1 | //===- X86ReplaceableInstrs.def ----------------------------------*- C++ -*-==// |
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 | // These are the replaceable SSE instructions. Some of these have Int variants |
10 | // that we don't include here. We don't want to replace instructions selected |
11 | // by intrinsics. |
12 | |
13 | #define ENTRY(A, B, C) {X86::A, X86::B, X86::C}, |
14 | static const uint16_t ReplaceableInstrs[][3] = { |
15 | // PackedSingle, PackedDouble, PackedInt |
16 | ENTRY(MOVAPSmr, MOVAPDmr, MOVDQAmr) |
17 | ENTRY(MOVAPSrm, MOVAPDrm, MOVDQArm) |
18 | ENTRY(MOVAPSrr, MOVAPDrr, MOVDQArr) |
19 | ENTRY(MOVUPSmr, MOVUPDmr, MOVDQUmr) |
20 | ENTRY(MOVUPSrm, MOVUPDrm, MOVDQUrm) |
21 | ENTRY(MOVLPSmr, MOVLPDmr, MOVPQI2QImr) |
22 | ENTRY(MOVSDmr, MOVSDmr, MOVPQI2QImr) |
23 | ENTRY(MOVSSmr, MOVSSmr, MOVPDI2DImr) |
24 | ENTRY(MOVSDrm, MOVSDrm, MOVQI2PQIrm) |
25 | ENTRY(MOVSDrm_alt, MOVSDrm_alt, MOVQI2PQIrm) |
26 | ENTRY(MOVSSrm, MOVSSrm, MOVDI2PDIrm) |
27 | ENTRY(MOVSSrm_alt, MOVSSrm_alt, MOVDI2PDIrm) |
28 | ENTRY(MOVNTPSmr, MOVNTPDmr, MOVNTDQmr) |
29 | ENTRY(ANDNPSrm, ANDNPDrm, PANDNrm) |
30 | ENTRY(ANDNPSrr, ANDNPDrr, PANDNrr) |
31 | ENTRY(ANDPSrm, ANDPDrm, PANDrm) |
32 | ENTRY(ANDPSrr, ANDPDrr, PANDrr) |
33 | ENTRY(ORPSrm, ORPDrm, PORrm) |
34 | ENTRY(ORPSrr, ORPDrr, PORrr) |
35 | ENTRY(XORPSrm, XORPDrm, PXORrm) |
36 | ENTRY(XORPSrr, XORPDrr, PXORrr) |
37 | ENTRY(UNPCKLPDrm, UNPCKLPDrm, PUNPCKLQDQrm) |
38 | ENTRY(MOVLHPSrr, UNPCKLPDrr, PUNPCKLQDQrr) |
39 | ENTRY(UNPCKHPDrm, UNPCKHPDrm, PUNPCKHQDQrm) |
40 | ENTRY(UNPCKHPDrr, UNPCKHPDrr, PUNPCKHQDQrr) |
41 | ENTRY(UNPCKLPSrm, UNPCKLPSrm, PUNPCKLDQrm) |
42 | ENTRY(UNPCKLPSrr, UNPCKLPSrr, PUNPCKLDQrr) |
43 | ENTRY(UNPCKHPSrm, UNPCKHPSrm, PUNPCKHDQrm) |
44 | ENTRY(UNPCKHPSrr, UNPCKHPSrr, PUNPCKHDQrr) |
45 | ENTRY(EXTRACTPSmr, EXTRACTPSmr, PEXTRDmr) |
46 | ENTRY(EXTRACTPSrr, EXTRACTPSrr, PEXTRDrr) |
47 | // AVX 128-bit support |
48 | ENTRY(VMOVAPSmr, VMOVAPDmr, VMOVDQAmr) |
49 | ENTRY(VMOVAPSrm, VMOVAPDrm, VMOVDQArm) |
50 | ENTRY(VMOVAPSrr, VMOVAPDrr, VMOVDQArr) |
51 | ENTRY(VMOVUPSmr, VMOVUPDmr, VMOVDQUmr) |
52 | ENTRY(VMOVUPSrm, VMOVUPDrm, VMOVDQUrm) |
53 | ENTRY(VMOVLPSmr, VMOVLPDmr, VMOVPQI2QImr) |
54 | ENTRY(VMOVSDmr, VMOVSDmr, VMOVPQI2QImr) |
55 | ENTRY(VMOVSSmr, VMOVSSmr, VMOVPDI2DImr) |
56 | ENTRY(VMOVSDrm, VMOVSDrm, VMOVQI2PQIrm) |
57 | ENTRY(VMOVSDrm_alt, VMOVSDrm_alt, VMOVQI2PQIrm) |
58 | ENTRY(VMOVSSrm, VMOVSSrm, VMOVDI2PDIrm) |
59 | ENTRY(VMOVSSrm_alt, VMOVSSrm_alt, VMOVDI2PDIrm) |
60 | ENTRY(VMOVNTPSmr, VMOVNTPDmr, VMOVNTDQmr) |
61 | ENTRY(VANDNPSrm, VANDNPDrm, VPANDNrm) |
62 | ENTRY(VANDNPSrr, VANDNPDrr, VPANDNrr) |
63 | ENTRY(VANDPSrm, VANDPDrm, VPANDrm) |
64 | ENTRY(VANDPSrr, VANDPDrr, VPANDrr) |
65 | ENTRY(VORPSrm, VORPDrm, VPORrm) |
66 | ENTRY(VORPSrr, VORPDrr, VPORrr) |
67 | ENTRY(VXORPSrm, VXORPDrm, VPXORrm) |
68 | ENTRY(VXORPSrr, VXORPDrr, VPXORrr) |
69 | ENTRY(VUNPCKLPDrm, VUNPCKLPDrm, VPUNPCKLQDQrm) |
70 | ENTRY(VMOVLHPSrr, VUNPCKLPDrr, VPUNPCKLQDQrr) |
71 | ENTRY(VUNPCKHPDrm, VUNPCKHPDrm, VPUNPCKHQDQrm) |
72 | ENTRY(VUNPCKHPDrr, VUNPCKHPDrr, VPUNPCKHQDQrr) |
73 | ENTRY(VUNPCKLPSrm, VUNPCKLPSrm, VPUNPCKLDQrm) |
74 | ENTRY(VUNPCKLPSrr, VUNPCKLPSrr, VPUNPCKLDQrr) |
75 | ENTRY(VUNPCKHPSrm, VUNPCKHPSrm, VPUNPCKHDQrm) |
76 | ENTRY(VUNPCKHPSrr, VUNPCKHPSrr, VPUNPCKHDQrr) |
77 | ENTRY(VEXTRACTPSmr, VEXTRACTPSmr, VPEXTRDmr) |
78 | ENTRY(VEXTRACTPSrr, VEXTRACTPSrr, VPEXTRDrr) |
79 | // AVX 256-bit support |
80 | ENTRY(VMOVAPSYmr, VMOVAPDYmr, VMOVDQAYmr) |
81 | ENTRY(VMOVAPSYrm, VMOVAPDYrm, VMOVDQAYrm) |
82 | ENTRY(VMOVAPSYrr, VMOVAPDYrr, VMOVDQAYrr) |
83 | ENTRY(VMOVUPSYmr, VMOVUPDYmr, VMOVDQUYmr) |
84 | ENTRY(VMOVUPSYrm, VMOVUPDYrm, VMOVDQUYrm) |
85 | ENTRY(VMOVNTPSYmr, VMOVNTPDYmr, VMOVNTDQYmr) |
86 | ENTRY(VPERMPSYrm, VPERMPSYrm, VPERMDYrm) |
87 | ENTRY(VPERMPSYrr, VPERMPSYrr, VPERMDYrr) |
88 | ENTRY(VPERMPDYmi, VPERMPDYmi, VPERMQYmi) |
89 | ENTRY(VPERMPDYri, VPERMPDYri, VPERMQYri) |
90 | // AVX512 support |
91 | ENTRY(VMOVLPSZ128mr, VMOVLPDZ128mr, VMOVPQI2QIZmr) |
92 | ENTRY(VMOVNTPSZ128mr, VMOVNTPDZ128mr, VMOVNTDQZ128mr) |
93 | ENTRY(VMOVNTPSZ256mr, VMOVNTPDZ256mr, VMOVNTDQZ256mr) |
94 | ENTRY(VMOVNTPSZmr, VMOVNTPDZmr, VMOVNTDQZmr) |
95 | ENTRY(VMOVSDZmr, VMOVSDZmr, VMOVPQI2QIZmr) |
96 | ENTRY(VMOVSSZmr, VMOVSSZmr, VMOVPDI2DIZmr) |
97 | ENTRY(VMOVSDZrm, VMOVSDZrm, VMOVQI2PQIZrm) |
98 | ENTRY(VMOVSDZrm_alt, VMOVSDZrm_alt, VMOVQI2PQIZrm) |
99 | ENTRY(VMOVSSZrm, VMOVSSZrm, VMOVDI2PDIZrm) |
100 | ENTRY(VMOVSSZrm_alt, VMOVSSZrm_alt, VMOVDI2PDIZrm) |
101 | ENTRY(VBROADCASTSSZ128rr, VBROADCASTSSZ128rr, VPBROADCASTDZ128rr) |
102 | ENTRY(VBROADCASTSSZ128rm, VBROADCASTSSZ128rm, VPBROADCASTDZ128rm) |
103 | ENTRY(VBROADCASTSSZ256rr, VBROADCASTSSZ256rr, VPBROADCASTDZ256rr) |
104 | ENTRY(VBROADCASTSSZ256rm, VBROADCASTSSZ256rm, VPBROADCASTDZ256rm) |
105 | ENTRY(VBROADCASTSSZrr, VBROADCASTSSZrr, VPBROADCASTDZrr) |
106 | ENTRY(VBROADCASTSSZrm, VBROADCASTSSZrm, VPBROADCASTDZrm) |
107 | ENTRY(VMOVDDUPZ128rr, VMOVDDUPZ128rr, VPBROADCASTQZ128rr) |
108 | ENTRY(VMOVDDUPZ128rm, VMOVDDUPZ128rm, VPBROADCASTQZ128rm) |
109 | ENTRY(VBROADCASTSDZ256rr, VBROADCASTSDZ256rr, VPBROADCASTQZ256rr) |
110 | ENTRY(VBROADCASTSDZ256rm, VBROADCASTSDZ256rm, VPBROADCASTQZ256rm) |
111 | ENTRY(VBROADCASTSDZrr, VBROADCASTSDZrr, VPBROADCASTQZrr) |
112 | ENTRY(VBROADCASTSDZrm, VBROADCASTSDZrm, VPBROADCASTQZrm) |
113 | ENTRY(VINSERTF32x4Zrr, VINSERTF32x4Zrr, VINSERTI32x4Zrr) |
114 | ENTRY(VINSERTF32x4Zrm, VINSERTF32x4Zrm, VINSERTI32x4Zrm) |
115 | ENTRY(VINSERTF32x8Zrr, VINSERTF32x8Zrr, VINSERTI32x8Zrr) |
116 | ENTRY(VINSERTF32x8Zrm, VINSERTF32x8Zrm, VINSERTI32x8Zrm) |
117 | ENTRY(VINSERTF64x2Zrr, VINSERTF64x2Zrr, VINSERTI64x2Zrr) |
118 | ENTRY(VINSERTF64x2Zrm, VINSERTF64x2Zrm, VINSERTI64x2Zrm) |
119 | ENTRY(VINSERTF64x4Zrr, VINSERTF64x4Zrr, VINSERTI64x4Zrr) |
120 | ENTRY(VINSERTF64x4Zrm, VINSERTF64x4Zrm, VINSERTI64x4Zrm) |
121 | ENTRY(VINSERTF32x4Z256rr, VINSERTF32x4Z256rr, VINSERTI32x4Z256rr) |
122 | ENTRY(VINSERTF32x4Z256rm, VINSERTF32x4Z256rm, VINSERTI32x4Z256rm) |
123 | ENTRY(VINSERTF64x2Z256rr, VINSERTF64x2Z256rr, VINSERTI64x2Z256rr) |
124 | ENTRY(VINSERTF64x2Z256rm, VINSERTF64x2Z256rm, VINSERTI64x2Z256rm) |
125 | ENTRY(VEXTRACTF32x4Zrr, VEXTRACTF32x4Zrr, VEXTRACTI32x4Zrr) |
126 | ENTRY(VEXTRACTF32x4Zmr, VEXTRACTF32x4Zmr, VEXTRACTI32x4Zmr) |
127 | ENTRY(VEXTRACTF32x8Zrr, VEXTRACTF32x8Zrr, VEXTRACTI32x8Zrr) |
128 | ENTRY(VEXTRACTF32x8Zmr, VEXTRACTF32x8Zmr, VEXTRACTI32x8Zmr) |
129 | ENTRY(VEXTRACTF64x2Zrr, VEXTRACTF64x2Zrr, VEXTRACTI64x2Zrr) |
130 | ENTRY(VEXTRACTF64x2Zmr, VEXTRACTF64x2Zmr, VEXTRACTI64x2Zmr) |
131 | ENTRY(VEXTRACTF64x4Zrr, VEXTRACTF64x4Zrr, VEXTRACTI64x4Zrr) |
132 | ENTRY(VEXTRACTF64x4Zmr, VEXTRACTF64x4Zmr, VEXTRACTI64x4Zmr) |
133 | ENTRY(VEXTRACTF32x4Z256rr, VEXTRACTF32x4Z256rr, VEXTRACTI32x4Z256rr) |
134 | ENTRY(VEXTRACTF32x4Z256mr, VEXTRACTF32x4Z256mr, VEXTRACTI32x4Z256mr) |
135 | ENTRY(VEXTRACTF64x2Z256rr, VEXTRACTF64x2Z256rr, VEXTRACTI64x2Z256rr) |
136 | ENTRY(VEXTRACTF64x2Z256mr, VEXTRACTF64x2Z256mr, VEXTRACTI64x2Z256mr) |
137 | ENTRY(VPERMILPSmi, VPERMILPSmi, VPSHUFDmi) |
138 | ENTRY(VPERMILPSri, VPERMILPSri, VPSHUFDri) |
139 | ENTRY(VPERMILPSZ128mi, VPERMILPSZ128mi, VPSHUFDZ128mi) |
140 | ENTRY(VPERMILPSZ128ri, VPERMILPSZ128ri, VPSHUFDZ128ri) |
141 | ENTRY(VPERMILPSZ256mi, VPERMILPSZ256mi, VPSHUFDZ256mi) |
142 | ENTRY(VPERMILPSZ256ri, VPERMILPSZ256ri, VPSHUFDZ256ri) |
143 | ENTRY(VPERMILPSZmi, VPERMILPSZmi, VPSHUFDZmi) |
144 | ENTRY(VPERMILPSZri, VPERMILPSZri, VPSHUFDZri) |
145 | ENTRY(VPERMPSZ256rm, VPERMPSZ256rm, VPERMDZ256rm) |
146 | ENTRY(VPERMPSZ256rr, VPERMPSZ256rr, VPERMDZ256rr) |
147 | ENTRY(VPERMPDZ256mi, VPERMPDZ256mi, VPERMQZ256mi) |
148 | ENTRY(VPERMPDZ256ri, VPERMPDZ256ri, VPERMQZ256ri) |
149 | ENTRY(VPERMPDZ256rm, VPERMPDZ256rm, VPERMQZ256rm) |
150 | ENTRY(VPERMPDZ256rr, VPERMPDZ256rr, VPERMQZ256rr) |
151 | ENTRY(VPERMPSZrm, VPERMPSZrm, VPERMDZrm) |
152 | ENTRY(VPERMPSZrr, VPERMPSZrr, VPERMDZrr) |
153 | ENTRY(VPERMPDZmi, VPERMPDZmi, VPERMQZmi) |
154 | ENTRY(VPERMPDZri, VPERMPDZri, VPERMQZri) |
155 | ENTRY(VPERMPDZrm, VPERMPDZrm, VPERMQZrm) |
156 | ENTRY(VPERMPDZrr, VPERMPDZrr, VPERMQZrr) |
157 | ENTRY(VUNPCKLPDZ256rm, VUNPCKLPDZ256rm, VPUNPCKLQDQZ256rm) |
158 | ENTRY(VUNPCKLPDZ256rr, VUNPCKLPDZ256rr, VPUNPCKLQDQZ256rr) |
159 | ENTRY(VUNPCKHPDZ256rm, VUNPCKHPDZ256rm, VPUNPCKHQDQZ256rm) |
160 | ENTRY(VUNPCKHPDZ256rr, VUNPCKHPDZ256rr, VPUNPCKHQDQZ256rr) |
161 | ENTRY(VUNPCKLPSZ256rm, VUNPCKLPSZ256rm, VPUNPCKLDQZ256rm) |
162 | ENTRY(VUNPCKLPSZ256rr, VUNPCKLPSZ256rr, VPUNPCKLDQZ256rr) |
163 | ENTRY(VUNPCKHPSZ256rm, VUNPCKHPSZ256rm, VPUNPCKHDQZ256rm) |
164 | ENTRY(VUNPCKHPSZ256rr, VUNPCKHPSZ256rr, VPUNPCKHDQZ256rr) |
165 | ENTRY(VUNPCKLPDZ128rm, VUNPCKLPDZ128rm, VPUNPCKLQDQZ128rm) |
166 | ENTRY(VMOVLHPSZrr, VUNPCKLPDZ128rr, VPUNPCKLQDQZ128rr) |
167 | ENTRY(VUNPCKHPDZ128rm, VUNPCKHPDZ128rm, VPUNPCKHQDQZ128rm) |
168 | ENTRY(VUNPCKHPDZ128rr, VUNPCKHPDZ128rr, VPUNPCKHQDQZ128rr) |
169 | ENTRY(VUNPCKLPSZ128rm, VUNPCKLPSZ128rm, VPUNPCKLDQZ128rm) |
170 | ENTRY(VUNPCKLPSZ128rr, VUNPCKLPSZ128rr, VPUNPCKLDQZ128rr) |
171 | ENTRY(VUNPCKHPSZ128rm, VUNPCKHPSZ128rm, VPUNPCKHDQZ128rm) |
172 | ENTRY(VUNPCKHPSZ128rr, VUNPCKHPSZ128rr, VPUNPCKHDQZ128rr) |
173 | ENTRY(VUNPCKLPDZrm, VUNPCKLPDZrm, VPUNPCKLQDQZrm) |
174 | ENTRY(VUNPCKLPDZrr, VUNPCKLPDZrr, VPUNPCKLQDQZrr) |
175 | ENTRY(VUNPCKHPDZrm, VUNPCKHPDZrm, VPUNPCKHQDQZrm) |
176 | ENTRY(VUNPCKHPDZrr, VUNPCKHPDZrr, VPUNPCKHQDQZrr) |
177 | ENTRY(VUNPCKLPSZrm, VUNPCKLPSZrm, VPUNPCKLDQZrm) |
178 | ENTRY(VUNPCKLPSZrr, VUNPCKLPSZrr, VPUNPCKLDQZrr) |
179 | ENTRY(VUNPCKHPSZrm, VUNPCKHPSZrm, VPUNPCKHDQZrm) |
180 | ENTRY(VUNPCKHPSZrr, VUNPCKHPSZrr, VPUNPCKHDQZrr) |
181 | ENTRY(VEXTRACTPSZmr, VEXTRACTPSZmr, VPEXTRDZmr) |
182 | ENTRY(VEXTRACTPSZrr, VEXTRACTPSZrr, VPEXTRDZrr) |
183 | }; |
184 | |
185 | static const uint16_t ReplaceableInstrsAVX2[][3] = { |
186 | // PackedSingle, PackedDouble, PackedInt |
187 | ENTRY(VANDNPSYrm, VANDNPDYrm, VPANDNYrm) |
188 | ENTRY(VANDNPSYrr, VANDNPDYrr, VPANDNYrr) |
189 | ENTRY(VANDPSYrm, VANDPDYrm, VPANDYrm) |
190 | ENTRY(VANDPSYrr, VANDPDYrr, VPANDYrr) |
191 | ENTRY(VORPSYrm, VORPDYrm, VPORYrm) |
192 | ENTRY(VORPSYrr, VORPDYrr, VPORYrr) |
193 | ENTRY(VXORPSYrm, VXORPDYrm, VPXORYrm) |
194 | ENTRY(VXORPSYrr, VXORPDYrr, VPXORYrr) |
195 | ENTRY(VPERM2F128rm, VPERM2F128rm, VPERM2I128rm) |
196 | ENTRY(VPERM2F128rr, VPERM2F128rr, VPERM2I128rr) |
197 | ENTRY(VBROADCASTSSrm, VBROADCASTSSrm, VPBROADCASTDrm) |
198 | ENTRY(VBROADCASTSSrr, VBROADCASTSSrr, VPBROADCASTDrr) |
199 | ENTRY(VMOVDDUPrm, VMOVDDUPrm, VPBROADCASTQrm) |
200 | ENTRY(VMOVDDUPrr, VMOVDDUPrr, VPBROADCASTQrr) |
201 | ENTRY(VBROADCASTSSYrr, VBROADCASTSSYrr, VPBROADCASTDYrr) |
202 | ENTRY(VBROADCASTSSYrm, VBROADCASTSSYrm, VPBROADCASTDYrm) |
203 | ENTRY(VBROADCASTSDYrr, VBROADCASTSDYrr, VPBROADCASTQYrr) |
204 | ENTRY(VBROADCASTSDYrm, VBROADCASTSDYrm, VPBROADCASTQYrm) |
205 | ENTRY(VBROADCASTF128rm, VBROADCASTF128rm, VBROADCASTI128rm) |
206 | ENTRY(VBLENDPSYrri, VBLENDPSYrri, VPBLENDDYrri) |
207 | ENTRY(VBLENDPSYrmi, VBLENDPSYrmi, VPBLENDDYrmi) |
208 | ENTRY(VPERMILPSYmi, VPERMILPSYmi, VPSHUFDYmi) |
209 | ENTRY(VPERMILPSYri, VPERMILPSYri, VPSHUFDYri) |
210 | ENTRY(VUNPCKLPDYrm, VUNPCKLPDYrm, VPUNPCKLQDQYrm) |
211 | ENTRY(VUNPCKLPDYrr, VUNPCKLPDYrr, VPUNPCKLQDQYrr) |
212 | ENTRY(VUNPCKHPDYrm, VUNPCKHPDYrm, VPUNPCKHQDQYrm) |
213 | ENTRY(VUNPCKHPDYrr, VUNPCKHPDYrr, VPUNPCKHQDQYrr) |
214 | ENTRY(VUNPCKLPSYrm, VUNPCKLPSYrm, VPUNPCKLDQYrm) |
215 | ENTRY(VUNPCKLPSYrr, VUNPCKLPSYrr, VPUNPCKLDQYrr) |
216 | ENTRY(VUNPCKHPSYrm, VUNPCKHPSYrm, VPUNPCKHDQYrm) |
217 | ENTRY(VUNPCKHPSYrr, VUNPCKHPSYrr, VPUNPCKHDQYrr) |
218 | }; |
219 | |
220 | static const uint16_t ReplaceableInstrsFP[][3] = { |
221 | // PackedSingle, PackedDouble |
222 | ENTRY(MOVLPSrm, MOVLPDrm, INSTRUCTION_LIST_END) |
223 | ENTRY(MOVHPSrm, MOVHPDrm, INSTRUCTION_LIST_END) |
224 | ENTRY(MOVHPSmr, MOVHPDmr, INSTRUCTION_LIST_END) |
225 | ENTRY(VMOVLPSrm, VMOVLPDrm, INSTRUCTION_LIST_END) |
226 | ENTRY(VMOVHPSrm, VMOVHPDrm, INSTRUCTION_LIST_END) |
227 | ENTRY(VMOVHPSmr, VMOVHPDmr, INSTRUCTION_LIST_END) |
228 | ENTRY(VMOVLPSZ128rm, VMOVLPDZ128rm, INSTRUCTION_LIST_END) |
229 | ENTRY(VMOVHPSZ128rm, VMOVHPDZ128rm, INSTRUCTION_LIST_END) |
230 | ENTRY(VMOVHPSZ128mr, VMOVHPDZ128mr, INSTRUCTION_LIST_END) |
231 | }; |
232 | |
233 | static const uint16_t [][3] = { |
234 | // PackedSingle, PackedDouble, PackedInt |
235 | ENTRY(VEXTRACTF128mr, VEXTRACTF128mr, VEXTRACTI128mr) |
236 | ENTRY(VEXTRACTF128rr, VEXTRACTF128rr, VEXTRACTI128rr) |
237 | ENTRY(VINSERTF128rm, VINSERTF128rm, VINSERTI128rm) |
238 | ENTRY(VINSERTF128rr, VINSERTF128rr, VINSERTI128rr) |
239 | }; |
240 | |
241 | // NOTE: These should only be used by the custom domain methods. |
242 | static const uint16_t ReplaceableBlendInstrs[][3] = { |
243 | //PackedSingle, PackedDouble, PackedInt |
244 | ENTRY(BLENDPSrmi, BLENDPDrmi, PBLENDWrmi) |
245 | ENTRY(BLENDPSrri, BLENDPDrri, PBLENDWrri) |
246 | ENTRY(VBLENDPSrmi, VBLENDPDrmi, VPBLENDWrmi) |
247 | ENTRY(VBLENDPSrri, VBLENDPDrri, VPBLENDWrri) |
248 | ENTRY(VBLENDPSYrmi, VBLENDPDYrmi, VPBLENDWYrmi) |
249 | ENTRY(VBLENDPSYrri, VBLENDPDYrri, VPBLENDWYrri) |
250 | }; |
251 | |
252 | static const uint16_t ReplaceableBlendAVX2Instrs[][3] = { |
253 | // PackedSingle, PackedDouble, PackedInt |
254 | ENTRY(VBLENDPSrmi, VBLENDPDrmi, VPBLENDDrmi) |
255 | ENTRY(VBLENDPSrri, VBLENDPDrri, VPBLENDDrri) |
256 | ENTRY(VBLENDPSYrmi, VBLENDPDYrmi, VPBLENDDYrmi) |
257 | ENTRY(VBLENDPSYrri, VBLENDPDYrri, VPBLENDDYrri) |
258 | }; |
259 | |
260 | #undef ENTRY |
261 | #define ENTRY(A, B, C, D) {X86::A, X86::B, X86::C, X86::D}, |
262 | static const uint16_t ReplaceableInstrsAVX512[][4] = { |
263 | // Two integer columns for 64-bit and 32-bit elements. |
264 | //PackedSingle, PackedDouble, PackedInt, PackedInt |
265 | ENTRY(VMOVAPSZ128mr, VMOVAPDZ128mr, VMOVDQA64Z128mr, VMOVDQA32Z128mr) |
266 | ENTRY(VMOVAPSZ128rm, VMOVAPDZ128rm, VMOVDQA64Z128rm, VMOVDQA32Z128rm) |
267 | ENTRY(VMOVAPSZ128rr, VMOVAPDZ128rr, VMOVDQA64Z128rr, VMOVDQA32Z128rr) |
268 | ENTRY(VMOVUPSZ128mr, VMOVUPDZ128mr, VMOVDQU64Z128mr, VMOVDQU32Z128mr) |
269 | ENTRY(VMOVUPSZ128rm, VMOVUPDZ128rm, VMOVDQU64Z128rm, VMOVDQU32Z128rm) |
270 | ENTRY(VMOVAPSZ256mr, VMOVAPDZ256mr, VMOVDQA64Z256mr, VMOVDQA32Z256mr) |
271 | ENTRY(VMOVAPSZ256rm, VMOVAPDZ256rm, VMOVDQA64Z256rm, VMOVDQA32Z256rm) |
272 | ENTRY(VMOVAPSZ256rr, VMOVAPDZ256rr, VMOVDQA64Z256rr, VMOVDQA32Z256rr) |
273 | ENTRY(VMOVUPSZ256mr, VMOVUPDZ256mr, VMOVDQU64Z256mr, VMOVDQU32Z256mr) |
274 | ENTRY(VMOVUPSZ256rm, VMOVUPDZ256rm, VMOVDQU64Z256rm, VMOVDQU32Z256rm) |
275 | ENTRY(VMOVAPSZmr, VMOVAPDZmr, VMOVDQA64Zmr, VMOVDQA32Zmr) |
276 | ENTRY(VMOVAPSZrm, VMOVAPDZrm, VMOVDQA64Zrm, VMOVDQA32Zrm) |
277 | ENTRY(VMOVAPSZrr, VMOVAPDZrr, VMOVDQA64Zrr, VMOVDQA32Zrr) |
278 | ENTRY(VMOVUPSZmr, VMOVUPDZmr, VMOVDQU64Zmr, VMOVDQU32Zmr) |
279 | ENTRY(VMOVUPSZrm, VMOVUPDZrm, VMOVDQU64Zrm, VMOVDQU32Zrm) |
280 | }; |
281 | |
282 | static const uint16_t ReplaceableInstrsAVX512DQ[][4] = { |
283 | // Two integer columns for 64-bit and 32-bit elements. |
284 | // PackedSingle, PackedDouble, PackedInt, PackedInt |
285 | ENTRY(VANDNPSZ128rm, VANDNPDZ128rm, VPANDNQZ128rm, VPANDNDZ128rm) |
286 | ENTRY(VANDNPSZ128rr, VANDNPDZ128rr, VPANDNQZ128rr, VPANDNDZ128rr) |
287 | ENTRY(VANDPSZ128rm, VANDPDZ128rm, VPANDQZ128rm, VPANDDZ128rm) |
288 | ENTRY(VANDPSZ128rr, VANDPDZ128rr, VPANDQZ128rr, VPANDDZ128rr) |
289 | ENTRY(VORPSZ128rm, VORPDZ128rm, VPORQZ128rm, VPORDZ128rm) |
290 | ENTRY(VORPSZ128rr, VORPDZ128rr, VPORQZ128rr, VPORDZ128rr) |
291 | ENTRY(VXORPSZ128rm, VXORPDZ128rm, VPXORQZ128rm, VPXORDZ128rm) |
292 | ENTRY(VXORPSZ128rr, VXORPDZ128rr, VPXORQZ128rr, VPXORDZ128rr) |
293 | ENTRY(VANDNPSZ256rm, VANDNPDZ256rm, VPANDNQZ256rm, VPANDNDZ256rm) |
294 | ENTRY(VANDNPSZ256rr, VANDNPDZ256rr, VPANDNQZ256rr, VPANDNDZ256rr) |
295 | ENTRY(VANDPSZ256rm, VANDPDZ256rm, VPANDQZ256rm, VPANDDZ256rm) |
296 | ENTRY(VANDPSZ256rr, VANDPDZ256rr, VPANDQZ256rr, VPANDDZ256rr) |
297 | ENTRY(VORPSZ256rm, VORPDZ256rm, VPORQZ256rm, VPORDZ256rm) |
298 | ENTRY(VORPSZ256rr, VORPDZ256rr, VPORQZ256rr, VPORDZ256rr) |
299 | ENTRY(VXORPSZ256rm, VXORPDZ256rm, VPXORQZ256rm, VPXORDZ256rm) |
300 | ENTRY(VXORPSZ256rr, VXORPDZ256rr, VPXORQZ256rr, VPXORDZ256rr) |
301 | ENTRY(VANDNPSZrm, VANDNPDZrm, VPANDNQZrm, VPANDNDZrm) |
302 | ENTRY(VANDNPSZrr, VANDNPDZrr, VPANDNQZrr, VPANDNDZrr) |
303 | ENTRY(VANDPSZrm, VANDPDZrm, VPANDQZrm, VPANDDZrm) |
304 | ENTRY(VANDPSZrr, VANDPDZrr, VPANDQZrr, VPANDDZrr) |
305 | ENTRY(VORPSZrm, VORPDZrm, VPORQZrm, VPORDZrm) |
306 | ENTRY(VORPSZrr, VORPDZrr, VPORQZrr, VPORDZrr) |
307 | ENTRY(VXORPSZrm, VXORPDZrm, VPXORQZrm, VPXORDZrm) |
308 | ENTRY(VXORPSZrr, VXORPDZrr, VPXORQZrr, VPXORDZrr) |
309 | }; |
310 | |
311 | static const uint16_t ReplaceableInstrsAVX512DQMasked[][4] = { |
312 | // Two integer columns for 64-bit and 32-bit elements. |
313 | // PackedSingle, PackedDouble, PackedInt, PackedInt |
314 | ENTRY(VANDNPSZ128rmk, VANDNPDZ128rmk, VPANDNQZ128rmk, VPANDNDZ128rmk) |
315 | ENTRY(VANDNPSZ128rmkz, VANDNPDZ128rmkz, VPANDNQZ128rmkz, VPANDNDZ128rmkz) |
316 | ENTRY(VANDNPSZ128rrk, VANDNPDZ128rrk, VPANDNQZ128rrk, VPANDNDZ128rrk) |
317 | ENTRY(VANDNPSZ128rrkz, VANDNPDZ128rrkz, VPANDNQZ128rrkz, VPANDNDZ128rrkz) |
318 | ENTRY(VANDPSZ128rmk, VANDPDZ128rmk, VPANDQZ128rmk, VPANDDZ128rmk) |
319 | ENTRY(VANDPSZ128rmkz, VANDPDZ128rmkz, VPANDQZ128rmkz, VPANDDZ128rmkz) |
320 | ENTRY(VANDPSZ128rrk, VANDPDZ128rrk, VPANDQZ128rrk, VPANDDZ128rrk) |
321 | ENTRY(VANDPSZ128rrkz, VANDPDZ128rrkz, VPANDQZ128rrkz, VPANDDZ128rrkz) |
322 | ENTRY(VORPSZ128rmk, VORPDZ128rmk, VPORQZ128rmk, VPORDZ128rmk) |
323 | ENTRY(VORPSZ128rmkz, VORPDZ128rmkz, VPORQZ128rmkz, VPORDZ128rmkz) |
324 | ENTRY(VORPSZ128rrk, VORPDZ128rrk, VPORQZ128rrk, VPORDZ128rrk) |
325 | ENTRY(VORPSZ128rrkz, VORPDZ128rrkz, VPORQZ128rrkz, VPORDZ128rrkz) |
326 | ENTRY(VXORPSZ128rmk, VXORPDZ128rmk, VPXORQZ128rmk, VPXORDZ128rmk) |
327 | ENTRY(VXORPSZ128rmkz, VXORPDZ128rmkz, VPXORQZ128rmkz, VPXORDZ128rmkz) |
328 | ENTRY(VXORPSZ128rrk, VXORPDZ128rrk, VPXORQZ128rrk, VPXORDZ128rrk) |
329 | ENTRY(VXORPSZ128rrkz, VXORPDZ128rrkz, VPXORQZ128rrkz, VPXORDZ128rrkz) |
330 | ENTRY(VANDNPSZ256rmk, VANDNPDZ256rmk, VPANDNQZ256rmk, VPANDNDZ256rmk) |
331 | ENTRY(VANDNPSZ256rmkz, VANDNPDZ256rmkz, VPANDNQZ256rmkz, VPANDNDZ256rmkz) |
332 | ENTRY(VANDNPSZ256rrk, VANDNPDZ256rrk, VPANDNQZ256rrk, VPANDNDZ256rrk) |
333 | ENTRY(VANDNPSZ256rrkz, VANDNPDZ256rrkz, VPANDNQZ256rrkz, VPANDNDZ256rrkz) |
334 | ENTRY(VANDPSZ256rmk, VANDPDZ256rmk, VPANDQZ256rmk, VPANDDZ256rmk) |
335 | ENTRY(VANDPSZ256rmkz, VANDPDZ256rmkz, VPANDQZ256rmkz, VPANDDZ256rmkz) |
336 | ENTRY(VANDPSZ256rrk, VANDPDZ256rrk, VPANDQZ256rrk, VPANDDZ256rrk) |
337 | ENTRY(VANDPSZ256rrkz, VANDPDZ256rrkz, VPANDQZ256rrkz, VPANDDZ256rrkz) |
338 | ENTRY(VORPSZ256rmk, VORPDZ256rmk, VPORQZ256rmk, VPORDZ256rmk) |
339 | ENTRY(VORPSZ256rmkz, VORPDZ256rmkz, VPORQZ256rmkz, VPORDZ256rmkz) |
340 | ENTRY(VORPSZ256rrk, VORPDZ256rrk, VPORQZ256rrk, VPORDZ256rrk) |
341 | ENTRY(VORPSZ256rrkz, VORPDZ256rrkz, VPORQZ256rrkz, VPORDZ256rrkz) |
342 | ENTRY(VXORPSZ256rmk, VXORPDZ256rmk, VPXORQZ256rmk, VPXORDZ256rmk) |
343 | ENTRY(VXORPSZ256rmkz, VXORPDZ256rmkz, VPXORQZ256rmkz, VPXORDZ256rmkz) |
344 | ENTRY(VXORPSZ256rrk, VXORPDZ256rrk, VPXORQZ256rrk, VPXORDZ256rrk) |
345 | ENTRY(VXORPSZ256rrkz, VXORPDZ256rrkz, VPXORQZ256rrkz, VPXORDZ256rrkz) |
346 | ENTRY(VANDNPSZrmk, VANDNPDZrmk, VPANDNQZrmk, VPANDNDZrmk) |
347 | ENTRY(VANDNPSZrmkz, VANDNPDZrmkz, VPANDNQZrmkz, VPANDNDZrmkz) |
348 | ENTRY(VANDNPSZrrk, VANDNPDZrrk, VPANDNQZrrk, VPANDNDZrrk) |
349 | ENTRY(VANDNPSZrrkz, VANDNPDZrrkz, VPANDNQZrrkz, VPANDNDZrrkz) |
350 | ENTRY(VANDPSZrmk, VANDPDZrmk, VPANDQZrmk, VPANDDZrmk) |
351 | ENTRY(VANDPSZrmkz, VANDPDZrmkz, VPANDQZrmkz, VPANDDZrmkz) |
352 | ENTRY(VANDPSZrrk, VANDPDZrrk, VPANDQZrrk, VPANDDZrrk) |
353 | ENTRY(VANDPSZrrkz, VANDPDZrrkz, VPANDQZrrkz, VPANDDZrrkz) |
354 | ENTRY(VORPSZrmk, VORPDZrmk, VPORQZrmk, VPORDZrmk) |
355 | ENTRY(VORPSZrmkz, VORPDZrmkz, VPORQZrmkz, VPORDZrmkz) |
356 | ENTRY(VORPSZrrk, VORPDZrrk, VPORQZrrk, VPORDZrrk) |
357 | ENTRY(VORPSZrrkz, VORPDZrrkz, VPORQZrrkz, VPORDZrrkz) |
358 | ENTRY(VXORPSZrmk, VXORPDZrmk, VPXORQZrmk, VPXORDZrmk) |
359 | ENTRY(VXORPSZrmkz, VXORPDZrmkz, VPXORQZrmkz, VPXORDZrmkz) |
360 | ENTRY(VXORPSZrrk, VXORPDZrrk, VPXORQZrrk, VPXORDZrrk) |
361 | ENTRY(VXORPSZrrkz, VXORPDZrrkz, VPXORQZrrkz, VPXORDZrrkz) |
362 | // Broadcast loads can be handled the same as masked operations to avoid |
363 | // changing element size. |
364 | ENTRY(VANDNPSZ128rmb, VANDNPDZ128rmb, VPANDNQZ128rmb, VPANDNDZ128rmb) |
365 | ENTRY(VANDPSZ128rmb, VANDPDZ128rmb, VPANDQZ128rmb, VPANDDZ128rmb) |
366 | ENTRY(VORPSZ128rmb, VORPDZ128rmb, VPORQZ128rmb, VPORDZ128rmb) |
367 | ENTRY(VXORPSZ128rmb, VXORPDZ128rmb, VPXORQZ128rmb, VPXORDZ128rmb) |
368 | ENTRY(VANDNPSZ256rmb, VANDNPDZ256rmb, VPANDNQZ256rmb, VPANDNDZ256rmb) |
369 | ENTRY(VANDPSZ256rmb, VANDPDZ256rmb, VPANDQZ256rmb, VPANDDZ256rmb) |
370 | ENTRY(VORPSZ256rmb, VORPDZ256rmb, VPORQZ256rmb, VPORDZ256rmb) |
371 | ENTRY(VXORPSZ256rmb, VXORPDZ256rmb, VPXORQZ256rmb, VPXORDZ256rmb) |
372 | ENTRY(VANDNPSZrmb, VANDNPDZrmb, VPANDNQZrmb, VPANDNDZrmb) |
373 | ENTRY(VANDPSZrmb, VANDPDZrmb, VPANDQZrmb, VPANDDZrmb) |
374 | ENTRY(VANDPSZrmb, VANDPDZrmb, VPANDQZrmb, VPANDDZrmb) |
375 | ENTRY(VORPSZrmb, VORPDZrmb, VPORQZrmb, VPORDZrmb) |
376 | ENTRY(VXORPSZrmb, VXORPDZrmb, VPXORQZrmb, VPXORDZrmb) |
377 | ENTRY(VANDNPSZ128rmbk, VANDNPDZ128rmbk, VPANDNQZ128rmbk, VPANDNDZ128rmbk) |
378 | ENTRY(VANDPSZ128rmbk, VANDPDZ128rmbk, VPANDQZ128rmbk, VPANDDZ128rmbk) |
379 | ENTRY(VORPSZ128rmbk, VORPDZ128rmbk, VPORQZ128rmbk, VPORDZ128rmbk) |
380 | ENTRY(VXORPSZ128rmbk, VXORPDZ128rmbk, VPXORQZ128rmbk, VPXORDZ128rmbk) |
381 | ENTRY(VANDNPSZ256rmbk, VANDNPDZ256rmbk, VPANDNQZ256rmbk, VPANDNDZ256rmbk) |
382 | ENTRY(VANDPSZ256rmbk, VANDPDZ256rmbk, VPANDQZ256rmbk, VPANDDZ256rmbk) |
383 | ENTRY(VORPSZ256rmbk, VORPDZ256rmbk, VPORQZ256rmbk, VPORDZ256rmbk) |
384 | ENTRY(VXORPSZ256rmbk, VXORPDZ256rmbk, VPXORQZ256rmbk, VPXORDZ256rmbk) |
385 | ENTRY(VANDNPSZrmbk, VANDNPDZrmbk, VPANDNQZrmbk, VPANDNDZrmbk) |
386 | ENTRY(VANDPSZrmbk, VANDPDZrmbk, VPANDQZrmbk, VPANDDZrmbk) |
387 | ENTRY(VANDPSZrmbk, VANDPDZrmbk, VPANDQZrmbk, VPANDDZrmbk) |
388 | ENTRY(VORPSZrmbk, VORPDZrmbk, VPORQZrmbk, VPORDZrmbk) |
389 | ENTRY(VXORPSZrmbk, VXORPDZrmbk, VPXORQZrmbk, VPXORDZrmbk) |
390 | ENTRY(VANDNPSZ128rmbkz, VANDNPDZ128rmbkz, VPANDNQZ128rmbkz, VPANDNDZ128rmbkz) |
391 | ENTRY(VANDPSZ128rmbkz, VANDPDZ128rmbkz, VPANDQZ128rmbkz, VPANDDZ128rmbkz) |
392 | ENTRY(VORPSZ128rmbkz, VORPDZ128rmbkz, VPORQZ128rmbkz, VPORDZ128rmbkz) |
393 | ENTRY(VXORPSZ128rmbkz, VXORPDZ128rmbkz, VPXORQZ128rmbkz, VPXORDZ128rmbkz) |
394 | ENTRY(VANDNPSZ256rmbkz, VANDNPDZ256rmbkz, VPANDNQZ256rmbkz, VPANDNDZ256rmbkz) |
395 | ENTRY(VANDPSZ256rmbkz, VANDPDZ256rmbkz, VPANDQZ256rmbkz, VPANDDZ256rmbkz) |
396 | ENTRY(VORPSZ256rmbkz, VORPDZ256rmbkz, VPORQZ256rmbkz, VPORDZ256rmbkz) |
397 | ENTRY(VXORPSZ256rmbkz, VXORPDZ256rmbkz, VPXORQZ256rmbkz, VPXORDZ256rmbkz) |
398 | ENTRY(VANDNPSZrmbkz, VANDNPDZrmbkz, VPANDNQZrmbkz, VPANDNDZrmbkz) |
399 | ENTRY(VANDPSZrmbkz, VANDPDZrmbkz, VPANDQZrmbkz, VPANDDZrmbkz) |
400 | ENTRY(VANDPSZrmbkz, VANDPDZrmbkz, VPANDQZrmbkz, VPANDDZrmbkz) |
401 | ENTRY(VORPSZrmbkz, VORPDZrmbkz, VPORQZrmbkz, VPORDZrmbkz) |
402 | ENTRY(VXORPSZrmbkz, VXORPDZrmbkz, VPXORQZrmbkz, VPXORDZrmbkz) |
403 | }; |
404 | |
405 | // Special table for changing EVEX logic instructions to VEX. |
406 | // TODO: Should we run EVEX->VEX earlier? |
407 | static const uint16_t ReplaceableCustomAVX512LogicInstrs[][4] = { |
408 | // Two integer columns for 64-bit and 32-bit elements. |
409 | // PackedSingle, PackedDouble, PackedInt, PackedInt |
410 | ENTRY(VANDNPSrm, VANDNPDrm, VPANDNQZ128rm, VPANDNDZ128rm) |
411 | ENTRY(VANDNPSrr, VANDNPDrr, VPANDNQZ128rr, VPANDNDZ128rr) |
412 | ENTRY(VANDPSrm, VANDPDrm, VPANDQZ128rm, VPANDDZ128rm) |
413 | ENTRY(VANDPSrr, VANDPDrr, VPANDQZ128rr, VPANDDZ128rr) |
414 | ENTRY(VORPSrm, VORPDrm, VPORQZ128rm, VPORDZ128rm) |
415 | ENTRY(VORPSrr, VORPDrr, VPORQZ128rr, VPORDZ128rr) |
416 | ENTRY(VXORPSrm, VXORPDrm, VPXORQZ128rm, VPXORDZ128rm) |
417 | ENTRY(VXORPSrr, VXORPDrr, VPXORQZ128rr, VPXORDZ128rr) |
418 | ENTRY(VANDNPSYrm, VANDNPDYrm, VPANDNQZ256rm, VPANDNDZ256rm) |
419 | ENTRY(VANDNPSYrr, VANDNPDYrr, VPANDNQZ256rr, VPANDNDZ256rr) |
420 | ENTRY(VANDPSYrm, VANDPDYrm, VPANDQZ256rm, VPANDDZ256rm) |
421 | ENTRY(VANDPSYrr, VANDPDYrr, VPANDQZ256rr, VPANDDZ256rr) |
422 | ENTRY(VORPSYrm, VORPDYrm, VPORQZ256rm, VPORDZ256rm) |
423 | ENTRY(VORPSYrr, VORPDYrr, VPORQZ256rr, VPORDZ256rr) |
424 | ENTRY(VXORPSYrm, VXORPDYrm, VPXORQZ256rm, VPXORDZ256rm) |
425 | ENTRY(VXORPSYrr, VXORPDYrr, VPXORQZ256rr, VPXORDZ256rr) |
426 | }; |
427 | |