1 //===-- X86InstrFMA.td - FMA Instruction Set ---------------*- tablegen -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file describes FMA (Fused Multiply-Add) instructions.
12 //===----------------------------------------------------------------------===//
14 //===----------------------------------------------------------------------===//
15 // FMA3 - Intel 3 operand Fused Multiply-Add instructions
16 //===----------------------------------------------------------------------===//
18 let Constraints = "$src1 = $dst" in {
19 multiclass fma3p_rm<bits<8> opc, string OpcodeStr,
20 PatFrag MemFrag128, PatFrag MemFrag256,
21 ValueType OpVT128, ValueType OpVT256,
22 bit IsRVariantCommutable = 0, bit IsMVariantCommutable = 0,
23 SDPatternOperator Op = null_frag> {
24 let usesCustomInserter = 1, isCommutable = IsRVariantCommutable in
25 def r : FMA3<opc, MRMSrcReg, (outs VR128:$dst),
26 (ins VR128:$src1, VR128:$src2, VR128:$src3),
28 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
29 [(set VR128:$dst, (OpVT128 (Op VR128:$src2,
30 VR128:$src1, VR128:$src3)))]>;
32 let mayLoad = 1, isCommutable = IsMVariantCommutable in
33 def m : FMA3<opc, MRMSrcMem, (outs VR128:$dst),
34 (ins VR128:$src1, VR128:$src2, f128mem:$src3),
36 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
37 [(set VR128:$dst, (OpVT128 (Op VR128:$src2, VR128:$src1,
38 (MemFrag128 addr:$src3))))]>;
40 let usesCustomInserter = 1, isCommutable = IsRVariantCommutable in
41 def rY : FMA3<opc, MRMSrcReg, (outs VR256:$dst),
42 (ins VR256:$src1, VR256:$src2, VR256:$src3),
44 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
45 [(set VR256:$dst, (OpVT256 (Op VR256:$src2, VR256:$src1,
46 VR256:$src3)))]>, VEX_L;
48 let mayLoad = 1, isCommutable = IsMVariantCommutable in
49 def mY : FMA3<opc, MRMSrcMem, (outs VR256:$dst),
50 (ins VR256:$src1, VR256:$src2, f256mem:$src3),
52 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
54 (OpVT256 (Op VR256:$src2, VR256:$src1,
55 (MemFrag256 addr:$src3))))]>, VEX_L;
57 } // Constraints = "$src1 = $dst"
59 multiclass fma3p_forms<bits<8> opc132, bits<8> opc213, bits<8> opc231,
60 string OpcodeStr, string PackTy,
61 PatFrag MemFrag128, PatFrag MemFrag256,
62 SDNode Op, ValueType OpTy128, ValueType OpTy256> {
63 // For 213, both the register and memory variant are commutable.
64 // Indeed, the commutable operands are 1 and 2 and both live in registers
66 defm r213 : fma3p_rm<opc213,
67 !strconcat(OpcodeStr, "213", PackTy),
68 MemFrag128, MemFrag256, OpTy128, OpTy256,
69 /* IsRVariantCommutable */ 1,
70 /* IsMVariantCommutable */ 1,
72 let neverHasSideEffects = 1 in {
73 defm r132 : fma3p_rm<opc132,
74 !strconcat(OpcodeStr, "132", PackTy),
75 MemFrag128, MemFrag256, OpTy128, OpTy256>;
76 // For 231, only the register variant is commutable.
77 // For the memory variant the folded operand must be in 3. Thus,
78 // in that case, it cannot be swapped with 2.
79 defm r231 : fma3p_rm<opc231,
80 !strconcat(OpcodeStr, "231", PackTy),
81 MemFrag128, MemFrag256, OpTy128, OpTy256,
82 /* IsRVariantCommutable */ 1,
83 /* IsMVariantCommutable */ 0>;
84 } // neverHasSideEffects = 1
88 let ExeDomain = SSEPackedSingle in {
89 defm VFMADDPS : fma3p_forms<0x98, 0xA8, 0xB8, "vfmadd", "ps", loadv4f32,
90 loadv8f32, X86Fmadd, v4f32, v8f32>;
91 defm VFMSUBPS : fma3p_forms<0x9A, 0xAA, 0xBA, "vfmsub", "ps", loadv4f32,
92 loadv8f32, X86Fmsub, v4f32, v8f32>;
93 defm VFMADDSUBPS : fma3p_forms<0x96, 0xA6, 0xB6, "vfmaddsub", "ps",
94 loadv4f32, loadv8f32, X86Fmaddsub,
96 defm VFMSUBADDPS : fma3p_forms<0x97, 0xA7, 0xB7, "vfmsubadd", "ps",
97 loadv4f32, loadv8f32, X86Fmsubadd,
101 let ExeDomain = SSEPackedDouble in {
102 defm VFMADDPD : fma3p_forms<0x98, 0xA8, 0xB8, "vfmadd", "pd", loadv2f64,
103 loadv4f64, X86Fmadd, v2f64, v4f64>, VEX_W;
104 defm VFMSUBPD : fma3p_forms<0x9A, 0xAA, 0xBA, "vfmsub", "pd", loadv2f64,
105 loadv4f64, X86Fmsub, v2f64, v4f64>, VEX_W;
106 defm VFMADDSUBPD : fma3p_forms<0x96, 0xA6, 0xB6, "vfmaddsub", "pd",
107 loadv2f64, loadv4f64, X86Fmaddsub,
108 v2f64, v4f64>, VEX_W;
109 defm VFMSUBADDPD : fma3p_forms<0x97, 0xA7, 0xB7, "vfmsubadd", "pd",
110 loadv2f64, loadv4f64, X86Fmsubadd,
111 v2f64, v4f64>, VEX_W;
114 // Fused Negative Multiply-Add
115 let ExeDomain = SSEPackedSingle in {
116 defm VFNMADDPS : fma3p_forms<0x9C, 0xAC, 0xBC, "vfnmadd", "ps", loadv4f32,
117 loadv8f32, X86Fnmadd, v4f32, v8f32>;
118 defm VFNMSUBPS : fma3p_forms<0x9E, 0xAE, 0xBE, "vfnmsub", "ps", loadv4f32,
119 loadv8f32, X86Fnmsub, v4f32, v8f32>;
121 let ExeDomain = SSEPackedDouble in {
122 defm VFNMADDPD : fma3p_forms<0x9C, 0xAC, 0xBC, "vfnmadd", "pd", loadv2f64,
123 loadv4f64, X86Fnmadd, v2f64, v4f64>, VEX_W;
124 defm VFNMSUBPD : fma3p_forms<0x9E, 0xAE, 0xBE, "vfnmsub", "pd",
125 loadv2f64, loadv4f64, X86Fnmsub, v2f64,
129 let Constraints = "$src1 = $dst" in {
130 multiclass fma3s_rm<bits<8> opc, string OpcodeStr, X86MemOperand x86memop,
131 RegisterClass RC, ValueType OpVT, PatFrag mem_frag,
132 bit IsRVariantCommutable = 0, bit IsMVariantCommutable = 0,
133 SDPatternOperator OpNode = null_frag> {
134 let usesCustomInserter = 1, isCommutable = IsRVariantCommutable in
135 def r : FMA3<opc, MRMSrcReg, (outs RC:$dst),
136 (ins RC:$src1, RC:$src2, RC:$src3),
137 !strconcat(OpcodeStr,
138 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
140 (OpVT (OpNode RC:$src2, RC:$src1, RC:$src3)))]>;
142 let mayLoad = 1, isCommutable = IsMVariantCommutable in
143 def m : FMA3<opc, MRMSrcMem, (outs RC:$dst),
144 (ins RC:$src1, RC:$src2, x86memop:$src3),
145 !strconcat(OpcodeStr,
146 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
148 (OpVT (OpNode RC:$src2, RC:$src1,
149 (mem_frag addr:$src3))))]>;
151 } // Constraints = "$src1 = $dst"
153 multiclass fma3s_forms<bits<8> opc132, bits<8> opc213, bits<8> opc231,
154 string OpStr, string PackTy, string PT2, Intrinsic Int,
155 SDNode OpNode, RegisterClass RC, ValueType OpVT,
156 X86MemOperand x86memop, Operand memop, PatFrag mem_frag,
157 ComplexPattern mem_cpat> {
158 let neverHasSideEffects = 1 in {
159 defm r132 : fma3s_rm<opc132, !strconcat(OpStr, "132", PackTy),
160 x86memop, RC, OpVT, mem_frag>;
161 // See the other defm of r231 for the explanation regarding the
163 defm r231 : fma3s_rm<opc231, !strconcat(OpStr, "231", PackTy),
164 x86memop, RC, OpVT, mem_frag,
165 /* IsRVariantCommutable */ 1,
166 /* IsMVariantCommutable */ 0>;
169 // See the other defm of r213 for the explanation regarding the
171 defm r213 : fma3s_rm<opc213, !strconcat(OpStr, "213", PackTy),
172 x86memop, RC, OpVT, mem_frag,
173 /* IsRVariantCommutable */ 1,
174 /* IsMVariantCommutable */ 1,
178 multiclass fma3s<bits<8> opc132, bits<8> opc213, bits<8> opc231,
179 string OpStr, Intrinsic IntF32, Intrinsic IntF64,
181 defm SS : fma3s_forms<opc132, opc213, opc231, OpStr, "ss", "SS", IntF32, OpNode,
182 FR32, f32, f32mem, ssmem, loadf32, sse_load_f32>;
183 defm SD : fma3s_forms<opc132, opc213, opc231, OpStr, "sd", "PD", IntF64, OpNode,
184 FR64, f64, f64mem, sdmem, loadf64, sse_load_f64>, VEX_W;
186 def : Pat<(IntF32 VR128:$src1, VR128:$src2, VR128:$src3),
188 (!cast<Instruction>(NAME#"SSr213r")
189 (COPY_TO_REGCLASS $src2, FR32),
190 (COPY_TO_REGCLASS $src1, FR32),
191 (COPY_TO_REGCLASS $src3, FR32)),
194 def : Pat<(IntF64 VR128:$src1, VR128:$src2, VR128:$src3),
196 (!cast<Instruction>(NAME#"SDr213r")
197 (COPY_TO_REGCLASS $src2, FR64),
198 (COPY_TO_REGCLASS $src1, FR64),
199 (COPY_TO_REGCLASS $src3, FR64)),
203 defm VFMADD : fma3s<0x99, 0xA9, 0xB9, "vfmadd", int_x86_fma_vfmadd_ss,
204 int_x86_fma_vfmadd_sd, X86Fmadd>, VEX_LIG;
205 defm VFMSUB : fma3s<0x9B, 0xAB, 0xBB, "vfmsub", int_x86_fma_vfmsub_ss,
206 int_x86_fma_vfmsub_sd, X86Fmsub>, VEX_LIG;
208 defm VFNMADD : fma3s<0x9D, 0xAD, 0xBD, "vfnmadd", int_x86_fma_vfnmadd_ss,
209 int_x86_fma_vfnmadd_sd, X86Fnmadd>, VEX_LIG;
210 defm VFNMSUB : fma3s<0x9F, 0xAF, 0xBF, "vfnmsub", int_x86_fma_vfnmsub_ss,
211 int_x86_fma_vfnmsub_sd, X86Fnmsub>, VEX_LIG;
214 //===----------------------------------------------------------------------===//
215 // FMA4 - AMD 4 operand Fused Multiply-Add instructions
216 //===----------------------------------------------------------------------===//
219 multiclass fma4s<bits<8> opc, string OpcodeStr, RegisterClass RC,
220 X86MemOperand x86memop, ValueType OpVT, SDNode OpNode,
222 let isCommutable = 1 in
223 def rr : FMA4<opc, MRMSrcReg, (outs RC:$dst),
224 (ins RC:$src1, RC:$src2, RC:$src3),
225 !strconcat(OpcodeStr,
226 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
228 (OpVT (OpNode RC:$src1, RC:$src2, RC:$src3)))]>, VEX_W, VEX_LIG, MemOp4;
229 def rm : FMA4<opc, MRMSrcMem, (outs RC:$dst),
230 (ins RC:$src1, RC:$src2, x86memop:$src3),
231 !strconcat(OpcodeStr,
232 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
233 [(set RC:$dst, (OpNode RC:$src1, RC:$src2,
234 (mem_frag addr:$src3)))]>, VEX_W, VEX_LIG, MemOp4;
235 def mr : FMA4<opc, MRMSrcMem, (outs RC:$dst),
236 (ins RC:$src1, x86memop:$src2, RC:$src3),
237 !strconcat(OpcodeStr,
238 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
240 (OpNode RC:$src1, (mem_frag addr:$src2), RC:$src3))]>, VEX_LIG;
242 let isCodeGenOnly = 1, ForceDisassemble = 1, hasSideEffects = 0 in
243 def rr_REV : FMA4<opc, MRMSrcReg, (outs RC:$dst),
244 (ins RC:$src1, RC:$src2, RC:$src3),
245 !strconcat(OpcodeStr,
246 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"), []>,
250 multiclass fma4s_int<bits<8> opc, string OpcodeStr, Operand memop,
251 ComplexPattern mem_cpat, Intrinsic Int> {
252 let isCodeGenOnly = 1 in {
253 let isCommutable = 1 in
254 def rr_Int : FMA4<opc, MRMSrcReg, (outs VR128:$dst),
255 (ins VR128:$src1, VR128:$src2, VR128:$src3),
256 !strconcat(OpcodeStr,
257 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
259 (Int VR128:$src1, VR128:$src2, VR128:$src3))]>, VEX_W, VEX_LIG, MemOp4;
260 def rm_Int : FMA4<opc, MRMSrcMem, (outs VR128:$dst),
261 (ins VR128:$src1, VR128:$src2, memop:$src3),
262 !strconcat(OpcodeStr,
263 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
264 [(set VR128:$dst, (Int VR128:$src1, VR128:$src2,
265 mem_cpat:$src3))]>, VEX_W, VEX_LIG, MemOp4;
266 def mr_Int : FMA4<opc, MRMSrcMem, (outs VR128:$dst),
267 (ins VR128:$src1, memop:$src2, VR128:$src3),
268 !strconcat(OpcodeStr,
269 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
271 (Int VR128:$src1, mem_cpat:$src2, VR128:$src3))]>, VEX_LIG;
272 } // isCodeGenOnly = 1
275 multiclass fma4p<bits<8> opc, string OpcodeStr, SDNode OpNode,
276 ValueType OpVT128, ValueType OpVT256,
277 PatFrag ld_frag128, PatFrag ld_frag256> {
278 let isCommutable = 1 in
279 def rr : FMA4<opc, MRMSrcReg, (outs VR128:$dst),
280 (ins VR128:$src1, VR128:$src2, VR128:$src3),
281 !strconcat(OpcodeStr,
282 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
284 (OpVT128 (OpNode VR128:$src1, VR128:$src2, VR128:$src3)))]>,
286 def rm : FMA4<opc, MRMSrcMem, (outs VR128:$dst),
287 (ins VR128:$src1, VR128:$src2, f128mem:$src3),
288 !strconcat(OpcodeStr,
289 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
290 [(set VR128:$dst, (OpNode VR128:$src1, VR128:$src2,
291 (ld_frag128 addr:$src3)))]>, VEX_W, MemOp4;
292 def mr : FMA4<opc, MRMSrcMem, (outs VR128:$dst),
293 (ins VR128:$src1, f128mem:$src2, VR128:$src3),
294 !strconcat(OpcodeStr,
295 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
297 (OpNode VR128:$src1, (ld_frag128 addr:$src2), VR128:$src3))]>;
298 let isCommutable = 1 in
299 def rrY : FMA4<opc, MRMSrcReg, (outs VR256:$dst),
300 (ins VR256:$src1, VR256:$src2, VR256:$src3),
301 !strconcat(OpcodeStr,
302 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
304 (OpVT256 (OpNode VR256:$src1, VR256:$src2, VR256:$src3)))]>,
305 VEX_W, MemOp4, VEX_L;
306 def rmY : FMA4<opc, MRMSrcMem, (outs VR256:$dst),
307 (ins VR256:$src1, VR256:$src2, f256mem:$src3),
308 !strconcat(OpcodeStr,
309 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
310 [(set VR256:$dst, (OpNode VR256:$src1, VR256:$src2,
311 (ld_frag256 addr:$src3)))]>, VEX_W, MemOp4, VEX_L;
312 def mrY : FMA4<opc, MRMSrcMem, (outs VR256:$dst),
313 (ins VR256:$src1, f256mem:$src2, VR256:$src3),
314 !strconcat(OpcodeStr,
315 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
316 [(set VR256:$dst, (OpNode VR256:$src1,
317 (ld_frag256 addr:$src2), VR256:$src3))]>, VEX_L;
319 let isCodeGenOnly = 1, ForceDisassemble = 1, hasSideEffects = 0 in {
320 def rr_REV : FMA4<opc, MRMSrcReg, (outs VR128:$dst),
321 (ins VR128:$src1, VR128:$src2, VR128:$src3),
322 !strconcat(OpcodeStr,
323 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"), []>;
324 def rrY_REV : FMA4<opc, MRMSrcReg, (outs VR256:$dst),
325 (ins VR256:$src1, VR256:$src2, VR256:$src3),
326 !strconcat(OpcodeStr,
327 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"), []>,
329 } // isCodeGenOnly = 1
332 defm VFMADDSS4 : fma4s<0x6A, "vfmaddss", FR32, f32mem, f32, X86Fmadd, loadf32>,
333 fma4s_int<0x6A, "vfmaddss", ssmem, sse_load_f32,
334 int_x86_fma_vfmadd_ss>;
335 defm VFMADDSD4 : fma4s<0x6B, "vfmaddsd", FR64, f64mem, f64, X86Fmadd, loadf64>,
336 fma4s_int<0x6B, "vfmaddsd", sdmem, sse_load_f64,
337 int_x86_fma_vfmadd_sd>;
338 defm VFMSUBSS4 : fma4s<0x6E, "vfmsubss", FR32, f32mem, f32, X86Fmsub, loadf32>,
339 fma4s_int<0x6E, "vfmsubss", ssmem, sse_load_f32,
340 int_x86_fma_vfmsub_ss>;
341 defm VFMSUBSD4 : fma4s<0x6F, "vfmsubsd", FR64, f64mem, f64, X86Fmsub, loadf64>,
342 fma4s_int<0x6F, "vfmsubsd", sdmem, sse_load_f64,
343 int_x86_fma_vfmsub_sd>;
344 defm VFNMADDSS4 : fma4s<0x7A, "vfnmaddss", FR32, f32mem, f32,
346 fma4s_int<0x7A, "vfnmaddss", ssmem, sse_load_f32,
347 int_x86_fma_vfnmadd_ss>;
348 defm VFNMADDSD4 : fma4s<0x7B, "vfnmaddsd", FR64, f64mem, f64,
350 fma4s_int<0x7B, "vfnmaddsd", sdmem, sse_load_f64,
351 int_x86_fma_vfnmadd_sd>;
352 defm VFNMSUBSS4 : fma4s<0x7E, "vfnmsubss", FR32, f32mem, f32,
354 fma4s_int<0x7E, "vfnmsubss", ssmem, sse_load_f32,
355 int_x86_fma_vfnmsub_ss>;
356 defm VFNMSUBSD4 : fma4s<0x7F, "vfnmsubsd", FR64, f64mem, f64,
358 fma4s_int<0x7F, "vfnmsubsd", sdmem, sse_load_f64,
359 int_x86_fma_vfnmsub_sd>;
361 let ExeDomain = SSEPackedSingle in {
362 defm VFMADDPS4 : fma4p<0x68, "vfmaddps", X86Fmadd, v4f32, v8f32,
363 loadv4f32, loadv8f32>;
364 defm VFMSUBPS4 : fma4p<0x6C, "vfmsubps", X86Fmsub, v4f32, v8f32,
365 loadv4f32, loadv8f32>;
366 defm VFNMADDPS4 : fma4p<0x78, "vfnmaddps", X86Fnmadd, v4f32, v8f32,
367 loadv4f32, loadv8f32>;
368 defm VFNMSUBPS4 : fma4p<0x7C, "vfnmsubps", X86Fnmsub, v4f32, v8f32,
369 loadv4f32, loadv8f32>;
370 defm VFMADDSUBPS4 : fma4p<0x5C, "vfmaddsubps", X86Fmaddsub, v4f32, v8f32,
371 loadv4f32, loadv8f32>;
372 defm VFMSUBADDPS4 : fma4p<0x5E, "vfmsubaddps", X86Fmsubadd, v4f32, v8f32,
373 loadv4f32, loadv8f32>;
376 let ExeDomain = SSEPackedDouble in {
377 defm VFMADDPD4 : fma4p<0x69, "vfmaddpd", X86Fmadd, v2f64, v4f64,
378 loadv2f64, loadv4f64>;
379 defm VFMSUBPD4 : fma4p<0x6D, "vfmsubpd", X86Fmsub, v2f64, v4f64,
380 loadv2f64, loadv4f64>;
381 defm VFNMADDPD4 : fma4p<0x79, "vfnmaddpd", X86Fnmadd, v2f64, v4f64,
382 loadv2f64, loadv4f64>;
383 defm VFNMSUBPD4 : fma4p<0x7D, "vfnmsubpd", X86Fnmsub, v2f64, v4f64,
384 loadv2f64, loadv4f64>;
385 defm VFMADDSUBPD4 : fma4p<0x5D, "vfmaddsubpd", X86Fmaddsub, v2f64, v4f64,
386 loadv2f64, loadv4f64>;
387 defm VFMSUBADDPD4 : fma4p<0x5F, "vfmsubaddpd", X86Fmsubadd, v2f64, v4f64,
388 loadv2f64, loadv4f64>;