xref: /freebsd/contrib/llvm-project/llvm/lib/Target/X86/X86InstrFMA.td (revision 5def4c47d4bd90b209b9b4a4ba9faec15846d8fd)
1//===-- X86InstrFMA.td - FMA Instruction Set ---------------*- 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 FMA (Fused Multiply-Add) instructions.
10//
11//===----------------------------------------------------------------------===//
12
13//===----------------------------------------------------------------------===//
14// FMA3 - Intel 3 operand Fused Multiply-Add instructions
15//===----------------------------------------------------------------------===//
16
17// For all FMA opcodes declared in fma3p_rm_* and fma3s_rm_* multiclasses
18// defined below, both the register and memory variants are commutable.
19// For the register form the commutable operands are 1, 2 and 3.
20// For the memory variant the folded operand must be in 3. Thus,
21// in that case, only the operands 1 and 2 can be swapped.
22// Commuting some of operands may require the opcode change.
23// FMA*213*:
24//   operands 1 and 2 (memory & register forms): *213* --> *213*(no changes);
25//   operands 1 and 3 (register forms only):     *213* --> *231*;
26//   operands 2 and 3 (register forms only):     *213* --> *132*.
27// FMA*132*:
28//   operands 1 and 2 (memory & register forms): *132* --> *231*;
29//   operands 1 and 3 (register forms only):     *132* --> *132*(no changes);
30//   operands 2 and 3 (register forms only):     *132* --> *213*.
31// FMA*231*:
32//   operands 1 and 2 (memory & register forms): *231* --> *132*;
33//   operands 1 and 3 (register forms only):     *231* --> *213*;
34//   operands 2 and 3 (register forms only):     *231* --> *231*(no changes).
35
36multiclass fma3p_rm_213<bits<8> opc, string OpcodeStr, RegisterClass RC,
37                        ValueType VT, X86MemOperand x86memop, PatFrag MemFrag,
38                        SDNode Op, X86FoldableSchedWrite sched> {
39  def r     : FMA3<opc, MRMSrcReg, (outs RC:$dst),
40                   (ins RC:$src1, RC:$src2, RC:$src3),
41                   !strconcat(OpcodeStr,
42                              "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
43                   [(set RC:$dst, (VT (Op RC:$src2, RC:$src1, RC:$src3)))]>,
44                   Sched<[sched]>;
45
46  let mayLoad = 1 in
47  def m     : FMA3<opc, MRMSrcMem, (outs RC:$dst),
48                   (ins RC:$src1, RC:$src2, x86memop:$src3),
49                   !strconcat(OpcodeStr,
50                              "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
51                   [(set RC:$dst, (VT (Op RC:$src2, RC:$src1,
52                                          (MemFrag addr:$src3))))]>,
53                   Sched<[sched.Folded, sched.ReadAfterFold, sched.ReadAfterFold]>;
54}
55
56multiclass fma3p_rm_231<bits<8> opc, string OpcodeStr, RegisterClass RC,
57                        ValueType VT, X86MemOperand x86memop, PatFrag MemFrag,
58                        SDNode Op, X86FoldableSchedWrite sched> {
59  let hasSideEffects = 0 in
60  def r     : FMA3<opc, MRMSrcReg, (outs RC:$dst),
61                   (ins RC:$src1, RC:$src2, RC:$src3),
62                   !strconcat(OpcodeStr,
63                              "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
64                   []>, Sched<[sched]>;
65
66  let mayLoad = 1 in
67  def m     : FMA3<opc, MRMSrcMem, (outs RC:$dst),
68                   (ins RC:$src1, RC:$src2, x86memop:$src3),
69                   !strconcat(OpcodeStr,
70                              "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
71                   [(set RC:$dst, (VT (Op RC:$src2, (MemFrag addr:$src3),
72                                          RC:$src1)))]>,
73                   Sched<[sched.Folded, sched.ReadAfterFold, sched.ReadAfterFold]>;
74}
75
76multiclass fma3p_rm_132<bits<8> opc, string OpcodeStr, RegisterClass RC,
77                        ValueType VT, X86MemOperand x86memop, PatFrag MemFrag,
78                        SDNode Op, X86FoldableSchedWrite sched> {
79  let hasSideEffects = 0 in
80  def r     : FMA3<opc, MRMSrcReg, (outs RC:$dst),
81                   (ins RC:$src1, RC:$src2, RC:$src3),
82                   !strconcat(OpcodeStr,
83                              "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
84                   []>, Sched<[sched]>;
85
86  // Pattern is 312 order so that the load is in a different place from the
87  // 213 and 231 patterns this helps tablegen's duplicate pattern detection.
88  let mayLoad = 1 in
89  def m     : FMA3<opc, MRMSrcMem, (outs RC:$dst),
90                   (ins RC:$src1, RC:$src2, x86memop:$src3),
91                   !strconcat(OpcodeStr,
92                              "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
93                   [(set RC:$dst, (VT (Op (MemFrag addr:$src3), RC:$src1,
94                                          RC:$src2)))]>,
95                   Sched<[sched.Folded, sched.ReadAfterFold, sched.ReadAfterFold]>;
96}
97
98let Constraints = "$src1 = $dst", hasSideEffects = 0, isCommutable = 1,
99    Uses = [MXCSR], mayRaiseFPException = 1 in
100multiclass fma3p_forms<bits<8> opc132, bits<8> opc213, bits<8> opc231,
101                       string OpcodeStr, string PackTy, string Suff,
102                       PatFrag MemFrag128, PatFrag MemFrag256,
103                       SDNode Op, ValueType OpTy128, ValueType OpTy256,
104                       X86SchedWriteWidths sched> {
105  defm NAME#213#Suff : fma3p_rm_213<opc213, !strconcat(OpcodeStr, "213", PackTy),
106                                    VR128, OpTy128, f128mem, MemFrag128, Op, sched.XMM>;
107  defm NAME#231#Suff : fma3p_rm_231<opc231, !strconcat(OpcodeStr, "231", PackTy),
108                                    VR128, OpTy128, f128mem, MemFrag128, Op, sched.XMM>;
109  defm NAME#132#Suff : fma3p_rm_132<opc132, !strconcat(OpcodeStr, "132", PackTy),
110                                    VR128, OpTy128, f128mem, MemFrag128, Op, sched.XMM>;
111
112  defm NAME#213#Suff#Y : fma3p_rm_213<opc213, !strconcat(OpcodeStr, "213", PackTy),
113                                      VR256, OpTy256, f256mem, MemFrag256, Op, sched.YMM>,
114                                      VEX_L;
115  defm NAME#231#Suff#Y : fma3p_rm_231<opc231, !strconcat(OpcodeStr, "231", PackTy),
116                                      VR256, OpTy256, f256mem, MemFrag256, Op, sched.YMM>,
117                                      VEX_L;
118  defm NAME#132#Suff#Y : fma3p_rm_132<opc132, !strconcat(OpcodeStr, "132", PackTy),
119                                      VR256, OpTy256, f256mem, MemFrag256, Op, sched.YMM>,
120                                      VEX_L;
121}
122
123// Fused Multiply-Add
124let ExeDomain = SSEPackedSingle in {
125  defm VFMADD    : fma3p_forms<0x98, 0xA8, 0xB8, "vfmadd", "ps", "PS",
126                               loadv4f32, loadv8f32, any_fma, v4f32, v8f32,
127                               SchedWriteFMA>;
128  defm VFMSUB    : fma3p_forms<0x9A, 0xAA, 0xBA, "vfmsub", "ps", "PS",
129                               loadv4f32, loadv8f32, X86any_Fmsub, v4f32, v8f32,
130                               SchedWriteFMA>;
131  defm VFMADDSUB : fma3p_forms<0x96, 0xA6, 0xB6, "vfmaddsub", "ps", "PS",
132                               loadv4f32, loadv8f32, X86Fmaddsub, v4f32, v8f32,
133                               SchedWriteFMA>;
134  defm VFMSUBADD : fma3p_forms<0x97, 0xA7, 0xB7, "vfmsubadd", "ps", "PS",
135                               loadv4f32, loadv8f32, X86Fmsubadd, v4f32, v8f32,
136                               SchedWriteFMA>;
137}
138
139let ExeDomain = SSEPackedDouble in {
140  defm VFMADD    : fma3p_forms<0x98, 0xA8, 0xB8, "vfmadd", "pd", "PD",
141                               loadv2f64, loadv4f64, any_fma, v2f64,
142                               v4f64, SchedWriteFMA>, VEX_W;
143  defm VFMSUB    : fma3p_forms<0x9A, 0xAA, 0xBA, "vfmsub", "pd", "PD",
144                               loadv2f64, loadv4f64, X86any_Fmsub, v2f64,
145                               v4f64, SchedWriteFMA>, VEX_W;
146  defm VFMADDSUB : fma3p_forms<0x96, 0xA6, 0xB6, "vfmaddsub", "pd", "PD",
147                               loadv2f64, loadv4f64, X86Fmaddsub,
148                               v2f64, v4f64, SchedWriteFMA>, VEX_W;
149  defm VFMSUBADD : fma3p_forms<0x97, 0xA7, 0xB7, "vfmsubadd", "pd", "PD",
150                               loadv2f64, loadv4f64, X86Fmsubadd,
151                               v2f64, v4f64, SchedWriteFMA>, VEX_W;
152}
153
154// Fused Negative Multiply-Add
155let ExeDomain = SSEPackedSingle in {
156  defm VFNMADD : fma3p_forms<0x9C, 0xAC, 0xBC, "vfnmadd", "ps", "PS", loadv4f32,
157                             loadv8f32, X86any_Fnmadd, v4f32, v8f32, SchedWriteFMA>;
158  defm VFNMSUB : fma3p_forms<0x9E, 0xAE, 0xBE, "vfnmsub", "ps", "PS", loadv4f32,
159                             loadv8f32, X86any_Fnmsub, v4f32, v8f32, SchedWriteFMA>;
160}
161let ExeDomain = SSEPackedDouble in {
162  defm VFNMADD : fma3p_forms<0x9C, 0xAC, 0xBC, "vfnmadd", "pd", "PD", loadv2f64,
163                             loadv4f64, X86any_Fnmadd, v2f64, v4f64, SchedWriteFMA>, VEX_W;
164  defm VFNMSUB : fma3p_forms<0x9E, 0xAE, 0xBE, "vfnmsub", "pd", "PD", loadv2f64,
165                             loadv4f64, X86any_Fnmsub, v2f64, v4f64, SchedWriteFMA>, VEX_W;
166}
167
168// All source register operands of FMA opcodes defined in fma3s_rm multiclass
169// can be commuted. In many cases such commute transformation requires an opcode
170// adjustment, for example, commuting the operands 1 and 2 in FMA*132 form
171// would require an opcode change to FMA*231:
172//     FMA*132* reg1, reg2, reg3; // reg1 * reg3 + reg2;
173//     -->
174//     FMA*231* reg2, reg1, reg3; // reg1 * reg3 + reg2;
175// Please see more detailed comment at the very beginning of the section
176// defining FMA3 opcodes above.
177multiclass fma3s_rm_213<bits<8> opc, string OpcodeStr,
178                        X86MemOperand x86memop, RegisterClass RC,
179                        SDPatternOperator OpNode,
180                        X86FoldableSchedWrite sched> {
181  def r : FMA3S<opc, MRMSrcReg, (outs RC:$dst),
182                (ins RC:$src1, RC:$src2, RC:$src3),
183                !strconcat(OpcodeStr,
184                           "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
185                [(set RC:$dst, (OpNode RC:$src2, RC:$src1, RC:$src3))]>,
186                Sched<[sched]>;
187
188  let mayLoad = 1 in
189  def m : FMA3S<opc, MRMSrcMem, (outs RC:$dst),
190                (ins RC:$src1, RC:$src2, x86memop:$src3),
191                !strconcat(OpcodeStr,
192                           "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
193                [(set RC:$dst,
194                  (OpNode RC:$src2, RC:$src1, (load addr:$src3)))]>,
195                Sched<[sched.Folded, sched.ReadAfterFold, sched.ReadAfterFold]>;
196}
197
198multiclass fma3s_rm_231<bits<8> opc, string OpcodeStr,
199                        X86MemOperand x86memop, RegisterClass RC,
200                        SDPatternOperator OpNode, X86FoldableSchedWrite sched> {
201  let hasSideEffects = 0 in
202  def r : FMA3S<opc, MRMSrcReg, (outs RC:$dst),
203                (ins RC:$src1, RC:$src2, RC:$src3),
204                !strconcat(OpcodeStr,
205                           "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
206                []>, Sched<[sched]>;
207
208  let mayLoad = 1 in
209  def m : FMA3S<opc, MRMSrcMem, (outs RC:$dst),
210                (ins RC:$src1, RC:$src2, x86memop:$src3),
211                !strconcat(OpcodeStr,
212                           "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
213                [(set RC:$dst,
214                  (OpNode RC:$src2, (load addr:$src3), RC:$src1))]>,
215                Sched<[sched.Folded, sched.ReadAfterFold, sched.ReadAfterFold]>;
216}
217
218multiclass fma3s_rm_132<bits<8> opc, string OpcodeStr,
219                        X86MemOperand x86memop, RegisterClass RC,
220                        SDPatternOperator OpNode, X86FoldableSchedWrite sched> {
221  let hasSideEffects = 0 in
222  def r : FMA3S<opc, MRMSrcReg, (outs RC:$dst),
223                (ins RC:$src1, RC:$src2, RC:$src3),
224                !strconcat(OpcodeStr,
225                           "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
226                []>, Sched<[sched]>;
227
228  // Pattern is 312 order so that the load is in a different place from the
229  // 213 and 231 patterns this helps tablegen's duplicate pattern detection.
230  let mayLoad = 1 in
231  def m : FMA3S<opc, MRMSrcMem, (outs RC:$dst),
232                (ins RC:$src1, RC:$src2, x86memop:$src3),
233                !strconcat(OpcodeStr,
234                           "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
235                [(set RC:$dst,
236                  (OpNode (load addr:$src3), RC:$src1, RC:$src2))]>,
237                Sched<[sched.Folded, sched.ReadAfterFold, sched.ReadAfterFold]>;
238}
239
240let Constraints = "$src1 = $dst", isCommutable = 1, isCodeGenOnly = 1,
241    hasSideEffects = 0, Uses = [MXCSR], mayRaiseFPException = 1 in
242multiclass fma3s_forms<bits<8> opc132, bits<8> opc213, bits<8> opc231,
243                       string OpStr, string PackTy, string Suff,
244                       SDNode OpNode, RegisterClass RC,
245                       X86MemOperand x86memop, X86FoldableSchedWrite sched> {
246  defm NAME#213#Suff : fma3s_rm_213<opc213, !strconcat(OpStr, "213", PackTy),
247                                    x86memop, RC, OpNode, sched>;
248  defm NAME#231#Suff : fma3s_rm_231<opc231, !strconcat(OpStr, "231", PackTy),
249                                    x86memop, RC, OpNode, sched>;
250  defm NAME#132#Suff : fma3s_rm_132<opc132, !strconcat(OpStr, "132", PackTy),
251                                    x86memop, RC, OpNode, sched>;
252}
253
254// These FMA*_Int instructions are defined specially for being used when
255// the scalar FMA intrinsics are lowered to machine instructions, and in that
256// sense, they are similar to existing ADD*_Int, SUB*_Int, MUL*_Int, etc.
257// instructions.
258//
259// All of the FMA*_Int opcodes are defined as commutable here.
260// Commuting the 2nd and 3rd source register operands of FMAs is quite trivial
261// and the corresponding optimizations have been developed.
262// Commuting the 1st operand of FMA*_Int requires some additional analysis,
263// the commute optimization is legal only if all users of FMA*_Int use only
264// the lowest element of the FMA*_Int instruction. Even though such analysis
265// may be not implemented yet we allow the routines doing the actual commute
266// transformation to decide if one or another instruction is commutable or not.
267let Constraints = "$src1 = $dst", isCommutable = 1, hasSideEffects = 0,
268    Uses = [MXCSR], mayRaiseFPException = 1 in
269multiclass fma3s_rm_int<bits<8> opc, string OpcodeStr,
270                        Operand memopr, RegisterClass RC,
271                        X86FoldableSchedWrite sched> {
272  def r_Int : FMA3S_Int<opc, MRMSrcReg, (outs RC:$dst),
273                        (ins RC:$src1, RC:$src2, RC:$src3),
274                        !strconcat(OpcodeStr,
275                                   "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
276                        []>, Sched<[sched]>;
277
278  let mayLoad = 1 in
279  def m_Int : FMA3S_Int<opc, MRMSrcMem, (outs RC:$dst),
280                        (ins RC:$src1, RC:$src2, memopr:$src3),
281                        !strconcat(OpcodeStr,
282                                   "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
283                        []>, Sched<[sched.Folded, sched.ReadAfterFold, sched.ReadAfterFold]>;
284}
285
286// The FMA 213 form is created for lowering of scalar FMA intrinsics
287// to machine instructions.
288// The FMA 132 form can trivially be get by commuting the 2nd and 3rd operands
289// of FMA 213 form.
290// The FMA 231 form can be get only by commuting the 1st operand of 213 or 132
291// forms and is possible only after special analysis of all uses of the initial
292// instruction. Such analysis do not exist yet and thus introducing the 231
293// form of FMA*_Int instructions is done using an optimistic assumption that
294// such analysis will be implemented eventually.
295multiclass fma3s_int_forms<bits<8> opc132, bits<8> opc213, bits<8> opc231,
296                           string OpStr, string PackTy, string Suff,
297                           RegisterClass RC, Operand memop,
298                           X86FoldableSchedWrite sched> {
299  defm NAME#132#Suff : fma3s_rm_int<opc132, !strconcat(OpStr, "132", PackTy),
300                                    memop, RC, sched>;
301  defm NAME#213#Suff : fma3s_rm_int<opc213, !strconcat(OpStr, "213", PackTy),
302                                    memop, RC, sched>;
303  defm NAME#231#Suff : fma3s_rm_int<opc231, !strconcat(OpStr, "231", PackTy),
304                                    memop, RC, sched>;
305}
306
307multiclass fma3s<bits<8> opc132, bits<8> opc213, bits<8> opc231,
308                 string OpStr, SDNode OpNode, X86FoldableSchedWrite sched> {
309  let ExeDomain = SSEPackedSingle in
310  defm NAME : fma3s_forms<opc132, opc213, opc231, OpStr, "ss", "SS", OpNode,
311                          FR32, f32mem, sched>,
312              fma3s_int_forms<opc132, opc213, opc231, OpStr, "ss", "SS",
313                              VR128, ssmem, sched>;
314
315  let ExeDomain = SSEPackedDouble in
316  defm NAME : fma3s_forms<opc132, opc213, opc231, OpStr, "sd", "SD", OpNode,
317                        FR64, f64mem, sched>,
318              fma3s_int_forms<opc132, opc213, opc231, OpStr, "sd", "SD",
319                              VR128, sdmem, sched>, VEX_W;
320}
321
322defm VFMADD : fma3s<0x99, 0xA9, 0xB9, "vfmadd", any_fma,
323                    SchedWriteFMA.Scl>, VEX_LIG;
324defm VFMSUB : fma3s<0x9B, 0xAB, 0xBB, "vfmsub", X86any_Fmsub,
325                    SchedWriteFMA.Scl>, VEX_LIG;
326
327defm VFNMADD : fma3s<0x9D, 0xAD, 0xBD, "vfnmadd", X86any_Fnmadd,
328                     SchedWriteFMA.Scl>, VEX_LIG;
329defm VFNMSUB : fma3s<0x9F, 0xAF, 0xBF, "vfnmsub", X86any_Fnmsub,
330                     SchedWriteFMA.Scl>, VEX_LIG;
331
332multiclass scalar_fma_patterns<SDNode Op, string Prefix, string Suffix,
333                               SDNode Move, ValueType VT, ValueType EltVT,
334                               RegisterClass RC, PatFrag mem_frag> {
335  let Predicates = [HasFMA, NoAVX512] in {
336    def : Pat<(VT (Move (VT VR128:$src1), (VT (scalar_to_vector
337                (Op RC:$src2,
338                    (EltVT (extractelt (VT VR128:$src1), (iPTR 0))),
339                    RC:$src3))))),
340              (!cast<Instruction>(Prefix#"213"#Suffix#"r_Int")
341               VR128:$src1, (VT (COPY_TO_REGCLASS RC:$src2, VR128)),
342               (VT (COPY_TO_REGCLASS RC:$src3, VR128)))>;
343
344    def : Pat<(VT (Move (VT VR128:$src1), (VT (scalar_to_vector
345                (Op RC:$src2, RC:$src3,
346                    (EltVT (extractelt (VT VR128:$src1), (iPTR 0)))))))),
347              (!cast<Instruction>(Prefix#"231"#Suffix#"r_Int")
348               VR128:$src1, (VT (COPY_TO_REGCLASS RC:$src2, VR128)),
349               (VT (COPY_TO_REGCLASS RC:$src3, VR128)))>;
350
351    def : Pat<(VT (Move (VT VR128:$src1), (VT (scalar_to_vector
352                (Op RC:$src2,
353                    (EltVT (extractelt (VT VR128:$src1), (iPTR 0))),
354                    (mem_frag addr:$src3)))))),
355              (!cast<Instruction>(Prefix#"213"#Suffix#"m_Int")
356               VR128:$src1, (VT (COPY_TO_REGCLASS RC:$src2, VR128)),
357               addr:$src3)>;
358
359    def : Pat<(VT (Move (VT VR128:$src1), (VT (scalar_to_vector
360                (Op (EltVT (extractelt (VT VR128:$src1), (iPTR 0))),
361                    (mem_frag addr:$src3), RC:$src2))))),
362              (!cast<Instruction>(Prefix#"132"#Suffix#"m_Int")
363               VR128:$src1, (VT (COPY_TO_REGCLASS RC:$src2, VR128)),
364               addr:$src3)>;
365
366    def : Pat<(VT (Move (VT VR128:$src1), (VT (scalar_to_vector
367                (Op RC:$src2, (mem_frag addr:$src3),
368                    (EltVT (extractelt (VT VR128:$src1), (iPTR 0)))))))),
369              (!cast<Instruction>(Prefix#"231"#Suffix#"m_Int")
370               VR128:$src1, (VT (COPY_TO_REGCLASS RC:$src2, VR128)),
371               addr:$src3)>;
372  }
373}
374
375defm : scalar_fma_patterns<any_fma, "VFMADD", "SS", X86Movss, v4f32, f32, FR32, loadf32>;
376defm : scalar_fma_patterns<X86any_Fmsub, "VFMSUB", "SS", X86Movss, v4f32, f32, FR32, loadf32>;
377defm : scalar_fma_patterns<X86any_Fnmadd, "VFNMADD", "SS", X86Movss, v4f32, f32, FR32, loadf32>;
378defm : scalar_fma_patterns<X86any_Fnmsub, "VFNMSUB", "SS", X86Movss, v4f32, f32, FR32, loadf32>;
379
380defm : scalar_fma_patterns<any_fma, "VFMADD", "SD", X86Movsd, v2f64, f64, FR64, loadf64>;
381defm : scalar_fma_patterns<X86any_Fmsub, "VFMSUB", "SD", X86Movsd, v2f64, f64, FR64, loadf64>;
382defm : scalar_fma_patterns<X86any_Fnmadd, "VFNMADD", "SD", X86Movsd, v2f64, f64, FR64, loadf64>;
383defm : scalar_fma_patterns<X86any_Fnmsub, "VFNMSUB", "SD", X86Movsd, v2f64, f64, FR64, loadf64>;
384
385//===----------------------------------------------------------------------===//
386// FMA4 - AMD 4 operand Fused Multiply-Add instructions
387//===----------------------------------------------------------------------===//
388
389let Uses = [MXCSR], mayRaiseFPException = 1 in
390multiclass fma4s<bits<8> opc, string OpcodeStr, RegisterClass RC,
391                 X86MemOperand x86memop, ValueType OpVT, SDNode OpNode,
392                 PatFrag mem_frag, X86FoldableSchedWrite sched> {
393  let isCommutable = 1 in
394  def rr : FMA4S<opc, MRMSrcRegOp4, (outs RC:$dst),
395           (ins RC:$src1, RC:$src2, RC:$src3),
396           !strconcat(OpcodeStr,
397           "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
398           [(set RC:$dst,
399             (OpVT (OpNode RC:$src1, RC:$src2, RC:$src3)))]>, VEX_W, VEX_LIG,
400           Sched<[sched]>;
401  def rm : FMA4S<opc, MRMSrcMemOp4, (outs RC:$dst),
402           (ins RC:$src1, RC:$src2, x86memop:$src3),
403           !strconcat(OpcodeStr,
404           "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
405           [(set RC:$dst, (OpNode RC:$src1, RC:$src2,
406                           (mem_frag addr:$src3)))]>, VEX_W, VEX_LIG,
407           Sched<[sched.Folded, sched.ReadAfterFold, sched.ReadAfterFold]>;
408  def mr : FMA4S<opc, MRMSrcMem, (outs RC:$dst),
409           (ins RC:$src1, x86memop:$src2, RC:$src3),
410           !strconcat(OpcodeStr,
411           "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
412           [(set RC:$dst,
413             (OpNode RC:$src1, (mem_frag addr:$src2), RC:$src3))]>, VEX_LIG,
414           Sched<[sched.Folded, sched.ReadAfterFold,
415                  // x86memop:$src2
416                  ReadDefault, ReadDefault, ReadDefault, ReadDefault,
417                  ReadDefault,
418                  // RC:$src3
419                  sched.ReadAfterFold]>;
420// For disassembler
421let isCodeGenOnly = 1, ForceDisassemble = 1, hasSideEffects = 0 in
422  def rr_REV : FMA4S<opc, MRMSrcReg, (outs RC:$dst),
423               (ins RC:$src1, RC:$src2, RC:$src3),
424               !strconcat(OpcodeStr,
425               "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"), []>,
426               VEX_LIG, FoldGenData<NAME#rr>, Sched<[sched]>;
427}
428
429multiclass fma4s_int<bits<8> opc, string OpcodeStr, Operand memop,
430                     ValueType VT, X86FoldableSchedWrite sched> {
431let isCodeGenOnly = 1, hasSideEffects = 0,
432    Uses = [MXCSR], mayRaiseFPException = 1 in {
433  def rr_Int : FMA4S_Int<opc, MRMSrcRegOp4, (outs VR128:$dst),
434               (ins VR128:$src1, VR128:$src2, VR128:$src3),
435               !strconcat(OpcodeStr,
436               "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
437               []>, VEX_W, VEX_LIG, Sched<[sched]>;
438  let mayLoad = 1 in
439  def rm_Int : FMA4S_Int<opc, MRMSrcMemOp4, (outs VR128:$dst),
440               (ins VR128:$src1, VR128:$src2, memop:$src3),
441               !strconcat(OpcodeStr,
442               "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
443               []>, VEX_W, VEX_LIG,
444               Sched<[sched.Folded, sched.ReadAfterFold, sched.ReadAfterFold]>;
445  let mayLoad = 1 in
446  def mr_Int : FMA4S_Int<opc, MRMSrcMem, (outs VR128:$dst),
447               (ins VR128:$src1, memop:$src2, VR128:$src3),
448               !strconcat(OpcodeStr,
449               "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
450               []>,
451               VEX_LIG, Sched<[sched.Folded, sched.ReadAfterFold,
452                               // memop:$src2
453                               ReadDefault, ReadDefault, ReadDefault,
454                               ReadDefault, ReadDefault,
455                               // VR128::$src3
456                               sched.ReadAfterFold]>;
457  def rr_Int_REV : FMA4S_Int<opc, MRMSrcReg, (outs VR128:$dst),
458               (ins VR128:$src1, VR128:$src2, VR128:$src3),
459               !strconcat(OpcodeStr,
460               "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
461               []>, VEX_LIG, FoldGenData<NAME#rr_Int>, Sched<[sched]>;
462} // isCodeGenOnly = 1
463}
464
465let Uses = [MXCSR], mayRaiseFPException = 1 in
466multiclass fma4p<bits<8> opc, string OpcodeStr, SDNode OpNode,
467                 ValueType OpVT128, ValueType OpVT256,
468                 PatFrag ld_frag128, PatFrag ld_frag256,
469                 X86SchedWriteWidths sched> {
470  let isCommutable = 1 in
471  def rr : FMA4<opc, MRMSrcRegOp4, (outs VR128:$dst),
472           (ins VR128:$src1, VR128:$src2, VR128:$src3),
473           !strconcat(OpcodeStr,
474           "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
475           [(set VR128:$dst,
476             (OpVT128 (OpNode VR128:$src1, VR128:$src2, VR128:$src3)))]>,
477           VEX_W, Sched<[sched.XMM]>;
478  def rm : FMA4<opc, MRMSrcMemOp4, (outs VR128:$dst),
479           (ins VR128:$src1, VR128:$src2, f128mem:$src3),
480           !strconcat(OpcodeStr,
481           "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
482           [(set VR128:$dst, (OpNode VR128:$src1, VR128:$src2,
483                              (ld_frag128 addr:$src3)))]>, VEX_W,
484           Sched<[sched.XMM.Folded, sched.XMM.ReadAfterFold, sched.XMM.ReadAfterFold]>;
485  def mr : FMA4<opc, MRMSrcMem, (outs VR128:$dst),
486           (ins VR128:$src1, f128mem:$src2, VR128:$src3),
487           !strconcat(OpcodeStr,
488           "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
489           [(set VR128:$dst,
490             (OpNode VR128:$src1, (ld_frag128 addr:$src2), VR128:$src3))]>,
491           Sched<[sched.XMM.Folded, sched.XMM.ReadAfterFold,
492                  // f128mem:$src2
493                  ReadDefault, ReadDefault, ReadDefault, ReadDefault,
494                  ReadDefault,
495                  // VR128::$src3
496                  sched.XMM.ReadAfterFold]>;
497  let isCommutable = 1 in
498  def Yrr : FMA4<opc, MRMSrcRegOp4, (outs VR256:$dst),
499           (ins VR256:$src1, VR256:$src2, VR256:$src3),
500           !strconcat(OpcodeStr,
501           "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
502           [(set VR256:$dst,
503             (OpVT256 (OpNode VR256:$src1, VR256:$src2, VR256:$src3)))]>,
504           VEX_W, VEX_L, Sched<[sched.YMM]>;
505  def Yrm : FMA4<opc, MRMSrcMemOp4, (outs VR256:$dst),
506           (ins VR256:$src1, VR256:$src2, f256mem:$src3),
507           !strconcat(OpcodeStr,
508           "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
509           [(set VR256:$dst, (OpNode VR256:$src1, VR256:$src2,
510                              (ld_frag256 addr:$src3)))]>, VEX_W, VEX_L,
511           Sched<[sched.YMM.Folded, sched.YMM.ReadAfterFold, sched.YMM.ReadAfterFold]>;
512  def Ymr : FMA4<opc, MRMSrcMem, (outs VR256:$dst),
513           (ins VR256:$src1, f256mem:$src2, VR256:$src3),
514           !strconcat(OpcodeStr,
515           "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
516           [(set VR256:$dst, (OpNode VR256:$src1,
517                              (ld_frag256 addr:$src2), VR256:$src3))]>, VEX_L,
518           Sched<[sched.YMM.Folded, sched.YMM.ReadAfterFold,
519                  // f256mem:$src2
520                  ReadDefault, ReadDefault, ReadDefault, ReadDefault,
521                  ReadDefault,
522                  // VR256::$src3
523                  sched.YMM.ReadAfterFold]>;
524// For disassembler
525let isCodeGenOnly = 1, ForceDisassemble = 1, hasSideEffects = 0 in {
526  def rr_REV : FMA4<opc, MRMSrcReg, (outs VR128:$dst),
527               (ins VR128:$src1, VR128:$src2, VR128:$src3),
528               !strconcat(OpcodeStr,
529               "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"), []>,
530               Sched<[sched.XMM]>, FoldGenData<NAME#rr>;
531  def Yrr_REV : FMA4<opc, MRMSrcReg, (outs VR256:$dst),
532                (ins VR256:$src1, VR256:$src2, VR256:$src3),
533                !strconcat(OpcodeStr,
534                "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"), []>,
535                VEX_L, Sched<[sched.YMM]>, FoldGenData<NAME#Yrr>;
536} // isCodeGenOnly = 1
537}
538
539let ExeDomain = SSEPackedSingle in {
540  // Scalar Instructions
541  defm VFMADDSS4  : fma4s<0x6A, "vfmaddss", FR32, f32mem, f32, any_fma, loadf32,
542                          SchedWriteFMA.Scl>,
543                    fma4s_int<0x6A, "vfmaddss", ssmem, v4f32,
544                              SchedWriteFMA.Scl>;
545  defm VFMSUBSS4  : fma4s<0x6E, "vfmsubss", FR32, f32mem, f32, X86any_Fmsub, loadf32,
546                          SchedWriteFMA.Scl>,
547                    fma4s_int<0x6E, "vfmsubss", ssmem, v4f32,
548                              SchedWriteFMA.Scl>;
549  defm VFNMADDSS4 : fma4s<0x7A, "vfnmaddss", FR32, f32mem, f32,
550                          X86any_Fnmadd, loadf32, SchedWriteFMA.Scl>,
551                    fma4s_int<0x7A, "vfnmaddss", ssmem, v4f32,
552                              SchedWriteFMA.Scl>;
553  defm VFNMSUBSS4 : fma4s<0x7E, "vfnmsubss", FR32, f32mem, f32,
554                          X86any_Fnmsub, loadf32, SchedWriteFMA.Scl>,
555                    fma4s_int<0x7E, "vfnmsubss", ssmem, v4f32,
556                              SchedWriteFMA.Scl>;
557  // Packed Instructions
558  defm VFMADDPS4    : fma4p<0x68, "vfmaddps", any_fma, v4f32, v8f32,
559                            loadv4f32, loadv8f32, SchedWriteFMA>;
560  defm VFMSUBPS4    : fma4p<0x6C, "vfmsubps", X86any_Fmsub, v4f32, v8f32,
561                            loadv4f32, loadv8f32, SchedWriteFMA>;
562  defm VFNMADDPS4   : fma4p<0x78, "vfnmaddps", X86any_Fnmadd, v4f32, v8f32,
563                            loadv4f32, loadv8f32, SchedWriteFMA>;
564  defm VFNMSUBPS4   : fma4p<0x7C, "vfnmsubps", X86any_Fnmsub, v4f32, v8f32,
565                            loadv4f32, loadv8f32, SchedWriteFMA>;
566  defm VFMADDSUBPS4 : fma4p<0x5C, "vfmaddsubps", X86Fmaddsub, v4f32, v8f32,
567                            loadv4f32, loadv8f32, SchedWriteFMA>;
568  defm VFMSUBADDPS4 : fma4p<0x5E, "vfmsubaddps", X86Fmsubadd, v4f32, v8f32,
569                            loadv4f32, loadv8f32, SchedWriteFMA>;
570}
571
572let ExeDomain = SSEPackedDouble in {
573  // Scalar Instructions
574  defm VFMADDSD4  : fma4s<0x6B, "vfmaddsd", FR64, f64mem, f64, any_fma, loadf64,
575                          SchedWriteFMA.Scl>,
576                    fma4s_int<0x6B, "vfmaddsd", sdmem, v2f64,
577                              SchedWriteFMA.Scl>;
578  defm VFMSUBSD4  : fma4s<0x6F, "vfmsubsd", FR64, f64mem, f64, X86any_Fmsub, loadf64,
579                          SchedWriteFMA.Scl>,
580                    fma4s_int<0x6F, "vfmsubsd", sdmem, v2f64,
581                              SchedWriteFMA.Scl>;
582  defm VFNMADDSD4 : fma4s<0x7B, "vfnmaddsd", FR64, f64mem, f64,
583                          X86any_Fnmadd, loadf64, SchedWriteFMA.Scl>,
584                    fma4s_int<0x7B, "vfnmaddsd", sdmem, v2f64,
585                              SchedWriteFMA.Scl>;
586  defm VFNMSUBSD4 : fma4s<0x7F, "vfnmsubsd", FR64, f64mem, f64,
587                          X86any_Fnmsub, loadf64, SchedWriteFMA.Scl>,
588                    fma4s_int<0x7F, "vfnmsubsd", sdmem, v2f64,
589                              SchedWriteFMA.Scl>;
590  // Packed Instructions
591  defm VFMADDPD4    : fma4p<0x69, "vfmaddpd", any_fma, v2f64, v4f64,
592                            loadv2f64, loadv4f64, SchedWriteFMA>;
593  defm VFMSUBPD4    : fma4p<0x6D, "vfmsubpd", X86any_Fmsub, v2f64, v4f64,
594                            loadv2f64, loadv4f64, SchedWriteFMA>;
595  defm VFNMADDPD4   : fma4p<0x79, "vfnmaddpd", X86any_Fnmadd, v2f64, v4f64,
596                            loadv2f64, loadv4f64, SchedWriteFMA>;
597  defm VFNMSUBPD4   : fma4p<0x7D, "vfnmsubpd", X86any_Fnmsub, v2f64, v4f64,
598                            loadv2f64, loadv4f64, SchedWriteFMA>;
599  defm VFMADDSUBPD4 : fma4p<0x5D, "vfmaddsubpd", X86Fmaddsub, v2f64, v4f64,
600                            loadv2f64, loadv4f64, SchedWriteFMA>;
601  defm VFMSUBADDPD4 : fma4p<0x5F, "vfmsubaddpd", X86Fmsubadd, v2f64, v4f64,
602                            loadv2f64, loadv4f64, SchedWriteFMA>;
603}
604
605multiclass scalar_fma4_patterns<SDNode Op, string Name,
606                               ValueType VT, ValueType EltVT,
607                               RegisterClass RC, PatFrag mem_frag> {
608  let Predicates = [HasFMA4] in {
609    def : Pat<(VT (X86vzmovl (VT (scalar_to_vector
610                                  (Op RC:$src1, RC:$src2, RC:$src3))))),
611              (!cast<Instruction>(Name#"rr_Int")
612               (VT (COPY_TO_REGCLASS RC:$src1, VR128)),
613               (VT (COPY_TO_REGCLASS RC:$src2, VR128)),
614               (VT (COPY_TO_REGCLASS RC:$src3, VR128)))>;
615
616    def : Pat<(VT (X86vzmovl (VT (scalar_to_vector
617                                  (Op RC:$src1, RC:$src2,
618                                      (mem_frag addr:$src3)))))),
619              (!cast<Instruction>(Name#"rm_Int")
620               (VT (COPY_TO_REGCLASS RC:$src1, VR128)),
621               (VT (COPY_TO_REGCLASS RC:$src2, VR128)), addr:$src3)>;
622
623    def : Pat<(VT (X86vzmovl (VT (scalar_to_vector
624                                  (Op RC:$src1, (mem_frag addr:$src2),
625                                      RC:$src3))))),
626              (!cast<Instruction>(Name#"mr_Int")
627               (VT (COPY_TO_REGCLASS RC:$src1, VR128)), addr:$src2,
628               (VT (COPY_TO_REGCLASS RC:$src3, VR128)))>;
629  }
630}
631
632defm : scalar_fma4_patterns<any_fma, "VFMADDSS4", v4f32, f32, FR32, loadf32>;
633defm : scalar_fma4_patterns<X86any_Fmsub, "VFMSUBSS4", v4f32, f32, FR32, loadf32>;
634defm : scalar_fma4_patterns<X86any_Fnmadd, "VFNMADDSS4", v4f32, f32, FR32, loadf32>;
635defm : scalar_fma4_patterns<X86any_Fnmsub, "VFNMSUBSS4", v4f32, f32, FR32, loadf32>;
636
637defm : scalar_fma4_patterns<any_fma, "VFMADDSD4", v2f64, f64, FR64, loadf64>;
638defm : scalar_fma4_patterns<X86any_Fmsub, "VFMSUBSD4", v2f64, f64, FR64, loadf64>;
639defm : scalar_fma4_patterns<X86any_Fnmadd, "VFNMADDSD4", v2f64, f64, FR64, loadf64>;
640defm : scalar_fma4_patterns<X86any_Fnmsub, "VFNMSUBSD4", v2f64, f64, FR64, loadf64>;
641