASS (KMOVBkk VK8:$mask), VK32)>;
349
350  def : Pat<(v64i1 (insert_subvector (v64i1 immAllZerosV),
351                                     (v8i1 VK8:$mask), (iPTR 0))),
352            (COPY_TO_REGCLASS (KMOVBkk VK8:$mask), VK64)>;
353}
354
355let Predicates = [HasBWI] in {
356  def : Pat<(v32i1 (insert_subvector (v32i1 immAllZerosV),
357                                     (v1i1 VK1:$mask), (iPTR 0))),
358            (KSHIFTRDri (KSHIFTLDri (COPY_TO_REGCLASS VK1:$mask, VK32),
359                                    (i8 31)), (i8 31))>;
360  def : Pat<(v32i1 (insert_subvector (v32i1 immAllZerosV),
361                                     (v2i1 VK2:$mask), (iPTR 0))),
362            (KSHIFTRDri (KSHIFTLDri (COPY_TO_REGCLASS VK2:$mask, VK32),
363                                    (i8 30)), (i8 30))>;
364  def : Pat<(v32i1 (insert_subvector (v32i1 immAllZerosV),
365                                     (v4i1 VK4:$mask), (iPTR 0))),
366            (KSHIFTRDri (KSHIFTLDri (COPY_TO_REGCLASS VK4:$mask, VK32),
367                                    (i8 28)), (i8 28))>;
368
369  def : Pat<(v64i1 (insert_subvector (v64i1 immAllZerosV),
370                                     (v1i1 VK1:$mask), (iPTR 0))),
371            (KSHIFTRQri (KSHIFTLQri (COPY_TO_REGCLASS VK1:$mask, VK64),
372                                    (i8 63)), (i8 63))>;
373  def : Pat<(v64i1 (insert_subvector (v64i1 immAllZerosV),
374                                     (v2i1 VK2:$mask), (iPTR 0))),
375            (KSHIFTRQri (KSHIFTLQri (COPY_TO_REGCLASS VK2:$mask, VK64),
376                                    (i8 62)), (i8 62))>;
377  def : Pat<(v64i1 (insert_subvector (v64i1 immAllZerosV),
378                                     (v4i1 VK4:$mask), (iPTR 0))),
379            (KSHIFTRQri (KSHIFTLQri (COPY_TO_REGCLASS VK4:$mask, VK64),
380                                    (i8 60)), (i8 60))>;
381}
382
383//===----------------------------------------------------------------------===//
384// Extra selection patterns for f128, f128mem
385
386// movaps is shorter than movdqa. movaps is in SSE and movdqa is in SSE2.
387let Predicates = [NoAVX] in {
388def : Pat<(alignedstore (f128 VR128:$src), addr:$dst),
389          (MOVAPSmr addr:$dst, VR128:$src)>;
390def : Pat<(store (f128 VR128:$src), addr:$dst),
391          (MOVUPSmr addr:$dst, VR128:$src)>;
392
393def : Pat<(alignedloadf128 addr:$src),
394          (MOVAPSrm addr:$src)>;
395def : Pat<(loadf128 addr:$src),
396          (MOVUPSrm addr:$src)>;
397}
398
399let Predicates = [HasAVX, NoVLX] in {
400def : Pat<(alignedstore (f128 VR128:$src), addr:$dst),
401          (VMOVAPSmr addr:$dst, VR128:$src)>;
402def : Pat<(store (f128 VR128:$src), addr:$dst),
403          (VMOVUPSmr addr:$dst, VR128:$src)>;
404
405def : Pat<(alignedloadf128 addr:$src),
406          (VMOVAPSrm addr:$src)>;
407def : Pat<(loadf128 addr:$src),
408          (VMOVUPSrm addr:$src)>;
409}
410
411let Predicates = [HasVLX] in {
412def : Pat<(alignedstore (f128 VR128X:$src), addr:$dst),
413          (VMOVAPSZ128mr addr:$dst, VR128X:$src)>;
414def : Pat<(store (f128 VR128X:$src), addr:$dst),
415          (VMOVUPSZ128mr addr:$dst, VR128X:$src)>;
416
417def : Pat<(alignedloadf128 addr:$src),
418          (VMOVAPSZ128rm addr:$src)>;
419def : Pat<(loadf128 addr:$src),
420          (VMOVUPSZ128rm addr:$src)>;
421}
422
423let Predicates = [UseSSE1] in {
424// andps is shorter than andpd or pand. andps is SSE and andpd/pand are in SSE2
425def : Pat<(f128 (X86fand VR128:$src1, (memopf128 addr:$src2))),
426          (ANDPSrm VR128:$src1, f128mem:$src2)>;
427
428def : Pat<(f128 (X86fand VR128:$src1, VR128:$src2)),
429          (ANDPSrr VR128:$src1, VR128:$src2)>;
430
431def : Pat<(f128 (X86for VR128:$src1, (memopf128 addr:$src2))),
432          (ORPSrm VR128:$src1, f128mem:$src2)>;
433
434def : Pat<(f128 (X86for VR128:$src1, VR128:$src2)),
435          (ORPSrr VR128:$src1, VR128:$src2)>;
436
437def : Pat<(f128 (X86fxor VR128:$src1, (memopf128 addr:$src2))),
438          (XORPSrm VR128:$src1, f128mem:$src2)>;
439
440def : Pat<(f128 (X86fxor VR128:$src1, VR128:$src2)),
441          (XORPSrr VR128:$src1, VR128:$src2)>;
442}
443
444let Predicates = [HasAVX, NoVLX] in {
445// andps is shorter than andpd or pand. andps is SSE and andpd/pand are in SSE2
446def : Pat<(f128 (X86fand VR128:$src1, (loadf128 addr:$src2))),
447          (VANDPSrm VR128:$src1, f128mem:$src2)>;
448
449def : Pat<(f128 (X86fand VR128:$src1, VR128:$src2)),
450          (VANDPSrr VR128:$src1, VR128:$src2)>;
451
452def : Pat<(f128 (X86for VR128:$src1, (loadf128 addr:$src2))),
453          (VORPSrm VR128:$src1, f128mem:$src2)>;
454
455def : Pat<(f128 (X86for VR128:$src1, VR128:$src2)),
456          (VORPSrr VR128:$src1, VR128:$src2)>;
457
458def : Pat<(f128 (X86fxor VR128:$src1, (loadf128 addr:$src2))),
459          (VXORPSrm VR128:$src1, f128mem:$src2)>;
460
461def : Pat<(f128 (X86fxor VR128:$src1, VR128:$src2)),
462          (VXORPSrr VR128:$src1, VR128:$src2)>;
463}
464
465let Predicates = [HasVLX] in {
466// andps is shorter than andpd or pand. andps is SSE and andpd/pand are in SSE2
467def : Pat<(f128 (X86fand VR128X:$src1, (loadf128 addr:$src2))),
468          (VANDPSZ128rm VR128X:$src1, f128mem:$src2)>;
469
470def : Pat<(f128 (X86fand VR128X:$src1, VR128X:$src2)),
471          (VANDPSZ128rr VR128X:$src1, VR128X:$src2)>;
472
473def : Pat<(f128 (X86for VR128X:$src1, (loadf128 addr:$src2))),
474          (VORPSZ128rm VR128X:$src1, f128mem:$src2)>;
475
476def : Pat<(f128 (X86for VR128X:$src1, VR128X:$src2)),
477          (VORPSZ128rr VR128X:$src1, VR128X:$src2)>;
478
479def : Pat<(f128 (X86fxor VR128X:$src1, (loadf128 addr:$src2))),
480          (VXORPSZ128rm VR128X:$src1, f128mem:$src2)>;
481
482def : Pat<(f128 (X86fxor VR128X:$src1, VR128X:$src2)),
483          (VXORPSZ128rr VR128X:$src1, VR128X:$src2)>;
484}
485