• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1//===-- X86InstrFMA.td - FMA Instruction Set ---------------*- tablegen -*-===//
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 describes FMA (Fused Multiply-Add) instructions.
11//
12//===----------------------------------------------------------------------===//
13
14//===----------------------------------------------------------------------===//
15// FMA3 - Intel 3 operand Fused Multiply-Add instructions
16//===----------------------------------------------------------------------===//
17
18// For all FMA opcodes declared in fma3p_rm and fma3s_rm milticlasses defined
19// below, both the register and memory variants are commutable.
20// For the register form the commutable operands are 1, 2 and 3.
21// For the memory variant the folded operand must be in 3. Thus,
22// in that case, only the operands 1 and 2 can be swapped.
23// Commuting some of operands may require the opcode change.
24// FMA*213*:
25//   operands 1 and 2 (memory & register forms): *213* --> *213*(no changes);
26//   operands 1 and 3 (register forms only):     *213* --> *231*;
27//   operands 2 and 3 (register forms only):     *213* --> *132*.
28// FMA*132*:
29//   operands 1 and 2 (memory & register forms): *132* --> *231*;
30//   operands 1 and 3 (register forms only):     *132* --> *132*(no changes);
31//   operands 2 and 3 (register forms only):     *132* --> *213*.
32// FMA*231*:
33//   operands 1 and 2 (memory & register forms): *231* --> *132*;
34//   operands 1 and 3 (register forms only):     *231* --> *213*;
35//   operands 2 and 3 (register forms only):     *231* --> *231*(no changes).
36
37let Constraints = "$src1 = $dst", hasSideEffects = 0, isCommutable = 1 in
38multiclass fma3p_rm<bits<8> opc, string OpcodeStr,
39                    PatFrag MemFrag128, PatFrag MemFrag256,
40                    ValueType OpVT128, ValueType OpVT256,
41                    SDPatternOperator Op = null_frag> {
42  let usesCustomInserter = 1 in
43  def r     : FMA3<opc, MRMSrcReg, (outs VR128:$dst),
44                   (ins VR128:$src1, VR128:$src2, VR128:$src3),
45                   !strconcat(OpcodeStr,
46                              "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
47                   [(set VR128:$dst, (OpVT128 (Op VR128:$src2,
48                                               VR128:$src1, VR128:$src3)))]>;
49
50  let mayLoad = 1 in
51  def m     : FMA3<opc, MRMSrcMem, (outs VR128:$dst),
52                   (ins VR128:$src1, VR128:$src2, f128mem:$src3),
53                   !strconcat(OpcodeStr,
54                              "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
55                   [(set VR128:$dst, (OpVT128 (Op VR128:$src2, VR128:$src1,
56                                               (MemFrag128 addr:$src3))))]>;
57
58  let usesCustomInserter = 1 in
59  def rY    : FMA3<opc, MRMSrcReg, (outs VR256:$dst),
60                   (ins VR256:$src1, VR256:$src2, VR256:$src3),
61                   !strconcat(OpcodeStr,
62                              "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
63                   [(set VR256:$dst, (OpVT256 (Op VR256:$src2, VR256:$src1,
64                                               VR256:$src3)))]>, VEX_L;
65
66  let mayLoad = 1 in
67  def mY    : FMA3<opc, MRMSrcMem, (outs VR256:$dst),
68                   (ins VR256:$src1, VR256:$src2, f256mem:$src3),
69                   !strconcat(OpcodeStr,
70                              "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
71                   [(set VR256:$dst,
72                     (OpVT256 (Op VR256:$src2, VR256:$src1,
73                               (MemFrag256 addr:$src3))))]>, VEX_L;
74}
75
76multiclass fma3p_forms<bits<8> opc132, bits<8> opc213, bits<8> opc231,
77                       string OpcodeStr, string PackTy,
78                       PatFrag MemFrag128, PatFrag MemFrag256,
79                       SDNode Op, ValueType OpTy128, ValueType OpTy256> {
80  defm r213 : fma3p_rm<opc213,
81                       !strconcat(OpcodeStr, "213", PackTy),
82                       MemFrag128, MemFrag256, OpTy128, OpTy256, Op>;
83  defm r132 : fma3p_rm<opc132,
84                       !strconcat(OpcodeStr, "132", PackTy),
85                       MemFrag128, MemFrag256, OpTy128, OpTy256>;
86  defm r231 : fma3p_rm<opc231,
87                       !strconcat(OpcodeStr, "231", PackTy),
88                       MemFrag128, MemFrag256, OpTy128, OpTy256>;
89}
90
91// Fused Multiply-Add
92let ExeDomain = SSEPackedSingle in {
93  defm VFMADDPS    : fma3p_forms<0x98, 0xA8, 0xB8, "vfmadd", "ps", loadv4f32,
94                                 loadv8f32, X86Fmadd, v4f32, v8f32>;
95  defm VFMSUBPS    : fma3p_forms<0x9A, 0xAA, 0xBA, "vfmsub", "ps", loadv4f32,
96                                 loadv8f32, X86Fmsub, v4f32, v8f32>;
97  defm VFMADDSUBPS : fma3p_forms<0x96, 0xA6, 0xB6, "vfmaddsub", "ps",
98                                 loadv4f32, loadv8f32, X86Fmaddsub,
99                                 v4f32, v8f32>;
100  defm VFMSUBADDPS : fma3p_forms<0x97, 0xA7, 0xB7, "vfmsubadd", "ps",
101                                 loadv4f32, loadv8f32, X86Fmsubadd,
102                                 v4f32, v8f32>;
103}
104
105let ExeDomain = SSEPackedDouble in {
106  defm VFMADDPD    : fma3p_forms<0x98, 0xA8, 0xB8, "vfmadd", "pd", loadv2f64,
107                                 loadv4f64, X86Fmadd, v2f64, v4f64>, VEX_W;
108  defm VFMSUBPD    : fma3p_forms<0x9A, 0xAA, 0xBA, "vfmsub", "pd", loadv2f64,
109                                 loadv4f64, X86Fmsub, v2f64, v4f64>, VEX_W;
110  defm VFMADDSUBPD : fma3p_forms<0x96, 0xA6, 0xB6, "vfmaddsub", "pd",
111                                 loadv2f64, loadv4f64, X86Fmaddsub,
112                                 v2f64, v4f64>, VEX_W;
113  defm VFMSUBADDPD : fma3p_forms<0x97, 0xA7, 0xB7, "vfmsubadd", "pd",
114                                 loadv2f64, loadv4f64, X86Fmsubadd,
115                                 v2f64, v4f64>, VEX_W;
116}
117
118// Fused Negative Multiply-Add
119let ExeDomain = SSEPackedSingle in {
120  defm VFNMADDPS : fma3p_forms<0x9C, 0xAC, 0xBC, "vfnmadd", "ps",  loadv4f32,
121                               loadv8f32, X86Fnmadd, v4f32, v8f32>;
122  defm VFNMSUBPS : fma3p_forms<0x9E, 0xAE, 0xBE, "vfnmsub", "ps",  loadv4f32,
123                               loadv8f32, X86Fnmsub, v4f32, v8f32>;
124}
125let ExeDomain = SSEPackedDouble in {
126  defm VFNMADDPD : fma3p_forms<0x9C, 0xAC, 0xBC, "vfnmadd", "pd", loadv2f64,
127                               loadv4f64, X86Fnmadd, v2f64, v4f64>, VEX_W;
128  defm VFNMSUBPD : fma3p_forms<0x9E, 0xAE, 0xBE, "vfnmsub", "pd",
129                               loadv2f64, loadv4f64, X86Fnmsub, v2f64,
130                               v4f64>, VEX_W;
131}
132
133// All source register operands of FMA opcodes defined in fma3s_rm multiclass
134// can be commuted. In many cases such commute transformation requres an opcode
135// adjustment, for example, commuting the operands 1 and 2 in FMA*132 form
136// would require an opcode change to FMA*231:
137//     FMA*132* reg1, reg2, reg3; // reg1 * reg3 + reg2;
138//     -->
139//     FMA*231* reg2, reg1, reg3; // reg1 * reg3 + reg2;
140// Please see more detailed comment at the very beginning of the section
141// defining FMA3 opcodes above.
142let Constraints = "$src1 = $dst", isCommutable = 1, hasSideEffects = 0 in
143multiclass fma3s_rm<bits<8> opc, string OpcodeStr,
144                    X86MemOperand x86memop, RegisterClass RC,
145                    SDPatternOperator OpNode = null_frag> {
146  let usesCustomInserter = 1 in
147  def r     : FMA3<opc, MRMSrcReg, (outs RC:$dst),
148                   (ins RC:$src1, RC:$src2, RC:$src3),
149                   !strconcat(OpcodeStr,
150                              "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
151                   [(set RC:$dst, (OpNode RC:$src2, RC:$src1, RC:$src3))]>;
152
153  let mayLoad = 1 in
154  def m     : FMA3<opc, MRMSrcMem, (outs RC:$dst),
155                   (ins RC:$src1, RC:$src2, x86memop:$src3),
156                   !strconcat(OpcodeStr,
157                              "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
158                   [(set RC:$dst,
159                     (OpNode RC:$src2, RC:$src1, (load addr:$src3)))]>;
160}
161
162// These FMA*_Int instructions are defined specially for being used when
163// the scalar FMA intrinsics are lowered to machine instructions, and in that
164// sense, they are similar to existing ADD*_Int, SUB*_Int, MUL*_Int, etc.
165// instructions.
166//
167// All of the FMA*_Int opcodes are defined as commutable here.
168// Commuting the 2nd and 3rd source register operands of FMAs is quite trivial
169// and the corresponding optimizations have been developed.
170// Commuting the 1st operand of FMA*_Int requires some additional analysis,
171// the commute optimization is legal only if all users of FMA*_Int use only
172// the lowest element of the FMA*_Int instruction. Even though such analysis
173// may be not implemented yet we allow the routines doing the actual commute
174// transformation to decide if one or another instruction is commutable or not.
175let Constraints = "$src1 = $dst", isCommutable = 1, isCodeGenOnly = 1,
176    hasSideEffects = 0 in
177multiclass fma3s_rm_int<bits<8> opc, string OpcodeStr,
178                        Operand memopr, RegisterClass RC> {
179  def r_Int : FMA3<opc, MRMSrcReg, (outs RC:$dst),
180                   (ins RC:$src1, RC:$src2, RC:$src3),
181                   !strconcat(OpcodeStr,
182                              "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
183                   []>;
184
185  let mayLoad = 1 in
186  def m_Int : FMA3<opc, MRMSrcMem, (outs RC:$dst),
187                   (ins RC:$src1, RC:$src2, memopr:$src3),
188                   !strconcat(OpcodeStr,
189                              "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
190                   []>;
191}
192
193multiclass fma3s_forms<bits<8> opc132, bits<8> opc213, bits<8> opc231,
194                       string OpStr, string PackTy,
195                       SDNode OpNode, RegisterClass RC,
196                       X86MemOperand x86memop> {
197  defm r132 : fma3s_rm<opc132, !strconcat(OpStr, "132", PackTy), x86memop, RC>;
198  defm r213 : fma3s_rm<opc213, !strconcat(OpStr, "213", PackTy), x86memop, RC,
199                       OpNode>;
200  defm r231 : fma3s_rm<opc231, !strconcat(OpStr, "231", PackTy), x86memop, RC>;
201}
202
203// The FMA 213 form is created for lowering of scalar FMA intrinscis
204// to machine instructions.
205// The FMA 132 form can trivially be get by commuting the 2nd and 3rd operands
206// of FMA 213 form.
207// The FMA 231 form can be get only by commuting the 1st operand of 213 or 132
208// forms and is possible only after special analysis of all uses of the initial
209// instruction. Such analysis do not exist yet and thus introducing the 231
210// form of FMA*_Int instructions is done using an optimistic assumption that
211// such analysis will be implemented eventually.
212multiclass fma3s_int_forms<bits<8> opc132, bits<8> opc213, bits<8> opc231,
213                           string OpStr, string PackTy,
214                           RegisterClass RC, Operand memop> {
215  defm r132 : fma3s_rm_int<opc132, !strconcat(OpStr, "132", PackTy),
216                           memop, RC>;
217  defm r213 : fma3s_rm_int<opc213, !strconcat(OpStr, "213", PackTy),
218                           memop, RC>;
219  defm r231 : fma3s_rm_int<opc231, !strconcat(OpStr, "231", PackTy),
220                           memop, RC>;
221}
222
223multiclass fma3s<bits<8> opc132, bits<8> opc213, bits<8> opc231,
224                 string OpStr, Intrinsic IntF32, Intrinsic IntF64,
225                 SDNode OpNode> {
226  let ExeDomain = SSEPackedSingle in
227  defm SS : fma3s_forms<opc132, opc213, opc231, OpStr, "ss", OpNode,
228                        FR32, f32mem>,
229            fma3s_int_forms<opc132, opc213, opc231, OpStr, "ss", VR128, ssmem>;
230
231  let ExeDomain = SSEPackedDouble in
232  defm SD : fma3s_forms<opc132, opc213, opc231, OpStr, "sd", OpNode,
233                        FR64, f64mem>,
234            fma3s_int_forms<opc132, opc213, opc231, OpStr, "sd", VR128, sdmem>,
235            VEX_W;
236
237  // These patterns use the 123 ordering, instead of 213, even though
238  // they match the intrinsic to the 213 version of the instruction.
239  // This is because src1 is tied to dest, and the scalar intrinsics
240  // require the pass-through values to come from the first source
241  // operand, not the second.
242  def : Pat<(IntF32 VR128:$src1, VR128:$src2, VR128:$src3),
243            (COPY_TO_REGCLASS(!cast<Instruction>(NAME#"SSr213r_Int")
244             $src1, $src2, $src3), VR128)>;
245
246  def : Pat<(IntF64 VR128:$src1, VR128:$src2, VR128:$src3),
247            (COPY_TO_REGCLASS(!cast<Instruction>(NAME#"SDr213r_Int")
248             $src1, $src2, $src3), VR128)>;
249}
250
251defm VFMADD : fma3s<0x99, 0xA9, 0xB9, "vfmadd", int_x86_fma_vfmadd_ss,
252                    int_x86_fma_vfmadd_sd, X86Fmadd>, VEX_LIG;
253defm VFMSUB : fma3s<0x9B, 0xAB, 0xBB, "vfmsub", int_x86_fma_vfmsub_ss,
254                    int_x86_fma_vfmsub_sd, X86Fmsub>, VEX_LIG;
255
256defm VFNMADD : fma3s<0x9D, 0xAD, 0xBD, "vfnmadd", int_x86_fma_vfnmadd_ss,
257                     int_x86_fma_vfnmadd_sd, X86Fnmadd>, VEX_LIG;
258defm VFNMSUB : fma3s<0x9F, 0xAF, 0xBF, "vfnmsub", int_x86_fma_vfnmsub_ss,
259                     int_x86_fma_vfnmsub_sd, X86Fnmsub>, VEX_LIG;
260
261
262//===----------------------------------------------------------------------===//
263// FMA4 - AMD 4 operand Fused Multiply-Add instructions
264//===----------------------------------------------------------------------===//
265
266
267multiclass fma4s<bits<8> opc, string OpcodeStr, RegisterClass RC,
268                 X86MemOperand x86memop, ValueType OpVT, SDNode OpNode,
269                 PatFrag mem_frag> {
270  let isCommutable = 1 in
271  def rr : FMA4<opc, MRMSrcReg, (outs RC:$dst),
272           (ins RC:$src1, RC:$src2, RC:$src3),
273           !strconcat(OpcodeStr,
274           "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
275           [(set RC:$dst,
276             (OpVT (OpNode RC:$src1, RC:$src2, RC:$src3)))]>, VEX_W, VEX_LIG, MemOp4;
277  def rm : FMA4<opc, MRMSrcMem, (outs RC:$dst),
278           (ins RC:$src1, RC:$src2, x86memop:$src3),
279           !strconcat(OpcodeStr,
280           "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
281           [(set RC:$dst, (OpNode RC:$src1, RC:$src2,
282                           (mem_frag addr:$src3)))]>, VEX_W, VEX_LIG, MemOp4;
283  def mr : FMA4<opc, MRMSrcMem, (outs RC:$dst),
284           (ins RC:$src1, x86memop:$src2, RC:$src3),
285           !strconcat(OpcodeStr,
286           "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
287           [(set RC:$dst,
288             (OpNode RC:$src1, (mem_frag addr:$src2), RC:$src3))]>, VEX_LIG;
289// For disassembler
290let isCodeGenOnly = 1, ForceDisassemble = 1, hasSideEffects = 0 in
291  def rr_REV : FMA4<opc, MRMSrcReg, (outs RC:$dst),
292               (ins RC:$src1, RC:$src2, RC:$src3),
293               !strconcat(OpcodeStr,
294               "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"), []>,
295               VEX_LIG;
296}
297
298multiclass fma4s_int<bits<8> opc, string OpcodeStr, Operand memop,
299                     ComplexPattern mem_cpat, Intrinsic Int> {
300let isCodeGenOnly = 1 in {
301  let isCommutable = 1 in
302  def rr_Int : FMA4<opc, MRMSrcReg, (outs VR128:$dst),
303               (ins VR128:$src1, VR128:$src2, VR128:$src3),
304               !strconcat(OpcodeStr,
305               "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
306               [(set VR128:$dst,
307                 (Int VR128:$src1, VR128:$src2, VR128:$src3))]>, VEX_W, VEX_LIG, MemOp4;
308  def rm_Int : FMA4<opc, MRMSrcMem, (outs VR128:$dst),
309               (ins VR128:$src1, VR128:$src2, memop:$src3),
310               !strconcat(OpcodeStr,
311               "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
312               [(set VR128:$dst, (Int VR128:$src1, VR128:$src2,
313                                  mem_cpat:$src3))]>, VEX_W, VEX_LIG, MemOp4;
314  def mr_Int : FMA4<opc, MRMSrcMem, (outs VR128:$dst),
315               (ins VR128:$src1, memop:$src2, VR128:$src3),
316               !strconcat(OpcodeStr,
317               "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
318               [(set VR128:$dst,
319                 (Int VR128:$src1, mem_cpat:$src2, VR128:$src3))]>, VEX_LIG;
320} // isCodeGenOnly = 1
321}
322
323multiclass fma4p<bits<8> opc, string OpcodeStr, SDNode OpNode,
324                 ValueType OpVT128, ValueType OpVT256,
325                 PatFrag ld_frag128, PatFrag ld_frag256> {
326  let isCommutable = 1 in
327  def rr : FMA4<opc, MRMSrcReg, (outs VR128:$dst),
328           (ins VR128:$src1, VR128:$src2, VR128:$src3),
329           !strconcat(OpcodeStr,
330           "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
331           [(set VR128:$dst,
332             (OpVT128 (OpNode VR128:$src1, VR128:$src2, VR128:$src3)))]>,
333           VEX_W, MemOp4;
334  def rm : FMA4<opc, MRMSrcMem, (outs VR128:$dst),
335           (ins VR128:$src1, VR128:$src2, f128mem:$src3),
336           !strconcat(OpcodeStr,
337           "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
338           [(set VR128:$dst, (OpNode VR128:$src1, VR128:$src2,
339                              (ld_frag128 addr:$src3)))]>, VEX_W, MemOp4;
340  def mr : FMA4<opc, MRMSrcMem, (outs VR128:$dst),
341           (ins VR128:$src1, f128mem:$src2, VR128:$src3),
342           !strconcat(OpcodeStr,
343           "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
344           [(set VR128:$dst,
345             (OpNode VR128:$src1, (ld_frag128 addr:$src2), VR128:$src3))]>;
346  let isCommutable = 1 in
347  def rrY : FMA4<opc, MRMSrcReg, (outs VR256:$dst),
348           (ins VR256:$src1, VR256:$src2, VR256:$src3),
349           !strconcat(OpcodeStr,
350           "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
351           [(set VR256:$dst,
352             (OpVT256 (OpNode VR256:$src1, VR256:$src2, VR256:$src3)))]>,
353           VEX_W, MemOp4, VEX_L;
354  def rmY : FMA4<opc, MRMSrcMem, (outs VR256:$dst),
355           (ins VR256:$src1, VR256:$src2, f256mem:$src3),
356           !strconcat(OpcodeStr,
357           "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
358           [(set VR256:$dst, (OpNode VR256:$src1, VR256:$src2,
359                              (ld_frag256 addr:$src3)))]>, VEX_W, MemOp4, VEX_L;
360  def mrY : FMA4<opc, MRMSrcMem, (outs VR256:$dst),
361           (ins VR256:$src1, f256mem:$src2, VR256:$src3),
362           !strconcat(OpcodeStr,
363           "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
364           [(set VR256:$dst, (OpNode VR256:$src1,
365                              (ld_frag256 addr:$src2), VR256:$src3))]>, VEX_L;
366// For disassembler
367let isCodeGenOnly = 1, ForceDisassemble = 1, hasSideEffects = 0 in {
368  def rr_REV : FMA4<opc, MRMSrcReg, (outs VR128:$dst),
369               (ins VR128:$src1, VR128:$src2, VR128:$src3),
370               !strconcat(OpcodeStr,
371               "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"), []>;
372  def rrY_REV : FMA4<opc, MRMSrcReg, (outs VR256:$dst),
373                (ins VR256:$src1, VR256:$src2, VR256:$src3),
374                !strconcat(OpcodeStr,
375                "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"), []>,
376                VEX_L;
377} // isCodeGenOnly = 1
378}
379
380let ExeDomain = SSEPackedSingle in {
381  // Scalar Instructions
382  defm VFMADDSS4  : fma4s<0x6A, "vfmaddss", FR32, f32mem, f32, X86Fmadd, loadf32>,
383                    fma4s_int<0x6A, "vfmaddss", ssmem, sse_load_f32,
384                              int_x86_fma_vfmadd_ss>;
385  defm VFMSUBSS4  : fma4s<0x6E, "vfmsubss", FR32, f32mem, f32, X86Fmsub, loadf32>,
386                    fma4s_int<0x6E, "vfmsubss", ssmem, sse_load_f32,
387                              int_x86_fma_vfmsub_ss>;
388  defm VFNMADDSS4 : fma4s<0x7A, "vfnmaddss", FR32, f32mem, f32,
389                          X86Fnmadd, loadf32>,
390                    fma4s_int<0x7A, "vfnmaddss", ssmem, sse_load_f32,
391                              int_x86_fma_vfnmadd_ss>;
392  defm VFNMSUBSS4 : fma4s<0x7E, "vfnmsubss", FR32, f32mem, f32,
393                          X86Fnmsub, loadf32>,
394                    fma4s_int<0x7E, "vfnmsubss", ssmem, sse_load_f32,
395                              int_x86_fma_vfnmsub_ss>;
396  // Packed Instructions
397  defm VFMADDPS4    : fma4p<0x68, "vfmaddps", X86Fmadd, v4f32, v8f32,
398                            loadv4f32, loadv8f32>;
399  defm VFMSUBPS4    : fma4p<0x6C, "vfmsubps", X86Fmsub, v4f32, v8f32,
400                            loadv4f32, loadv8f32>;
401  defm VFNMADDPS4   : fma4p<0x78, "vfnmaddps", X86Fnmadd, v4f32, v8f32,
402                            loadv4f32, loadv8f32>;
403  defm VFNMSUBPS4   : fma4p<0x7C, "vfnmsubps", X86Fnmsub, v4f32, v8f32,
404                            loadv4f32, loadv8f32>;
405  defm VFMADDSUBPS4 : fma4p<0x5C, "vfmaddsubps", X86Fmaddsub, v4f32, v8f32,
406                            loadv4f32, loadv8f32>;
407  defm VFMSUBADDPS4 : fma4p<0x5E, "vfmsubaddps", X86Fmsubadd, v4f32, v8f32,
408                            loadv4f32, loadv8f32>;
409}
410
411let ExeDomain = SSEPackedDouble in {
412  // Scalar Instructions
413  defm VFMADDSD4  : fma4s<0x6B, "vfmaddsd", FR64, f64mem, f64, X86Fmadd, loadf64>,
414                    fma4s_int<0x6B, "vfmaddsd", sdmem, sse_load_f64,
415                              int_x86_fma_vfmadd_sd>;
416  defm VFMSUBSD4  : fma4s<0x6F, "vfmsubsd", FR64, f64mem, f64, X86Fmsub, loadf64>,
417                    fma4s_int<0x6F, "vfmsubsd", sdmem, sse_load_f64,
418                              int_x86_fma_vfmsub_sd>;
419  defm VFNMADDSD4 : fma4s<0x7B, "vfnmaddsd", FR64, f64mem, f64,
420                          X86Fnmadd, loadf64>,
421                    fma4s_int<0x7B, "vfnmaddsd", sdmem, sse_load_f64,
422                              int_x86_fma_vfnmadd_sd>;
423  defm VFNMSUBSD4 : fma4s<0x7F, "vfnmsubsd", FR64, f64mem, f64,
424                          X86Fnmsub, loadf64>,
425                    fma4s_int<0x7F, "vfnmsubsd", sdmem, sse_load_f64,
426                              int_x86_fma_vfnmsub_sd>;
427  // Packed Instructions
428  defm VFMADDPD4    : fma4p<0x69, "vfmaddpd", X86Fmadd, v2f64, v4f64,
429                            loadv2f64, loadv4f64>;
430  defm VFMSUBPD4    : fma4p<0x6D, "vfmsubpd", X86Fmsub, v2f64, v4f64,
431                            loadv2f64, loadv4f64>;
432  defm VFNMADDPD4   : fma4p<0x79, "vfnmaddpd", X86Fnmadd, v2f64, v4f64,
433                            loadv2f64, loadv4f64>;
434  defm VFNMSUBPD4   : fma4p<0x7D, "vfnmsubpd", X86Fnmsub, v2f64, v4f64,
435                            loadv2f64, loadv4f64>;
436  defm VFMADDSUBPD4 : fma4p<0x5D, "vfmaddsubpd", X86Fmaddsub, v2f64, v4f64,
437                            loadv2f64, loadv4f64>;
438  defm VFMSUBADDPD4 : fma4p<0x5F, "vfmsubaddpd", X86Fmsubadd, v2f64, v4f64,
439                            loadv2f64, loadv4f64>;
440}
441
442