xref: /freebsd/contrib/llvm-project/llvm/lib/Target/X86/X86InstrVecCompiler.td (revision e0c4386e7e71d93b0edc0c8fa156263fc4a8b0b6)
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 = [HasVLX] in {
134  defm : subvec_zero_lowering<"APDZ128", VR128X, v4f64, v2f64, sub_xmm>;
135  defm : subvec_zero_lowering<"APSZ128", VR128X, v8f32, v4f32, sub_xmm>;
136  defm : subvec_zero_lowering<"DQA64Z128", VR128X, v4i64, v2i64, sub_xmm>;
137  defm : subvec_zero_lowering<"DQA64Z128", VR128X, v8i32, v4i32, sub_xmm>;
138  defm : subvec_zero_lowering<"DQA64Z128", VR128X, v16i16, v8i16, sub_xmm>;
139  defm : subvec_zero_lowering<"DQA64Z128", VR128X, v32i8, v16i8, sub_xmm>;
140
141  defm : subvec_zero_lowering<"APDZ128", VR128X, v8f64, v2f64, sub_xmm>;
142  defm : subvec_zero_lowering<"APSZ128", VR128X, v16f32, v4f32, sub_xmm>;
143  defm : subvec_zero_lowering<"DQA64Z128", VR128X, v8i64, v2i64, sub_xmm>;
144  defm : subvec_zero_lowering<"DQA64Z128", VR128X, v16i32, v4i32, sub_xmm>;
145  defm : subvec_zero_lowering<"DQA64Z128", VR128X, v32i16, v8i16, sub_xmm>;
146  defm : subvec_zero_lowering<"DQA64Z128", VR128X, v64i8, v16i8, sub_xmm>;
147
148  defm : subvec_zero_lowering<"APDZ256", VR256X, v8f64, v4f64, sub_ymm>;
149  defm : subvec_zero_lowering<"APSZ256", VR256X, v16f32, v8f32, sub_ymm>;
150  defm : subvec_zero_lowering<"DQA64Z256", VR256X, v8i64, v4i64, sub_ymm>;
151  defm : subvec_zero_lowering<"DQA64Z256", VR256X, v16i32, v8i32, sub_ymm>;
152  defm : subvec_zero_lowering<"DQA64Z256", VR256X, v32i16, v16i16, sub_ymm>;
153  defm : subvec_zero_lowering<"DQA64Z256", VR256X, v64i8, v32i8, sub_ymm>;
154}
155
156let Predicates = [HasAVX512, NoVLX] in {
157  defm : subvec_zero_lowering<"APD", VR128, v8f64, v2f64, sub_xmm>;
158  defm : subvec_zero_lowering<"APS", VR128, v16f32, v4f32, sub_xmm>;
159  defm : subvec_zero_lowering<"DQA", VR128, v8i64, v2i64, sub_xmm>;
160  defm : subvec_zero_lowering<"DQA", VR128, v16i32, v4i32, sub_xmm>;
161  defm : subvec_zero_lowering<"DQA", VR128, v32i16, v8i16, sub_xmm>;
162  defm : subvec_zero_lowering<"DQA", VR128, v64i8, v16i8, sub_xmm>;
163
164  defm : subvec_zero_lowering<"APDY", VR256, v8f64, v4f64, sub_ymm>;
165  defm : subvec_zero_lowering<"APSY", VR256, v16f32, v8f32, sub_ymm>;
166  defm : subvec_zero_lowering<"DQAY", VR256, v8i64, v4i64, sub_ymm>;
167  defm : subvec_zero_lowering<"DQAY", VR256, v16i32, v8i32, sub_ymm>;
168  defm : subvec_zero_lowering<"DQAY", VR256, v32i16, v16i16, sub_ymm>;
169  defm : subvec_zero_lowering<"DQAY", VR256, v64i8, v32i8, sub_ymm>;
170}
171
172let Predicates = [HasFP16, HasVLX] in {
173  defm : subvec_zero_lowering<"APSZ128", VR128X, v16f16, v8f16, sub_xmm>;
174  defm : subvec_zero_lowering<"APSZ128", VR128X, v32f16, v8f16, sub_xmm>;
175  defm : subvec_zero_lowering<"APSZ256", VR256X, v32f16, v16f16, sub_ymm>;
176}
177
178class maskzeroupper<ValueType vt, RegisterClass RC> :
179  PatLeaf<(vt RC:$src), [{
180    return isMaskZeroExtended(N);
181  }]>;
182
183def maskzeroupperv1i1  : maskzeroupper<v1i1,  VK1>;
184def maskzeroupperv2i1  : maskzeroupper<v2i1,  VK2>;
185def maskzeroupperv4i1  : maskzeroupper<v4i1,  VK4>;
186def maskzeroupperv8i1  : maskzeroupper<v8i1,  VK8>;
187def maskzeroupperv16i1 : maskzeroupper<v16i1, VK16>;
188def maskzeroupperv32i1 : maskzeroupper<v32i1, VK32>;
189
190// The patterns determine if we can depend on the upper bits of a mask register
191// being zeroed by the previous operation so that we can skip explicit
192// zeroing.
193let Predicates = [HasBWI] in {
194  def : Pat<(v32i1 (insert_subvector (v32i1 immAllZerosV),
195                                     maskzeroupperv1i1:$src, (iPTR 0))),
196            (COPY_TO_REGCLASS VK1:$src, VK32)>;
197  def : Pat<(v32i1 (insert_subvector (v32i1 immAllZerosV),
198                                     maskzeroupperv8i1:$src, (iPTR 0))),
199            (COPY_TO_REGCLASS VK8:$src, VK32)>;
200  def : Pat<(v32i1 (insert_subvector (v32i1 immAllZerosV),
201                                     maskzeroupperv16i1:$src, (iPTR 0))),
202            (COPY_TO_REGCLASS VK16:$src, VK32)>;
203
204  def : Pat<(v64i1 (insert_subvector (v64i1 immAllZerosV),
205                                     maskzeroupperv1i1:$src, (iPTR 0))),
206            (COPY_TO_REGCLASS VK1:$src, VK64)>;
207  def : Pat<(v64i1 (insert_subvector (v64i1 immAllZerosV),
208                                     maskzeroupperv8i1:$src, (iPTR 0))),
209            (COPY_TO_REGCLASS VK8:$src, VK64)>;
210  def : Pat<(v64i1 (insert_subvector (v64i1 immAllZerosV),
211                                     maskzeroupperv16i1:$src, (iPTR 0))),
212            (COPY_TO_REGCLASS VK16:$src, VK64)>;
213  def : Pat<(v64i1 (insert_subvector (v64i1 immAllZerosV),
214                                     maskzeroupperv32i1:$src, (iPTR 0))),
215            (COPY_TO_REGCLASS VK32:$src, VK64)>;
216}
217
218let Predicates = [HasAVX512] in {
219  def : Pat<(v16i1 (insert_subvector (v16i1 immAllZerosV),
220                                     maskzeroupperv1i1:$src, (iPTR 0))),
221            (COPY_TO_REGCLASS VK1:$src, VK16)>;
222  def : Pat<(v16i1 (insert_subvector (v16i1 immAllZerosV),
223                                     maskzeroupperv8i1:$src, (iPTR 0))),
224            (COPY_TO_REGCLASS VK8:$src, VK16)>;
225}
226
227let Predicates = [HasDQI] in {
228  def : Pat<(v8i1 (insert_subvector (v8i1 immAllZerosV),
229                                    maskzeroupperv1i1:$src, (iPTR 0))),
230            (COPY_TO_REGCLASS VK1:$src, VK8)>;
231}
232
233let Predicates = [HasVLX, HasDQI] in {
234  def : Pat<(v8i1 (insert_subvector (v8i1 immAllZerosV),
235                                    maskzeroupperv2i1:$src, (iPTR 0))),
236            (COPY_TO_REGCLASS VK2:$src, VK8)>;
237  def : Pat<(v8i1 (insert_subvector (v8i1 immAllZerosV),
238                                    maskzeroupperv4i1:$src, (iPTR 0))),
239            (COPY_TO_REGCLASS VK4:$src, VK8)>;
240}
241
242let Predicates = [HasVLX] in {
243  def : Pat<(v16i1 (insert_subvector (v16i1 immAllZerosV),
244                                     maskzeroupperv2i1:$src, (iPTR 0))),
245            (COPY_TO_REGCLASS VK2:$src, VK16)>;
246  def : Pat<(v16i1 (insert_subvector (v16i1 immAllZerosV),
247                                     maskzeroupperv4i1:$src, (iPTR 0))),
248            (COPY_TO_REGCLASS VK4:$src, VK16)>;
249}
250
251let Predicates = [HasBWI, HasVLX] in {
252  def : Pat<(v32i1 (insert_subvector (v32i1 immAllZerosV),
253                                     maskzeroupperv2i1:$src, (iPTR 0))),
254            (COPY_TO_REGCLASS VK2:$src, VK32)>;
255  def : Pat<(v32i1 (insert_subvector (v32i1 immAllZerosV),
256                                     maskzeroupperv4i1:$src, (iPTR 0))),
257            (COPY_TO_REGCLASS VK4:$src, VK32)>;
258  def : Pat<(v64i1 (insert_subvector (v64i1 immAllZerosV),
259                                     maskzeroupperv2i1:$src, (iPTR 0))),
260            (COPY_TO_REGCLASS VK2:$src, VK64)>;
261  def : Pat<(v64i1 (insert_subvector (v64i1 immAllZerosV),
262                                     maskzeroupperv4i1:$src, (iPTR 0))),
263            (COPY_TO_REGCLASS VK4:$src, VK64)>;
264}
265
266// If the bits are not zero we have to fall back to explicitly zeroing by
267// using shifts.
268let Predicates = [HasAVX512] in {
269  def : Pat<(v16i1 (insert_subvector (v16i1 immAllZerosV),
270                                     (v1i1 VK1:$mask), (iPTR 0))),
271            (KSHIFTRWri (KSHIFTLWri (COPY_TO_REGCLASS VK1:$mask, VK16),
272                                    (i8 15)), (i8 15))>;
273
274  def : Pat<(v16i1 (insert_subvector (v16i1 immAllZerosV),
275                                     (v2i1 VK2:$mask), (iPTR 0))),
276            (KSHIFTRWri (KSHIFTLWri (COPY_TO_REGCLASS VK2:$mask, VK16),
277                                    (i8 14)), (i8 14))>;
278
279  def : Pat<(v16i1 (insert_subvector (v16i1 immAllZerosV),
280                                     (v4i1 VK4:$mask), (iPTR 0))),
281            (KSHIFTRWri (KSHIFTLWri (COPY_TO_REGCLASS VK4:$mask, VK16),
282                                    (i8 12)), (i8 12))>;
283}
284
285let Predicates = [HasAVX512, NoDQI] in {
286  def : Pat<(v16i1 (insert_subvector (v16i1 immAllZerosV),
287                                     (v8i1 VK8:$mask), (iPTR 0))),
288            (KSHIFTRWri (KSHIFTLWri (COPY_TO_REGCLASS VK8:$mask, VK16),
289                                    (i8 8)), (i8 8))>;
290}
291
292let Predicates = [HasDQI] in {
293  def : Pat<(v16i1 (insert_subvector (v16i1 immAllZerosV),
294                                     (v8i1 VK8:$mask), (iPTR 0))),
295            (COPY_TO_REGCLASS (KMOVBkk VK8:$mask), VK16)>;
296
297  def : Pat<(v8i1 (insert_subvector (v8i1 immAllZerosV),
298                                    (v1i1 VK1:$mask), (iPTR 0))),
299            (KSHIFTRBri (KSHIFTLBri (COPY_TO_REGCLASS VK1:$mask, VK8),
300                                    (i8 7)), (i8 7))>;
301  def : Pat<(v8i1 (insert_subvector (v8i1 immAllZerosV),
302                                    (v2i1 VK2:$mask), (iPTR 0))),
303            (KSHIFTRBri (KSHIFTLBri (COPY_TO_REGCLASS VK2:$mask, VK8),
304                                    (i8 6)), (i8 6))>;
305  def : Pat<(v8i1 (insert_subvector (v8i1 immAllZerosV),
306                                    (v4i1 VK4:$mask), (iPTR 0))),
307            (KSHIFTRBri (KSHIFTLBri (COPY_TO_REGCLASS VK4:$mask, VK8),
308                                    (i8 4)), (i8 4))>;
309}
310
311let Predicates = [HasBWI] in {
312  def : Pat<(v32i1 (insert_subvector (v32i1 immAllZerosV),
313                                     (v16i1 VK16:$mask), (iPTR 0))),
314            (COPY_TO_REGCLASS (KMOVWkk VK16:$mask), VK32)>;
315
316  def : Pat<(v64i1 (insert_subvector (v64i1 immAllZerosV),
317                                     (v16i1 VK16:$mask), (iPTR 0))),
318            (COPY_TO_REGCLASS (KMOVWkk VK16:$mask), VK64)>;
319  def : Pat<(v64i1 (insert_subvector (v64i1 immAllZerosV),
320                                     (v32i1 VK32:$mask), (iPTR 0))),
321            (COPY_TO_REGCLASS (KMOVDkk VK32:$mask), VK64)>;
322}
323
324let Predicates = [HasBWI, NoDQI] in {
325  def : Pat<(v32i1 (insert_subvector (v32i1 immAllZerosV),
326                                     (v8i1 VK8:$mask), (iPTR 0))),
327            (KSHIFTRDri (KSHIFTLDri (COPY_TO_REGCLASS VK8:$mask, VK32),
328                                    (i8 24)), (i8 24))>;
329
330  def : Pat<(v64i1 (insert_subvector (v64i1 immAllZerosV),
331                                     (v8i1 VK8:$mask), (iPTR 0))),
332            (KSHIFTRQri (KSHIFTLQri (COPY_TO_REGCLASS VK8:$mask, VK64),
333                                    (i8 56)), (i8 56))>;
334}
335
336let Predicates = [HasBWI, HasDQI] in {
337  def : Pat<(v32i1 (insert_subvector (v32i1 immAllZerosV),
338                                     (v8i1 VK8:$mask), (iPTR 0))),
339            (COPY_TO_REGCLASS (KMOVBkk VK8:$mask), VK32)>;
340
341  def : Pat<(v64i1 (insert_subvector (v64i1 immAllZerosV),
342                                     (v8i1 VK8:$mask), (iPTR 0))),
343            (COPY_TO_REGCLASS (KMOVBkk VK8:$mask), VK64)>;
344}
345
346let Predicates = [HasBWI] in {
347  def : Pat<(v32i1 (insert_subvector (v32i1 immAllZerosV),
348                                     (v1i1 VK1:$mask), (iPTR 0))),
349            (KSHIFTRDri (KSHIFTLDri (COPY_TO_REGCLASS VK1:$mask, VK32),
350                                    (i8 31)), (i8 31))>;
351  def : Pat<(v32i1 (insert_subvector (v32i1 immAllZerosV),
352                                     (v2i1 VK2:$mask), (iPTR 0))),
353            (KSHIFTRDri (KSHIFTLDri (COPY_TO_REGCLASS VK2:$mask, VK32),
354                                    (i8 30)), (i8 30))>;
355  def : Pat<(v32i1 (insert_subvector (v32i1 immAllZerosV),
356                                     (v4i1 VK4:$mask), (iPTR 0))),
357            (KSHIFTRDri (KSHIFTLDri (COPY_TO_REGCLASS VK4:$mask, VK32),
358                                    (i8 28)), (i8 28))>;
359
360  def : Pat<(v64i1 (insert_subvector (v64i1 immAllZerosV),
361                                     (v1i1 VK1:$mask), (iPTR 0))),
362            (KSHIFTRQri (KSHIFTLQri (COPY_TO_REGCLASS VK1:$mask, VK64),
363                                    (i8 63)), (i8 63))>;
364  def : Pat<(v64i1 (insert_subvector (v64i1 immAllZerosV),
365                                     (v2i1 VK2:$mask), (iPTR 0))),
366            (KSHIFTRQri (KSHIFTLQri (COPY_TO_REGCLASS VK2:$mask, VK64),
367                                    (i8 62)), (i8 62))>;
368  def : Pat<(v64i1 (insert_subvector (v64i1 immAllZerosV),
369                                     (v4i1 VK4:$mask), (iPTR 0))),
370            (KSHIFTRQri (KSHIFTLQri (COPY_TO_REGCLASS VK4:$mask, VK64),
371                                    (i8 60)), (i8 60))>;
372}
373
374//===----------------------------------------------------------------------===//
375// Extra selection patterns for f128, f128mem
376
377// movaps is shorter than movdqa. movaps is in SSE and movdqa is in SSE2.
378let Predicates = [NoAVX] in {
379def : Pat<(alignedstore (f128 VR128:$src), addr:$dst),
380          (MOVAPSmr addr:$dst, VR128:$src)>;
381def : Pat<(store (f128 VR128:$src), addr:$dst),
382          (MOVUPSmr addr:$dst, VR128:$src)>;
383
384def : Pat<(alignedloadf128 addr:$src),
385          (MOVAPSrm addr:$src)>;
386def : Pat<(loadf128 addr:$src),
387          (MOVUPSrm addr:$src)>;
388}
389
390let Predicates = [HasAVX, NoVLX] in {
391def : Pat<(alignedstore (f128 VR128:$src), addr:$dst),
392          (VMOVAPSmr addr:$dst, VR128:$src)>;
393def : Pat<(store (f128 VR128:$src), addr:$dst),
394          (VMOVUPSmr addr:$dst, VR128:$src)>;
395
396def : Pat<(alignedloadf128 addr:$src),
397          (VMOVAPSrm addr:$src)>;
398def : Pat<(loadf128 addr:$src),
399          (VMOVUPSrm addr:$src)>;
400}
401
402let Predicates = [HasVLX] in {
403def : Pat<(alignedstore (f128 VR128X:$src), addr:$dst),
404          (VMOVAPSZ128mr addr:$dst, VR128X:$src)>;
405def : Pat<(store (f128 VR128X:$src), addr:$dst),
406          (VMOVUPSZ128mr addr:$dst, VR128X:$src)>;
407
408def : Pat<(alignedloadf128 addr:$src),
409          (VMOVAPSZ128rm addr:$src)>;
410def : Pat<(loadf128 addr:$src),
411          (VMOVUPSZ128rm addr:$src)>;
412}
413
414let Predicates = [UseSSE1] in {
415// andps is shorter than andpd or pand. andps is SSE and andpd/pand are in SSE2
416def : Pat<(f128 (X86fand VR128:$src1, (memopf128 addr:$src2))),
417          (ANDPSrm VR128:$src1, f128mem:$src2)>;
418
419def : Pat<(f128 (X86fand VR128:$src1, VR128:$src2)),
420          (ANDPSrr VR128:$src1, VR128:$src2)>;
421
422def : Pat<(f128 (X86for VR128:$src1, (memopf128 addr:$src2))),
423          (ORPSrm VR128:$src1, f128mem:$src2)>;
424
425def : Pat<(f128 (X86for VR128:$src1, VR128:$src2)),
426          (ORPSrr VR128:$src1, VR128:$src2)>;
427
428def : Pat<(f128 (X86fxor VR128:$src1, (memopf128 addr:$src2))),
429          (XORPSrm VR128:$src1, f128mem:$src2)>;
430
431def : Pat<(f128 (X86fxor VR128:$src1, VR128:$src2)),
432          (XORPSrr VR128:$src1, VR128:$src2)>;
433}
434
435let Predicates = [HasAVX, NoVLX] in {
436// andps is shorter than andpd or pand. andps is SSE and andpd/pand are in SSE2
437def : Pat<(f128 (X86fand VR128:$src1, (loadf128 addr:$src2))),
438          (VANDPSrm VR128:$src1, f128mem:$src2)>;
439
440def : Pat<(f128 (X86fand VR128:$src1, VR128:$src2)),
441          (VANDPSrr VR128:$src1, VR128:$src2)>;
442
443def : Pat<(f128 (X86for VR128:$src1, (loadf128 addr:$src2))),
444          (VORPSrm VR128:$src1, f128mem:$src2)>;
445
446def : Pat<(f128 (X86for VR128:$src1, VR128:$src2)),
447          (VORPSrr VR128:$src1, VR128:$src2)>;
448
449def : Pat<(f128 (X86fxor VR128:$src1, (loadf128 addr:$src2))),
450          (VXORPSrm VR128:$src1, f128mem:$src2)>;
451
452def : Pat<(f128 (X86fxor VR128:$src1, VR128:$src2)),
453          (VXORPSrr VR128:$src1, VR128:$src2)>;
454}
455
456let Predicates = [HasVLX] in {
457// andps is shorter than andpd or pand. andps is SSE and andpd/pand are in SSE2
458def : Pat<(f128 (X86fand VR128X:$src1, (loadf128 addr:$src2))),
459          (VANDPSZ128rm VR128X:$src1, f128mem:$src2)>;
460
461def : Pat<(f128 (X86fand VR128X:$src1, VR128X:$src2)),
462          (VANDPSZ128rr VR128X:$src1, VR128X:$src2)>;
463
464def : Pat<(f128 (X86for VR128X:$src1, (loadf128 addr:$src2))),
465          (VORPSZ128rm VR128X:$src1, f128mem:$src2)>;
466
467def : Pat<(f128 (X86for VR128X:$src1, VR128X:$src2)),
468          (VORPSZ128rr VR128X:$src1, VR128X:$src2)>;
469
470def : Pat<(f128 (X86fxor VR128X:$src1, (loadf128 addr:$src2))),
471          (VXORPSZ128rm VR128X:$src1, f128mem:$src2)>;
472
473def : Pat<(f128 (X86fxor VR128X:$src1, VR128X:$src2)),
474          (VXORPSZ128rr VR128X:$src1, VR128X:$src2)>;
475}
476