1 //===-- X86InstrFMA3Info.cpp - X86 FMA3 Instruction Information -----------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file contains the implementation of the classes providing information
11 // about existing X86 FMA3 opcodes, classifying and grouping them.
12 //
13 //===----------------------------------------------------------------------===//
14 
15 #include "X86InstrFMA3Info.h"
16 #include "X86InstrInfo.h"
17 #include "llvm/Support/ManagedStatic.h"
18 #include "llvm/Support/Threading.h"
19 #include <cassert>
20 #include <cstdint>
21 
22 using namespace llvm;
23 
24 /// This flag is used in the method llvm::call_once() used below to make the
25 /// initialization of the map 'OpcodeToGroup' thread safe.
26 static llvm::once_flag InitGroupsOnceFlag;
27 
28 static ManagedStatic<X86InstrFMA3Info> X86InstrFMA3InfoObj;
29 X86InstrFMA3Info *X86InstrFMA3Info::getX86InstrFMA3Info() {
30   return &*X86InstrFMA3InfoObj;
31 }
32 
33 #define FMA3BASE(R132, R213, R231, M132, M213, M231, Attrs)                    \
34   { { R132, R213, R231 }, { M132, M213, M231 }, Attrs },
35 
36 #define FMA3RMA(R132, R213, R231, M132, M213, M231, Attrs)                     \
37   FMA3BASE(X86::R132, X86::R213, X86::R231, X86::M132, X86::M213, X86::M231, Attrs)
38 
39 #define FMA3RM(R132, R213, R231, M132, M213, M231)                             \
40   FMA3RMA(R132, R213, R231, M132, M213, M231, 0)
41 
42 #define FMA3RA(R132, R213, R231, Attrs)                                        \
43   FMA3BASE(X86::R132, X86::R213, X86::R231, 0, 0, 0, Attrs)
44 
45 #define FMA3R(R132, R213, R231)                                                \
46   FMA3RA(R132, R213, R231, 0)
47 
48 #define FMA3MA(M132, M213, M231, Attrs)                                        \
49   FMA3BASE(0, 0, 0, X86::M132, X86::M213, X86::M231, Attrs)
50 
51 #define FMA3M(M132, M213, M231)                                                \
52   FMA3MA(M132, M213, M231, 0)
53 
54 #define FMA3_AVX2_VECTOR_GROUP(Name)                                           \
55   FMA3RM(Name##132PSr, Name##213PSr, Name##231PSr,                             \
56          Name##132PSm, Name##213PSm, Name##231PSm)                             \
57   FMA3RM(Name##132PDr, Name##213PDr, Name##231PDr,                             \
58          Name##132PDm, Name##213PDm, Name##231PDm)                             \
59   FMA3RM(Name##132PSYr, Name##213PSYr, Name##231PSYr,                          \
60          Name##132PSYm, Name##213PSYm, Name##231PSYm)                          \
61   FMA3RM(Name##132PDYr, Name##213PDYr, Name##231PDYr,                          \
62          Name##132PDYm, Name##213PDYm, Name##231PDYm)
63 
64 #define FMA3_AVX2_SCALAR_GROUP(Name)                                           \
65   FMA3RM(Name##132SSr, Name##213SSr, Name##231SSr,                             \
66          Name##132SSm, Name##213SSm, Name##231SSm)                             \
67   FMA3RM(Name##132SDr, Name##213SDr, Name##231SDr,                             \
68          Name##132SDm, Name##213SDm, Name##231SDm)                             \
69   FMA3RMA(Name##132SSr_Int, Name##213SSr_Int, Name##231SSr_Int,                \
70           Name##132SSm_Int, Name##213SSm_Int, Name##231SSm_Int,                \
71           X86InstrFMA3Group::X86FMA3Intrinsic)                                 \
72   FMA3RMA(Name##132SDr_Int, Name##213SDr_Int, Name##231SDr_Int,                \
73           Name##132SDm_Int, Name##213SDm_Int, Name##231SDm_Int,                \
74           X86InstrFMA3Group::X86FMA3Intrinsic)
75 
76 #define FMA3_AVX2_FULL_GROUP(Name)                                             \
77   FMA3_AVX2_VECTOR_GROUP(Name)                                                 \
78   FMA3_AVX2_SCALAR_GROUP(Name)
79 
80 #define FMA3_AVX512_VECTOR_GROUP(Name)                                         \
81   FMA3RM(Name##132PSZ128r, Name##213PSZ128r, Name##231PSZ128r,                 \
82          Name##132PSZ128m, Name##213PSZ128m, Name##231PSZ128m)                 \
83   FMA3RM(Name##132PDZ128r, Name##213PDZ128r, Name##231PDZ128r,                 \
84          Name##132PDZ128m, Name##213PDZ128m, Name##231PDZ128m)                 \
85   FMA3RM(Name##132PSZ256r, Name##213PSZ256r, Name##231PSZ256r,                 \
86          Name##132PSZ256m, Name##213PSZ256m, Name##231PSZ256m)                 \
87   FMA3RM(Name##132PDZ256r, Name##213PDZ256r, Name##231PDZ256r,                 \
88          Name##132PDZ256m, Name##213PDZ256m, Name##231PDZ256m)                 \
89   FMA3RM(Name##132PSZr,    Name##213PSZr,    Name##231PSZr,                    \
90          Name##132PSZm,    Name##213PSZm,    Name##231PSZm)                    \
91   FMA3RM(Name##132PDZr,    Name##213PDZr,    Name##231PDZr,                    \
92          Name##132PDZm,    Name##213PDZm,    Name##231PDZm)                    \
93   FMA3RMA(Name##132PSZ128rk, Name##213PSZ128rk, Name##231PSZ128rk,             \
94           Name##132PSZ128mk, Name##213PSZ128mk, Name##231PSZ128mk,             \
95           X86InstrFMA3Group::X86FMA3KMergeMasked)                              \
96   FMA3RMA(Name##132PDZ128rk, Name##213PDZ128rk, Name##231PDZ128rk,             \
97           Name##132PDZ128mk, Name##213PDZ128mk, Name##231PDZ128mk,             \
98           X86InstrFMA3Group::X86FMA3KMergeMasked)                              \
99   FMA3RMA(Name##132PSZ256rk, Name##213PSZ256rk, Name##231PSZ256rk,             \
100           Name##132PSZ256mk, Name##213PSZ256mk, Name##231PSZ256mk,             \
101           X86InstrFMA3Group::X86FMA3KMergeMasked)                              \
102   FMA3RMA(Name##132PDZ256rk, Name##213PDZ256rk, Name##231PDZ256rk,             \
103           Name##132PDZ256mk, Name##213PDZ256mk, Name##231PDZ256mk,             \
104           X86InstrFMA3Group::X86FMA3KMergeMasked)                              \
105   FMA3RMA(Name##132PSZrk,    Name##213PSZrk,    Name##231PSZrk,                \
106           Name##132PSZmk,    Name##213PSZmk,    Name##231PSZmk,                \
107           X86InstrFMA3Group::X86FMA3KMergeMasked)                              \
108   FMA3RMA(Name##132PDZrk,    Name##213PDZrk,    Name##231PDZrk,                \
109           Name##132PDZmk,    Name##213PDZmk,    Name##231PDZmk,                \
110           X86InstrFMA3Group::X86FMA3KMergeMasked)                              \
111   FMA3RMA(Name##132PSZ128rkz, Name##213PSZ128rkz, Name##231PSZ128rkz,          \
112           Name##132PSZ128mkz, Name##213PSZ128mkz, Name##231PSZ128mkz,          \
113           X86InstrFMA3Group::X86FMA3KZeroMasked)                               \
114   FMA3RMA(Name##132PDZ128rkz, Name##213PDZ128rkz, Name##231PDZ128rkz,          \
115           Name##132PDZ128mkz, Name##213PDZ128mkz, Name##231PDZ128mkz,          \
116           X86InstrFMA3Group::X86FMA3KZeroMasked)                               \
117   FMA3RMA(Name##132PSZ256rkz, Name##213PSZ256rkz, Name##231PSZ256rkz,          \
118           Name##132PSZ256mkz, Name##213PSZ256mkz, Name##231PSZ256mkz,          \
119           X86InstrFMA3Group::X86FMA3KZeroMasked)                               \
120   FMA3RMA(Name##132PDZ256rkz, Name##213PDZ256rkz, Name##231PDZ256rkz,          \
121           Name##132PDZ256mkz, Name##213PDZ256mkz, Name##231PDZ256mkz,          \
122           X86InstrFMA3Group::X86FMA3KZeroMasked)                               \
123   FMA3RMA(Name##132PSZrkz,    Name##213PSZrkz,    Name##231PSZrkz,             \
124           Name##132PSZmkz,    Name##213PSZmkz,    Name##231PSZmkz,             \
125           X86InstrFMA3Group::X86FMA3KZeroMasked)                               \
126   FMA3RMA(Name##132PDZrkz,    Name##213PDZrkz,    Name##231PDZrkz,             \
127           Name##132PDZmkz,    Name##213PDZmkz,    Name##231PDZmkz,             \
128           X86InstrFMA3Group::X86FMA3KZeroMasked)                               \
129   FMA3R(Name##132PSZrb, Name##213PSZrb, Name##231PSZrb)                        \
130   FMA3R(Name##132PDZrb, Name##213PDZrb, Name##231PDZrb)                        \
131   FMA3RA(Name##132PSZrbk, Name##213PSZrbk, Name##231PSZrbk,                    \
132          X86InstrFMA3Group::X86FMA3KMergeMasked)                               \
133   FMA3RA(Name##132PDZrbk, Name##213PDZrbk, Name##231PDZrbk,                    \
134          X86InstrFMA3Group::X86FMA3KMergeMasked)                               \
135   FMA3RA(Name##132PSZrbkz, Name##213PSZrbkz, Name##231PSZrbkz,                 \
136          X86InstrFMA3Group::X86FMA3KZeroMasked)                                \
137   FMA3RA(Name##132PDZrbkz, Name##213PDZrbkz, Name##231PDZrbkz,                 \
138          X86InstrFMA3Group::X86FMA3KZeroMasked)                                \
139   FMA3M(Name##132PSZ128mb, Name##213PSZ128mb, Name##231PSZ128mb)               \
140   FMA3M(Name##132PDZ128mb, Name##213PDZ128mb, Name##231PDZ128mb)               \
141   FMA3M(Name##132PSZ256mb, Name##213PSZ256mb, Name##231PSZ256mb)               \
142   FMA3M(Name##132PDZ256mb, Name##213PDZ256mb, Name##231PDZ256mb)               \
143   FMA3M(Name##132PSZmb, Name##213PSZmb, Name##231PSZmb)                        \
144   FMA3M(Name##132PDZmb, Name##213PDZmb, Name##231PDZmb)                        \
145   FMA3MA(Name##132PSZ128mbk, Name##213PSZ128mbk, Name##231PSZ128mbk,           \
146          X86InstrFMA3Group::X86FMA3KMergeMasked)                               \
147   FMA3MA(Name##132PDZ128mbk, Name##213PDZ128mbk, Name##231PDZ128mbk,           \
148          X86InstrFMA3Group::X86FMA3KMergeMasked)                               \
149   FMA3MA(Name##132PSZ256mbk, Name##213PSZ256mbk, Name##231PSZ256mbk,           \
150          X86InstrFMA3Group::X86FMA3KMergeMasked)                               \
151   FMA3MA(Name##132PDZ256mbk, Name##213PDZ256mbk, Name##231PDZ256mbk,           \
152          X86InstrFMA3Group::X86FMA3KMergeMasked)                               \
153   FMA3MA(Name##132PSZmbk,    Name##213PSZmbk,    Name##231PSZmbk,              \
154          X86InstrFMA3Group::X86FMA3KMergeMasked)                               \
155   FMA3MA(Name##132PDZmbk,    Name##213PDZmbk,    Name##231PDZmbk,              \
156          X86InstrFMA3Group::X86FMA3KMergeMasked)                               \
157   FMA3MA(Name##132PSZ128mbkz, Name##213PSZ128mbkz, Name##231PSZ128mbkz,        \
158          X86InstrFMA3Group::X86FMA3KZeroMasked)                                \
159   FMA3MA(Name##132PDZ128mbkz, Name##213PDZ128mbkz, Name##231PDZ128mbkz,        \
160          X86InstrFMA3Group::X86FMA3KZeroMasked)                                \
161   FMA3MA(Name##132PSZ256mbkz, Name##213PSZ256mbkz, Name##231PSZ256mbkz,        \
162          X86InstrFMA3Group::X86FMA3KZeroMasked)                                \
163   FMA3MA(Name##132PDZ256mbkz, Name##213PDZ256mbkz, Name##231PDZ256mbkz,        \
164          X86InstrFMA3Group::X86FMA3KZeroMasked)                                \
165   FMA3MA(Name##132PSZmbkz, Name##213PSZmbkz, Name##231PSZmbkz,                 \
166          X86InstrFMA3Group::X86FMA3KZeroMasked)                                \
167   FMA3MA(Name##132PDZmbkz, Name##213PDZmbkz, Name##231PDZmbkz,                 \
168          X86InstrFMA3Group::X86FMA3KZeroMasked)
169 
170 #define FMA3_AVX512_SCALAR_GROUP(Name)                                         \
171   FMA3RM(Name##132SSZr,      Name##213SSZr,     Name##231SSZr,                 \
172          Name##132SSZm,      Name##213SSZm,     Name##231SSZm)                 \
173   FMA3RM(Name##132SDZr,      Name##213SDZr,     Name##231SDZr,                 \
174          Name##132SDZm,      Name##213SDZm,     Name##231SDZm)                 \
175   FMA3RMA(Name##132SSZr_Int, Name##213SSZr_Int, Name##231SSZr_Int,             \
176           Name##132SSZm_Int, Name##213SSZm_Int, Name##231SSZm_Int,             \
177           X86InstrFMA3Group::X86FMA3Intrinsic)                                 \
178   FMA3RMA(Name##132SDZr_Int, Name##213SDZr_Int, Name##231SDZr_Int,             \
179           Name##132SDZm_Int, Name##213SDZm_Int, Name##231SDZm_Int,             \
180           X86InstrFMA3Group::X86FMA3Intrinsic)                                 \
181   FMA3RMA(Name##132SSZr_Intk, Name##213SSZr_Intk, Name##231SSZr_Intk,          \
182           Name##132SSZm_Intk, Name##213SSZm_Intk, Name##231SSZm_Intk,          \
183           X86InstrFMA3Group::X86FMA3Intrinsic |                                \
184               X86InstrFMA3Group::X86FMA3KMergeMasked)                          \
185   FMA3RMA(Name##132SDZr_Intk, Name##213SDZr_Intk, Name##231SDZr_Intk,          \
186           Name##132SDZm_Intk, Name##213SDZm_Intk, Name##231SDZm_Intk,          \
187           X86InstrFMA3Group::X86FMA3Intrinsic |                                \
188               X86InstrFMA3Group::X86FMA3KMergeMasked)                          \
189   FMA3RMA(Name##132SSZr_Intkz, Name##213SSZr_Intkz, Name##231SSZr_Intkz,       \
190           Name##132SSZm_Intkz, Name##213SSZm_Intkz, Name##231SSZm_Intkz,       \
191           X86InstrFMA3Group::X86FMA3Intrinsic |                                \
192               X86InstrFMA3Group::X86FMA3KZeroMasked)                           \
193   FMA3RMA(Name##132SDZr_Intkz, Name##213SDZr_Intkz, Name##231SDZr_Intkz,       \
194           Name##132SDZm_Intkz, Name##213SDZm_Intkz, Name##231SDZm_Intkz,       \
195           X86InstrFMA3Group::X86FMA3Intrinsic |                                \
196               X86InstrFMA3Group::X86FMA3KZeroMasked)                           \
197   FMA3RA(Name##132SSZrb_Int, Name##213SSZrb_Int, Name##231SSZrb_Int,           \
198          X86InstrFMA3Group::X86FMA3Intrinsic)                                  \
199   FMA3RA(Name##132SDZrb_Int, Name##213SDZrb_Int, Name##231SDZrb_Int,           \
200          X86InstrFMA3Group::X86FMA3Intrinsic)                                  \
201   FMA3RA(Name##132SSZrb_Intk, Name##213SSZrb_Intk, Name##231SSZrb_Intk,        \
202          X86InstrFMA3Group::X86FMA3Intrinsic |                                 \
203              X86InstrFMA3Group::X86FMA3KMergeMasked)                           \
204   FMA3RA(Name##132SDZrb_Intk, Name##213SDZrb_Intk, Name##231SDZrb_Intk,        \
205          X86InstrFMA3Group::X86FMA3Intrinsic |                                 \
206              X86InstrFMA3Group::X86FMA3KMergeMasked)                           \
207   FMA3RA(Name##132SSZrb_Intkz, Name##213SSZrb_Intkz, Name##231SSZrb_Intkz,     \
208          X86InstrFMA3Group::X86FMA3Intrinsic |                                 \
209              X86InstrFMA3Group::X86FMA3KZeroMasked)                            \
210   FMA3RA(Name##132SDZrb_Intkz, Name##213SDZrb_Intkz, Name##231SDZrb_Intkz,     \
211          X86InstrFMA3Group::X86FMA3Intrinsic |                                 \
212              X86InstrFMA3Group::X86FMA3KZeroMasked)
213 
214 #define FMA3_AVX512_FULL_GROUP(Name)                                           \
215   FMA3_AVX512_VECTOR_GROUP(Name)                                               \
216   FMA3_AVX512_SCALAR_GROUP(Name)
217 
218 static const X86InstrFMA3Group Groups[] = {
219   FMA3_AVX2_FULL_GROUP(VFMADD)
220   FMA3_AVX2_FULL_GROUP(VFMSUB)
221   FMA3_AVX2_FULL_GROUP(VFNMADD)
222   FMA3_AVX2_FULL_GROUP(VFNMSUB)
223 
224   FMA3_AVX2_VECTOR_GROUP(VFMADDSUB)
225   FMA3_AVX2_VECTOR_GROUP(VFMSUBADD)
226 
227   FMA3_AVX512_FULL_GROUP(VFMADD)
228   FMA3_AVX512_FULL_GROUP(VFMSUB)
229   FMA3_AVX512_FULL_GROUP(VFNMADD)
230   FMA3_AVX512_FULL_GROUP(VFNMSUB)
231 
232   FMA3_AVX512_VECTOR_GROUP(VFMADDSUB)
233   FMA3_AVX512_VECTOR_GROUP(VFMSUBADD)
234 };
235 
236 void X86InstrFMA3Info::initGroupsOnceImpl() {
237   for (const X86InstrFMA3Group &G : Groups) {
238     if (G.RegOpcodes[0])
239       OpcodeToGroup[G.RegOpcodes[0]] = &G;
240     if (G.RegOpcodes[1])
241       OpcodeToGroup[G.RegOpcodes[1]] = &G;
242     if (G.RegOpcodes[2])
243       OpcodeToGroup[G.RegOpcodes[2]] = &G;
244     if (G.MemOpcodes[0])
245       OpcodeToGroup[G.MemOpcodes[0]] = &G;
246     if (G.MemOpcodes[1])
247       OpcodeToGroup[G.MemOpcodes[1]] = &G;
248     if (G.MemOpcodes[2])
249       OpcodeToGroup[G.MemOpcodes[2]] = &G;
250   }
251 }
252 
253 void X86InstrFMA3Info::initGroupsOnce() {
254   llvm::call_once(InitGroupsOnceFlag,
255                   []() { getX86InstrFMA3Info()->initGroupsOnceImpl(); });
256 }
257