1 //====- X86InstrSSE.td - Describe the X86 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 the X86 SSE instruction set, defining the instructions,
11 // and properties of the instructions which are needed for code generation,
12 // machine code emission, and analysis.
14 //===----------------------------------------------------------------------===//
17 //===----------------------------------------------------------------------===//
18 // SSE 1 & 2 Instructions Classes
19 //===----------------------------------------------------------------------===//
21 /// sse12_fp_scalar - SSE 1 & 2 scalar instructions class
22 multiclass sse12_fp_scalar<bits<8> opc, string OpcodeStr, SDNode OpNode,
23 RegisterClass RC, X86MemOperand x86memop,
25 let isCommutable = 1 in {
26 def rr : SI<opc, MRMSrcReg, (outs RC:$dst), (ins RC:$src1, RC:$src2),
28 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
29 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
30 [(set RC:$dst, (OpNode RC:$src1, RC:$src2))]>;
32 def rm : SI<opc, MRMSrcMem, (outs RC:$dst), (ins RC:$src1, x86memop:$src2),
34 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
35 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
36 [(set RC:$dst, (OpNode RC:$src1, (load addr:$src2)))]>;
39 /// sse12_fp_scalar_int - SSE 1 & 2 scalar instructions intrinsics class
40 multiclass sse12_fp_scalar_int<bits<8> opc, string OpcodeStr, RegisterClass RC,
41 string asm, string SSEVer, string FPSizeStr,
42 Operand memopr, ComplexPattern mem_cpat,
44 def rr_Int : SI<opc, MRMSrcReg, (outs RC:$dst), (ins RC:$src1, RC:$src2),
46 !strconcat(asm, "\t{$src2, $dst|$dst, $src2}"),
47 !strconcat(asm, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
48 [(set RC:$dst, (!cast<Intrinsic>(
49 !strconcat("int_x86_sse", SSEVer, "_", OpcodeStr, FPSizeStr))
50 RC:$src1, RC:$src2))]>;
51 def rm_Int : SI<opc, MRMSrcMem, (outs RC:$dst), (ins RC:$src1, memopr:$src2),
53 !strconcat(asm, "\t{$src2, $dst|$dst, $src2}"),
54 !strconcat(asm, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
55 [(set RC:$dst, (!cast<Intrinsic>(!strconcat("int_x86_sse",
56 SSEVer, "_", OpcodeStr, FPSizeStr))
57 RC:$src1, mem_cpat:$src2))]>;
60 /// sse12_fp_packed - SSE 1 & 2 packed instructions class
61 multiclass sse12_fp_packed<bits<8> opc, string OpcodeStr, SDNode OpNode,
62 RegisterClass RC, ValueType vt,
63 X86MemOperand x86memop, PatFrag mem_frag,
64 Domain d, bit Is2Addr = 1> {
65 let isCommutable = 1 in
66 def rr : PI<opc, MRMSrcReg, (outs RC:$dst), (ins RC:$src1, RC:$src2),
68 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
69 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
70 [(set RC:$dst, (vt (OpNode RC:$src1, RC:$src2)))], d>;
72 def rm : PI<opc, MRMSrcMem, (outs RC:$dst), (ins RC:$src1, x86memop:$src2),
74 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
75 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
76 [(set RC:$dst, (OpNode RC:$src1, (mem_frag addr:$src2)))], d>;
79 /// sse12_fp_packed_logical_rm - SSE 1 & 2 packed instructions class
80 multiclass sse12_fp_packed_logical_rm<bits<8> opc, RegisterClass RC, Domain d,
81 string OpcodeStr, X86MemOperand x86memop,
82 list<dag> pat_rr, list<dag> pat_rm,
84 let isCommutable = 1 in
85 def rr : PI<opc, MRMSrcReg, (outs RC:$dst), (ins RC:$src1, RC:$src2),
87 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
88 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
90 def rm : PI<opc, MRMSrcMem, (outs RC:$dst), (ins RC:$src1, x86memop:$src2),
92 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
93 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
97 /// sse12_fp_packed_int - SSE 1 & 2 packed instructions intrinsics class
98 multiclass sse12_fp_packed_int<bits<8> opc, string OpcodeStr, RegisterClass RC,
99 string asm, string SSEVer, string FPSizeStr,
100 X86MemOperand x86memop, PatFrag mem_frag,
101 Domain d, bit Is2Addr = 1> {
102 def rr_Int : PI<opc, MRMSrcReg, (outs RC:$dst), (ins RC:$src1, RC:$src2),
104 !strconcat(asm, "\t{$src2, $dst|$dst, $src2}"),
105 !strconcat(asm, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
106 [(set RC:$dst, (!cast<Intrinsic>(
107 !strconcat("int_x86_", SSEVer, "_", OpcodeStr, FPSizeStr))
108 RC:$src1, RC:$src2))], d>;
109 def rm_Int : PI<opc, MRMSrcMem, (outs RC:$dst), (ins RC:$src1,x86memop:$src2),
111 !strconcat(asm, "\t{$src2, $dst|$dst, $src2}"),
112 !strconcat(asm, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
113 [(set RC:$dst, (!cast<Intrinsic>(
114 !strconcat("int_x86_", SSEVer, "_", OpcodeStr, FPSizeStr))
115 RC:$src1, (mem_frag addr:$src2)))], d>;
118 //===----------------------------------------------------------------------===//
119 // SSE 1 & 2 - Move Instructions
120 //===----------------------------------------------------------------------===//
122 class sse12_move_rr<RegisterClass RC, ValueType vt, string asm> :
123 SI<0x10, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, RC:$src2), asm,
124 [(set (vt VR128:$dst), (movl VR128:$src1, (scalar_to_vector RC:$src2)))]>;
126 // Loading from memory automatically zeroing upper bits.
127 class sse12_move_rm<RegisterClass RC, X86MemOperand x86memop,
128 PatFrag mem_pat, string OpcodeStr> :
129 SI<0x10, MRMSrcMem, (outs RC:$dst), (ins x86memop:$src),
130 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
131 [(set RC:$dst, (mem_pat addr:$src))]>;
133 // Move Instructions. Register-to-register movss/movsd is not used for FR32/64
134 // register copies because it's a partial register update; FsMOVAPSrr/FsMOVAPDrr
135 // is used instead. Register-to-register movss/movsd is not modeled as an
136 // INSERT_SUBREG because INSERT_SUBREG requires that the insert be implementable
137 // in terms of a copy, and just mentioned, we don't use movss/movsd for copies.
138 def VMOVSSrr : sse12_move_rr<FR32, v4f32,
139 "movss\t{$src2, $src1, $dst|$dst, $src1, $src2}">, XS, VEX_4V;
140 def VMOVSDrr : sse12_move_rr<FR64, v2f64,
141 "movsd\t{$src2, $src1, $dst|$dst, $src1, $src2}">, XD, VEX_4V;
143 let canFoldAsLoad = 1, isReMaterializable = 1 in {
144 def VMOVSSrm : sse12_move_rm<FR32, f32mem, loadf32, "movss">, XS, VEX;
146 let AddedComplexity = 20 in
147 def VMOVSDrm : sse12_move_rm<FR64, f64mem, loadf64, "movsd">, XD, VEX;
150 let Constraints = "$src1 = $dst" in {
151 def MOVSSrr : sse12_move_rr<FR32, v4f32,
152 "movss\t{$src2, $dst|$dst, $src2}">, XS;
153 def MOVSDrr : sse12_move_rr<FR64, v2f64,
154 "movsd\t{$src2, $dst|$dst, $src2}">, XD;
157 let canFoldAsLoad = 1, isReMaterializable = 1 in {
158 def MOVSSrm : sse12_move_rm<FR32, f32mem, loadf32, "movss">, XS;
160 let AddedComplexity = 20 in
161 def MOVSDrm : sse12_move_rm<FR64, f64mem, loadf64, "movsd">, XD;
164 let AddedComplexity = 15 in {
165 // Extract the low 32-bit value from one vector and insert it into another.
166 def : Pat<(v4f32 (movl VR128:$src1, VR128:$src2)),
167 (MOVSSrr (v4f32 VR128:$src1),
168 (EXTRACT_SUBREG (v4f32 VR128:$src2), sub_ss))>;
169 // Extract the low 64-bit value from one vector and insert it into another.
170 def : Pat<(v2f64 (movl VR128:$src1, VR128:$src2)),
171 (MOVSDrr (v2f64 VR128:$src1),
172 (EXTRACT_SUBREG (v2f64 VR128:$src2), sub_sd))>;
175 // Implicitly promote a 32-bit scalar to a vector.
176 def : Pat<(v4f32 (scalar_to_vector FR32:$src)),
177 (INSERT_SUBREG (v4f32 (IMPLICIT_DEF)), FR32:$src, sub_ss)>;
178 // Implicitly promote a 64-bit scalar to a vector.
179 def : Pat<(v2f64 (scalar_to_vector FR64:$src)),
180 (INSERT_SUBREG (v2f64 (IMPLICIT_DEF)), FR64:$src, sub_sd)>;
181 // Implicitly promote a 32-bit scalar to a vector.
182 def : Pat<(v8f32 (scalar_to_vector FR32:$src)),
183 (INSERT_SUBREG (v8f32 (IMPLICIT_DEF)), FR32:$src, sub_ss)>;
184 // Implicitly promote a 64-bit scalar to a vector.
185 def : Pat<(v4f64 (scalar_to_vector FR64:$src)),
186 (INSERT_SUBREG (v4f64 (IMPLICIT_DEF)), FR64:$src, sub_sd)>;
188 let AddedComplexity = 20 in {
189 let Predicates = [HasSSE1] in {
190 // MOVSSrm zeros the high parts of the register; represent this
191 // with SUBREG_TO_REG.
192 def : Pat<(v4f32 (X86vzmovl (v4f32 (scalar_to_vector (loadf32 addr:$src))))),
193 (SUBREG_TO_REG (i32 0), (MOVSSrm addr:$src), sub_ss)>;
194 def : Pat<(v4f32 (scalar_to_vector (loadf32 addr:$src))),
195 (SUBREG_TO_REG (i32 0), (MOVSSrm addr:$src), sub_ss)>;
196 def : Pat<(v4f32 (X86vzmovl (loadv4f32 addr:$src))),
197 (SUBREG_TO_REG (i32 0), (MOVSSrm addr:$src), sub_ss)>;
199 let Predicates = [HasSSE2] in {
200 // MOVSDrm zeros the high parts of the register; represent this
201 // with SUBREG_TO_REG.
202 def : Pat<(v2f64 (X86vzmovl (v2f64 (scalar_to_vector (loadf64 addr:$src))))),
203 (SUBREG_TO_REG (i64 0), (MOVSDrm addr:$src), sub_sd)>;
204 def : Pat<(v2f64 (scalar_to_vector (loadf64 addr:$src))),
205 (SUBREG_TO_REG (i64 0), (MOVSDrm addr:$src), sub_sd)>;
206 def : Pat<(v2f64 (X86vzmovl (loadv2f64 addr:$src))),
207 (SUBREG_TO_REG (i64 0), (MOVSDrm addr:$src), sub_sd)>;
208 def : Pat<(v2f64 (X86vzmovl (bc_v2f64 (loadv4f32 addr:$src)))),
209 (SUBREG_TO_REG (i64 0), (MOVSDrm addr:$src), sub_sd)>;
210 def : Pat<(v2f64 (X86vzload addr:$src)),
211 (SUBREG_TO_REG (i64 0), (MOVSDrm addr:$src), sub_sd)>;
215 let AddedComplexity = 20, Predicates = [HasAVX] in {
216 // MOVSSrm zeros the high parts of the register; represent this
217 // with SUBREG_TO_REG. The AVX versions also write: DST[255:128] <- 0
218 def : Pat<(v4f32 (X86vzmovl (v4f32 (scalar_to_vector (loadf32 addr:$src))))),
219 (SUBREG_TO_REG (i32 0), (VMOVSSrm addr:$src), sub_ss)>;
220 def : Pat<(v4f32 (scalar_to_vector (loadf32 addr:$src))),
221 (SUBREG_TO_REG (i32 0), (VMOVSSrm addr:$src), sub_ss)>;
222 def : Pat<(v4f32 (X86vzmovl (loadv4f32 addr:$src))),
223 (SUBREG_TO_REG (i32 0), (VMOVSSrm addr:$src), sub_ss)>;
224 // MOVSDrm zeros the high parts of the register; represent this
225 // with SUBREG_TO_REG. The AVX versions also write: DST[255:128] <- 0
226 def : Pat<(v2f64 (X86vzmovl (v2f64 (scalar_to_vector (loadf64 addr:$src))))),
227 (SUBREG_TO_REG (i64 0), (VMOVSDrm addr:$src), sub_sd)>;
228 def : Pat<(v2f64 (scalar_to_vector (loadf64 addr:$src))),
229 (SUBREG_TO_REG (i64 0), (VMOVSDrm addr:$src), sub_sd)>;
230 def : Pat<(v2f64 (X86vzmovl (loadv2f64 addr:$src))),
231 (SUBREG_TO_REG (i64 0), (VMOVSDrm addr:$src), sub_sd)>;
232 def : Pat<(v2f64 (X86vzmovl (bc_v2f64 (loadv4f32 addr:$src)))),
233 (SUBREG_TO_REG (i64 0), (VMOVSDrm addr:$src), sub_sd)>;
234 def : Pat<(v2f64 (X86vzload addr:$src)),
235 (SUBREG_TO_REG (i64 0), (VMOVSDrm addr:$src), sub_sd)>;
236 // Represent the same patterns above but in the form they appear for
238 def : Pat<(v8f32 (X86vzmovl (insert_subvector undef,
239 (v4f32 (scalar_to_vector (loadf32 addr:$src))), (i32 0)))),
240 (SUBREG_TO_REG (i32 0), (VMOVSSrm addr:$src), sub_ss)>;
241 def : Pat<(v4f64 (X86vzmovl (insert_subvector undef,
242 (v2f64 (scalar_to_vector (loadf64 addr:$src))), (i32 0)))),
243 (SUBREG_TO_REG (i32 0), (VMOVSDrm addr:$src), sub_sd)>;
246 // Store scalar value to memory.
247 def MOVSSmr : SSI<0x11, MRMDestMem, (outs), (ins f32mem:$dst, FR32:$src),
248 "movss\t{$src, $dst|$dst, $src}",
249 [(store FR32:$src, addr:$dst)]>;
250 def MOVSDmr : SDI<0x11, MRMDestMem, (outs), (ins f64mem:$dst, FR64:$src),
251 "movsd\t{$src, $dst|$dst, $src}",
252 [(store FR64:$src, addr:$dst)]>;
254 def VMOVSSmr : SI<0x11, MRMDestMem, (outs), (ins f32mem:$dst, FR32:$src),
255 "movss\t{$src, $dst|$dst, $src}",
256 [(store FR32:$src, addr:$dst)]>, XS, VEX;
257 def VMOVSDmr : SI<0x11, MRMDestMem, (outs), (ins f64mem:$dst, FR64:$src),
258 "movsd\t{$src, $dst|$dst, $src}",
259 [(store FR64:$src, addr:$dst)]>, XD, VEX;
261 // Extract and store.
262 def : Pat<(store (f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
265 (EXTRACT_SUBREG (v4f32 VR128:$src), sub_ss))>;
266 def : Pat<(store (f64 (vector_extract (v2f64 VR128:$src), (iPTR 0))),
269 (EXTRACT_SUBREG (v2f64 VR128:$src), sub_sd))>;
271 // Move Aligned/Unaligned floating point values
272 multiclass sse12_mov_packed<bits<8> opc, RegisterClass RC,
273 X86MemOperand x86memop, PatFrag ld_frag,
274 string asm, Domain d,
275 bit IsReMaterializable = 1> {
276 let neverHasSideEffects = 1 in
277 def rr : PI<opc, MRMSrcReg, (outs RC:$dst), (ins RC:$src),
278 !strconcat(asm, "\t{$src, $dst|$dst, $src}"), [], d>;
279 let canFoldAsLoad = 1, isReMaterializable = IsReMaterializable in
280 def rm : PI<opc, MRMSrcMem, (outs RC:$dst), (ins x86memop:$src),
281 !strconcat(asm, "\t{$src, $dst|$dst, $src}"),
282 [(set RC:$dst, (ld_frag addr:$src))], d>;
285 defm VMOVAPS : sse12_mov_packed<0x28, VR128, f128mem, alignedloadv4f32,
286 "movaps", SSEPackedSingle>, VEX;
287 defm VMOVAPD : sse12_mov_packed<0x28, VR128, f128mem, alignedloadv2f64,
288 "movapd", SSEPackedDouble>, OpSize, VEX;
289 defm VMOVUPS : sse12_mov_packed<0x10, VR128, f128mem, loadv4f32,
290 "movups", SSEPackedSingle>, VEX;
291 defm VMOVUPD : sse12_mov_packed<0x10, VR128, f128mem, loadv2f64,
292 "movupd", SSEPackedDouble, 0>, OpSize, VEX;
294 defm VMOVAPSY : sse12_mov_packed<0x28, VR256, f256mem, alignedloadv8f32,
295 "movaps", SSEPackedSingle>, VEX;
296 defm VMOVAPDY : sse12_mov_packed<0x28, VR256, f256mem, alignedloadv4f64,
297 "movapd", SSEPackedDouble>, OpSize, VEX;
298 defm VMOVUPSY : sse12_mov_packed<0x10, VR256, f256mem, loadv8f32,
299 "movups", SSEPackedSingle>, VEX;
300 defm VMOVUPDY : sse12_mov_packed<0x10, VR256, f256mem, loadv4f64,
301 "movupd", SSEPackedDouble, 0>, OpSize, VEX;
302 defm MOVAPS : sse12_mov_packed<0x28, VR128, f128mem, alignedloadv4f32,
303 "movaps", SSEPackedSingle>, TB;
304 defm MOVAPD : sse12_mov_packed<0x28, VR128, f128mem, alignedloadv2f64,
305 "movapd", SSEPackedDouble>, TB, OpSize;
306 defm MOVUPS : sse12_mov_packed<0x10, VR128, f128mem, loadv4f32,
307 "movups", SSEPackedSingle>, TB;
308 defm MOVUPD : sse12_mov_packed<0x10, VR128, f128mem, loadv2f64,
309 "movupd", SSEPackedDouble, 0>, TB, OpSize;
311 def VMOVAPSmr : VPSI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
312 "movaps\t{$src, $dst|$dst, $src}",
313 [(alignedstore (v4f32 VR128:$src), addr:$dst)]>, VEX;
314 def VMOVAPDmr : VPDI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
315 "movapd\t{$src, $dst|$dst, $src}",
316 [(alignedstore (v2f64 VR128:$src), addr:$dst)]>, VEX;
317 def VMOVUPSmr : VPSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
318 "movups\t{$src, $dst|$dst, $src}",
319 [(store (v4f32 VR128:$src), addr:$dst)]>, VEX;
320 def VMOVUPDmr : VPDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
321 "movupd\t{$src, $dst|$dst, $src}",
322 [(store (v2f64 VR128:$src), addr:$dst)]>, VEX;
323 def VMOVAPSYmr : VPSI<0x29, MRMDestMem, (outs), (ins f256mem:$dst, VR256:$src),
324 "movaps\t{$src, $dst|$dst, $src}",
325 [(alignedstore (v8f32 VR256:$src), addr:$dst)]>, VEX;
326 def VMOVAPDYmr : VPDI<0x29, MRMDestMem, (outs), (ins f256mem:$dst, VR256:$src),
327 "movapd\t{$src, $dst|$dst, $src}",
328 [(alignedstore (v4f64 VR256:$src), addr:$dst)]>, VEX;
329 def VMOVUPSYmr : VPSI<0x11, MRMDestMem, (outs), (ins f256mem:$dst, VR256:$src),
330 "movups\t{$src, $dst|$dst, $src}",
331 [(store (v8f32 VR256:$src), addr:$dst)]>, VEX;
332 def VMOVUPDYmr : VPDI<0x11, MRMDestMem, (outs), (ins f256mem:$dst, VR256:$src),
333 "movupd\t{$src, $dst|$dst, $src}",
334 [(store (v4f64 VR256:$src), addr:$dst)]>, VEX;
336 def : Pat<(int_x86_avx_loadu_ps_256 addr:$src), (VMOVUPSYrm addr:$src)>;
337 def : Pat<(int_x86_avx_storeu_ps_256 addr:$dst, VR256:$src),
338 (VMOVUPSYmr addr:$dst, VR256:$src)>;
340 def : Pat<(int_x86_avx_loadu_pd_256 addr:$src), (VMOVUPDYrm addr:$src)>;
341 def : Pat<(int_x86_avx_storeu_pd_256 addr:$dst, VR256:$src),
342 (VMOVUPDYmr addr:$dst, VR256:$src)>;
344 def MOVAPSmr : PSI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
345 "movaps\t{$src, $dst|$dst, $src}",
346 [(alignedstore (v4f32 VR128:$src), addr:$dst)]>;
347 def MOVAPDmr : PDI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
348 "movapd\t{$src, $dst|$dst, $src}",
349 [(alignedstore (v2f64 VR128:$src), addr:$dst)]>;
350 def MOVUPSmr : PSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
351 "movups\t{$src, $dst|$dst, $src}",
352 [(store (v4f32 VR128:$src), addr:$dst)]>;
353 def MOVUPDmr : PDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
354 "movupd\t{$src, $dst|$dst, $src}",
355 [(store (v2f64 VR128:$src), addr:$dst)]>;
357 // Intrinsic forms of MOVUPS/D load and store
358 def VMOVUPSmr_Int : VPSI<0x11, MRMDestMem, (outs),
359 (ins f128mem:$dst, VR128:$src),
360 "movups\t{$src, $dst|$dst, $src}",
361 [(int_x86_sse_storeu_ps addr:$dst, VR128:$src)]>, VEX;
362 def VMOVUPDmr_Int : VPDI<0x11, MRMDestMem, (outs),
363 (ins f128mem:$dst, VR128:$src),
364 "movupd\t{$src, $dst|$dst, $src}",
365 [(int_x86_sse2_storeu_pd addr:$dst, VR128:$src)]>, VEX;
367 def MOVUPSmr_Int : PSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
368 "movups\t{$src, $dst|$dst, $src}",
369 [(int_x86_sse_storeu_ps addr:$dst, VR128:$src)]>;
370 def MOVUPDmr_Int : PDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
371 "movupd\t{$src, $dst|$dst, $src}",
372 [(int_x86_sse2_storeu_pd addr:$dst, VR128:$src)]>;
374 // Move Low/High packed floating point values
375 multiclass sse12_mov_hilo_packed<bits<8>opc, RegisterClass RC,
376 PatFrag mov_frag, string base_opc,
378 def PSrm : PI<opc, MRMSrcMem,
379 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
380 !strconcat(base_opc, "s", asm_opr),
383 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2))))))],
384 SSEPackedSingle>, TB;
386 def PDrm : PI<opc, MRMSrcMem,
387 (outs RC:$dst), (ins RC:$src1, f64mem:$src2),
388 !strconcat(base_opc, "d", asm_opr),
389 [(set RC:$dst, (v2f64 (mov_frag RC:$src1,
390 (scalar_to_vector (loadf64 addr:$src2)))))],
391 SSEPackedDouble>, TB, OpSize;
394 let AddedComplexity = 20 in {
395 defm VMOVL : sse12_mov_hilo_packed<0x12, VR128, movlp, "movlp",
396 "\t{$src2, $src1, $dst|$dst, $src1, $src2}">, VEX_4V;
397 defm VMOVH : sse12_mov_hilo_packed<0x16, VR128, movlhps, "movhp",
398 "\t{$src2, $src1, $dst|$dst, $src1, $src2}">, VEX_4V;
400 let Constraints = "$src1 = $dst", AddedComplexity = 20 in {
401 defm MOVL : sse12_mov_hilo_packed<0x12, VR128, movlp, "movlp",
402 "\t{$src2, $dst|$dst, $src2}">;
403 defm MOVH : sse12_mov_hilo_packed<0x16, VR128, movlhps, "movhp",
404 "\t{$src2, $dst|$dst, $src2}">;
407 def VMOVLPSmr : VPSI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
408 "movlps\t{$src, $dst|$dst, $src}",
409 [(store (f64 (vector_extract (bc_v2f64 (v4f32 VR128:$src)),
410 (iPTR 0))), addr:$dst)]>, VEX;
411 def VMOVLPDmr : VPDI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
412 "movlpd\t{$src, $dst|$dst, $src}",
413 [(store (f64 (vector_extract (v2f64 VR128:$src),
414 (iPTR 0))), addr:$dst)]>, VEX;
415 def MOVLPSmr : PSI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
416 "movlps\t{$src, $dst|$dst, $src}",
417 [(store (f64 (vector_extract (bc_v2f64 (v4f32 VR128:$src)),
418 (iPTR 0))), addr:$dst)]>;
419 def MOVLPDmr : PDI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
420 "movlpd\t{$src, $dst|$dst, $src}",
421 [(store (f64 (vector_extract (v2f64 VR128:$src),
422 (iPTR 0))), addr:$dst)]>;
424 // v2f64 extract element 1 is always custom lowered to unpack high to low
425 // and extract element 0 so the non-store version isn't too horrible.
426 def VMOVHPSmr : VPSI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
427 "movhps\t{$src, $dst|$dst, $src}",
428 [(store (f64 (vector_extract
429 (unpckh (bc_v2f64 (v4f32 VR128:$src)),
430 (undef)), (iPTR 0))), addr:$dst)]>,
432 def VMOVHPDmr : VPDI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
433 "movhpd\t{$src, $dst|$dst, $src}",
434 [(store (f64 (vector_extract
435 (v2f64 (unpckh VR128:$src, (undef))),
436 (iPTR 0))), addr:$dst)]>,
438 def MOVHPSmr : PSI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
439 "movhps\t{$src, $dst|$dst, $src}",
440 [(store (f64 (vector_extract
441 (unpckh (bc_v2f64 (v4f32 VR128:$src)),
442 (undef)), (iPTR 0))), addr:$dst)]>;
443 def MOVHPDmr : PDI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
444 "movhpd\t{$src, $dst|$dst, $src}",
445 [(store (f64 (vector_extract
446 (v2f64 (unpckh VR128:$src, (undef))),
447 (iPTR 0))), addr:$dst)]>;
449 let AddedComplexity = 20 in {
450 def VMOVLHPSrr : VPSI<0x16, MRMSrcReg, (outs VR128:$dst),
451 (ins VR128:$src1, VR128:$src2),
452 "movlhps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
454 (v4f32 (movlhps VR128:$src1, VR128:$src2)))]>,
456 def VMOVHLPSrr : VPSI<0x12, MRMSrcReg, (outs VR128:$dst),
457 (ins VR128:$src1, VR128:$src2),
458 "movhlps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
460 (v4f32 (movhlps VR128:$src1, VR128:$src2)))]>,
463 let Constraints = "$src1 = $dst", AddedComplexity = 20 in {
464 def MOVLHPSrr : PSI<0x16, MRMSrcReg, (outs VR128:$dst),
465 (ins VR128:$src1, VR128:$src2),
466 "movlhps\t{$src2, $dst|$dst, $src2}",
468 (v4f32 (movlhps VR128:$src1, VR128:$src2)))]>;
469 def MOVHLPSrr : PSI<0x12, MRMSrcReg, (outs VR128:$dst),
470 (ins VR128:$src1, VR128:$src2),
471 "movhlps\t{$src2, $dst|$dst, $src2}",
473 (v4f32 (movhlps VR128:$src1, VR128:$src2)))]>;
476 def : Pat<(movlhps VR128:$src1, (bc_v4i32 (v2i64 (X86vzload addr:$src2)))),
477 (MOVHPSrm (v4i32 VR128:$src1), addr:$src2)>;
478 let AddedComplexity = 20 in {
479 def : Pat<(v4f32 (movddup VR128:$src, (undef))),
480 (MOVLHPSrr (v4f32 VR128:$src), (v4f32 VR128:$src))>;
481 def : Pat<(v2i64 (movddup VR128:$src, (undef))),
482 (MOVLHPSrr (v2i64 VR128:$src), (v2i64 VR128:$src))>;
485 //===----------------------------------------------------------------------===//
486 // SSE 1 & 2 - Conversion Instructions
487 //===----------------------------------------------------------------------===//
489 multiclass sse12_cvt_s<bits<8> opc, RegisterClass SrcRC, RegisterClass DstRC,
490 SDNode OpNode, X86MemOperand x86memop, PatFrag ld_frag,
492 def rr : SI<opc, MRMSrcReg, (outs DstRC:$dst), (ins SrcRC:$src), asm,
493 [(set DstRC:$dst, (OpNode SrcRC:$src))]>;
494 def rm : SI<opc, MRMSrcMem, (outs DstRC:$dst), (ins x86memop:$src), asm,
495 [(set DstRC:$dst, (OpNode (ld_frag addr:$src)))]>;
498 multiclass sse12_cvt_s_np<bits<8> opc, RegisterClass SrcRC, RegisterClass DstRC,
499 X86MemOperand x86memop, string asm> {
500 def rr : SI<opc, MRMSrcReg, (outs DstRC:$dst), (ins SrcRC:$src), asm,
502 def rm : SI<opc, MRMSrcMem, (outs DstRC:$dst), (ins x86memop:$src), asm,
506 multiclass sse12_cvt_p<bits<8> opc, RegisterClass SrcRC, RegisterClass DstRC,
507 SDNode OpNode, X86MemOperand x86memop, PatFrag ld_frag,
508 string asm, Domain d> {
509 def rr : PI<opc, MRMSrcReg, (outs DstRC:$dst), (ins SrcRC:$src), asm,
510 [(set DstRC:$dst, (OpNode SrcRC:$src))], d>;
511 def rm : PI<opc, MRMSrcMem, (outs DstRC:$dst), (ins x86memop:$src), asm,
512 [(set DstRC:$dst, (OpNode (ld_frag addr:$src)))], d>;
515 multiclass sse12_vcvt_avx<bits<8> opc, RegisterClass SrcRC, RegisterClass DstRC,
516 X86MemOperand x86memop, string asm> {
517 def rr : SI<opc, MRMSrcReg, (outs DstRC:$dst), (ins DstRC:$src1, SrcRC:$src),
518 !strconcat(asm,"\t{$src, $src1, $dst|$dst, $src1, $src}"), []>;
519 def rm : SI<opc, MRMSrcMem, (outs DstRC:$dst),
520 (ins DstRC:$src1, x86memop:$src),
521 !strconcat(asm,"\t{$src, $src1, $dst|$dst, $src1, $src}"), []>;
524 defm VCVTTSS2SI : sse12_cvt_s<0x2C, FR32, GR32, fp_to_sint, f32mem, loadf32,
525 "cvttss2si\t{$src, $dst|$dst, $src}">, XS, VEX;
526 defm VCVTTSS2SI64 : sse12_cvt_s<0x2C, FR32, GR64, fp_to_sint, f32mem, loadf32,
527 "cvttss2si\t{$src, $dst|$dst, $src}">, XS, VEX,
529 defm VCVTTSD2SI : sse12_cvt_s<0x2C, FR64, GR32, fp_to_sint, f64mem, loadf64,
530 "cvttsd2si\t{$src, $dst|$dst, $src}">, XD, VEX;
531 defm VCVTTSD2SI64 : sse12_cvt_s<0x2C, FR64, GR64, fp_to_sint, f64mem, loadf64,
532 "cvttsd2si\t{$src, $dst|$dst, $src}">, XD,
535 // The assembler can recognize rr 64-bit instructions by seeing a rxx
536 // register, but the same isn't true when only using memory operands,
537 // provide other assembly "l" and "q" forms to address this explicitly
538 // where appropriate to do so.
539 defm VCVTSI2SS : sse12_vcvt_avx<0x2A, GR32, FR32, i32mem, "cvtsi2ss">, XS,
541 defm VCVTSI2SS64 : sse12_vcvt_avx<0x2A, GR64, FR32, i64mem, "cvtsi2ss{q}">, XS,
543 defm VCVTSI2SD : sse12_vcvt_avx<0x2A, GR32, FR64, i32mem, "cvtsi2sd">, XD,
545 defm VCVTSI2SDL : sse12_vcvt_avx<0x2A, GR32, FR64, i32mem, "cvtsi2sd{l}">, XD,
547 defm VCVTSI2SD64 : sse12_vcvt_avx<0x2A, GR64, FR64, i64mem, "cvtsi2sd{q}">, XD,
550 let Predicates = [HasAVX] in {
551 def : Pat<(f32 (sint_to_fp (loadi32 addr:$src))),
552 (VCVTSI2SSrm (f32 (IMPLICIT_DEF)), addr:$src)>;
553 def : Pat<(f32 (sint_to_fp (loadi64 addr:$src))),
554 (VCVTSI2SS64rm (f32 (IMPLICIT_DEF)), addr:$src)>;
555 def : Pat<(f64 (sint_to_fp (loadi32 addr:$src))),
556 (VCVTSI2SDrm (f64 (IMPLICIT_DEF)), addr:$src)>;
557 def : Pat<(f64 (sint_to_fp (loadi64 addr:$src))),
558 (VCVTSI2SD64rm (f64 (IMPLICIT_DEF)), addr:$src)>;
560 def : Pat<(f32 (sint_to_fp GR32:$src)),
561 (VCVTSI2SSrr (f32 (IMPLICIT_DEF)), GR32:$src)>;
562 def : Pat<(f32 (sint_to_fp GR64:$src)),
563 (VCVTSI2SS64rr (f32 (IMPLICIT_DEF)), GR64:$src)>;
564 def : Pat<(f64 (sint_to_fp GR32:$src)),
565 (VCVTSI2SDrr (f64 (IMPLICIT_DEF)), GR32:$src)>;
566 def : Pat<(f64 (sint_to_fp GR64:$src)),
567 (VCVTSI2SD64rr (f64 (IMPLICIT_DEF)), GR64:$src)>;
570 defm CVTTSS2SI : sse12_cvt_s<0x2C, FR32, GR32, fp_to_sint, f32mem, loadf32,
571 "cvttss2si\t{$src, $dst|$dst, $src}">, XS;
572 defm CVTTSS2SI64 : sse12_cvt_s<0x2C, FR32, GR64, fp_to_sint, f32mem, loadf32,
573 "cvttss2si{q}\t{$src, $dst|$dst, $src}">, XS, REX_W;
574 defm CVTTSD2SI : sse12_cvt_s<0x2C, FR64, GR32, fp_to_sint, f64mem, loadf64,
575 "cvttsd2si\t{$src, $dst|$dst, $src}">, XD;
576 defm CVTTSD2SI64 : sse12_cvt_s<0x2C, FR64, GR64, fp_to_sint, f64mem, loadf64,
577 "cvttsd2si{q}\t{$src, $dst|$dst, $src}">, XD, REX_W;
578 defm CVTSI2SS : sse12_cvt_s<0x2A, GR32, FR32, sint_to_fp, i32mem, loadi32,
579 "cvtsi2ss\t{$src, $dst|$dst, $src}">, XS;
580 defm CVTSI2SS64 : sse12_cvt_s<0x2A, GR64, FR32, sint_to_fp, i64mem, loadi64,
581 "cvtsi2ss{q}\t{$src, $dst|$dst, $src}">, XS, REX_W;
582 defm CVTSI2SD : sse12_cvt_s<0x2A, GR32, FR64, sint_to_fp, i32mem, loadi32,
583 "cvtsi2sd\t{$src, $dst|$dst, $src}">, XD;
584 defm CVTSI2SD64 : sse12_cvt_s<0x2A, GR64, FR64, sint_to_fp, i64mem, loadi64,
585 "cvtsi2sd{q}\t{$src, $dst|$dst, $src}">, XD, REX_W;
587 // Conversion Instructions Intrinsics - Match intrinsics which expect MM
588 // and/or XMM operand(s).
590 multiclass sse12_cvt_sint<bits<8> opc, RegisterClass SrcRC, RegisterClass DstRC,
591 Intrinsic Int, X86MemOperand x86memop, PatFrag ld_frag,
593 def rr : SI<opc, MRMSrcReg, (outs DstRC:$dst), (ins SrcRC:$src),
594 !strconcat(asm, "\t{$src, $dst|$dst, $src}"),
595 [(set DstRC:$dst, (Int SrcRC:$src))]>;
596 def rm : SI<opc, MRMSrcMem, (outs DstRC:$dst), (ins x86memop:$src),
597 !strconcat(asm, "\t{$src, $dst|$dst, $src}"),
598 [(set DstRC:$dst, (Int (ld_frag addr:$src)))]>;
601 multiclass sse12_cvt_sint_3addr<bits<8> opc, RegisterClass SrcRC,
602 RegisterClass DstRC, Intrinsic Int, X86MemOperand x86memop,
603 PatFrag ld_frag, string asm, bit Is2Addr = 1> {
604 def rr : SI<opc, MRMSrcReg, (outs DstRC:$dst), (ins DstRC:$src1, SrcRC:$src2),
606 !strconcat(asm, "\t{$src2, $dst|$dst, $src2}"),
607 !strconcat(asm, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
608 [(set DstRC:$dst, (Int DstRC:$src1, SrcRC:$src2))]>;
609 def rm : SI<opc, MRMSrcMem, (outs DstRC:$dst),
610 (ins DstRC:$src1, x86memop:$src2),
612 !strconcat(asm, "\t{$src2, $dst|$dst, $src2}"),
613 !strconcat(asm, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
614 [(set DstRC:$dst, (Int DstRC:$src1, (ld_frag addr:$src2)))]>;
617 defm Int_VCVTSD2SI : sse12_cvt_sint<0x2D, VR128, GR32, int_x86_sse2_cvtsd2si,
618 f128mem, load, "cvtsd2si">, XD, VEX;
619 defm Int_VCVTSD2SI64 : sse12_cvt_sint<0x2D, VR128, GR64,
620 int_x86_sse2_cvtsd2si64, f128mem, load, "cvtsd2si">,
623 // FIXME: The asm matcher has a hack to ignore instructions with _Int and Int_
624 // Get rid of this hack or rename the intrinsics, there are several
625 // intructions that only match with the intrinsic form, why create duplicates
626 // to let them be recognized by the assembler?
627 defm VCVTSD2SI_alt : sse12_cvt_s_np<0x2D, FR64, GR32, f64mem,
628 "cvtsd2si\t{$src, $dst|$dst, $src}">, XD, VEX;
629 defm VCVTSD2SI64 : sse12_cvt_s_np<0x2D, FR64, GR64, f64mem,
630 "cvtsd2si\t{$src, $dst|$dst, $src}">, XD, VEX, VEX_W;
631 defm CVTSD2SI : sse12_cvt_sint<0x2D, VR128, GR32, int_x86_sse2_cvtsd2si,
632 f128mem, load, "cvtsd2si{l}">, XD;
633 defm CVTSD2SI64 : sse12_cvt_sint<0x2D, VR128, GR64, int_x86_sse2_cvtsd2si64,
634 f128mem, load, "cvtsd2si{q}">, XD, REX_W;
637 defm Int_VCVTSI2SS : sse12_cvt_sint_3addr<0x2A, GR32, VR128,
638 int_x86_sse_cvtsi2ss, i32mem, loadi32, "cvtsi2ss", 0>, XS, VEX_4V;
639 defm Int_VCVTSI2SS64 : sse12_cvt_sint_3addr<0x2A, GR64, VR128,
640 int_x86_sse_cvtsi642ss, i64mem, loadi64, "cvtsi2ss", 0>, XS, VEX_4V,
642 defm Int_VCVTSI2SD : sse12_cvt_sint_3addr<0x2A, GR32, VR128,
643 int_x86_sse2_cvtsi2sd, i32mem, loadi32, "cvtsi2sd", 0>, XD, VEX_4V;
644 defm Int_VCVTSI2SD64 : sse12_cvt_sint_3addr<0x2A, GR64, VR128,
645 int_x86_sse2_cvtsi642sd, i64mem, loadi64, "cvtsi2sd", 0>, XD,
648 let Constraints = "$src1 = $dst" in {
649 defm Int_CVTSI2SS : sse12_cvt_sint_3addr<0x2A, GR32, VR128,
650 int_x86_sse_cvtsi2ss, i32mem, loadi32,
652 defm Int_CVTSI2SS64 : sse12_cvt_sint_3addr<0x2A, GR64, VR128,
653 int_x86_sse_cvtsi642ss, i64mem, loadi64,
654 "cvtsi2ss{q}">, XS, REX_W;
655 defm Int_CVTSI2SD : sse12_cvt_sint_3addr<0x2A, GR32, VR128,
656 int_x86_sse2_cvtsi2sd, i32mem, loadi32,
658 defm Int_CVTSI2SD64 : sse12_cvt_sint_3addr<0x2A, GR64, VR128,
659 int_x86_sse2_cvtsi642sd, i64mem, loadi64,
660 "cvtsi2sd">, XD, REX_W;
665 // Aliases for intrinsics
666 defm Int_VCVTTSS2SI : sse12_cvt_sint<0x2C, VR128, GR32, int_x86_sse_cvttss2si,
667 f32mem, load, "cvttss2si">, XS, VEX;
668 defm Int_VCVTTSS2SI64 : sse12_cvt_sint<0x2C, VR128, GR64,
669 int_x86_sse_cvttss2si64, f32mem, load,
670 "cvttss2si">, XS, VEX, VEX_W;
671 defm Int_VCVTTSD2SI : sse12_cvt_sint<0x2C, VR128, GR32, int_x86_sse2_cvttsd2si,
672 f128mem, load, "cvttsd2si">, XD, VEX;
673 defm Int_VCVTTSD2SI64 : sse12_cvt_sint<0x2C, VR128, GR64,
674 int_x86_sse2_cvttsd2si64, f128mem, load,
675 "cvttsd2si">, XD, VEX, VEX_W;
676 defm Int_CVTTSS2SI : sse12_cvt_sint<0x2C, VR128, GR32, int_x86_sse_cvttss2si,
677 f32mem, load, "cvttss2si">, XS;
678 defm Int_CVTTSS2SI64 : sse12_cvt_sint<0x2C, VR128, GR64,
679 int_x86_sse_cvttss2si64, f32mem, load,
680 "cvttss2si{q}">, XS, REX_W;
681 defm Int_CVTTSD2SI : sse12_cvt_sint<0x2C, VR128, GR32, int_x86_sse2_cvttsd2si,
682 f128mem, load, "cvttsd2si">, XD;
683 defm Int_CVTTSD2SI64 : sse12_cvt_sint<0x2C, VR128, GR64,
684 int_x86_sse2_cvttsd2si64, f128mem, load,
685 "cvttsd2si{q}">, XD, REX_W;
687 let Pattern = []<dag> in {
688 defm VCVTSS2SI : sse12_cvt_s<0x2D, FR32, GR32, undef, f32mem, load,
689 "cvtss2si{l}\t{$src, $dst|$dst, $src}">, XS, VEX;
690 defm VCVTSS2SI64 : sse12_cvt_s<0x2D, FR32, GR64, undef, f32mem, load,
691 "cvtss2si\t{$src, $dst|$dst, $src}">, XS, VEX,
693 defm VCVTDQ2PS : sse12_cvt_p<0x5B, VR128, VR128, undef, i128mem, load,
694 "cvtdq2ps\t{$src, $dst|$dst, $src}",
695 SSEPackedSingle>, TB, VEX;
696 defm VCVTDQ2PSY : sse12_cvt_p<0x5B, VR256, VR256, undef, i256mem, load,
697 "cvtdq2ps\t{$src, $dst|$dst, $src}",
698 SSEPackedSingle>, TB, VEX;
701 let Pattern = []<dag> in {
702 defm CVTSS2SI : sse12_cvt_s<0x2D, FR32, GR32, undef, f32mem, load /*dummy*/,
703 "cvtss2si{l}\t{$src, $dst|$dst, $src}">, XS;
704 defm CVTSS2SI64 : sse12_cvt_s<0x2D, FR32, GR64, undef, f32mem, load /*dummy*/,
705 "cvtss2si{q}\t{$src, $dst|$dst, $src}">, XS, REX_W;
706 defm CVTDQ2PS : sse12_cvt_p<0x5B, VR128, VR128, undef, i128mem, load /*dummy*/,
707 "cvtdq2ps\t{$src, $dst|$dst, $src}",
708 SSEPackedSingle>, TB; /* PD SSE3 form is avaiable */
711 let Predicates = [HasSSE1] in {
712 def : Pat<(int_x86_sse_cvtss2si VR128:$src),
713 (CVTSS2SIrr (EXTRACT_SUBREG (v4f32 VR128:$src), sub_ss))>;
714 def : Pat<(int_x86_sse_cvtss2si (load addr:$src)),
715 (CVTSS2SIrm addr:$src)>;
716 def : Pat<(int_x86_sse_cvtss2si64 VR128:$src),
717 (CVTSS2SI64rr (EXTRACT_SUBREG (v4f32 VR128:$src), sub_ss))>;
718 def : Pat<(int_x86_sse_cvtss2si64 (load addr:$src)),
719 (CVTSS2SI64rm addr:$src)>;
722 let Predicates = [HasAVX] in {
723 def : Pat<(int_x86_sse_cvtss2si VR128:$src),
724 (VCVTSS2SIrr (EXTRACT_SUBREG (v4f32 VR128:$src), sub_ss))>;
725 def : Pat<(int_x86_sse_cvtss2si (load addr:$src)),
726 (VCVTSS2SIrm addr:$src)>;
727 def : Pat<(int_x86_sse_cvtss2si64 VR128:$src),
728 (VCVTSS2SI64rr (EXTRACT_SUBREG (v4f32 VR128:$src), sub_ss))>;
729 def : Pat<(int_x86_sse_cvtss2si64 (load addr:$src)),
730 (VCVTSS2SI64rm addr:$src)>;
735 // Convert scalar double to scalar single
736 def VCVTSD2SSrr : VSDI<0x5A, MRMSrcReg, (outs FR32:$dst),
737 (ins FR64:$src1, FR64:$src2),
738 "cvtsd2ss\t{$src2, $src1, $dst|$dst, $src1, $src2}", []>,
740 def VCVTSD2SSrm : I<0x5A, MRMSrcMem, (outs FR32:$dst),
741 (ins FR64:$src1, f64mem:$src2),
742 "vcvtsd2ss\t{$src2, $src1, $dst|$dst, $src1, $src2}",
743 []>, XD, Requires<[HasAVX, OptForSize]>, VEX_4V;
744 def : Pat<(f32 (fround FR64:$src)), (VCVTSD2SSrr FR64:$src, FR64:$src)>,
747 def CVTSD2SSrr : SDI<0x5A, MRMSrcReg, (outs FR32:$dst), (ins FR64:$src),
748 "cvtsd2ss\t{$src, $dst|$dst, $src}",
749 [(set FR32:$dst, (fround FR64:$src))]>;
750 def CVTSD2SSrm : I<0x5A, MRMSrcMem, (outs FR32:$dst), (ins f64mem:$src),
751 "cvtsd2ss\t{$src, $dst|$dst, $src}",
752 [(set FR32:$dst, (fround (loadf64 addr:$src)))]>, XD,
753 Requires<[HasSSE2, OptForSize]>;
755 defm Int_VCVTSD2SS: sse12_cvt_sint_3addr<0x5A, VR128, VR128,
756 int_x86_sse2_cvtsd2ss, f64mem, load, "cvtsd2ss", 0>,
758 let Constraints = "$src1 = $dst" in
759 defm Int_CVTSD2SS: sse12_cvt_sint_3addr<0x5A, VR128, VR128,
760 int_x86_sse2_cvtsd2ss, f64mem, load, "cvtsd2ss">, XS;
762 // Convert scalar single to scalar double
763 // SSE2 instructions with XS prefix
764 def VCVTSS2SDrr : I<0x5A, MRMSrcReg, (outs FR64:$dst),
765 (ins FR32:$src1, FR32:$src2),
766 "vcvtss2sd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
767 []>, XS, Requires<[HasAVX]>, VEX_4V;
768 def VCVTSS2SDrm : I<0x5A, MRMSrcMem, (outs FR64:$dst),
769 (ins FR32:$src1, f32mem:$src2),
770 "vcvtss2sd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
771 []>, XS, VEX_4V, Requires<[HasAVX, OptForSize]>;
773 let Predicates = [HasAVX] in {
774 def : Pat<(f64 (fextend FR32:$src)),
775 (VCVTSS2SDrr FR32:$src, FR32:$src)>;
776 def : Pat<(fextend (loadf32 addr:$src)),
777 (VCVTSS2SDrm (f32 (IMPLICIT_DEF)), addr:$src)>;
778 def : Pat<(extloadf32 addr:$src),
779 (VCVTSS2SDrm (f32 (IMPLICIT_DEF)), addr:$src)>;
782 def CVTSS2SDrr : I<0x5A, MRMSrcReg, (outs FR64:$dst), (ins FR32:$src),
783 "cvtss2sd\t{$src, $dst|$dst, $src}",
784 [(set FR64:$dst, (fextend FR32:$src))]>, XS,
786 def CVTSS2SDrm : I<0x5A, MRMSrcMem, (outs FR64:$dst), (ins f32mem:$src),
787 "cvtss2sd\t{$src, $dst|$dst, $src}",
788 [(set FR64:$dst, (extloadf32 addr:$src))]>, XS,
789 Requires<[HasSSE2, OptForSize]>;
791 def Int_VCVTSS2SDrr: I<0x5A, MRMSrcReg,
792 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
793 "vcvtss2sd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
794 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
795 VR128:$src2))]>, XS, VEX_4V,
797 def Int_VCVTSS2SDrm: I<0x5A, MRMSrcMem,
798 (outs VR128:$dst), (ins VR128:$src1, f32mem:$src2),
799 "vcvtss2sd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
800 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
801 (load addr:$src2)))]>, XS, VEX_4V,
803 let Constraints = "$src1 = $dst" in { // SSE2 instructions with XS prefix
804 def Int_CVTSS2SDrr: I<0x5A, MRMSrcReg,
805 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
806 "cvtss2sd\t{$src2, $dst|$dst, $src2}",
807 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
810 def Int_CVTSS2SDrm: I<0x5A, MRMSrcMem,
811 (outs VR128:$dst), (ins VR128:$src1, f32mem:$src2),
812 "cvtss2sd\t{$src2, $dst|$dst, $src2}",
813 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
814 (load addr:$src2)))]>, XS,
818 def : Pat<(extloadf32 addr:$src),
819 (CVTSS2SDrr (MOVSSrm addr:$src))>,
820 Requires<[HasSSE2, OptForSpeed]>;
822 // Convert doubleword to packed single/double fp
823 // SSE2 instructions without OpSize prefix
824 def Int_VCVTDQ2PSrr : I<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
825 "vcvtdq2ps\t{$src, $dst|$dst, $src}",
826 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps VR128:$src))]>,
827 TB, VEX, Requires<[HasAVX]>;
828 def Int_VCVTDQ2PSrm : I<0x5B, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
829 "vcvtdq2ps\t{$src, $dst|$dst, $src}",
830 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps
831 (bitconvert (memopv2i64 addr:$src))))]>,
832 TB, VEX, Requires<[HasAVX]>;
833 def Int_CVTDQ2PSrr : I<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
834 "cvtdq2ps\t{$src, $dst|$dst, $src}",
835 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps VR128:$src))]>,
836 TB, Requires<[HasSSE2]>;
837 def Int_CVTDQ2PSrm : I<0x5B, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
838 "cvtdq2ps\t{$src, $dst|$dst, $src}",
839 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps
840 (bitconvert (memopv2i64 addr:$src))))]>,
841 TB, Requires<[HasSSE2]>;
843 // FIXME: why the non-intrinsic version is described as SSE3?
844 // SSE2 instructions with XS prefix
845 def Int_VCVTDQ2PDrr : I<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
846 "vcvtdq2pd\t{$src, $dst|$dst, $src}",
847 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd VR128:$src))]>,
848 XS, VEX, Requires<[HasAVX]>;
849 def Int_VCVTDQ2PDrm : I<0xE6, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
850 "vcvtdq2pd\t{$src, $dst|$dst, $src}",
851 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd
852 (bitconvert (memopv2i64 addr:$src))))]>,
853 XS, VEX, Requires<[HasAVX]>;
854 def Int_CVTDQ2PDrr : I<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
855 "cvtdq2pd\t{$src, $dst|$dst, $src}",
856 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd VR128:$src))]>,
857 XS, Requires<[HasSSE2]>;
858 def Int_CVTDQ2PDrm : I<0xE6, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
859 "cvtdq2pd\t{$src, $dst|$dst, $src}",
860 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd
861 (bitconvert (memopv2i64 addr:$src))))]>,
862 XS, Requires<[HasSSE2]>;
865 // Convert packed single/double fp to doubleword
866 def VCVTPS2DQrr : VPDI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
867 "cvtps2dq\t{$src, $dst|$dst, $src}", []>, VEX;
868 def VCVTPS2DQrm : VPDI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
869 "cvtps2dq\t{$src, $dst|$dst, $src}", []>, VEX;
870 def VCVTPS2DQYrr : VPDI<0x5B, MRMSrcReg, (outs VR256:$dst), (ins VR256:$src),
871 "cvtps2dq\t{$src, $dst|$dst, $src}", []>, VEX;
872 def VCVTPS2DQYrm : VPDI<0x5B, MRMSrcMem, (outs VR256:$dst), (ins f256mem:$src),
873 "cvtps2dq\t{$src, $dst|$dst, $src}", []>, VEX;
874 def CVTPS2DQrr : PDI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
875 "cvtps2dq\t{$src, $dst|$dst, $src}", []>;
876 def CVTPS2DQrm : PDI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
877 "cvtps2dq\t{$src, $dst|$dst, $src}", []>;
879 def Int_VCVTPS2DQrr : VPDI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
880 "cvtps2dq\t{$src, $dst|$dst, $src}",
881 [(set VR128:$dst, (int_x86_sse2_cvtps2dq VR128:$src))]>,
883 def Int_VCVTPS2DQrm : VPDI<0x5B, MRMSrcMem, (outs VR128:$dst),
885 "cvtps2dq\t{$src, $dst|$dst, $src}",
886 [(set VR128:$dst, (int_x86_sse2_cvtps2dq
887 (memop addr:$src)))]>, VEX;
888 def Int_CVTPS2DQrr : PDI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
889 "cvtps2dq\t{$src, $dst|$dst, $src}",
890 [(set VR128:$dst, (int_x86_sse2_cvtps2dq VR128:$src))]>;
891 def Int_CVTPS2DQrm : PDI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
892 "cvtps2dq\t{$src, $dst|$dst, $src}",
893 [(set VR128:$dst, (int_x86_sse2_cvtps2dq
894 (memop addr:$src)))]>;
896 // SSE2 packed instructions with XD prefix
897 def Int_VCVTPD2DQrr : I<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
898 "vcvtpd2dq\t{$src, $dst|$dst, $src}",
899 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq VR128:$src))]>,
900 XD, VEX, Requires<[HasAVX]>;
901 def Int_VCVTPD2DQrm : I<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
902 "vcvtpd2dq\t{$src, $dst|$dst, $src}",
903 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq
904 (memop addr:$src)))]>,
905 XD, VEX, Requires<[HasAVX]>;
906 def Int_CVTPD2DQrr : I<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
907 "cvtpd2dq\t{$src, $dst|$dst, $src}",
908 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq VR128:$src))]>,
909 XD, Requires<[HasSSE2]>;
910 def Int_CVTPD2DQrm : I<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
911 "cvtpd2dq\t{$src, $dst|$dst, $src}",
912 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq
913 (memop addr:$src)))]>,
914 XD, Requires<[HasSSE2]>;
917 // Convert with truncation packed single/double fp to doubleword
918 // SSE2 packed instructions with XS prefix
919 def VCVTTPS2DQrr : VSSI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
920 "cvttps2dq\t{$src, $dst|$dst, $src}", []>, VEX;
921 def VCVTTPS2DQrm : VSSI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
922 "cvttps2dq\t{$src, $dst|$dst, $src}", []>, VEX;
923 def VCVTTPS2DQYrr : VSSI<0x5B, MRMSrcReg, (outs VR256:$dst), (ins VR256:$src),
924 "cvttps2dq\t{$src, $dst|$dst, $src}", []>, VEX;
925 def VCVTTPS2DQYrm : VSSI<0x5B, MRMSrcMem, (outs VR256:$dst), (ins f256mem:$src),
926 "cvttps2dq\t{$src, $dst|$dst, $src}", []>, VEX;
927 def CVTTPS2DQrr : SSI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
928 "cvttps2dq\t{$src, $dst|$dst, $src}",
930 (int_x86_sse2_cvttps2dq VR128:$src))]>;
931 def CVTTPS2DQrm : SSI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
932 "cvttps2dq\t{$src, $dst|$dst, $src}",
934 (int_x86_sse2_cvttps2dq (memop addr:$src)))]>;
936 def Int_VCVTTPS2DQrr : I<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
937 "vcvttps2dq\t{$src, $dst|$dst, $src}",
939 (int_x86_sse2_cvttps2dq VR128:$src))]>,
940 XS, VEX, Requires<[HasAVX]>;
941 def Int_VCVTTPS2DQrm : I<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
942 "vcvttps2dq\t{$src, $dst|$dst, $src}",
943 [(set VR128:$dst, (int_x86_sse2_cvttps2dq
944 (memop addr:$src)))]>,
945 XS, VEX, Requires<[HasAVX]>;
947 def : Pat<(v4f32 (sint_to_fp (v4i32 VR128:$src))),
948 (Int_CVTDQ2PSrr VR128:$src)>, Requires<[HasSSE2]>;
949 def : Pat<(v4i32 (fp_to_sint (v4f32 VR128:$src))),
950 (CVTTPS2DQrr VR128:$src)>, Requires<[HasSSE2]>;
952 def : Pat<(v4f32 (sint_to_fp (v4i32 VR128:$src))),
953 (Int_VCVTDQ2PSrr VR128:$src)>, Requires<[HasAVX]>;
954 def : Pat<(v4i32 (fp_to_sint (v4f32 VR128:$src))),
955 (VCVTTPS2DQrr VR128:$src)>, Requires<[HasAVX]>;
956 def : Pat<(v8f32 (sint_to_fp (v8i32 VR256:$src))),
957 (VCVTDQ2PSYrr VR256:$src)>, Requires<[HasAVX]>;
958 def : Pat<(v8i32 (fp_to_sint (v8f32 VR256:$src))),
959 (VCVTTPS2DQYrr VR256:$src)>, Requires<[HasAVX]>;
961 def Int_VCVTTPD2DQrr : VPDI<0xE6, MRMSrcReg, (outs VR128:$dst),
963 "cvttpd2dq\t{$src, $dst|$dst, $src}",
964 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq VR128:$src))]>,
966 def Int_VCVTTPD2DQrm : VPDI<0xE6, MRMSrcMem, (outs VR128:$dst),
968 "cvttpd2dq\t{$src, $dst|$dst, $src}",
969 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq
970 (memop addr:$src)))]>, VEX;
971 def CVTTPD2DQrr : PDI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
972 "cvttpd2dq\t{$src, $dst|$dst, $src}",
973 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq VR128:$src))]>;
974 def CVTTPD2DQrm : PDI<0xE6, MRMSrcMem, (outs VR128:$dst),(ins f128mem:$src),
975 "cvttpd2dq\t{$src, $dst|$dst, $src}",
976 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq
977 (memop addr:$src)))]>;
979 // The assembler can recognize rr 256-bit instructions by seeing a ymm
980 // register, but the same isn't true when using memory operands instead.
981 // Provide other assembly rr and rm forms to address this explicitly.
982 def VCVTTPD2DQrr : VPDI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
983 "cvttpd2dq\t{$src, $dst|$dst, $src}", []>, VEX;
984 def VCVTTPD2DQXrYr : VPDI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR256:$src),
985 "cvttpd2dq\t{$src, $dst|$dst, $src}", []>, VEX;
988 def VCVTTPD2DQXrr : VPDI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
989 "cvttpd2dqx\t{$src, $dst|$dst, $src}", []>, VEX;
990 def VCVTTPD2DQXrm : VPDI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
991 "cvttpd2dqx\t{$src, $dst|$dst, $src}", []>, VEX;
994 def VCVTTPD2DQYrr : VPDI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR256:$src),
995 "cvttpd2dqy\t{$src, $dst|$dst, $src}", []>, VEX;
996 def VCVTTPD2DQYrm : VPDI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f256mem:$src),
997 "cvttpd2dqy\t{$src, $dst|$dst, $src}", []>, VEX, VEX_L;
999 // Convert packed single to packed double
1000 let Predicates = [HasAVX] in {
1001 // SSE2 instructions without OpSize prefix
1002 def VCVTPS2PDrr : I<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1003 "vcvtps2pd\t{$src, $dst|$dst, $src}", []>, VEX;
1004 def VCVTPS2PDrm : I<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
1005 "vcvtps2pd\t{$src, $dst|$dst, $src}", []>, VEX;
1006 def VCVTPS2PDYrr : I<0x5A, MRMSrcReg, (outs VR256:$dst), (ins VR128:$src),
1007 "vcvtps2pd\t{$src, $dst|$dst, $src}", []>, VEX;
1008 def VCVTPS2PDYrm : I<0x5A, MRMSrcMem, (outs VR256:$dst), (ins f128mem:$src),
1009 "vcvtps2pd\t{$src, $dst|$dst, $src}", []>, VEX;
1011 def CVTPS2PDrr : I<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1012 "cvtps2pd\t{$src, $dst|$dst, $src}", []>, TB;
1013 def CVTPS2PDrm : I<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
1014 "cvtps2pd\t{$src, $dst|$dst, $src}", []>, TB;
1016 def Int_VCVTPS2PDrr : I<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1017 "vcvtps2pd\t{$src, $dst|$dst, $src}",
1018 [(set VR128:$dst, (int_x86_sse2_cvtps2pd VR128:$src))]>,
1019 VEX, Requires<[HasAVX]>;
1020 def Int_VCVTPS2PDrm : I<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
1021 "vcvtps2pd\t{$src, $dst|$dst, $src}",
1022 [(set VR128:$dst, (int_x86_sse2_cvtps2pd
1023 (load addr:$src)))]>,
1024 VEX, Requires<[HasAVX]>;
1025 def Int_CVTPS2PDrr : I<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1026 "cvtps2pd\t{$src, $dst|$dst, $src}",
1027 [(set VR128:$dst, (int_x86_sse2_cvtps2pd VR128:$src))]>,
1028 TB, Requires<[HasSSE2]>;
1029 def Int_CVTPS2PDrm : I<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
1030 "cvtps2pd\t{$src, $dst|$dst, $src}",
1031 [(set VR128:$dst, (int_x86_sse2_cvtps2pd
1032 (load addr:$src)))]>,
1033 TB, Requires<[HasSSE2]>;
1035 // Convert packed double to packed single
1036 // The assembler can recognize rr 256-bit instructions by seeing a ymm
1037 // register, but the same isn't true when using memory operands instead.
1038 // Provide other assembly rr and rm forms to address this explicitly.
1039 def VCVTPD2PSrr : VPDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1040 "cvtpd2ps\t{$src, $dst|$dst, $src}", []>, VEX;
1041 def VCVTPD2PSXrYr : VPDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR256:$src),
1042 "cvtpd2ps\t{$src, $dst|$dst, $src}", []>, VEX;
1045 def VCVTPD2PSXrr : VPDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1046 "cvtpd2psx\t{$src, $dst|$dst, $src}", []>, VEX;
1047 def VCVTPD2PSXrm : VPDI<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1048 "cvtpd2psx\t{$src, $dst|$dst, $src}", []>, VEX;
1051 def VCVTPD2PSYrr : VPDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR256:$src),
1052 "cvtpd2psy\t{$src, $dst|$dst, $src}", []>, VEX;
1053 def VCVTPD2PSYrm : VPDI<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f256mem:$src),
1054 "cvtpd2psy\t{$src, $dst|$dst, $src}", []>, VEX, VEX_L;
1055 def CVTPD2PSrr : PDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1056 "cvtpd2ps\t{$src, $dst|$dst, $src}", []>;
1057 def CVTPD2PSrm : PDI<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1058 "cvtpd2ps\t{$src, $dst|$dst, $src}", []>;
1061 def Int_VCVTPD2PSrr : VPDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1062 "cvtpd2ps\t{$src, $dst|$dst, $src}",
1063 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps VR128:$src))]>;
1064 def Int_VCVTPD2PSrm : VPDI<0x5A, MRMSrcMem, (outs VR128:$dst),
1066 "cvtpd2ps\t{$src, $dst|$dst, $src}",
1067 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps
1068 (memop addr:$src)))]>;
1069 def Int_CVTPD2PSrr : PDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1070 "cvtpd2ps\t{$src, $dst|$dst, $src}",
1071 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps VR128:$src))]>;
1072 def Int_CVTPD2PSrm : PDI<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1073 "cvtpd2ps\t{$src, $dst|$dst, $src}",
1074 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps
1075 (memop addr:$src)))]>;
1077 // AVX 256-bit register conversion intrinsics
1078 // FIXME: Migrate SSE conversion intrinsics matching to use patterns as below
1079 // whenever possible to avoid declaring two versions of each one.
1080 def : Pat<(int_x86_avx_cvtdq2_ps_256 VR256:$src),
1081 (VCVTDQ2PSYrr VR256:$src)>;
1082 def : Pat<(int_x86_avx_cvtdq2_ps_256 (memopv8i32 addr:$src)),
1083 (VCVTDQ2PSYrm addr:$src)>;
1085 def : Pat<(int_x86_avx_cvt_pd2_ps_256 VR256:$src),
1086 (VCVTPD2PSYrr VR256:$src)>;
1087 def : Pat<(int_x86_avx_cvt_pd2_ps_256 (memopv4f64 addr:$src)),
1088 (VCVTPD2PSYrm addr:$src)>;
1090 def : Pat<(int_x86_avx_cvt_ps2dq_256 VR256:$src),
1091 (VCVTPS2DQYrr VR256:$src)>;
1092 def : Pat<(int_x86_avx_cvt_ps2dq_256 (memopv8f32 addr:$src)),
1093 (VCVTPS2DQYrm addr:$src)>;
1095 def : Pat<(int_x86_avx_cvt_ps2_pd_256 VR128:$src),
1096 (VCVTPS2PDYrr VR128:$src)>;
1097 def : Pat<(int_x86_avx_cvt_ps2_pd_256 (memopv4f32 addr:$src)),
1098 (VCVTPS2PDYrm addr:$src)>;
1100 def : Pat<(int_x86_avx_cvtt_pd2dq_256 VR256:$src),
1101 (VCVTTPD2DQYrr VR256:$src)>;
1102 def : Pat<(int_x86_avx_cvtt_pd2dq_256 (memopv4f64 addr:$src)),
1103 (VCVTTPD2DQYrm addr:$src)>;
1105 def : Pat<(int_x86_avx_cvtt_ps2dq_256 VR256:$src),
1106 (VCVTTPS2DQYrr VR256:$src)>;
1107 def : Pat<(int_x86_avx_cvtt_ps2dq_256 (memopv8f32 addr:$src)),
1108 (VCVTTPS2DQYrm addr:$src)>;
1110 // Match fround and fextend for 128/256-bit conversions
1111 def : Pat<(v4f32 (fround (v4f64 VR256:$src))),
1112 (VCVTPD2PSYrr VR256:$src)>;
1113 def : Pat<(v4f32 (fround (loadv4f64 addr:$src))),
1114 (VCVTPD2PSYrm addr:$src)>;
1116 def : Pat<(v4f64 (fextend (v4f32 VR128:$src))),
1117 (VCVTPS2PDYrr VR128:$src)>;
1118 def : Pat<(v4f64 (fextend (loadv4f32 addr:$src))),
1119 (VCVTPS2PDYrm addr:$src)>;
1121 //===----------------------------------------------------------------------===//
1122 // SSE 1 & 2 - Compare Instructions
1123 //===----------------------------------------------------------------------===//
1125 // sse12_cmp_scalar - sse 1 & 2 compare scalar instructions
1126 multiclass sse12_cmp_scalar<RegisterClass RC, X86MemOperand x86memop,
1127 string asm, string asm_alt> {
1128 let isAsmParserOnly = 1 in {
1129 def rr : SIi8<0xC2, MRMSrcReg,
1130 (outs RC:$dst), (ins RC:$src1, RC:$src, SSECC:$cc),
1133 def rm : SIi8<0xC2, MRMSrcMem,
1134 (outs RC:$dst), (ins RC:$src1, x86memop:$src, SSECC:$cc),
1138 // Accept explicit immediate argument form instead of comparison code.
1139 def rr_alt : SIi8<0xC2, MRMSrcReg,
1140 (outs RC:$dst), (ins RC:$src1, RC:$src, i8imm:$src2),
1143 def rm_alt : SIi8<0xC2, MRMSrcMem,
1144 (outs RC:$dst), (ins RC:$src1, x86memop:$src, i8imm:$src2),
1148 let neverHasSideEffects = 1 in {
1149 defm VCMPSS : sse12_cmp_scalar<FR32, f32mem,
1150 "cmp${cc}ss\t{$src, $src1, $dst|$dst, $src1, $src}",
1151 "cmpss\t{$src2, $src, $src1, $dst|$dst, $src1, $src, $src2}">,
1153 defm VCMPSD : sse12_cmp_scalar<FR64, f64mem,
1154 "cmp${cc}sd\t{$src, $src1, $dst|$dst, $src1, $src}",
1155 "cmpsd\t{$src2, $src, $src1, $dst|$dst, $src1, $src, $src2}">,
1159 let Constraints = "$src1 = $dst" in {
1160 def CMPSSrr : SIi8<0xC2, MRMSrcReg,
1161 (outs FR32:$dst), (ins FR32:$src1, FR32:$src2, SSECC:$cc),
1162 "cmp${cc}ss\t{$src2, $dst|$dst, $src2}",
1163 [(set FR32:$dst, (X86cmpss (f32 FR32:$src1), FR32:$src2, imm:$cc))]>, XS;
1164 def CMPSSrm : SIi8<0xC2, MRMSrcMem,
1165 (outs FR32:$dst), (ins FR32:$src1, f32mem:$src2, SSECC:$cc),
1166 "cmp${cc}ss\t{$src2, $dst|$dst, $src2}",
1167 [(set FR32:$dst, (X86cmpss (f32 FR32:$src1), (loadf32 addr:$src2), imm:$cc))]>, XS;
1168 def CMPSDrr : SIi8<0xC2, MRMSrcReg,
1169 (outs FR64:$dst), (ins FR64:$src1, FR64:$src2, SSECC:$cc),
1170 "cmp${cc}sd\t{$src2, $dst|$dst, $src2}",
1171 [(set FR64:$dst, (X86cmpsd (f64 FR64:$src1), FR64:$src2, imm:$cc))]>, XD;
1172 def CMPSDrm : SIi8<0xC2, MRMSrcMem,
1173 (outs FR64:$dst), (ins FR64:$src1, f64mem:$src2, SSECC:$cc),
1174 "cmp${cc}sd\t{$src2, $dst|$dst, $src2}",
1175 [(set FR64:$dst, (X86cmpsd (f64 FR64:$src1), (loadf64 addr:$src2), imm:$cc))]>, XD;
1177 let Constraints = "$src1 = $dst", neverHasSideEffects = 1 in {
1178 def CMPSSrr_alt : SIi8<0xC2, MRMSrcReg,
1179 (outs FR32:$dst), (ins FR32:$src1, FR32:$src, i8imm:$src2),
1180 "cmpss\t{$src2, $src, $dst|$dst, $src, $src2}", []>, XS;
1181 def CMPSSrm_alt : SIi8<0xC2, MRMSrcMem,
1182 (outs FR32:$dst), (ins FR32:$src1, f32mem:$src, i8imm:$src2),
1183 "cmpss\t{$src2, $src, $dst|$dst, $src, $src2}", []>, XS;
1184 def CMPSDrr_alt : SIi8<0xC2, MRMSrcReg,
1185 (outs FR64:$dst), (ins FR64:$src1, FR64:$src, i8imm:$src2),
1186 "cmpsd\t{$src2, $src, $dst|$dst, $src, $src2}", []>, XD;
1187 def CMPSDrm_alt : SIi8<0xC2, MRMSrcMem,
1188 (outs FR64:$dst), (ins FR64:$src1, f64mem:$src, i8imm:$src2),
1189 "cmpsd\t{$src2, $src, $dst|$dst, $src, $src2}", []>, XD;
1192 multiclass sse12_cmp_scalar_int<RegisterClass RC, X86MemOperand x86memop,
1193 Intrinsic Int, string asm> {
1194 def rr : SIi8<0xC2, MRMSrcReg, (outs VR128:$dst),
1195 (ins VR128:$src1, VR128:$src, SSECC:$cc), asm,
1196 [(set VR128:$dst, (Int VR128:$src1,
1197 VR128:$src, imm:$cc))]>;
1198 def rm : SIi8<0xC2, MRMSrcMem, (outs VR128:$dst),
1199 (ins VR128:$src1, f32mem:$src, SSECC:$cc), asm,
1200 [(set VR128:$dst, (Int VR128:$src1,
1201 (load addr:$src), imm:$cc))]>;
1204 // Aliases to match intrinsics which expect XMM operand(s).
1205 defm Int_VCMPSS : sse12_cmp_scalar_int<VR128, f32mem, int_x86_sse_cmp_ss,
1206 "cmp${cc}ss\t{$src, $src1, $dst|$dst, $src1, $src}">,
1208 defm Int_VCMPSD : sse12_cmp_scalar_int<VR128, f64mem, int_x86_sse2_cmp_sd,
1209 "cmp${cc}sd\t{$src, $src1, $dst|$dst, $src1, $src}">,
1211 let Constraints = "$src1 = $dst" in {
1212 defm Int_CMPSS : sse12_cmp_scalar_int<VR128, f32mem, int_x86_sse_cmp_ss,
1213 "cmp${cc}ss\t{$src, $dst|$dst, $src}">, XS;
1214 defm Int_CMPSD : sse12_cmp_scalar_int<VR128, f64mem, int_x86_sse2_cmp_sd,
1215 "cmp${cc}sd\t{$src, $dst|$dst, $src}">, XD;
1219 // sse12_ord_cmp - Unordered/Ordered scalar fp compare and set EFLAGS
1220 multiclass sse12_ord_cmp<bits<8> opc, RegisterClass RC, SDNode OpNode,
1221 ValueType vt, X86MemOperand x86memop,
1222 PatFrag ld_frag, string OpcodeStr, Domain d> {
1223 def rr: PI<opc, MRMSrcReg, (outs), (ins RC:$src1, RC:$src2),
1224 !strconcat(OpcodeStr, "\t{$src2, $src1|$src1, $src2}"),
1225 [(set EFLAGS, (OpNode (vt RC:$src1), RC:$src2))], d>;
1226 def rm: PI<opc, MRMSrcMem, (outs), (ins RC:$src1, x86memop:$src2),
1227 !strconcat(OpcodeStr, "\t{$src2, $src1|$src1, $src2}"),
1228 [(set EFLAGS, (OpNode (vt RC:$src1),
1229 (ld_frag addr:$src2)))], d>;
1232 let Defs = [EFLAGS] in {
1233 defm VUCOMISS : sse12_ord_cmp<0x2E, FR32, X86cmp, f32, f32mem, loadf32,
1234 "ucomiss", SSEPackedSingle>, VEX;
1235 defm VUCOMISD : sse12_ord_cmp<0x2E, FR64, X86cmp, f64, f64mem, loadf64,
1236 "ucomisd", SSEPackedDouble>, OpSize, VEX;
1237 let Pattern = []<dag> in {
1238 defm VCOMISS : sse12_ord_cmp<0x2F, VR128, undef, v4f32, f128mem, load,
1239 "comiss", SSEPackedSingle>, VEX;
1240 defm VCOMISD : sse12_ord_cmp<0x2F, VR128, undef, v2f64, f128mem, load,
1241 "comisd", SSEPackedDouble>, OpSize, VEX;
1244 defm Int_VUCOMISS : sse12_ord_cmp<0x2E, VR128, X86ucomi, v4f32, f128mem,
1245 load, "ucomiss", SSEPackedSingle>, VEX;
1246 defm Int_VUCOMISD : sse12_ord_cmp<0x2E, VR128, X86ucomi, v2f64, f128mem,
1247 load, "ucomisd", SSEPackedDouble>, OpSize, VEX;
1249 defm Int_VCOMISS : sse12_ord_cmp<0x2F, VR128, X86comi, v4f32, f128mem,
1250 load, "comiss", SSEPackedSingle>, VEX;
1251 defm Int_VCOMISD : sse12_ord_cmp<0x2F, VR128, X86comi, v2f64, f128mem,
1252 load, "comisd", SSEPackedDouble>, OpSize, VEX;
1253 defm UCOMISS : sse12_ord_cmp<0x2E, FR32, X86cmp, f32, f32mem, loadf32,
1254 "ucomiss", SSEPackedSingle>, TB;
1255 defm UCOMISD : sse12_ord_cmp<0x2E, FR64, X86cmp, f64, f64mem, loadf64,
1256 "ucomisd", SSEPackedDouble>, TB, OpSize;
1258 let Pattern = []<dag> in {
1259 defm COMISS : sse12_ord_cmp<0x2F, VR128, undef, v4f32, f128mem, load,
1260 "comiss", SSEPackedSingle>, TB;
1261 defm COMISD : sse12_ord_cmp<0x2F, VR128, undef, v2f64, f128mem, load,
1262 "comisd", SSEPackedDouble>, TB, OpSize;
1265 defm Int_UCOMISS : sse12_ord_cmp<0x2E, VR128, X86ucomi, v4f32, f128mem,
1266 load, "ucomiss", SSEPackedSingle>, TB;
1267 defm Int_UCOMISD : sse12_ord_cmp<0x2E, VR128, X86ucomi, v2f64, f128mem,
1268 load, "ucomisd", SSEPackedDouble>, TB, OpSize;
1270 defm Int_COMISS : sse12_ord_cmp<0x2F, VR128, X86comi, v4f32, f128mem, load,
1271 "comiss", SSEPackedSingle>, TB;
1272 defm Int_COMISD : sse12_ord_cmp<0x2F, VR128, X86comi, v2f64, f128mem, load,
1273 "comisd", SSEPackedDouble>, TB, OpSize;
1274 } // Defs = [EFLAGS]
1276 // sse12_cmp_packed - sse 1 & 2 compared packed instructions
1277 multiclass sse12_cmp_packed<RegisterClass RC, X86MemOperand x86memop,
1278 Intrinsic Int, string asm, string asm_alt,
1280 let isAsmParserOnly = 1 in {
1281 def rri : PIi8<0xC2, MRMSrcReg,
1282 (outs RC:$dst), (ins RC:$src1, RC:$src, SSECC:$cc), asm,
1283 [(set RC:$dst, (Int RC:$src1, RC:$src, imm:$cc))], d>;
1284 def rmi : PIi8<0xC2, MRMSrcMem,
1285 (outs RC:$dst), (ins RC:$src1, f128mem:$src, SSECC:$cc), asm,
1286 [(set RC:$dst, (Int RC:$src1, (memop addr:$src), imm:$cc))], d>;
1289 // Accept explicit immediate argument form instead of comparison code.
1290 def rri_alt : PIi8<0xC2, MRMSrcReg,
1291 (outs RC:$dst), (ins RC:$src1, RC:$src, i8imm:$src2),
1293 def rmi_alt : PIi8<0xC2, MRMSrcMem,
1294 (outs RC:$dst), (ins RC:$src1, f128mem:$src, i8imm:$src2),
1298 defm VCMPPS : sse12_cmp_packed<VR128, f128mem, int_x86_sse_cmp_ps,
1299 "cmp${cc}ps\t{$src, $src1, $dst|$dst, $src1, $src}",
1300 "cmpps\t{$src2, $src, $src1, $dst|$dst, $src1, $src, $src2}",
1301 SSEPackedSingle>, VEX_4V;
1302 defm VCMPPD : sse12_cmp_packed<VR128, f128mem, int_x86_sse2_cmp_pd,
1303 "cmp${cc}pd\t{$src, $src1, $dst|$dst, $src1, $src}",
1304 "cmppd\t{$src2, $src, $src1, $dst|$dst, $src1, $src, $src2}",
1305 SSEPackedDouble>, OpSize, VEX_4V;
1306 defm VCMPPSY : sse12_cmp_packed<VR256, f256mem, int_x86_avx_cmp_ps_256,
1307 "cmp${cc}ps\t{$src, $src1, $dst|$dst, $src1, $src}",
1308 "cmpps\t{$src2, $src, $src1, $dst|$dst, $src1, $src, $src2}",
1309 SSEPackedSingle>, VEX_4V;
1310 defm VCMPPDY : sse12_cmp_packed<VR256, f256mem, int_x86_avx_cmp_pd_256,
1311 "cmp${cc}pd\t{$src, $src1, $dst|$dst, $src1, $src}",
1312 "cmppd\t{$src2, $src, $src1, $dst|$dst, $src1, $src, $src2}",
1313 SSEPackedDouble>, OpSize, VEX_4V;
1314 let Constraints = "$src1 = $dst" in {
1315 defm CMPPS : sse12_cmp_packed<VR128, f128mem, int_x86_sse_cmp_ps,
1316 "cmp${cc}ps\t{$src, $dst|$dst, $src}",
1317 "cmpps\t{$src2, $src, $dst|$dst, $src, $src2}",
1318 SSEPackedSingle>, TB;
1319 defm CMPPD : sse12_cmp_packed<VR128, f128mem, int_x86_sse2_cmp_pd,
1320 "cmp${cc}pd\t{$src, $dst|$dst, $src}",
1321 "cmppd\t{$src2, $src, $dst|$dst, $src, $src2}",
1322 SSEPackedDouble>, TB, OpSize;
1325 let Predicates = [HasSSE1] in {
1326 def : Pat<(v4i32 (X86cmpps (v4f32 VR128:$src1), VR128:$src2, imm:$cc)),
1327 (CMPPSrri (v4f32 VR128:$src1), (v4f32 VR128:$src2), imm:$cc)>;
1328 def : Pat<(v4i32 (X86cmpps (v4f32 VR128:$src1), (memop addr:$src2), imm:$cc)),
1329 (CMPPSrmi (v4f32 VR128:$src1), addr:$src2, imm:$cc)>;
1332 let Predicates = [HasSSE2] in {
1333 def : Pat<(v2i64 (X86cmppd (v2f64 VR128:$src1), VR128:$src2, imm:$cc)),
1334 (CMPPDrri VR128:$src1, VR128:$src2, imm:$cc)>;
1335 def : Pat<(v2i64 (X86cmppd (v2f64 VR128:$src1), (memop addr:$src2), imm:$cc)),
1336 (CMPPDrmi VR128:$src1, addr:$src2, imm:$cc)>;
1339 let Predicates = [HasAVX] in {
1340 def : Pat<(v4i32 (X86cmpps (v4f32 VR128:$src1), VR128:$src2, imm:$cc)),
1341 (VCMPPSrri (v4f32 VR128:$src1), (v4f32 VR128:$src2), imm:$cc)>;
1342 def : Pat<(v4i32 (X86cmpps (v4f32 VR128:$src1), (memop addr:$src2), imm:$cc)),
1343 (VCMPPSrmi (v4f32 VR128:$src1), addr:$src2, imm:$cc)>;
1344 def : Pat<(v2i64 (X86cmppd (v2f64 VR128:$src1), VR128:$src2, imm:$cc)),
1345 (VCMPPDrri VR128:$src1, VR128:$src2, imm:$cc)>;
1346 def : Pat<(v2i64 (X86cmppd (v2f64 VR128:$src1), (memop addr:$src2), imm:$cc)),
1347 (VCMPPDrmi VR128:$src1, addr:$src2, imm:$cc)>;
1349 def : Pat<(v8i32 (X86cmpps (v8f32 VR256:$src1), VR256:$src2, imm:$cc)),
1350 (VCMPPSYrri (v8f32 VR256:$src1), (v8f32 VR256:$src2), imm:$cc)>;
1351 def : Pat<(v8i32 (X86cmpps (v8f32 VR256:$src1), (memop addr:$src2), imm:$cc)),
1352 (VCMPPSYrmi (v8f32 VR256:$src1), addr:$src2, imm:$cc)>;
1353 def : Pat<(v4i64 (X86cmppd (v4f64 VR256:$src1), VR256:$src2, imm:$cc)),
1354 (VCMPPDYrri VR256:$src1, VR256:$src2, imm:$cc)>;
1355 def : Pat<(v4i64 (X86cmppd (v4f64 VR256:$src1), (memop addr:$src2), imm:$cc)),
1356 (VCMPPDYrmi VR256:$src1, addr:$src2, imm:$cc)>;
1359 //===----------------------------------------------------------------------===//
1360 // SSE 1 & 2 - Shuffle Instructions
1361 //===----------------------------------------------------------------------===//
1363 /// sse12_shuffle - sse 1 & 2 shuffle instructions
1364 multiclass sse12_shuffle<RegisterClass RC, X86MemOperand x86memop,
1365 ValueType vt, string asm, PatFrag mem_frag,
1366 Domain d, bit IsConvertibleToThreeAddress = 0> {
1367 def rmi : PIi8<0xC6, MRMSrcMem, (outs RC:$dst),
1368 (ins RC:$src1, f128mem:$src2, i8imm:$src3), asm,
1369 [(set RC:$dst, (vt (shufp:$src3
1370 RC:$src1, (mem_frag addr:$src2))))], d>;
1371 let isConvertibleToThreeAddress = IsConvertibleToThreeAddress in
1372 def rri : PIi8<0xC6, MRMSrcReg, (outs RC:$dst),
1373 (ins RC:$src1, RC:$src2, i8imm:$src3), asm,
1375 (vt (shufp:$src3 RC:$src1, RC:$src2)))], d>;
1378 defm VSHUFPS : sse12_shuffle<VR128, f128mem, v4f32,
1379 "shufps\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
1380 memopv4f32, SSEPackedSingle>, TB, VEX_4V;
1381 defm VSHUFPSY : sse12_shuffle<VR256, f256mem, v8f32,
1382 "shufps\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
1383 memopv8f32, SSEPackedSingle>, TB, VEX_4V;
1384 defm VSHUFPD : sse12_shuffle<VR128, f128mem, v2f64,
1385 "shufpd\t{$src3, $src2, $src1, $dst|$dst, $src2, $src2, $src3}",
1386 memopv2f64, SSEPackedDouble>, TB, OpSize, VEX_4V;
1387 defm VSHUFPDY : sse12_shuffle<VR256, f256mem, v4f64,
1388 "shufpd\t{$src3, $src2, $src1, $dst|$dst, $src2, $src2, $src3}",
1389 memopv4f64, SSEPackedDouble>, TB, OpSize, VEX_4V;
1391 let Constraints = "$src1 = $dst" in {
1392 defm SHUFPS : sse12_shuffle<VR128, f128mem, v4f32,
1393 "shufps\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1394 memopv4f32, SSEPackedSingle, 1 /* cvt to pshufd */>,
1396 defm SHUFPD : sse12_shuffle<VR128, f128mem, v2f64,
1397 "shufpd\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1398 memopv2f64, SSEPackedDouble>, TB, OpSize;
1401 //===----------------------------------------------------------------------===//
1402 // SSE 1 & 2 - Unpack Instructions
1403 //===----------------------------------------------------------------------===//
1405 /// sse12_unpack_interleave - sse 1 & 2 unpack and interleave
1406 multiclass sse12_unpack_interleave<bits<8> opc, PatFrag OpNode, ValueType vt,
1407 PatFrag mem_frag, RegisterClass RC,
1408 X86MemOperand x86memop, string asm,
1410 def rr : PI<opc, MRMSrcReg,
1411 (outs RC:$dst), (ins RC:$src1, RC:$src2),
1413 (vt (OpNode RC:$src1, RC:$src2)))], d>;
1414 def rm : PI<opc, MRMSrcMem,
1415 (outs RC:$dst), (ins RC:$src1, x86memop:$src2),
1417 (vt (OpNode RC:$src1,
1418 (mem_frag addr:$src2))))], d>;
1421 let AddedComplexity = 10 in {
1422 defm VUNPCKHPS: sse12_unpack_interleave<0x15, unpckh, v4f32, memopv4f32,
1423 VR128, f128mem, "unpckhps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1424 SSEPackedSingle>, VEX_4V;
1425 defm VUNPCKHPD: sse12_unpack_interleave<0x15, unpckh, v2f64, memopv2f64,
1426 VR128, f128mem, "unpckhpd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1427 SSEPackedDouble>, OpSize, VEX_4V;
1428 defm VUNPCKLPS: sse12_unpack_interleave<0x14, unpckl, v4f32, memopv4f32,
1429 VR128, f128mem, "unpcklps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1430 SSEPackedSingle>, VEX_4V;
1431 defm VUNPCKLPD: sse12_unpack_interleave<0x14, unpckl, v2f64, memopv2f64,
1432 VR128, f128mem, "unpcklpd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1433 SSEPackedDouble>, OpSize, VEX_4V;
1435 defm VUNPCKHPSY: sse12_unpack_interleave<0x15, unpckh, v8f32, memopv8f32,
1436 VR256, f256mem, "unpckhps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1437 SSEPackedSingle>, VEX_4V;
1438 defm VUNPCKHPDY: sse12_unpack_interleave<0x15, unpckh, v4f64, memopv4f64,
1439 VR256, f256mem, "unpckhpd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1440 SSEPackedDouble>, OpSize, VEX_4V;
1441 defm VUNPCKLPSY: sse12_unpack_interleave<0x14, unpckl, v8f32, memopv8f32,
1442 VR256, f256mem, "unpcklps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1443 SSEPackedSingle>, VEX_4V;
1444 defm VUNPCKLPDY: sse12_unpack_interleave<0x14, unpckl, v4f64, memopv4f64,
1445 VR256, f256mem, "unpcklpd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1446 SSEPackedDouble>, OpSize, VEX_4V;
1448 let Constraints = "$src1 = $dst" in {
1449 defm UNPCKHPS: sse12_unpack_interleave<0x15, unpckh, v4f32, memopv4f32,
1450 VR128, f128mem, "unpckhps\t{$src2, $dst|$dst, $src2}",
1451 SSEPackedSingle>, TB;
1452 defm UNPCKHPD: sse12_unpack_interleave<0x15, unpckh, v2f64, memopv2f64,
1453 VR128, f128mem, "unpckhpd\t{$src2, $dst|$dst, $src2}",
1454 SSEPackedDouble>, TB, OpSize;
1455 defm UNPCKLPS: sse12_unpack_interleave<0x14, unpckl, v4f32, memopv4f32,
1456 VR128, f128mem, "unpcklps\t{$src2, $dst|$dst, $src2}",
1457 SSEPackedSingle>, TB;
1458 defm UNPCKLPD: sse12_unpack_interleave<0x14, unpckl, v2f64, memopv2f64,
1459 VR128, f128mem, "unpcklpd\t{$src2, $dst|$dst, $src2}",
1460 SSEPackedDouble>, TB, OpSize;
1461 } // Constraints = "$src1 = $dst"
1462 } // AddedComplexity
1464 //===----------------------------------------------------------------------===//
1465 // SSE 1 & 2 - Extract Floating-Point Sign mask
1466 //===----------------------------------------------------------------------===//
1468 /// sse12_extr_sign_mask - sse 1 & 2 unpack and interleave
1469 multiclass sse12_extr_sign_mask<RegisterClass RC, Intrinsic Int, string asm,
1471 def rr32 : PI<0x50, MRMSrcReg, (outs GR32:$dst), (ins RC:$src),
1472 !strconcat(asm, "\t{$src, $dst|$dst, $src}"),
1473 [(set GR32:$dst, (Int RC:$src))], d>;
1474 def rr64 : PI<0x50, MRMSrcReg, (outs GR64:$dst), (ins RC:$src),
1475 !strconcat(asm, "\t{$src, $dst|$dst, $src}"), [], d>, REX_W;
1478 defm MOVMSKPS : sse12_extr_sign_mask<VR128, int_x86_sse_movmsk_ps, "movmskps",
1479 SSEPackedSingle>, TB;
1480 defm MOVMSKPD : sse12_extr_sign_mask<VR128, int_x86_sse2_movmsk_pd, "movmskpd",
1481 SSEPackedDouble>, TB, OpSize;
1484 def MOVMSKPDrr32_alt : PI<0x50, MRMSrcReg, (outs GR32:$dst), (ins FR64:$src),
1485 "movmskpd\t{$src, $dst|$dst, $src}",
1486 [(set GR32:$dst, (X86fgetsign FR64:$src))], SSEPackedDouble>, TB,
1488 def MOVMSKPDrr64_alt : PI<0x50, MRMSrcReg, (outs GR64:$dst), (ins FR64:$src),
1489 "movmskpd\t{$src, $dst|$dst, $src}",
1490 [(set GR64:$dst, (X86fgetsign FR64:$src))], SSEPackedDouble>, TB,
1492 def MOVMSKPSrr32_alt : PI<0x50, MRMSrcReg, (outs GR32:$dst), (ins FR32:$src),
1493 "movmskps\t{$src, $dst|$dst, $src}",
1494 [(set GR32:$dst, (X86fgetsign FR32:$src))], SSEPackedSingle>, TB;
1495 def MOVMSKPSrr64_alt : PI<0x50, MRMSrcReg, (outs GR64:$dst), (ins FR32:$src),
1496 "movmskps\t{$src, $dst|$dst, $src}",
1497 [(set GR64:$dst, (X86fgetsign FR32:$src))], SSEPackedSingle>, TB;
1499 let Predicates = [HasAVX] in {
1500 defm VMOVMSKPS : sse12_extr_sign_mask<VR128, int_x86_sse_movmsk_ps,
1501 "movmskps", SSEPackedSingle>, TB, VEX;
1502 defm VMOVMSKPD : sse12_extr_sign_mask<VR128, int_x86_sse2_movmsk_pd,
1503 "movmskpd", SSEPackedDouble>, TB, OpSize,
1505 defm VMOVMSKPSY : sse12_extr_sign_mask<VR256, int_x86_avx_movmsk_ps_256,
1506 "movmskps", SSEPackedSingle>, TB, VEX;
1507 defm VMOVMSKPDY : sse12_extr_sign_mask<VR256, int_x86_avx_movmsk_pd_256,
1508 "movmskpd", SSEPackedDouble>, TB, OpSize,
1512 def VMOVMSKPSr64r : PI<0x50, MRMSrcReg, (outs GR64:$dst), (ins VR128:$src),
1513 "movmskps\t{$src, $dst|$dst, $src}", [], SSEPackedSingle>, VEX;
1514 def VMOVMSKPDr64r : PI<0x50, MRMSrcReg, (outs GR64:$dst), (ins VR128:$src),
1515 "movmskpd\t{$src, $dst|$dst, $src}", [], SSEPackedDouble>, OpSize,
1517 def VMOVMSKPSYr64r : PI<0x50, MRMSrcReg, (outs GR64:$dst), (ins VR256:$src),
1518 "movmskps\t{$src, $dst|$dst, $src}", [], SSEPackedSingle>, VEX;
1519 def VMOVMSKPDYr64r : PI<0x50, MRMSrcReg, (outs GR64:$dst), (ins VR256:$src),
1520 "movmskpd\t{$src, $dst|$dst, $src}", [], SSEPackedDouble>, OpSize,
1524 //===----------------------------------------------------------------------===//
1525 // SSE 1 & 2 - Misc aliasing of packed SSE 1 & 2 instructions
1526 //===----------------------------------------------------------------------===//
1528 // Aliases of packed SSE1 & SSE2 instructions for scalar use. These all have
1529 // names that start with 'Fs'.
1531 // Alias instructions that map fld0 to pxor for sse.
1532 let isReMaterializable = 1, isAsCheapAsAMove = 1, isCodeGenOnly = 1,
1533 canFoldAsLoad = 1 in {
1534 // FIXME: Set encoding to pseudo!
1535 def FsFLD0SS : I<0xEF, MRMInitReg, (outs FR32:$dst), (ins), "",
1536 [(set FR32:$dst, fp32imm0)]>,
1537 Requires<[HasSSE1]>, TB, OpSize;
1538 def FsFLD0SD : I<0xEF, MRMInitReg, (outs FR64:$dst), (ins), "",
1539 [(set FR64:$dst, fpimm0)]>,
1540 Requires<[HasSSE2]>, TB, OpSize;
1541 def VFsFLD0SS : I<0xEF, MRMInitReg, (outs FR32:$dst), (ins), "",
1542 [(set FR32:$dst, fp32imm0)]>,
1543 Requires<[HasAVX]>, TB, OpSize, VEX_4V;
1544 def VFsFLD0SD : I<0xEF, MRMInitReg, (outs FR64:$dst), (ins), "",
1545 [(set FR64:$dst, fpimm0)]>,
1546 Requires<[HasAVX]>, TB, OpSize, VEX_4V;
1549 // Alias instruction to do FR32 or FR64 reg-to-reg copy using movaps. Upper
1550 // bits are disregarded.
1551 let neverHasSideEffects = 1 in {
1552 def FsMOVAPSrr : PSI<0x28, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
1553 "movaps\t{$src, $dst|$dst, $src}", []>;
1554 def FsMOVAPDrr : PDI<0x28, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1555 "movapd\t{$src, $dst|$dst, $src}", []>;
1558 // Alias instruction to load FR32 or FR64 from f128mem using movaps. Upper
1559 // bits are disregarded.
1560 let canFoldAsLoad = 1, isReMaterializable = 1 in {
1561 def FsMOVAPSrm : PSI<0x28, MRMSrcMem, (outs FR32:$dst), (ins f128mem:$src),
1562 "movaps\t{$src, $dst|$dst, $src}",
1563 [(set FR32:$dst, (alignedloadfsf32 addr:$src))]>;
1564 def FsMOVAPDrm : PDI<0x28, MRMSrcMem, (outs FR64:$dst), (ins f128mem:$src),
1565 "movapd\t{$src, $dst|$dst, $src}",
1566 [(set FR64:$dst, (alignedloadfsf64 addr:$src))]>;
1569 //===----------------------------------------------------------------------===//
1570 // SSE 1 & 2 - Logical Instructions
1571 //===----------------------------------------------------------------------===//
1573 /// sse12_fp_alias_pack_logical - SSE 1 & 2 aliased packed FP logical ops
1575 multiclass sse12_fp_alias_pack_logical<bits<8> opc, string OpcodeStr,
1577 defm V#NAME#PS : sse12_fp_packed<opc, !strconcat(OpcodeStr, "ps"), OpNode,
1578 FR32, f32, f128mem, memopfsf32, SSEPackedSingle, 0>, VEX_4V;
1580 defm V#NAME#PD : sse12_fp_packed<opc, !strconcat(OpcodeStr, "pd"), OpNode,
1581 FR64, f64, f128mem, memopfsf64, SSEPackedDouble, 0>, OpSize, VEX_4V;
1583 let Constraints = "$src1 = $dst" in {
1584 defm PS : sse12_fp_packed<opc, !strconcat(OpcodeStr, "ps"), OpNode, FR32,
1585 f32, f128mem, memopfsf32, SSEPackedSingle>, TB;
1587 defm PD : sse12_fp_packed<opc, !strconcat(OpcodeStr, "pd"), OpNode, FR64,
1588 f64, f128mem, memopfsf64, SSEPackedDouble>, TB, OpSize;
1592 // Alias bitwise logical operations using SSE logical ops on packed FP values.
1593 let mayLoad = 0 in {
1594 defm FsAND : sse12_fp_alias_pack_logical<0x54, "and", X86fand>;
1595 defm FsOR : sse12_fp_alias_pack_logical<0x56, "or", X86for>;
1596 defm FsXOR : sse12_fp_alias_pack_logical<0x57, "xor", X86fxor>;
1599 let neverHasSideEffects = 1, Pattern = []<dag>, isCommutable = 0 in
1600 defm FsANDN : sse12_fp_alias_pack_logical<0x55, "andn", undef>;
1602 /// sse12_fp_packed_logical - SSE 1 & 2 packed FP logical ops
1604 multiclass sse12_fp_packed_logical<bits<8> opc, string OpcodeStr,
1606 let Pattern = []<dag> in {
1607 defm V#NAME#PS : sse12_fp_packed_logical_rm<opc, VR128, SSEPackedSingle,
1608 !strconcat(OpcodeStr, "ps"), f128mem,
1609 [(set VR128:$dst, (v2i64 (OpNode VR128:$src1, VR128:$src2)))],
1610 [(set VR128:$dst, (OpNode (bc_v2i64 (v4f32 VR128:$src1)),
1611 (memopv2i64 addr:$src2)))], 0>, VEX_4V;
1613 defm V#NAME#PD : sse12_fp_packed_logical_rm<opc, VR128, SSEPackedDouble,
1614 !strconcat(OpcodeStr, "pd"), f128mem,
1615 [(set VR128:$dst, (OpNode (bc_v2i64 (v2f64 VR128:$src1)),
1616 (bc_v2i64 (v2f64 VR128:$src2))))],
1617 [(set VR128:$dst, (OpNode (bc_v2i64 (v2f64 VR128:$src1)),
1618 (memopv2i64 addr:$src2)))], 0>,
1621 let Constraints = "$src1 = $dst" in {
1622 defm PS : sse12_fp_packed_logical_rm<opc, VR128, SSEPackedSingle,
1623 !strconcat(OpcodeStr, "ps"), f128mem,
1624 [(set VR128:$dst, (v2i64 (OpNode VR128:$src1, VR128:$src2)))],
1625 [(set VR128:$dst, (OpNode (bc_v2i64 (v4f32 VR128:$src1)),
1626 (memopv2i64 addr:$src2)))]>, TB;
1628 defm PD : sse12_fp_packed_logical_rm<opc, VR128, SSEPackedDouble,
1629 !strconcat(OpcodeStr, "pd"), f128mem,
1630 [(set VR128:$dst, (OpNode (bc_v2i64 (v2f64 VR128:$src1)),
1631 (bc_v2i64 (v2f64 VR128:$src2))))],
1632 [(set VR128:$dst, (OpNode (bc_v2i64 (v2f64 VR128:$src1)),
1633 (memopv2i64 addr:$src2)))]>, TB, OpSize;
1637 /// sse12_fp_packed_logical_y - AVX 256-bit SSE 1 & 2 logical ops forms
1639 multiclass sse12_fp_packed_logical_y<bits<8> opc, string OpcodeStr,
1641 defm PSY : sse12_fp_packed_logical_rm<opc, VR256, SSEPackedSingle,
1642 !strconcat(OpcodeStr, "ps"), f256mem,
1643 [(set VR256:$dst, (v4i64 (OpNode VR256:$src1, VR256:$src2)))],
1644 [(set VR256:$dst, (OpNode (bc_v4i64 (v8f32 VR256:$src1)),
1645 (memopv4i64 addr:$src2)))], 0>, VEX_4V;
1647 defm PDY : sse12_fp_packed_logical_rm<opc, VR256, SSEPackedDouble,
1648 !strconcat(OpcodeStr, "pd"), f256mem,
1649 [(set VR256:$dst, (OpNode (bc_v4i64 (v4f64 VR256:$src1)),
1650 (bc_v4i64 (v4f64 VR256:$src2))))],
1651 [(set VR256:$dst, (OpNode (bc_v4i64 (v4f64 VR256:$src1)),
1652 (memopv4i64 addr:$src2)))], 0>,
1656 // AVX 256-bit packed logical ops forms
1657 defm VAND : sse12_fp_packed_logical_y<0x54, "and", and>;
1658 defm VOR : sse12_fp_packed_logical_y<0x56, "or", or>;
1659 defm VXOR : sse12_fp_packed_logical_y<0x57, "xor", xor>;
1660 defm VANDN : sse12_fp_packed_logical_y<0x55, "andn", X86andnp>;
1662 defm AND : sse12_fp_packed_logical<0x54, "and", and>;
1663 defm OR : sse12_fp_packed_logical<0x56, "or", or>;
1664 defm XOR : sse12_fp_packed_logical<0x57, "xor", xor>;
1665 let isCommutable = 0 in
1666 defm ANDN : sse12_fp_packed_logical<0x55, "andn", X86andnp>;
1668 //===----------------------------------------------------------------------===//
1669 // SSE 1 & 2 - Arithmetic Instructions
1670 //===----------------------------------------------------------------------===//
1672 /// basic_sse12_fp_binop_xxx - SSE 1 & 2 binops come in both scalar and
1675 /// In addition, we also have a special variant of the scalar form here to
1676 /// represent the associated intrinsic operation. This form is unlike the
1677 /// plain scalar form, in that it takes an entire vector (instead of a scalar)
1678 /// and leaves the top elements unmodified (therefore these cannot be commuted).
1680 /// These three forms can each be reg+reg or reg+mem.
1683 /// FIXME: once all 256-bit intrinsics are matched, cleanup and refactor those
1685 multiclass basic_sse12_fp_binop_s<bits<8> opc, string OpcodeStr, SDNode OpNode,
1687 defm SS : sse12_fp_scalar<opc, !strconcat(OpcodeStr, "ss"),
1688 OpNode, FR32, f32mem, Is2Addr>, XS;
1689 defm SD : sse12_fp_scalar<opc, !strconcat(OpcodeStr, "sd"),
1690 OpNode, FR64, f64mem, Is2Addr>, XD;
1693 multiclass basic_sse12_fp_binop_p<bits<8> opc, string OpcodeStr, SDNode OpNode,
1695 let mayLoad = 0 in {
1696 defm PS : sse12_fp_packed<opc, !strconcat(OpcodeStr, "ps"), OpNode, VR128,
1697 v4f32, f128mem, memopv4f32, SSEPackedSingle, Is2Addr>, TB;
1698 defm PD : sse12_fp_packed<opc, !strconcat(OpcodeStr, "pd"), OpNode, VR128,
1699 v2f64, f128mem, memopv2f64, SSEPackedDouble, Is2Addr>, TB, OpSize;
1703 multiclass basic_sse12_fp_binop_p_y<bits<8> opc, string OpcodeStr,
1705 let mayLoad = 0 in {
1706 defm PSY : sse12_fp_packed<opc, !strconcat(OpcodeStr, "ps"), OpNode, VR256,
1707 v8f32, f256mem, memopv8f32, SSEPackedSingle, 0>, TB;
1708 defm PDY : sse12_fp_packed<opc, !strconcat(OpcodeStr, "pd"), OpNode, VR256,
1709 v4f64, f256mem, memopv4f64, SSEPackedDouble, 0>, TB, OpSize;
1713 multiclass basic_sse12_fp_binop_s_int<bits<8> opc, string OpcodeStr,
1715 defm SS : sse12_fp_scalar_int<opc, OpcodeStr, VR128,
1716 !strconcat(OpcodeStr, "ss"), "", "_ss", ssmem, sse_load_f32, Is2Addr>, XS;
1717 defm SD : sse12_fp_scalar_int<opc, OpcodeStr, VR128,
1718 !strconcat(OpcodeStr, "sd"), "2", "_sd", sdmem, sse_load_f64, Is2Addr>, XD;
1721 multiclass basic_sse12_fp_binop_p_int<bits<8> opc, string OpcodeStr,
1723 defm PS : sse12_fp_packed_int<opc, OpcodeStr, VR128,
1724 !strconcat(OpcodeStr, "ps"), "sse", "_ps", f128mem, memopv4f32,
1725 SSEPackedSingle, Is2Addr>, TB;
1727 defm PD : sse12_fp_packed_int<opc, OpcodeStr, VR128,
1728 !strconcat(OpcodeStr, "pd"), "sse2", "_pd", f128mem, memopv2f64,
1729 SSEPackedDouble, Is2Addr>, TB, OpSize;
1732 multiclass basic_sse12_fp_binop_p_y_int<bits<8> opc, string OpcodeStr> {
1733 defm PSY : sse12_fp_packed_int<opc, OpcodeStr, VR256,
1734 !strconcat(OpcodeStr, "ps"), "avx", "_ps_256", f256mem, memopv8f32,
1735 SSEPackedSingle, 0>, TB;
1737 defm PDY : sse12_fp_packed_int<opc, OpcodeStr, VR256,
1738 !strconcat(OpcodeStr, "pd"), "avx", "_pd_256", f256mem, memopv4f64,
1739 SSEPackedDouble, 0>, TB, OpSize;
1742 // Binary Arithmetic instructions
1743 defm VADD : basic_sse12_fp_binop_s<0x58, "add", fadd, 0>,
1744 basic_sse12_fp_binop_s_int<0x58, "add", 0>,
1745 basic_sse12_fp_binop_p<0x58, "add", fadd, 0>,
1746 basic_sse12_fp_binop_p_y<0x58, "add", fadd>, VEX_4V;
1747 defm VMUL : basic_sse12_fp_binop_s<0x59, "mul", fmul, 0>,
1748 basic_sse12_fp_binop_s_int<0x59, "mul", 0>,
1749 basic_sse12_fp_binop_p<0x59, "mul", fmul, 0>,
1750 basic_sse12_fp_binop_p_y<0x59, "mul", fmul>, VEX_4V;
1752 let isCommutable = 0 in {
1753 defm VSUB : basic_sse12_fp_binop_s<0x5C, "sub", fsub, 0>,
1754 basic_sse12_fp_binop_s_int<0x5C, "sub", 0>,
1755 basic_sse12_fp_binop_p<0x5C, "sub", fsub, 0>,
1756 basic_sse12_fp_binop_p_y<0x5C, "sub", fsub>, VEX_4V;
1757 defm VDIV : basic_sse12_fp_binop_s<0x5E, "div", fdiv, 0>,
1758 basic_sse12_fp_binop_s_int<0x5E, "div", 0>,
1759 basic_sse12_fp_binop_p<0x5E, "div", fdiv, 0>,
1760 basic_sse12_fp_binop_p_y<0x5E, "div", fdiv>, VEX_4V;
1761 defm VMAX : basic_sse12_fp_binop_s<0x5F, "max", X86fmax, 0>,
1762 basic_sse12_fp_binop_s_int<0x5F, "max", 0>,
1763 basic_sse12_fp_binop_p<0x5F, "max", X86fmax, 0>,
1764 basic_sse12_fp_binop_p_int<0x5F, "max", 0>,
1765 basic_sse12_fp_binop_p_y<0x5F, "max", X86fmax>,
1766 basic_sse12_fp_binop_p_y_int<0x5F, "max">, VEX_4V;
1767 defm VMIN : basic_sse12_fp_binop_s<0x5D, "min", X86fmin, 0>,
1768 basic_sse12_fp_binop_s_int<0x5D, "min", 0>,
1769 basic_sse12_fp_binop_p<0x5D, "min", X86fmin, 0>,
1770 basic_sse12_fp_binop_p_int<0x5D, "min", 0>,
1771 basic_sse12_fp_binop_p_y_int<0x5D, "min">,
1772 basic_sse12_fp_binop_p_y<0x5D, "min", X86fmin>, VEX_4V;
1775 let Constraints = "$src1 = $dst" in {
1776 defm ADD : basic_sse12_fp_binop_s<0x58, "add", fadd>,
1777 basic_sse12_fp_binop_p<0x58, "add", fadd>,
1778 basic_sse12_fp_binop_s_int<0x58, "add">;
1779 defm MUL : basic_sse12_fp_binop_s<0x59, "mul", fmul>,
1780 basic_sse12_fp_binop_p<0x59, "mul", fmul>,
1781 basic_sse12_fp_binop_s_int<0x59, "mul">;
1783 let isCommutable = 0 in {
1784 defm SUB : basic_sse12_fp_binop_s<0x5C, "sub", fsub>,
1785 basic_sse12_fp_binop_p<0x5C, "sub", fsub>,
1786 basic_sse12_fp_binop_s_int<0x5C, "sub">;
1787 defm DIV : basic_sse12_fp_binop_s<0x5E, "div", fdiv>,
1788 basic_sse12_fp_binop_p<0x5E, "div", fdiv>,
1789 basic_sse12_fp_binop_s_int<0x5E, "div">;
1790 defm MAX : basic_sse12_fp_binop_s<0x5F, "max", X86fmax>,
1791 basic_sse12_fp_binop_p<0x5F, "max", X86fmax>,
1792 basic_sse12_fp_binop_s_int<0x5F, "max">,
1793 basic_sse12_fp_binop_p_int<0x5F, "max">;
1794 defm MIN : basic_sse12_fp_binop_s<0x5D, "min", X86fmin>,
1795 basic_sse12_fp_binop_p<0x5D, "min", X86fmin>,
1796 basic_sse12_fp_binop_s_int<0x5D, "min">,
1797 basic_sse12_fp_binop_p_int<0x5D, "min">;
1802 /// In addition, we also have a special variant of the scalar form here to
1803 /// represent the associated intrinsic operation. This form is unlike the
1804 /// plain scalar form, in that it takes an entire vector (instead of a
1805 /// scalar) and leaves the top elements undefined.
1807 /// And, we have a special variant form for a full-vector intrinsic form.
1809 /// sse1_fp_unop_s - SSE1 unops in scalar form.
1810 multiclass sse1_fp_unop_s<bits<8> opc, string OpcodeStr,
1811 SDNode OpNode, Intrinsic F32Int> {
1812 def SSr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
1813 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
1814 [(set FR32:$dst, (OpNode FR32:$src))]>;
1815 // For scalar unary operations, fold a load into the operation
1816 // only in OptForSize mode. It eliminates an instruction, but it also
1817 // eliminates a whole-register clobber (the load), so it introduces a
1818 // partial register update condition.
1819 def SSm : I<opc, MRMSrcMem, (outs FR32:$dst), (ins f32mem:$src),
1820 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
1821 [(set FR32:$dst, (OpNode (load addr:$src)))]>, XS,
1822 Requires<[HasSSE1, OptForSize]>;
1823 def SSr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1824 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
1825 [(set VR128:$dst, (F32Int VR128:$src))]>;
1826 def SSm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst), (ins ssmem:$src),
1827 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
1828 [(set VR128:$dst, (F32Int sse_load_f32:$src))]>;
1831 /// sse1_fp_unop_s_avx - AVX SSE1 unops in scalar form.
1832 multiclass sse1_fp_unop_s_avx<bits<8> opc, string OpcodeStr,
1833 SDNode OpNode, Intrinsic F32Int> {
1834 def SSr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
1835 !strconcat(OpcodeStr,
1836 "ss\t{$src2, $src1, $dst|$dst, $src1, $src2}"), []>;
1837 def SSm : I<opc, MRMSrcMem, (outs FR32:$dst), (ins FR32:$src1, f32mem:$src2),
1838 !strconcat(OpcodeStr,
1839 "ss\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
1840 []>, XS, Requires<[HasAVX, OptForSize]>;
1841 def SSr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1842 !strconcat(OpcodeStr,
1843 "ss\t{$src, $dst, $dst|$dst, $dst, $src}"),
1844 [(set VR128:$dst, (F32Int VR128:$src))]>;
1845 def SSm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst), (ins ssmem:$src),
1846 !strconcat(OpcodeStr,
1847 "ss\t{$src, $dst, $dst|$dst, $dst, $src}"),
1848 [(set VR128:$dst, (F32Int sse_load_f32:$src))]>;
1851 /// sse1_fp_unop_p - SSE1 unops in packed form.
1852 multiclass sse1_fp_unop_p<bits<8> opc, string OpcodeStr, SDNode OpNode> {
1853 def PSr : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1854 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
1855 [(set VR128:$dst, (v4f32 (OpNode VR128:$src)))]>;
1856 def PSm : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1857 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
1858 [(set VR128:$dst, (OpNode (memopv4f32 addr:$src)))]>;
1861 /// sse1_fp_unop_p_y - AVX 256-bit SSE1 unops in packed form.
1862 multiclass sse1_fp_unop_p_y<bits<8> opc, string OpcodeStr, SDNode OpNode> {
1863 def PSYr : PSI<opc, MRMSrcReg, (outs VR256:$dst), (ins VR256:$src),
1864 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
1865 [(set VR256:$dst, (v8f32 (OpNode VR256:$src)))]>;
1866 def PSYm : PSI<opc, MRMSrcMem, (outs VR256:$dst), (ins f256mem:$src),
1867 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
1868 [(set VR256:$dst, (OpNode (memopv8f32 addr:$src)))]>;
1871 /// sse1_fp_unop_p_int - SSE1 intrinsics unops in packed forms.
1872 multiclass sse1_fp_unop_p_int<bits<8> opc, string OpcodeStr,
1873 Intrinsic V4F32Int> {
1874 def PSr_Int : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1875 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
1876 [(set VR128:$dst, (V4F32Int VR128:$src))]>;
1877 def PSm_Int : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1878 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
1879 [(set VR128:$dst, (V4F32Int (memopv4f32 addr:$src)))]>;
1882 /// sse1_fp_unop_p_y_int - AVX 256-bit intrinsics unops in packed forms.
1883 multiclass sse1_fp_unop_p_y_int<bits<8> opc, string OpcodeStr,
1884 Intrinsic V4F32Int> {
1885 def PSYr_Int : PSI<opc, MRMSrcReg, (outs VR256:$dst), (ins VR256:$src),
1886 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
1887 [(set VR256:$dst, (V4F32Int VR256:$src))]>;
1888 def PSYm_Int : PSI<opc, MRMSrcMem, (outs VR256:$dst), (ins f256mem:$src),
1889 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
1890 [(set VR256:$dst, (V4F32Int (memopv8f32 addr:$src)))]>;
1893 /// sse2_fp_unop_s - SSE2 unops in scalar form.
1894 multiclass sse2_fp_unop_s<bits<8> opc, string OpcodeStr,
1895 SDNode OpNode, Intrinsic F64Int> {
1896 def SDr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1897 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1898 [(set FR64:$dst, (OpNode FR64:$src))]>;
1899 // See the comments in sse1_fp_unop_s for why this is OptForSize.
1900 def SDm : I<opc, MRMSrcMem, (outs FR64:$dst), (ins f64mem:$src),
1901 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1902 [(set FR64:$dst, (OpNode (load addr:$src)))]>, XD,
1903 Requires<[HasSSE2, OptForSize]>;
1904 def SDr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1905 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1906 [(set VR128:$dst, (F64Int VR128:$src))]>;
1907 def SDm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst), (ins sdmem:$src),
1908 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1909 [(set VR128:$dst, (F64Int sse_load_f64:$src))]>;
1912 /// sse2_fp_unop_s_avx - AVX SSE2 unops in scalar form.
1913 multiclass sse2_fp_unop_s_avx<bits<8> opc, string OpcodeStr,
1914 SDNode OpNode, Intrinsic F64Int> {
1915 def SDr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
1916 !strconcat(OpcodeStr,
1917 "sd\t{$src2, $src1, $dst|$dst, $src1, $src2}"), []>;
1918 def SDm : SDI<opc, MRMSrcMem, (outs FR64:$dst),
1919 (ins FR64:$src1, f64mem:$src2),
1920 !strconcat(OpcodeStr,
1921 "sd\t{$src2, $src1, $dst|$dst, $src1, $src2}"), []>;
1922 def SDr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1923 !strconcat(OpcodeStr, "sd\t{$src, $dst, $dst|$dst, $dst, $src}"),
1924 [(set VR128:$dst, (F64Int VR128:$src))]>;
1925 def SDm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst), (ins sdmem:$src),
1926 !strconcat(OpcodeStr, "sd\t{$src, $dst, $dst|$dst, $dst, $src}"),
1927 [(set VR128:$dst, (F64Int sse_load_f64:$src))]>;
1930 /// sse2_fp_unop_p - SSE2 unops in vector forms.
1931 multiclass sse2_fp_unop_p<bits<8> opc, string OpcodeStr,
1933 def PDr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1934 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1935 [(set VR128:$dst, (v2f64 (OpNode VR128:$src)))]>;
1936 def PDm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1937 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1938 [(set VR128:$dst, (OpNode (memopv2f64 addr:$src)))]>;
1941 /// sse2_fp_unop_p_y - AVX SSE2 256-bit unops in vector forms.
1942 multiclass sse2_fp_unop_p_y<bits<8> opc, string OpcodeStr, SDNode OpNode> {
1943 def PDYr : PDI<opc, MRMSrcReg, (outs VR256:$dst), (ins VR256:$src),
1944 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1945 [(set VR256:$dst, (v4f64 (OpNode VR256:$src)))]>;
1946 def PDYm : PDI<opc, MRMSrcMem, (outs VR256:$dst), (ins f256mem:$src),
1947 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1948 [(set VR256:$dst, (OpNode (memopv4f64 addr:$src)))]>;
1951 /// sse2_fp_unop_p_int - SSE2 intrinsic unops in vector forms.
1952 multiclass sse2_fp_unop_p_int<bits<8> opc, string OpcodeStr,
1953 Intrinsic V2F64Int> {
1954 def PDr_Int : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1955 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1956 [(set VR128:$dst, (V2F64Int VR128:$src))]>;
1957 def PDm_Int : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1958 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1959 [(set VR128:$dst, (V2F64Int (memopv2f64 addr:$src)))]>;
1962 /// sse2_fp_unop_p_y_int - AVX 256-bit intrinsic unops in vector forms.
1963 multiclass sse2_fp_unop_p_y_int<bits<8> opc, string OpcodeStr,
1964 Intrinsic V2F64Int> {
1965 def PDYr_Int : PDI<opc, MRMSrcReg, (outs VR256:$dst), (ins VR256:$src),
1966 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1967 [(set VR256:$dst, (V2F64Int VR256:$src))]>;
1968 def PDYm_Int : PDI<opc, MRMSrcMem, (outs VR256:$dst), (ins f256mem:$src),
1969 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1970 [(set VR256:$dst, (V2F64Int (memopv4f64 addr:$src)))]>;
1973 let Predicates = [HasAVX] in {
1975 defm VSQRT : sse1_fp_unop_s_avx<0x51, "vsqrt", fsqrt, int_x86_sse_sqrt_ss>,
1976 sse2_fp_unop_s_avx<0x51, "vsqrt", fsqrt, int_x86_sse2_sqrt_sd>,
1979 defm VSQRT : sse1_fp_unop_p<0x51, "vsqrt", fsqrt>,
1980 sse2_fp_unop_p<0x51, "vsqrt", fsqrt>,
1981 sse1_fp_unop_p_y<0x51, "vsqrt", fsqrt>,
1982 sse2_fp_unop_p_y<0x51, "vsqrt", fsqrt>,
1983 sse1_fp_unop_p_int<0x51, "vsqrt", int_x86_sse_sqrt_ps>,
1984 sse2_fp_unop_p_int<0x51, "vsqrt", int_x86_sse2_sqrt_pd>,
1985 sse1_fp_unop_p_y_int<0x51, "vsqrt", int_x86_avx_sqrt_ps_256>,
1986 sse2_fp_unop_p_y_int<0x51, "vsqrt", int_x86_avx_sqrt_pd_256>,
1989 // Reciprocal approximations. Note that these typically require refinement
1990 // in order to obtain suitable precision.
1991 defm VRSQRT : sse1_fp_unop_s_avx<0x52, "vrsqrt", X86frsqrt,
1992 int_x86_sse_rsqrt_ss>, VEX_4V;
1993 defm VRSQRT : sse1_fp_unop_p<0x52, "vrsqrt", X86frsqrt>,
1994 sse1_fp_unop_p_y<0x52, "vrsqrt", X86frsqrt>,
1995 sse1_fp_unop_p_y_int<0x52, "vrsqrt", int_x86_avx_rsqrt_ps_256>,
1996 sse1_fp_unop_p_int<0x52, "vrsqrt", int_x86_sse_rsqrt_ps>, VEX;
1998 defm VRCP : sse1_fp_unop_s_avx<0x53, "vrcp", X86frcp, int_x86_sse_rcp_ss>,
2000 defm VRCP : sse1_fp_unop_p<0x53, "vrcp", X86frcp>,
2001 sse1_fp_unop_p_y<0x53, "vrcp", X86frcp>,
2002 sse1_fp_unop_p_y_int<0x53, "vrcp", int_x86_avx_rcp_ps_256>,
2003 sse1_fp_unop_p_int<0x53, "vrcp", int_x86_sse_rcp_ps>, VEX;
2006 def : Pat<(f32 (fsqrt FR32:$src)),
2007 (VSQRTSSr (f32 (IMPLICIT_DEF)), FR32:$src)>, Requires<[HasAVX]>;
2008 def : Pat<(f64 (fsqrt FR64:$src)),
2009 (VSQRTSDr (f64 (IMPLICIT_DEF)), FR64:$src)>, Requires<[HasAVX]>;
2010 def : Pat<(f64 (fsqrt (load addr:$src))),
2011 (VSQRTSDm (f64 (IMPLICIT_DEF)), addr:$src)>,
2012 Requires<[HasAVX, OptForSize]>;
2013 def : Pat<(f32 (fsqrt (load addr:$src))),
2014 (VSQRTSSm (f32 (IMPLICIT_DEF)), addr:$src)>,
2015 Requires<[HasAVX, OptForSize]>;
2018 defm SQRT : sse1_fp_unop_s<0x51, "sqrt", fsqrt, int_x86_sse_sqrt_ss>,
2019 sse1_fp_unop_p<0x51, "sqrt", fsqrt>,
2020 sse1_fp_unop_p_int<0x51, "sqrt", int_x86_sse_sqrt_ps>,
2021 sse2_fp_unop_s<0x51, "sqrt", fsqrt, int_x86_sse2_sqrt_sd>,
2022 sse2_fp_unop_p<0x51, "sqrt", fsqrt>,
2023 sse2_fp_unop_p_int<0x51, "sqrt", int_x86_sse2_sqrt_pd>;
2025 // Reciprocal approximations. Note that these typically require refinement
2026 // in order to obtain suitable precision.
2027 defm RSQRT : sse1_fp_unop_s<0x52, "rsqrt", X86frsqrt, int_x86_sse_rsqrt_ss>,
2028 sse1_fp_unop_p<0x52, "rsqrt", X86frsqrt>,
2029 sse1_fp_unop_p_int<0x52, "rsqrt", int_x86_sse_rsqrt_ps>;
2030 defm RCP : sse1_fp_unop_s<0x53, "rcp", X86frcp, int_x86_sse_rcp_ss>,
2031 sse1_fp_unop_p<0x53, "rcp", X86frcp>,
2032 sse1_fp_unop_p_int<0x53, "rcp", int_x86_sse_rcp_ps>;
2034 // There is no f64 version of the reciprocal approximation instructions.
2036 //===----------------------------------------------------------------------===//
2037 // SSE 1 & 2 - Non-temporal stores
2038 //===----------------------------------------------------------------------===//
2040 let AddedComplexity = 400 in { // Prefer non-temporal versions
2041 def VMOVNTPSmr : VPSI<0x2B, MRMDestMem, (outs),
2042 (ins f128mem:$dst, VR128:$src),
2043 "movntps\t{$src, $dst|$dst, $src}",
2044 [(alignednontemporalstore (v4f32 VR128:$src),
2046 def VMOVNTPDmr : VPDI<0x2B, MRMDestMem, (outs),
2047 (ins f128mem:$dst, VR128:$src),
2048 "movntpd\t{$src, $dst|$dst, $src}",
2049 [(alignednontemporalstore (v2f64 VR128:$src),
2051 def VMOVNTDQ_64mr : VPDI<0xE7, MRMDestMem, (outs),
2052 (ins f128mem:$dst, VR128:$src),
2053 "movntdq\t{$src, $dst|$dst, $src}",
2054 [(alignednontemporalstore (v2f64 VR128:$src),
2057 let ExeDomain = SSEPackedInt in
2058 def VMOVNTDQmr : VPDI<0xE7, MRMDestMem, (outs),
2059 (ins f128mem:$dst, VR128:$src),
2060 "movntdq\t{$src, $dst|$dst, $src}",
2061 [(alignednontemporalstore (v4f32 VR128:$src),
2064 def : Pat<(alignednontemporalstore (v2i64 VR128:$src), addr:$dst),
2065 (VMOVNTDQmr addr:$dst, VR128:$src)>, Requires<[HasAVX]>;
2067 def VMOVNTPSYmr : VPSI<0x2B, MRMDestMem, (outs),
2068 (ins f256mem:$dst, VR256:$src),
2069 "movntps\t{$src, $dst|$dst, $src}",
2070 [(alignednontemporalstore (v8f32 VR256:$src),
2072 def VMOVNTPDYmr : VPDI<0x2B, MRMDestMem, (outs),
2073 (ins f256mem:$dst, VR256:$src),
2074 "movntpd\t{$src, $dst|$dst, $src}",
2075 [(alignednontemporalstore (v4f64 VR256:$src),
2077 def VMOVNTDQY_64mr : VPDI<0xE7, MRMDestMem, (outs),
2078 (ins f256mem:$dst, VR256:$src),
2079 "movntdq\t{$src, $dst|$dst, $src}",
2080 [(alignednontemporalstore (v4f64 VR256:$src),
2082 let ExeDomain = SSEPackedInt in
2083 def VMOVNTDQYmr : VPDI<0xE7, MRMDestMem, (outs),
2084 (ins f256mem:$dst, VR256:$src),
2085 "movntdq\t{$src, $dst|$dst, $src}",
2086 [(alignednontemporalstore (v8f32 VR256:$src),
2090 def : Pat<(int_x86_avx_movnt_dq_256 addr:$dst, VR256:$src),
2091 (VMOVNTDQYmr addr:$dst, VR256:$src)>;
2092 def : Pat<(int_x86_avx_movnt_pd_256 addr:$dst, VR256:$src),
2093 (VMOVNTPDYmr addr:$dst, VR256:$src)>;
2094 def : Pat<(int_x86_avx_movnt_ps_256 addr:$dst, VR256:$src),
2095 (VMOVNTPSYmr addr:$dst, VR256:$src)>;
2097 let AddedComplexity = 400 in { // Prefer non-temporal versions
2098 def MOVNTPSmr : PSI<0x2B, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
2099 "movntps\t{$src, $dst|$dst, $src}",
2100 [(alignednontemporalstore (v4f32 VR128:$src), addr:$dst)]>;
2101 def MOVNTPDmr : PDI<0x2B, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
2102 "movntpd\t{$src, $dst|$dst, $src}",
2103 [(alignednontemporalstore(v2f64 VR128:$src), addr:$dst)]>;
2105 def MOVNTDQ_64mr : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
2106 "movntdq\t{$src, $dst|$dst, $src}",
2107 [(alignednontemporalstore (v2f64 VR128:$src), addr:$dst)]>;
2109 let ExeDomain = SSEPackedInt in
2110 def MOVNTDQmr : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
2111 "movntdq\t{$src, $dst|$dst, $src}",
2112 [(alignednontemporalstore (v4f32 VR128:$src), addr:$dst)]>;
2114 def : Pat<(alignednontemporalstore (v2i64 VR128:$src), addr:$dst),
2115 (MOVNTDQmr addr:$dst, VR128:$src)>;
2117 // There is no AVX form for instructions below this point
2118 def MOVNTImr : I<0xC3, MRMDestMem, (outs), (ins i32mem:$dst, GR32:$src),
2119 "movnti{l}\t{$src, $dst|$dst, $src}",
2120 [(nontemporalstore (i32 GR32:$src), addr:$dst)]>,
2121 TB, Requires<[HasSSE2]>;
2122 def MOVNTI_64mr : RI<0xC3, MRMDestMem, (outs), (ins i64mem:$dst, GR64:$src),
2123 "movnti{q}\t{$src, $dst|$dst, $src}",
2124 [(nontemporalstore (i64 GR64:$src), addr:$dst)]>,
2125 TB, Requires<[HasSSE2]>;
2128 //===----------------------------------------------------------------------===//
2129 // SSE 1 & 2 - Misc Instructions (No AVX form)
2130 //===----------------------------------------------------------------------===//
2132 // Prefetch intrinsic.
2133 def PREFETCHT0 : PSI<0x18, MRM1m, (outs), (ins i8mem:$src),
2134 "prefetcht0\t$src", [(prefetch addr:$src, imm, (i32 3), (i32 1))]>;
2135 def PREFETCHT1 : PSI<0x18, MRM2m, (outs), (ins i8mem:$src),
2136 "prefetcht1\t$src", [(prefetch addr:$src, imm, (i32 2), (i32 1))]>;
2137 def PREFETCHT2 : PSI<0x18, MRM3m, (outs), (ins i8mem:$src),
2138 "prefetcht2\t$src", [(prefetch addr:$src, imm, (i32 1), (i32 1))]>;
2139 def PREFETCHNTA : PSI<0x18, MRM0m, (outs), (ins i8mem:$src),
2140 "prefetchnta\t$src", [(prefetch addr:$src, imm, (i32 0), (i32 1))]>;
2142 // Load, store, and memory fence
2143 def SFENCE : I<0xAE, MRM_F8, (outs), (ins), "sfence", [(int_x86_sse_sfence)]>,
2144 TB, Requires<[HasSSE1]>;
2145 def : Pat<(X86SFence), (SFENCE)>;
2147 // Alias instructions that map zero vector to pxor / xorp* for sse.
2148 // We set canFoldAsLoad because this can be converted to a constant-pool
2149 // load of an all-zeros value if folding it would be beneficial.
2150 // FIXME: Change encoding to pseudo! This is blocked right now by the x86
2151 // JIT implementation, it does not expand the instructions below like
2152 // X86MCInstLower does.
2153 let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
2154 isCodeGenOnly = 1 in {
2155 def V_SET0PS : PSI<0x57, MRMInitReg, (outs VR128:$dst), (ins), "",
2156 [(set VR128:$dst, (v4f32 immAllZerosV))]>;
2157 def V_SET0PD : PDI<0x57, MRMInitReg, (outs VR128:$dst), (ins), "",
2158 [(set VR128:$dst, (v2f64 immAllZerosV))]>;
2159 let ExeDomain = SSEPackedInt in
2160 def V_SET0PI : PDI<0xEF, MRMInitReg, (outs VR128:$dst), (ins), "",
2161 [(set VR128:$dst, (v4i32 immAllZerosV))]>;
2164 // The same as done above but for AVX. The 128-bit versions are the
2165 // same, but re-encoded. The 256-bit does not support PI version, and
2166 // doesn't need it because on sandy bridge the register is set to zero
2167 // at the rename stage without using any execution unit, so SET0PSY
2168 // and SET0PDY can be used for vector int instructions without penalty
2169 // FIXME: Change encoding to pseudo! This is blocked right now by the x86
2170 // JIT implementatioan, it does not expand the instructions below like
2171 // X86MCInstLower does.
2172 let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
2173 isCodeGenOnly = 1, Predicates = [HasAVX] in {
2174 def AVX_SET0PS : PSI<0x57, MRMInitReg, (outs VR128:$dst), (ins), "",
2175 [(set VR128:$dst, (v4f32 immAllZerosV))]>, VEX_4V;
2176 def AVX_SET0PD : PDI<0x57, MRMInitReg, (outs VR128:$dst), (ins), "",
2177 [(set VR128:$dst, (v2f64 immAllZerosV))]>, VEX_4V;
2178 def AVX_SET0PSY : PSI<0x57, MRMInitReg, (outs VR256:$dst), (ins), "",
2179 [(set VR256:$dst, (v8f32 immAllZerosV))]>, VEX_4V;
2180 def AVX_SET0PDY : PDI<0x57, MRMInitReg, (outs VR256:$dst), (ins), "",
2181 [(set VR256:$dst, (v4f64 immAllZerosV))]>, VEX_4V;
2182 let ExeDomain = SSEPackedInt in
2183 def AVX_SET0PI : PDI<0xEF, MRMInitReg, (outs VR128:$dst), (ins), "",
2184 [(set VR128:$dst, (v4i32 immAllZerosV))]>;
2187 def : Pat<(v2i64 immAllZerosV), (V_SET0PI)>;
2188 def : Pat<(v8i16 immAllZerosV), (V_SET0PI)>;
2189 def : Pat<(v16i8 immAllZerosV), (V_SET0PI)>;
2191 def : Pat<(f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
2192 (f32 (EXTRACT_SUBREG (v4f32 VR128:$src), sub_ss))>;
2194 // AVX has no support for 256-bit integer instructions, but since the 128-bit
2195 // VPXOR instruction writes zero to its upper part, it's safe build zeros.
2196 def : Pat<(v8i32 immAllZerosV), (SUBREG_TO_REG (i32 0), (AVX_SET0PI), sub_xmm)>;
2197 def : Pat<(bc_v8i32 (v8f32 immAllZerosV)),
2198 (SUBREG_TO_REG (i32 0), (AVX_SET0PI), sub_xmm)>;
2200 def : Pat<(v4i64 immAllZerosV), (SUBREG_TO_REG (i64 0), (AVX_SET0PI), sub_xmm)>;
2201 def : Pat<(bc_v4i64 (v8f32 immAllZerosV)),
2202 (SUBREG_TO_REG (i64 0), (AVX_SET0PI), sub_xmm)>;
2204 //===----------------------------------------------------------------------===//
2205 // SSE 1 & 2 - Load/Store XCSR register
2206 //===----------------------------------------------------------------------===//
2208 def VLDMXCSR : VPSI<0xAE, MRM2m, (outs), (ins i32mem:$src),
2209 "ldmxcsr\t$src", [(int_x86_sse_ldmxcsr addr:$src)]>, VEX;
2210 def VSTMXCSR : VPSI<0xAE, MRM3m, (outs), (ins i32mem:$dst),
2211 "stmxcsr\t$dst", [(int_x86_sse_stmxcsr addr:$dst)]>, VEX;
2213 def LDMXCSR : PSI<0xAE, MRM2m, (outs), (ins i32mem:$src),
2214 "ldmxcsr\t$src", [(int_x86_sse_ldmxcsr addr:$src)]>;
2215 def STMXCSR : PSI<0xAE, MRM3m, (outs), (ins i32mem:$dst),
2216 "stmxcsr\t$dst", [(int_x86_sse_stmxcsr addr:$dst)]>;
2218 //===---------------------------------------------------------------------===//
2219 // SSE2 - Move Aligned/Unaligned Packed Integer Instructions
2220 //===---------------------------------------------------------------------===//
2222 let ExeDomain = SSEPackedInt in { // SSE integer instructions
2224 let neverHasSideEffects = 1 in {
2225 def VMOVDQArr : VPDI<0x6F, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2226 "movdqa\t{$src, $dst|$dst, $src}", []>, VEX;
2227 def VMOVDQAYrr : VPDI<0x6F, MRMSrcReg, (outs VR256:$dst), (ins VR256:$src),
2228 "movdqa\t{$src, $dst|$dst, $src}", []>, VEX;
2230 def VMOVDQUrr : VPDI<0x6F, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2231 "movdqu\t{$src, $dst|$dst, $src}", []>, XS, VEX;
2232 def VMOVDQUYrr : VPDI<0x6F, MRMSrcReg, (outs VR256:$dst), (ins VR256:$src),
2233 "movdqu\t{$src, $dst|$dst, $src}", []>, XS, VEX;
2235 let canFoldAsLoad = 1, mayLoad = 1 in {
2236 def VMOVDQArm : VPDI<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
2237 "movdqa\t{$src, $dst|$dst, $src}", []>, VEX;
2238 def VMOVDQAYrm : VPDI<0x6F, MRMSrcMem, (outs VR256:$dst), (ins i256mem:$src),
2239 "movdqa\t{$src, $dst|$dst, $src}", []>, VEX;
2240 let Predicates = [HasAVX] in {
2241 def VMOVDQUrm : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
2242 "vmovdqu\t{$src, $dst|$dst, $src}",[]>, XS, VEX;
2243 def VMOVDQUYrm : I<0x6F, MRMSrcMem, (outs VR256:$dst), (ins i256mem:$src),
2244 "vmovdqu\t{$src, $dst|$dst, $src}",[]>, XS, VEX;
2248 let mayStore = 1 in {
2249 def VMOVDQAmr : VPDI<0x7F, MRMDestMem, (outs),
2250 (ins i128mem:$dst, VR128:$src),
2251 "movdqa\t{$src, $dst|$dst, $src}", []>, VEX;
2252 def VMOVDQAYmr : VPDI<0x7F, MRMDestMem, (outs),
2253 (ins i256mem:$dst, VR256:$src),
2254 "movdqa\t{$src, $dst|$dst, $src}", []>, VEX;
2255 let Predicates = [HasAVX] in {
2256 def VMOVDQUmr : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
2257 "vmovdqu\t{$src, $dst|$dst, $src}",[]>, XS, VEX;
2258 def VMOVDQUYmr : I<0x7F, MRMDestMem, (outs), (ins i256mem:$dst, VR256:$src),
2259 "vmovdqu\t{$src, $dst|$dst, $src}",[]>, XS, VEX;
2263 let neverHasSideEffects = 1 in
2264 def MOVDQArr : PDI<0x6F, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2265 "movdqa\t{$src, $dst|$dst, $src}", []>;
2267 def MOVDQUrr : I<0x6F, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2268 "movdqu\t{$src, $dst|$dst, $src}",
2269 []>, XS, Requires<[HasSSE2]>;
2271 let canFoldAsLoad = 1, mayLoad = 1 in {
2272 def MOVDQArm : PDI<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
2273 "movdqa\t{$src, $dst|$dst, $src}",
2274 [/*(set VR128:$dst, (alignedloadv2i64 addr:$src))*/]>;
2275 def MOVDQUrm : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
2276 "movdqu\t{$src, $dst|$dst, $src}",
2277 [/*(set VR128:$dst, (loadv2i64 addr:$src))*/]>,
2278 XS, Requires<[HasSSE2]>;
2281 let mayStore = 1 in {
2282 def MOVDQAmr : PDI<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
2283 "movdqa\t{$src, $dst|$dst, $src}",
2284 [/*(alignedstore (v2i64 VR128:$src), addr:$dst)*/]>;
2285 def MOVDQUmr : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
2286 "movdqu\t{$src, $dst|$dst, $src}",
2287 [/*(store (v2i64 VR128:$src), addr:$dst)*/]>,
2288 XS, Requires<[HasSSE2]>;
2291 // Intrinsic forms of MOVDQU load and store
2292 def VMOVDQUmr_Int : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
2293 "vmovdqu\t{$src, $dst|$dst, $src}",
2294 [(int_x86_sse2_storeu_dq addr:$dst, VR128:$src)]>,
2295 XS, VEX, Requires<[HasAVX]>;
2297 def MOVDQUmr_Int : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
2298 "movdqu\t{$src, $dst|$dst, $src}",
2299 [(int_x86_sse2_storeu_dq addr:$dst, VR128:$src)]>,
2300 XS, Requires<[HasSSE2]>;
2302 } // ExeDomain = SSEPackedInt
2304 def : Pat<(int_x86_avx_loadu_dq_256 addr:$src), (VMOVDQUYrm addr:$src)>;
2305 def : Pat<(int_x86_avx_storeu_dq_256 addr:$dst, VR256:$src),
2306 (VMOVDQUYmr addr:$dst, VR256:$src)>;
2308 //===---------------------------------------------------------------------===//
2309 // SSE2 - Packed Integer Arithmetic Instructions
2310 //===---------------------------------------------------------------------===//
2312 let ExeDomain = SSEPackedInt in { // SSE integer instructions
2314 multiclass PDI_binop_rm_int<bits<8> opc, string OpcodeStr, Intrinsic IntId,
2315 bit IsCommutable = 0, bit Is2Addr = 1> {
2316 let isCommutable = IsCommutable in
2317 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
2318 (ins VR128:$src1, VR128:$src2),
2320 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2321 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
2322 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]>;
2323 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
2324 (ins VR128:$src1, i128mem:$src2),
2326 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2327 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
2328 [(set VR128:$dst, (IntId VR128:$src1,
2329 (bitconvert (memopv2i64 addr:$src2))))]>;
2332 multiclass PDI_binop_rmi_int<bits<8> opc, bits<8> opc2, Format ImmForm,
2333 string OpcodeStr, Intrinsic IntId,
2334 Intrinsic IntId2, bit Is2Addr = 1> {
2335 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
2336 (ins VR128:$src1, VR128:$src2),
2338 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2339 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
2340 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]>;
2341 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
2342 (ins VR128:$src1, i128mem:$src2),
2344 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2345 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
2346 [(set VR128:$dst, (IntId VR128:$src1,
2347 (bitconvert (memopv2i64 addr:$src2))))]>;
2348 def ri : PDIi8<opc2, ImmForm, (outs VR128:$dst),
2349 (ins VR128:$src1, i32i8imm:$src2),
2351 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2352 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
2353 [(set VR128:$dst, (IntId2 VR128:$src1, (i32 imm:$src2)))]>;
2356 /// PDI_binop_rm - Simple SSE2 binary operator.
2357 multiclass PDI_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
2358 ValueType OpVT, bit IsCommutable = 0, bit Is2Addr = 1> {
2359 let isCommutable = IsCommutable in
2360 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
2361 (ins VR128:$src1, VR128:$src2),
2363 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2364 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
2365 [(set VR128:$dst, (OpVT (OpNode VR128:$src1, VR128:$src2)))]>;
2366 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
2367 (ins VR128:$src1, i128mem:$src2),
2369 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2370 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
2371 [(set VR128:$dst, (OpVT (OpNode VR128:$src1,
2372 (bitconvert (memopv2i64 addr:$src2)))))]>;
2375 /// PDI_binop_rm_v2i64 - Simple SSE2 binary operator whose type is v2i64.
2377 /// FIXME: we could eliminate this and use PDI_binop_rm instead if tblgen knew
2378 /// to collapse (bitconvert VT to VT) into its operand.
2380 multiclass PDI_binop_rm_v2i64<bits<8> opc, string OpcodeStr, SDNode OpNode,
2381 bit IsCommutable = 0, bit Is2Addr = 1> {
2382 let isCommutable = IsCommutable in
2383 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
2384 (ins VR128:$src1, VR128:$src2),
2386 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2387 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
2388 [(set VR128:$dst, (v2i64 (OpNode VR128:$src1, VR128:$src2)))]>;
2389 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
2390 (ins VR128:$src1, i128mem:$src2),
2392 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2393 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
2394 [(set VR128:$dst, (OpNode VR128:$src1, (memopv2i64 addr:$src2)))]>;
2397 } // ExeDomain = SSEPackedInt
2399 // 128-bit Integer Arithmetic
2401 let Predicates = [HasAVX] in {
2402 defm VPADDB : PDI_binop_rm<0xFC, "vpaddb", add, v16i8, 1, 0 /*3addr*/>, VEX_4V;
2403 defm VPADDW : PDI_binop_rm<0xFD, "vpaddw", add, v8i16, 1, 0>, VEX_4V;
2404 defm VPADDD : PDI_binop_rm<0xFE, "vpaddd", add, v4i32, 1, 0>, VEX_4V;
2405 defm VPADDQ : PDI_binop_rm_v2i64<0xD4, "vpaddq", add, 1, 0>, VEX_4V;
2406 defm VPMULLW : PDI_binop_rm<0xD5, "vpmullw", mul, v8i16, 1, 0>, VEX_4V;
2407 defm VPSUBB : PDI_binop_rm<0xF8, "vpsubb", sub, v16i8, 0, 0>, VEX_4V;
2408 defm VPSUBW : PDI_binop_rm<0xF9, "vpsubw", sub, v8i16, 0, 0>, VEX_4V;
2409 defm VPSUBD : PDI_binop_rm<0xFA, "vpsubd", sub, v4i32, 0, 0>, VEX_4V;
2410 defm VPSUBQ : PDI_binop_rm_v2i64<0xFB, "vpsubq", sub, 0, 0>, VEX_4V;
2413 defm VPSUBSB : PDI_binop_rm_int<0xE8, "vpsubsb" , int_x86_sse2_psubs_b, 0, 0>,
2415 defm VPSUBSW : PDI_binop_rm_int<0xE9, "vpsubsw" , int_x86_sse2_psubs_w, 0, 0>,
2417 defm VPSUBUSB : PDI_binop_rm_int<0xD8, "vpsubusb", int_x86_sse2_psubus_b, 0, 0>,
2419 defm VPSUBUSW : PDI_binop_rm_int<0xD9, "vpsubusw", int_x86_sse2_psubus_w, 0, 0>,
2421 defm VPADDSB : PDI_binop_rm_int<0xEC, "vpaddsb" , int_x86_sse2_padds_b, 1, 0>,
2423 defm VPADDSW : PDI_binop_rm_int<0xED, "vpaddsw" , int_x86_sse2_padds_w, 1, 0>,
2425 defm VPADDUSB : PDI_binop_rm_int<0xDC, "vpaddusb", int_x86_sse2_paddus_b, 1, 0>,
2427 defm VPADDUSW : PDI_binop_rm_int<0xDD, "vpaddusw", int_x86_sse2_paddus_w, 1, 0>,
2429 defm VPMULHUW : PDI_binop_rm_int<0xE4, "vpmulhuw", int_x86_sse2_pmulhu_w, 1, 0>,
2431 defm VPMULHW : PDI_binop_rm_int<0xE5, "vpmulhw" , int_x86_sse2_pmulh_w, 1, 0>,
2433 defm VPMULUDQ : PDI_binop_rm_int<0xF4, "vpmuludq", int_x86_sse2_pmulu_dq, 1, 0>,
2435 defm VPMADDWD : PDI_binop_rm_int<0xF5, "vpmaddwd", int_x86_sse2_pmadd_wd, 1, 0>,
2437 defm VPAVGB : PDI_binop_rm_int<0xE0, "vpavgb", int_x86_sse2_pavg_b, 1, 0>,
2439 defm VPAVGW : PDI_binop_rm_int<0xE3, "vpavgw", int_x86_sse2_pavg_w, 1, 0>,
2441 defm VPMINUB : PDI_binop_rm_int<0xDA, "vpminub", int_x86_sse2_pminu_b, 1, 0>,
2443 defm VPMINSW : PDI_binop_rm_int<0xEA, "vpminsw", int_x86_sse2_pmins_w, 1, 0>,
2445 defm VPMAXUB : PDI_binop_rm_int<0xDE, "vpmaxub", int_x86_sse2_pmaxu_b, 1, 0>,
2447 defm VPMAXSW : PDI_binop_rm_int<0xEE, "vpmaxsw", int_x86_sse2_pmaxs_w, 1, 0>,
2449 defm VPSADBW : PDI_binop_rm_int<0xF6, "vpsadbw", int_x86_sse2_psad_bw, 1, 0>,
2453 let Constraints = "$src1 = $dst" in {
2454 defm PADDB : PDI_binop_rm<0xFC, "paddb", add, v16i8, 1>;
2455 defm PADDW : PDI_binop_rm<0xFD, "paddw", add, v8i16, 1>;
2456 defm PADDD : PDI_binop_rm<0xFE, "paddd", add, v4i32, 1>;
2457 defm PADDQ : PDI_binop_rm_v2i64<0xD4, "paddq", add, 1>;
2458 defm PMULLW : PDI_binop_rm<0xD5, "pmullw", mul, v8i16, 1>;
2459 defm PSUBB : PDI_binop_rm<0xF8, "psubb", sub, v16i8>;
2460 defm PSUBW : PDI_binop_rm<0xF9, "psubw", sub, v8i16>;
2461 defm PSUBD : PDI_binop_rm<0xFA, "psubd", sub, v4i32>;
2462 defm PSUBQ : PDI_binop_rm_v2i64<0xFB, "psubq", sub>;
2465 defm PSUBSB : PDI_binop_rm_int<0xE8, "psubsb" , int_x86_sse2_psubs_b>;
2466 defm PSUBSW : PDI_binop_rm_int<0xE9, "psubsw" , int_x86_sse2_psubs_w>;
2467 defm PSUBUSB : PDI_binop_rm_int<0xD8, "psubusb", int_x86_sse2_psubus_b>;
2468 defm PSUBUSW : PDI_binop_rm_int<0xD9, "psubusw", int_x86_sse2_psubus_w>;
2469 defm PADDSB : PDI_binop_rm_int<0xEC, "paddsb" , int_x86_sse2_padds_b, 1>;
2470 defm PADDSW : PDI_binop_rm_int<0xED, "paddsw" , int_x86_sse2_padds_w, 1>;
2471 defm PADDUSB : PDI_binop_rm_int<0xDC, "paddusb", int_x86_sse2_paddus_b, 1>;
2472 defm PADDUSW : PDI_binop_rm_int<0xDD, "paddusw", int_x86_sse2_paddus_w, 1>;
2473 defm PMULHUW : PDI_binop_rm_int<0xE4, "pmulhuw", int_x86_sse2_pmulhu_w, 1>;
2474 defm PMULHW : PDI_binop_rm_int<0xE5, "pmulhw" , int_x86_sse2_pmulh_w, 1>;
2475 defm PMULUDQ : PDI_binop_rm_int<0xF4, "pmuludq", int_x86_sse2_pmulu_dq, 1>;
2476 defm PMADDWD : PDI_binop_rm_int<0xF5, "pmaddwd", int_x86_sse2_pmadd_wd, 1>;
2477 defm PAVGB : PDI_binop_rm_int<0xE0, "pavgb", int_x86_sse2_pavg_b, 1>;
2478 defm PAVGW : PDI_binop_rm_int<0xE3, "pavgw", int_x86_sse2_pavg_w, 1>;
2479 defm PMINUB : PDI_binop_rm_int<0xDA, "pminub", int_x86_sse2_pminu_b, 1>;
2480 defm PMINSW : PDI_binop_rm_int<0xEA, "pminsw", int_x86_sse2_pmins_w, 1>;
2481 defm PMAXUB : PDI_binop_rm_int<0xDE, "pmaxub", int_x86_sse2_pmaxu_b, 1>;
2482 defm PMAXSW : PDI_binop_rm_int<0xEE, "pmaxsw", int_x86_sse2_pmaxs_w, 1>;
2483 defm PSADBW : PDI_binop_rm_int<0xF6, "psadbw", int_x86_sse2_psad_bw, 1>;
2485 } // Constraints = "$src1 = $dst"
2487 //===---------------------------------------------------------------------===//
2488 // SSE2 - Packed Integer Logical Instructions
2489 //===---------------------------------------------------------------------===//
2491 let Predicates = [HasAVX] in {
2492 defm VPSLLW : PDI_binop_rmi_int<0xF1, 0x71, MRM6r, "vpsllw",
2493 int_x86_sse2_psll_w, int_x86_sse2_pslli_w, 0>,
2495 defm VPSLLD : PDI_binop_rmi_int<0xF2, 0x72, MRM6r, "vpslld",
2496 int_x86_sse2_psll_d, int_x86_sse2_pslli_d, 0>,
2498 defm VPSLLQ : PDI_binop_rmi_int<0xF3, 0x73, MRM6r, "vpsllq",
2499 int_x86_sse2_psll_q, int_x86_sse2_pslli_q, 0>,
2502 defm VPSRLW : PDI_binop_rmi_int<0xD1, 0x71, MRM2r, "vpsrlw",
2503 int_x86_sse2_psrl_w, int_x86_sse2_psrli_w, 0>,
2505 defm VPSRLD : PDI_binop_rmi_int<0xD2, 0x72, MRM2r, "vpsrld",
2506 int_x86_sse2_psrl_d, int_x86_sse2_psrli_d, 0>,
2508 defm VPSRLQ : PDI_binop_rmi_int<0xD3, 0x73, MRM2r, "vpsrlq",
2509 int_x86_sse2_psrl_q, int_x86_sse2_psrli_q, 0>,
2512 defm VPSRAW : PDI_binop_rmi_int<0xE1, 0x71, MRM4r, "vpsraw",
2513 int_x86_sse2_psra_w, int_x86_sse2_psrai_w, 0>,
2515 defm VPSRAD : PDI_binop_rmi_int<0xE2, 0x72, MRM4r, "vpsrad",
2516 int_x86_sse2_psra_d, int_x86_sse2_psrai_d, 0>,
2519 defm VPAND : PDI_binop_rm_v2i64<0xDB, "vpand", and, 1, 0>, VEX_4V;
2520 defm VPOR : PDI_binop_rm_v2i64<0xEB, "vpor" , or, 1, 0>, VEX_4V;
2521 defm VPXOR : PDI_binop_rm_v2i64<0xEF, "vpxor", xor, 1, 0>, VEX_4V;
2523 let ExeDomain = SSEPackedInt in {
2524 let neverHasSideEffects = 1 in {
2525 // 128-bit logical shifts.
2526 def VPSLLDQri : PDIi8<0x73, MRM7r,
2527 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
2528 "vpslldq\t{$src2, $src1, $dst|$dst, $src1, $src2}", []>,
2530 def VPSRLDQri : PDIi8<0x73, MRM3r,
2531 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
2532 "vpsrldq\t{$src2, $src1, $dst|$dst, $src1, $src2}", []>,
2534 // PSRADQri doesn't exist in SSE[1-3].
2536 def VPANDNrr : PDI<0xDF, MRMSrcReg,
2537 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2538 "vpandn\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2539 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
2540 VR128:$src2)))]>, VEX_4V;
2542 def VPANDNrm : PDI<0xDF, MRMSrcMem,
2543 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2544 "vpandn\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2545 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
2546 (memopv2i64 addr:$src2))))]>,
2551 let Constraints = "$src1 = $dst" in {
2552 defm PSLLW : PDI_binop_rmi_int<0xF1, 0x71, MRM6r, "psllw",
2553 int_x86_sse2_psll_w, int_x86_sse2_pslli_w>;
2554 defm PSLLD : PDI_binop_rmi_int<0xF2, 0x72, MRM6r, "pslld",
2555 int_x86_sse2_psll_d, int_x86_sse2_pslli_d>;
2556 defm PSLLQ : PDI_binop_rmi_int<0xF3, 0x73, MRM6r, "psllq",
2557 int_x86_sse2_psll_q, int_x86_sse2_pslli_q>;
2559 defm PSRLW : PDI_binop_rmi_int<0xD1, 0x71, MRM2r, "psrlw",
2560 int_x86_sse2_psrl_w, int_x86_sse2_psrli_w>;
2561 defm PSRLD : PDI_binop_rmi_int<0xD2, 0x72, MRM2r, "psrld",
2562 int_x86_sse2_psrl_d, int_x86_sse2_psrli_d>;
2563 defm PSRLQ : PDI_binop_rmi_int<0xD3, 0x73, MRM2r, "psrlq",
2564 int_x86_sse2_psrl_q, int_x86_sse2_psrli_q>;
2566 defm PSRAW : PDI_binop_rmi_int<0xE1, 0x71, MRM4r, "psraw",
2567 int_x86_sse2_psra_w, int_x86_sse2_psrai_w>;
2568 defm PSRAD : PDI_binop_rmi_int<0xE2, 0x72, MRM4r, "psrad",
2569 int_x86_sse2_psra_d, int_x86_sse2_psrai_d>;
2571 defm PAND : PDI_binop_rm_v2i64<0xDB, "pand", and, 1>;
2572 defm POR : PDI_binop_rm_v2i64<0xEB, "por" , or, 1>;
2573 defm PXOR : PDI_binop_rm_v2i64<0xEF, "pxor", xor, 1>;
2575 let ExeDomain = SSEPackedInt in {
2576 let neverHasSideEffects = 1 in {
2577 // 128-bit logical shifts.
2578 def PSLLDQri : PDIi8<0x73, MRM7r,
2579 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
2580 "pslldq\t{$src2, $dst|$dst, $src2}", []>;
2581 def PSRLDQri : PDIi8<0x73, MRM3r,
2582 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
2583 "psrldq\t{$src2, $dst|$dst, $src2}", []>;
2584 // PSRADQri doesn't exist in SSE[1-3].
2586 def PANDNrr : PDI<0xDF, MRMSrcReg,
2587 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2588 "pandn\t{$src2, $dst|$dst, $src2}", []>;
2590 def PANDNrm : PDI<0xDF, MRMSrcMem,
2591 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2592 "pandn\t{$src2, $dst|$dst, $src2}", []>;
2594 } // Constraints = "$src1 = $dst"
2596 let Predicates = [HasAVX] in {
2597 def : Pat<(int_x86_sse2_psll_dq VR128:$src1, imm:$src2),
2598 (v2i64 (VPSLLDQri VR128:$src1, (BYTE_imm imm:$src2)))>;
2599 def : Pat<(int_x86_sse2_psrl_dq VR128:$src1, imm:$src2),
2600 (v2i64 (VPSRLDQri VR128:$src1, (BYTE_imm imm:$src2)))>;
2601 def : Pat<(int_x86_sse2_psll_dq_bs VR128:$src1, imm:$src2),
2602 (v2i64 (VPSLLDQri VR128:$src1, imm:$src2))>;
2603 def : Pat<(int_x86_sse2_psrl_dq_bs VR128:$src1, imm:$src2),
2604 (v2i64 (VPSRLDQri VR128:$src1, imm:$src2))>;
2605 def : Pat<(v2f64 (X86fsrl VR128:$src1, i32immSExt8:$src2)),
2606 (v2f64 (VPSRLDQri VR128:$src1, (BYTE_imm imm:$src2)))>;
2608 // Shift up / down and insert zero's.
2609 def : Pat<(v2i64 (X86vshl VR128:$src, (i8 imm:$amt))),
2610 (v2i64 (VPSLLDQri VR128:$src, (BYTE_imm imm:$amt)))>;
2611 def : Pat<(v2i64 (X86vshr VR128:$src, (i8 imm:$amt))),
2612 (v2i64 (VPSRLDQri VR128:$src, (BYTE_imm imm:$amt)))>;
2615 let Predicates = [HasSSE2] in {
2616 def : Pat<(int_x86_sse2_psll_dq VR128:$src1, imm:$src2),
2617 (v2i64 (PSLLDQri VR128:$src1, (BYTE_imm imm:$src2)))>;
2618 def : Pat<(int_x86_sse2_psrl_dq VR128:$src1, imm:$src2),
2619 (v2i64 (PSRLDQri VR128:$src1, (BYTE_imm imm:$src2)))>;
2620 def : Pat<(int_x86_sse2_psll_dq_bs VR128:$src1, imm:$src2),
2621 (v2i64 (PSLLDQri VR128:$src1, imm:$src2))>;
2622 def : Pat<(int_x86_sse2_psrl_dq_bs VR128:$src1, imm:$src2),
2623 (v2i64 (PSRLDQri VR128:$src1, imm:$src2))>;
2624 def : Pat<(v2f64 (X86fsrl VR128:$src1, i32immSExt8:$src2)),
2625 (v2f64 (PSRLDQri VR128:$src1, (BYTE_imm imm:$src2)))>;
2627 // Shift up / down and insert zero's.
2628 def : Pat<(v2i64 (X86vshl VR128:$src, (i8 imm:$amt))),
2629 (v2i64 (PSLLDQri VR128:$src, (BYTE_imm imm:$amt)))>;
2630 def : Pat<(v2i64 (X86vshr VR128:$src, (i8 imm:$amt))),
2631 (v2i64 (PSRLDQri VR128:$src, (BYTE_imm imm:$amt)))>;
2634 //===---------------------------------------------------------------------===//
2635 // SSE2 - Packed Integer Comparison Instructions
2636 //===---------------------------------------------------------------------===//
2638 let Predicates = [HasAVX] in {
2639 defm VPCMPEQB : PDI_binop_rm_int<0x74, "vpcmpeqb", int_x86_sse2_pcmpeq_b, 1,
2641 defm VPCMPEQW : PDI_binop_rm_int<0x75, "vpcmpeqw", int_x86_sse2_pcmpeq_w, 1,
2643 defm VPCMPEQD : PDI_binop_rm_int<0x76, "vpcmpeqd", int_x86_sse2_pcmpeq_d, 1,
2645 defm VPCMPGTB : PDI_binop_rm_int<0x64, "vpcmpgtb", int_x86_sse2_pcmpgt_b, 0,
2647 defm VPCMPGTW : PDI_binop_rm_int<0x65, "vpcmpgtw", int_x86_sse2_pcmpgt_w, 0,
2649 defm VPCMPGTD : PDI_binop_rm_int<0x66, "vpcmpgtd", int_x86_sse2_pcmpgt_d, 0,
2653 let Constraints = "$src1 = $dst" in {
2654 defm PCMPEQB : PDI_binop_rm_int<0x74, "pcmpeqb", int_x86_sse2_pcmpeq_b, 1>;
2655 defm PCMPEQW : PDI_binop_rm_int<0x75, "pcmpeqw", int_x86_sse2_pcmpeq_w, 1>;
2656 defm PCMPEQD : PDI_binop_rm_int<0x76, "pcmpeqd", int_x86_sse2_pcmpeq_d, 1>;
2657 defm PCMPGTB : PDI_binop_rm_int<0x64, "pcmpgtb", int_x86_sse2_pcmpgt_b>;
2658 defm PCMPGTW : PDI_binop_rm_int<0x65, "pcmpgtw", int_x86_sse2_pcmpgt_w>;
2659 defm PCMPGTD : PDI_binop_rm_int<0x66, "pcmpgtd", int_x86_sse2_pcmpgt_d>;
2660 } // Constraints = "$src1 = $dst"
2662 def : Pat<(v16i8 (X86pcmpeqb VR128:$src1, VR128:$src2)),
2663 (PCMPEQBrr VR128:$src1, VR128:$src2)>;
2664 def : Pat<(v16i8 (X86pcmpeqb VR128:$src1, (memop addr:$src2))),
2665 (PCMPEQBrm VR128:$src1, addr:$src2)>;
2666 def : Pat<(v8i16 (X86pcmpeqw VR128:$src1, VR128:$src2)),
2667 (PCMPEQWrr VR128:$src1, VR128:$src2)>;
2668 def : Pat<(v8i16 (X86pcmpeqw VR128:$src1, (memop addr:$src2))),
2669 (PCMPEQWrm VR128:$src1, addr:$src2)>;
2670 def : Pat<(v4i32 (X86pcmpeqd VR128:$src1, VR128:$src2)),
2671 (PCMPEQDrr VR128:$src1, VR128:$src2)>;
2672 def : Pat<(v4i32 (X86pcmpeqd VR128:$src1, (memop addr:$src2))),
2673 (PCMPEQDrm VR128:$src1, addr:$src2)>;
2675 def : Pat<(v16i8 (X86pcmpgtb VR128:$src1, VR128:$src2)),
2676 (PCMPGTBrr VR128:$src1, VR128:$src2)>;
2677 def : Pat<(v16i8 (X86pcmpgtb VR128:$src1, (memop addr:$src2))),
2678 (PCMPGTBrm VR128:$src1, addr:$src2)>;
2679 def : Pat<(v8i16 (X86pcmpgtw VR128:$src1, VR128:$src2)),
2680 (PCMPGTWrr VR128:$src1, VR128:$src2)>;
2681 def : Pat<(v8i16 (X86pcmpgtw VR128:$src1, (memop addr:$src2))),
2682 (PCMPGTWrm VR128:$src1, addr:$src2)>;
2683 def : Pat<(v4i32 (X86pcmpgtd VR128:$src1, VR128:$src2)),
2684 (PCMPGTDrr VR128:$src1, VR128:$src2)>;
2685 def : Pat<(v4i32 (X86pcmpgtd VR128:$src1, (memop addr:$src2))),
2686 (PCMPGTDrm VR128:$src1, addr:$src2)>;
2688 //===---------------------------------------------------------------------===//
2689 // SSE2 - Packed Integer Pack Instructions
2690 //===---------------------------------------------------------------------===//
2692 let Predicates = [HasAVX] in {
2693 defm VPACKSSWB : PDI_binop_rm_int<0x63, "vpacksswb", int_x86_sse2_packsswb_128,
2695 defm VPACKSSDW : PDI_binop_rm_int<0x6B, "vpackssdw", int_x86_sse2_packssdw_128,
2697 defm VPACKUSWB : PDI_binop_rm_int<0x67, "vpackuswb", int_x86_sse2_packuswb_128,
2701 let Constraints = "$src1 = $dst" in {
2702 defm PACKSSWB : PDI_binop_rm_int<0x63, "packsswb", int_x86_sse2_packsswb_128>;
2703 defm PACKSSDW : PDI_binop_rm_int<0x6B, "packssdw", int_x86_sse2_packssdw_128>;
2704 defm PACKUSWB : PDI_binop_rm_int<0x67, "packuswb", int_x86_sse2_packuswb_128>;
2705 } // Constraints = "$src1 = $dst"
2707 //===---------------------------------------------------------------------===//
2708 // SSE2 - Packed Integer Shuffle Instructions
2709 //===---------------------------------------------------------------------===//
2711 let ExeDomain = SSEPackedInt in {
2712 multiclass sse2_pshuffle<string OpcodeStr, ValueType vt, PatFrag pshuf_frag,
2714 def ri : Ii8<0x70, MRMSrcReg,
2715 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
2716 !strconcat(OpcodeStr,
2717 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
2718 [(set VR128:$dst, (vt (pshuf_frag:$src2 VR128:$src1,
2720 def mi : Ii8<0x70, MRMSrcMem,
2721 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
2722 !strconcat(OpcodeStr,
2723 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
2724 [(set VR128:$dst, (vt (pshuf_frag:$src2
2725 (bc_frag (memopv2i64 addr:$src1)),
2728 } // ExeDomain = SSEPackedInt
2730 let Predicates = [HasAVX] in {
2731 let AddedComplexity = 5 in
2732 defm VPSHUFD : sse2_pshuffle<"vpshufd", v4i32, pshufd, bc_v4i32>, OpSize,
2735 // SSE2 with ImmT == Imm8 and XS prefix.
2736 defm VPSHUFHW : sse2_pshuffle<"vpshufhw", v8i16, pshufhw, bc_v8i16>, XS,
2739 // SSE2 with ImmT == Imm8 and XD prefix.
2740 defm VPSHUFLW : sse2_pshuffle<"vpshuflw", v8i16, pshuflw, bc_v8i16>, XD,
2744 let Predicates = [HasSSE2] in {
2745 let AddedComplexity = 5 in
2746 defm PSHUFD : sse2_pshuffle<"pshufd", v4i32, pshufd, bc_v4i32>, TB, OpSize;
2748 // SSE2 with ImmT == Imm8 and XS prefix.
2749 defm PSHUFHW : sse2_pshuffle<"pshufhw", v8i16, pshufhw, bc_v8i16>, XS;
2751 // SSE2 with ImmT == Imm8 and XD prefix.
2752 defm PSHUFLW : sse2_pshuffle<"pshuflw", v8i16, pshuflw, bc_v8i16>, XD;
2755 //===---------------------------------------------------------------------===//
2756 // SSE2 - Packed Integer Unpack Instructions
2757 //===---------------------------------------------------------------------===//
2759 let ExeDomain = SSEPackedInt in {
2760 multiclass sse2_unpack<bits<8> opc, string OpcodeStr, ValueType vt,
2761 SDNode OpNode, PatFrag bc_frag, bit Is2Addr = 1> {
2762 def rr : PDI<opc, MRMSrcReg,
2763 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2765 !strconcat(OpcodeStr,"\t{$src2, $dst|$dst, $src2}"),
2766 !strconcat(OpcodeStr,"\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
2767 [(set VR128:$dst, (vt (OpNode VR128:$src1, VR128:$src2)))]>;
2768 def rm : PDI<opc, MRMSrcMem,
2769 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2771 !strconcat(OpcodeStr,"\t{$src2, $dst|$dst, $src2}"),
2772 !strconcat(OpcodeStr,"\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
2773 [(set VR128:$dst, (OpNode VR128:$src1,
2774 (bc_frag (memopv2i64
2778 let Predicates = [HasAVX] in {
2779 defm VPUNPCKLBW : sse2_unpack<0x60, "vpunpcklbw", v16i8, X86Punpcklbw,
2780 bc_v16i8, 0>, VEX_4V;
2781 defm VPUNPCKLWD : sse2_unpack<0x61, "vpunpcklwd", v8i16, X86Punpcklwd,
2782 bc_v8i16, 0>, VEX_4V;
2783 defm VPUNPCKLDQ : sse2_unpack<0x62, "vpunpckldq", v4i32, X86Punpckldq,
2784 bc_v4i32, 0>, VEX_4V;
2786 /// FIXME: we could eliminate this and use sse2_unpack instead if tblgen
2787 /// knew to collapse (bitconvert VT to VT) into its operand.
2788 def VPUNPCKLQDQrr : PDI<0x6C, MRMSrcReg,
2789 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2790 "vpunpcklqdq\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2791 [(set VR128:$dst, (v2i64 (X86Punpcklqdq VR128:$src1,
2792 VR128:$src2)))]>, VEX_4V;
2793 def VPUNPCKLQDQrm : PDI<0x6C, MRMSrcMem,
2794 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2795 "vpunpcklqdq\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2796 [(set VR128:$dst, (v2i64 (X86Punpcklqdq VR128:$src1,
2797 (memopv2i64 addr:$src2))))]>, VEX_4V;
2799 defm VPUNPCKHBW : sse2_unpack<0x68, "vpunpckhbw", v16i8, X86Punpckhbw,
2800 bc_v16i8, 0>, VEX_4V;
2801 defm VPUNPCKHWD : sse2_unpack<0x69, "vpunpckhwd", v8i16, X86Punpckhwd,
2802 bc_v8i16, 0>, VEX_4V;
2803 defm VPUNPCKHDQ : sse2_unpack<0x6A, "vpunpckhdq", v4i32, X86Punpckhdq,
2804 bc_v4i32, 0>, VEX_4V;
2806 /// FIXME: we could eliminate this and use sse2_unpack instead if tblgen
2807 /// knew to collapse (bitconvert VT to VT) into its operand.
2808 def VPUNPCKHQDQrr : PDI<0x6D, MRMSrcReg,
2809 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2810 "vpunpckhqdq\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2811 [(set VR128:$dst, (v2i64 (X86Punpckhqdq VR128:$src1,
2812 VR128:$src2)))]>, VEX_4V;
2813 def VPUNPCKHQDQrm : PDI<0x6D, MRMSrcMem,
2814 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2815 "vpunpckhqdq\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2816 [(set VR128:$dst, (v2i64 (X86Punpckhqdq VR128:$src1,
2817 (memopv2i64 addr:$src2))))]>, VEX_4V;
2820 let Constraints = "$src1 = $dst" in {
2821 defm PUNPCKLBW : sse2_unpack<0x60, "punpcklbw", v16i8, X86Punpcklbw, bc_v16i8>;
2822 defm PUNPCKLWD : sse2_unpack<0x61, "punpcklwd", v8i16, X86Punpcklwd, bc_v8i16>;
2823 defm PUNPCKLDQ : sse2_unpack<0x62, "punpckldq", v4i32, X86Punpckldq, bc_v4i32>;
2825 /// FIXME: we could eliminate this and use sse2_unpack instead if tblgen
2826 /// knew to collapse (bitconvert VT to VT) into its operand.
2827 def PUNPCKLQDQrr : PDI<0x6C, MRMSrcReg,
2828 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2829 "punpcklqdq\t{$src2, $dst|$dst, $src2}",
2831 (v2i64 (X86Punpcklqdq VR128:$src1, VR128:$src2)))]>;
2832 def PUNPCKLQDQrm : PDI<0x6C, MRMSrcMem,
2833 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2834 "punpcklqdq\t{$src2, $dst|$dst, $src2}",
2836 (v2i64 (X86Punpcklqdq VR128:$src1,
2837 (memopv2i64 addr:$src2))))]>;
2839 defm PUNPCKHBW : sse2_unpack<0x68, "punpckhbw", v16i8, X86Punpckhbw, bc_v16i8>;
2840 defm PUNPCKHWD : sse2_unpack<0x69, "punpckhwd", v8i16, X86Punpckhwd, bc_v8i16>;
2841 defm PUNPCKHDQ : sse2_unpack<0x6A, "punpckhdq", v4i32, X86Punpckhdq, bc_v4i32>;
2843 /// FIXME: we could eliminate this and use sse2_unpack instead if tblgen
2844 /// knew to collapse (bitconvert VT to VT) into its operand.
2845 def PUNPCKHQDQrr : PDI<0x6D, MRMSrcReg,
2846 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2847 "punpckhqdq\t{$src2, $dst|$dst, $src2}",
2849 (v2i64 (X86Punpckhqdq VR128:$src1, VR128:$src2)))]>;
2850 def PUNPCKHQDQrm : PDI<0x6D, MRMSrcMem,
2851 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2852 "punpckhqdq\t{$src2, $dst|$dst, $src2}",
2854 (v2i64 (X86Punpckhqdq VR128:$src1,
2855 (memopv2i64 addr:$src2))))]>;
2858 } // ExeDomain = SSEPackedInt
2860 //===---------------------------------------------------------------------===//
2861 // SSE2 - Packed Integer Extract and Insert
2862 //===---------------------------------------------------------------------===//
2864 let ExeDomain = SSEPackedInt in {
2865 multiclass sse2_pinsrw<bit Is2Addr = 1> {
2866 def rri : Ii8<0xC4, MRMSrcReg,
2867 (outs VR128:$dst), (ins VR128:$src1,
2868 GR32:$src2, i32i8imm:$src3),
2870 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2871 "vpinsrw\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
2873 (X86pinsrw VR128:$src1, GR32:$src2, imm:$src3))]>;
2874 def rmi : Ii8<0xC4, MRMSrcMem,
2875 (outs VR128:$dst), (ins VR128:$src1,
2876 i16mem:$src2, i32i8imm:$src3),
2878 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2879 "vpinsrw\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
2881 (X86pinsrw VR128:$src1, (extloadi16 addr:$src2),
2886 let Predicates = [HasAVX] in
2887 def VPEXTRWri : Ii8<0xC5, MRMSrcReg,
2888 (outs GR32:$dst), (ins VR128:$src1, i32i8imm:$src2),
2889 "vpextrw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2890 [(set GR32:$dst, (X86pextrw (v8i16 VR128:$src1),
2891 imm:$src2))]>, OpSize, VEX;
2892 def PEXTRWri : PDIi8<0xC5, MRMSrcReg,
2893 (outs GR32:$dst), (ins VR128:$src1, i32i8imm:$src2),
2894 "pextrw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2895 [(set GR32:$dst, (X86pextrw (v8i16 VR128:$src1),
2899 let Predicates = [HasAVX] in {
2900 defm VPINSRW : sse2_pinsrw<0>, OpSize, VEX_4V;
2901 def VPINSRWrr64i : Ii8<0xC4, MRMSrcReg, (outs VR128:$dst),
2902 (ins VR128:$src1, GR64:$src2, i32i8imm:$src3),
2903 "vpinsrw\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
2904 []>, OpSize, VEX_4V;
2907 let Constraints = "$src1 = $dst" in
2908 defm PINSRW : sse2_pinsrw, TB, OpSize, Requires<[HasSSE2]>;
2910 } // ExeDomain = SSEPackedInt
2912 //===---------------------------------------------------------------------===//
2913 // SSE2 - Packed Mask Creation
2914 //===---------------------------------------------------------------------===//
2916 let ExeDomain = SSEPackedInt in {
2918 def VPMOVMSKBrr : VPDI<0xD7, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
2919 "pmovmskb\t{$src, $dst|$dst, $src}",
2920 [(set GR32:$dst, (int_x86_sse2_pmovmskb_128 VR128:$src))]>, VEX;
2921 def VPMOVMSKBr64r : VPDI<0xD7, MRMSrcReg, (outs GR64:$dst), (ins VR128:$src),
2922 "pmovmskb\t{$src, $dst|$dst, $src}", []>, VEX;
2923 def PMOVMSKBrr : PDI<0xD7, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
2924 "pmovmskb\t{$src, $dst|$dst, $src}",
2925 [(set GR32:$dst, (int_x86_sse2_pmovmskb_128 VR128:$src))]>;
2927 } // ExeDomain = SSEPackedInt
2929 //===---------------------------------------------------------------------===//
2930 // SSE2 - Conditional Store
2931 //===---------------------------------------------------------------------===//
2933 let ExeDomain = SSEPackedInt in {
2936 def VMASKMOVDQU : VPDI<0xF7, MRMSrcReg, (outs),
2937 (ins VR128:$src, VR128:$mask),
2938 "maskmovdqu\t{$mask, $src|$src, $mask}",
2939 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, EDI)]>, VEX;
2941 def VMASKMOVDQU64 : VPDI<0xF7, MRMSrcReg, (outs),
2942 (ins VR128:$src, VR128:$mask),
2943 "maskmovdqu\t{$mask, $src|$src, $mask}",
2944 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, RDI)]>, VEX;
2947 def MASKMOVDQU : PDI<0xF7, MRMSrcReg, (outs), (ins VR128:$src, VR128:$mask),
2948 "maskmovdqu\t{$mask, $src|$src, $mask}",
2949 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, EDI)]>;
2951 def MASKMOVDQU64 : PDI<0xF7, MRMSrcReg, (outs), (ins VR128:$src, VR128:$mask),
2952 "maskmovdqu\t{$mask, $src|$src, $mask}",
2953 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, RDI)]>;
2955 } // ExeDomain = SSEPackedInt
2957 //===---------------------------------------------------------------------===//
2958 // SSE2 - Move Doubleword
2959 //===---------------------------------------------------------------------===//
2961 //===---------------------------------------------------------------------===//
2962 // Move Int Doubleword to Packed Double Int
2964 def VMOVDI2PDIrr : VPDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
2965 "movd\t{$src, $dst|$dst, $src}",
2967 (v4i32 (scalar_to_vector GR32:$src)))]>, VEX;
2968 def VMOVDI2PDIrm : VPDI<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
2969 "movd\t{$src, $dst|$dst, $src}",
2971 (v4i32 (scalar_to_vector (loadi32 addr:$src))))]>,
2973 def VMOV64toPQIrr : VRPDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR64:$src),
2974 "mov{d|q}\t{$src, $dst|$dst, $src}",
2976 (v2i64 (scalar_to_vector GR64:$src)))]>, VEX;
2977 def VMOV64toSDrr : VRPDI<0x6E, MRMSrcReg, (outs FR64:$dst), (ins GR64:$src),
2978 "mov{d|q}\t{$src, $dst|$dst, $src}",
2979 [(set FR64:$dst, (bitconvert GR64:$src))]>, VEX;
2981 def MOVDI2PDIrr : PDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
2982 "movd\t{$src, $dst|$dst, $src}",
2984 (v4i32 (scalar_to_vector GR32:$src)))]>;
2985 def MOVDI2PDIrm : PDI<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
2986 "movd\t{$src, $dst|$dst, $src}",
2988 (v4i32 (scalar_to_vector (loadi32 addr:$src))))]>;
2989 def MOV64toPQIrr : RPDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR64:$src),
2990 "mov{d|q}\t{$src, $dst|$dst, $src}",
2992 (v2i64 (scalar_to_vector GR64:$src)))]>;
2993 def MOV64toSDrr : RPDI<0x6E, MRMSrcReg, (outs FR64:$dst), (ins GR64:$src),
2994 "mov{d|q}\t{$src, $dst|$dst, $src}",
2995 [(set FR64:$dst, (bitconvert GR64:$src))]>;
2997 //===---------------------------------------------------------------------===//
2998 // Move Int Doubleword to Single Scalar
3000 def VMOVDI2SSrr : VPDI<0x6E, MRMSrcReg, (outs FR32:$dst), (ins GR32:$src),
3001 "movd\t{$src, $dst|$dst, $src}",
3002 [(set FR32:$dst, (bitconvert GR32:$src))]>, VEX;
3004 def VMOVDI2SSrm : VPDI<0x6E, MRMSrcMem, (outs FR32:$dst), (ins i32mem:$src),
3005 "movd\t{$src, $dst|$dst, $src}",
3006 [(set FR32:$dst, (bitconvert (loadi32 addr:$src)))]>,
3008 def MOVDI2SSrr : PDI<0x6E, MRMSrcReg, (outs FR32:$dst), (ins GR32:$src),
3009 "movd\t{$src, $dst|$dst, $src}",
3010 [(set FR32:$dst, (bitconvert GR32:$src))]>;
3012 def MOVDI2SSrm : PDI<0x6E, MRMSrcMem, (outs FR32:$dst), (ins i32mem:$src),
3013 "movd\t{$src, $dst|$dst, $src}",
3014 [(set FR32:$dst, (bitconvert (loadi32 addr:$src)))]>;
3016 //===---------------------------------------------------------------------===//
3017 // Move Packed Doubleword Int to Packed Double Int
3019 def VMOVPDI2DIrr : VPDI<0x7E, MRMDestReg, (outs GR32:$dst), (ins VR128:$src),
3020 "movd\t{$src, $dst|$dst, $src}",
3021 [(set GR32:$dst, (vector_extract (v4i32 VR128:$src),
3023 def VMOVPDI2DImr : VPDI<0x7E, MRMDestMem, (outs),
3024 (ins i32mem:$dst, VR128:$src),
3025 "movd\t{$src, $dst|$dst, $src}",
3026 [(store (i32 (vector_extract (v4i32 VR128:$src),
3027 (iPTR 0))), addr:$dst)]>, VEX;
3028 def MOVPDI2DIrr : PDI<0x7E, MRMDestReg, (outs GR32:$dst), (ins VR128:$src),
3029 "movd\t{$src, $dst|$dst, $src}",
3030 [(set GR32:$dst, (vector_extract (v4i32 VR128:$src),
3032 def MOVPDI2DImr : PDI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, VR128:$src),
3033 "movd\t{$src, $dst|$dst, $src}",
3034 [(store (i32 (vector_extract (v4i32 VR128:$src),
3035 (iPTR 0))), addr:$dst)]>;
3037 def MOVPQIto64rr : RPDI<0x7E, MRMDestReg, (outs GR64:$dst), (ins VR128:$src),
3038 "mov{d|q}\t{$src, $dst|$dst, $src}",
3039 [(set GR64:$dst, (vector_extract (v2i64 VR128:$src),
3041 def MOV64toSDrm : S3SI<0x7E, MRMSrcMem, (outs FR64:$dst), (ins i64mem:$src),
3042 "movq\t{$src, $dst|$dst, $src}",
3043 [(set FR64:$dst, (bitconvert (loadi64 addr:$src)))]>;
3045 def MOVSDto64rr : RPDI<0x7E, MRMDestReg, (outs GR64:$dst), (ins FR64:$src),
3046 "mov{d|q}\t{$src, $dst|$dst, $src}",
3047 [(set GR64:$dst, (bitconvert FR64:$src))]>;
3048 def MOVSDto64mr : RPDI<0x7E, MRMDestMem, (outs), (ins i64mem:$dst, FR64:$src),
3049 "movq\t{$src, $dst|$dst, $src}",
3050 [(store (i64 (bitconvert FR64:$src)), addr:$dst)]>;
3052 //===---------------------------------------------------------------------===//
3053 // Move Scalar Single to Double Int
3055 def VMOVSS2DIrr : VPDI<0x7E, MRMDestReg, (outs GR32:$dst), (ins FR32:$src),
3056 "movd\t{$src, $dst|$dst, $src}",
3057 [(set GR32:$dst, (bitconvert FR32:$src))]>, VEX;
3058 def VMOVSS2DImr : VPDI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, FR32:$src),
3059 "movd\t{$src, $dst|$dst, $src}",
3060 [(store (i32 (bitconvert FR32:$src)), addr:$dst)]>, VEX;
3061 def MOVSS2DIrr : PDI<0x7E, MRMDestReg, (outs GR32:$dst), (ins FR32:$src),
3062 "movd\t{$src, $dst|$dst, $src}",
3063 [(set GR32:$dst, (bitconvert FR32:$src))]>;
3064 def MOVSS2DImr : PDI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, FR32:$src),
3065 "movd\t{$src, $dst|$dst, $src}",
3066 [(store (i32 (bitconvert FR32:$src)), addr:$dst)]>;
3068 //===---------------------------------------------------------------------===//
3069 // Patterns and instructions to describe movd/movq to XMM register zero-extends
3071 let AddedComplexity = 15 in {
3072 def VMOVZDI2PDIrr : VPDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
3073 "movd\t{$src, $dst|$dst, $src}",
3074 [(set VR128:$dst, (v4i32 (X86vzmovl
3075 (v4i32 (scalar_to_vector GR32:$src)))))]>,
3077 def VMOVZQI2PQIrr : VPDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR64:$src),
3078 "mov{d|q}\t{$src, $dst|$dst, $src}", // X86-64 only
3079 [(set VR128:$dst, (v2i64 (X86vzmovl
3080 (v2i64 (scalar_to_vector GR64:$src)))))]>,
3083 let AddedComplexity = 15 in {
3084 def MOVZDI2PDIrr : PDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
3085 "movd\t{$src, $dst|$dst, $src}",
3086 [(set VR128:$dst, (v4i32 (X86vzmovl
3087 (v4i32 (scalar_to_vector GR32:$src)))))]>;
3088 def MOVZQI2PQIrr : RPDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR64:$src),
3089 "mov{d|q}\t{$src, $dst|$dst, $src}", // X86-64 only
3090 [(set VR128:$dst, (v2i64 (X86vzmovl
3091 (v2i64 (scalar_to_vector GR64:$src)))))]>;
3094 let AddedComplexity = 20 in {
3095 def VMOVZDI2PDIrm : VPDI<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
3096 "movd\t{$src, $dst|$dst, $src}",
3098 (v4i32 (X86vzmovl (v4i32 (scalar_to_vector
3099 (loadi32 addr:$src))))))]>,
3101 def MOVZDI2PDIrm : PDI<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
3102 "movd\t{$src, $dst|$dst, $src}",
3104 (v4i32 (X86vzmovl (v4i32 (scalar_to_vector
3105 (loadi32 addr:$src))))))]>;
3107 def : Pat<(v4i32 (X86vzmovl (loadv4i32 addr:$src))),
3108 (MOVZDI2PDIrm addr:$src)>;
3109 def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv4f32 addr:$src)))),
3110 (MOVZDI2PDIrm addr:$src)>;
3111 def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv2i64 addr:$src)))),
3112 (MOVZDI2PDIrm addr:$src)>;
3115 // AVX 128-bit movd/movq instruction write zeros in the high 128-bit part.
3116 // Use regular 128-bit instructions to match 256-bit scalar_to_vec+zext.
3117 def : Pat<(v8i32 (X86vzmovl (insert_subvector undef,
3118 (v4i32 (scalar_to_vector GR32:$src)), (i32 0)))),
3119 (SUBREG_TO_REG (i32 0), (VMOVZDI2PDIrr GR32:$src), sub_xmm)>;
3120 def : Pat<(v4i64 (X86vzmovl (insert_subvector undef,
3121 (v2i64 (scalar_to_vector GR64:$src)), (i32 0)))),
3122 (SUBREG_TO_REG (i64 0), (VMOVZQI2PQIrr GR64:$src), sub_xmm)>;
3124 // These are the correct encodings of the instructions so that we know how to
3125 // read correct assembly, even though we continue to emit the wrong ones for
3126 // compatibility with Darwin's buggy assembler.
3127 def : InstAlias<"movq\t{$src, $dst|$dst, $src}",
3128 (MOV64toPQIrr VR128:$dst, GR64:$src), 0>;
3129 def : InstAlias<"movq\t{$src, $dst|$dst, $src}",
3130 (MOV64toSDrr FR64:$dst, GR64:$src), 0>;
3131 def : InstAlias<"movq\t{$src, $dst|$dst, $src}",
3132 (MOVPQIto64rr GR64:$dst, VR128:$src), 0>;
3133 def : InstAlias<"movq\t{$src, $dst|$dst, $src}",
3134 (MOVSDto64rr GR64:$dst, FR64:$src), 0>;
3135 def : InstAlias<"movq\t{$src, $dst|$dst, $src}",
3136 (VMOVZQI2PQIrr VR128:$dst, GR64:$src), 0>;
3137 def : InstAlias<"movq\t{$src, $dst|$dst, $src}",
3138 (MOVZQI2PQIrr VR128:$dst, GR64:$src), 0>;
3140 //===---------------------------------------------------------------------===//
3141 // SSE2 - Move Quadword
3142 //===---------------------------------------------------------------------===//
3144 //===---------------------------------------------------------------------===//
3145 // Move Quadword Int to Packed Quadword Int
3147 def VMOVQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
3148 "vmovq\t{$src, $dst|$dst, $src}",
3150 (v2i64 (scalar_to_vector (loadi64 addr:$src))))]>, XS,
3151 VEX, Requires<[HasAVX]>;
3152 def MOVQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
3153 "movq\t{$src, $dst|$dst, $src}",
3155 (v2i64 (scalar_to_vector (loadi64 addr:$src))))]>, XS,
3156 Requires<[HasSSE2]>; // SSE2 instruction with XS Prefix
3158 //===---------------------------------------------------------------------===//
3159 // Move Packed Quadword Int to Quadword Int
3161 def VMOVPQI2QImr : VPDI<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
3162 "movq\t{$src, $dst|$dst, $src}",
3163 [(store (i64 (vector_extract (v2i64 VR128:$src),
3164 (iPTR 0))), addr:$dst)]>, VEX;
3165 def MOVPQI2QImr : PDI<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
3166 "movq\t{$src, $dst|$dst, $src}",
3167 [(store (i64 (vector_extract (v2i64 VR128:$src),
3168 (iPTR 0))), addr:$dst)]>;
3170 def : Pat<(f64 (vector_extract (v2f64 VR128:$src), (iPTR 0))),
3171 (f64 (EXTRACT_SUBREG (v2f64 VR128:$src), sub_sd))>;
3173 //===---------------------------------------------------------------------===//
3174 // Store / copy lower 64-bits of a XMM register.
3176 def VMOVLQ128mr : VPDI<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
3177 "movq\t{$src, $dst|$dst, $src}",
3178 [(int_x86_sse2_storel_dq addr:$dst, VR128:$src)]>, VEX;
3179 def MOVLQ128mr : PDI<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
3180 "movq\t{$src, $dst|$dst, $src}",
3181 [(int_x86_sse2_storel_dq addr:$dst, VR128:$src)]>;
3183 let AddedComplexity = 20 in
3184 def VMOVZQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
3185 "vmovq\t{$src, $dst|$dst, $src}",
3187 (v2i64 (X86vzmovl (v2i64 (scalar_to_vector
3188 (loadi64 addr:$src))))))]>,
3189 XS, VEX, Requires<[HasAVX]>;
3191 let AddedComplexity = 20 in {
3192 def MOVZQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
3193 "movq\t{$src, $dst|$dst, $src}",
3195 (v2i64 (X86vzmovl (v2i64 (scalar_to_vector
3196 (loadi64 addr:$src))))))]>,
3197 XS, Requires<[HasSSE2]>;
3199 def : Pat<(v2i64 (X86vzmovl (loadv2i64 addr:$src))),
3200 (MOVZQI2PQIrm addr:$src)>;
3201 def : Pat<(v2i64 (X86vzmovl (bc_v2i64 (loadv4f32 addr:$src)))),
3202 (MOVZQI2PQIrm addr:$src)>;
3203 def : Pat<(v2i64 (X86vzload addr:$src)), (MOVZQI2PQIrm addr:$src)>;
3206 //===---------------------------------------------------------------------===//
3207 // Moving from XMM to XMM and clear upper 64 bits. Note, there is a bug in
3208 // IA32 document. movq xmm1, xmm2 does clear the high bits.
3210 let AddedComplexity = 15 in
3211 def VMOVZPQILo2PQIrr : I<0x7E, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3212 "vmovq\t{$src, $dst|$dst, $src}",
3213 [(set VR128:$dst, (v2i64 (X86vzmovl (v2i64 VR128:$src))))]>,
3214 XS, VEX, Requires<[HasAVX]>;
3215 let AddedComplexity = 15 in
3216 def MOVZPQILo2PQIrr : I<0x7E, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3217 "movq\t{$src, $dst|$dst, $src}",
3218 [(set VR128:$dst, (v2i64 (X86vzmovl (v2i64 VR128:$src))))]>,
3219 XS, Requires<[HasSSE2]>;
3221 let AddedComplexity = 20 in
3222 def VMOVZPQILo2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
3223 "vmovq\t{$src, $dst|$dst, $src}",
3224 [(set VR128:$dst, (v2i64 (X86vzmovl
3225 (loadv2i64 addr:$src))))]>,
3226 XS, VEX, Requires<[HasAVX]>;
3227 let AddedComplexity = 20 in {
3228 def MOVZPQILo2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
3229 "movq\t{$src, $dst|$dst, $src}",
3230 [(set VR128:$dst, (v2i64 (X86vzmovl
3231 (loadv2i64 addr:$src))))]>,
3232 XS, Requires<[HasSSE2]>;
3234 def : Pat<(v2i64 (X86vzmovl (bc_v2i64 (loadv4i32 addr:$src)))),
3235 (MOVZPQILo2PQIrm addr:$src)>;
3238 // Instructions to match in the assembler
3239 def VMOVQs64rr : VPDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR64:$src),
3240 "movq\t{$src, $dst|$dst, $src}", []>, VEX, VEX_W;
3241 def VMOVQd64rr : VPDI<0x7E, MRMDestReg, (outs GR64:$dst), (ins VR128:$src),
3242 "movq\t{$src, $dst|$dst, $src}", []>, VEX, VEX_W;
3243 // Recognize "movd" with GR64 destination, but encode as a "movq"
3244 def VMOVQd64rr_alt : VPDI<0x7E, MRMDestReg, (outs GR64:$dst), (ins VR128:$src),
3245 "movd\t{$src, $dst|$dst, $src}", []>, VEX, VEX_W;
3247 // Instructions for the disassembler
3248 // xr = XMM register
3251 let Predicates = [HasAVX] in
3252 def VMOVQxrxr: I<0x7E, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3253 "vmovq\t{$src, $dst|$dst, $src}", []>, VEX, XS;
3254 def MOVQxrxr : I<0x7E, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3255 "movq\t{$src, $dst|$dst, $src}", []>, XS;
3257 //===---------------------------------------------------------------------===//
3258 // SSE2 - Misc Instructions
3259 //===---------------------------------------------------------------------===//
3262 def CLFLUSH : I<0xAE, MRM7m, (outs), (ins i8mem:$src),
3263 "clflush\t$src", [(int_x86_sse2_clflush addr:$src)]>,
3264 TB, Requires<[HasSSE2]>;
3266 // Load, store, and memory fence
3267 def LFENCE : I<0xAE, MRM_E8, (outs), (ins),
3268 "lfence", [(int_x86_sse2_lfence)]>, TB, Requires<[HasSSE2]>;
3269 def MFENCE : I<0xAE, MRM_F0, (outs), (ins),
3270 "mfence", [(int_x86_sse2_mfence)]>, TB, Requires<[HasSSE2]>;
3271 def : Pat<(X86LFence), (LFENCE)>;
3272 def : Pat<(X86MFence), (MFENCE)>;
3275 // Pause. This "instruction" is encoded as "rep; nop", so even though it
3276 // was introduced with SSE2, it's backward compatible.
3277 def PAUSE : I<0x90, RawFrm, (outs), (ins), "pause", []>, REP;
3279 // Alias instructions that map zero vector to pxor / xorp* for sse.
3280 // We set canFoldAsLoad because this can be converted to a constant-pool
3281 // load of an all-ones value if folding it would be beneficial.
3282 // FIXME: Change encoding to pseudo! This is blocked right now by the x86
3283 // JIT implementation, it does not expand the instructions below like
3284 // X86MCInstLower does.
3285 let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
3286 isCodeGenOnly = 1, ExeDomain = SSEPackedInt in
3287 def V_SETALLONES : PDI<0x76, MRMInitReg, (outs VR128:$dst), (ins), "",
3288 [(set VR128:$dst, (v4i32 immAllOnesV))]>;
3289 let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
3290 isCodeGenOnly = 1, ExeDomain = SSEPackedInt, Predicates = [HasAVX] in
3291 def AVX_SETALLONES : PDI<0x76, MRMInitReg, (outs VR128:$dst), (ins), "",
3292 [(set VR128:$dst, (v4i32 immAllOnesV))]>, VEX_4V;
3294 //===---------------------------------------------------------------------===//
3295 // SSE3 - Conversion Instructions
3296 //===---------------------------------------------------------------------===//
3298 // Convert Packed Double FP to Packed DW Integers
3299 let Predicates = [HasAVX] in {
3300 // The assembler can recognize rr 256-bit instructions by seeing a ymm
3301 // register, but the same isn't true when using memory operands instead.
3302 // Provide other assembly rr and rm forms to address this explicitly.
3303 def VCVTPD2DQrr : S3DI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3304 "vcvtpd2dq\t{$src, $dst|$dst, $src}", []>, VEX;
3305 def VCVTPD2DQXrYr : S3DI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR256:$src),
3306 "vcvtpd2dq\t{$src, $dst|$dst, $src}", []>, VEX;
3309 def VCVTPD2DQXrr : S3DI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3310 "vcvtpd2dqx\t{$src, $dst|$dst, $src}", []>, VEX;
3311 def VCVTPD2DQXrm : S3DI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
3312 "vcvtpd2dqx\t{$src, $dst|$dst, $src}", []>, VEX;
3315 def VCVTPD2DQYrr : S3DI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR256:$src),
3316 "vcvtpd2dqy\t{$src, $dst|$dst, $src}", []>, VEX;
3317 def VCVTPD2DQYrm : S3DI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f256mem:$src),
3318 "vcvtpd2dqy\t{$src, $dst|$dst, $src}", []>, VEX, VEX_L;
3321 def CVTPD2DQrm : S3DI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
3322 "cvtpd2dq\t{$src, $dst|$dst, $src}", []>;
3323 def CVTPD2DQrr : S3DI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3324 "cvtpd2dq\t{$src, $dst|$dst, $src}", []>;
3326 def : Pat<(v4i32 (fp_to_sint (v4f64 VR256:$src))),
3327 (VCVTPD2DQYrr VR256:$src)>;
3328 def : Pat<(v4i32 (fp_to_sint (memopv4f64 addr:$src))),
3329 (VCVTPD2DQYrm addr:$src)>;
3331 // Convert Packed DW Integers to Packed Double FP
3332 let Predicates = [HasAVX] in {
3333 def VCVTDQ2PDrm : S3SI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
3334 "vcvtdq2pd\t{$src, $dst|$dst, $src}", []>, VEX;
3335 def VCVTDQ2PDrr : S3SI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3336 "vcvtdq2pd\t{$src, $dst|$dst, $src}", []>, VEX;
3337 def VCVTDQ2PDYrm : S3SI<0xE6, MRMSrcMem, (outs VR256:$dst), (ins f128mem:$src),
3338 "vcvtdq2pd\t{$src, $dst|$dst, $src}", []>, VEX;
3339 def VCVTDQ2PDYrr : S3SI<0xE6, MRMSrcReg, (outs VR256:$dst), (ins VR128:$src),
3340 "vcvtdq2pd\t{$src, $dst|$dst, $src}", []>, VEX;
3343 def CVTDQ2PDrm : S3SI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
3344 "cvtdq2pd\t{$src, $dst|$dst, $src}", []>;
3345 def CVTDQ2PDrr : S3SI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3346 "cvtdq2pd\t{$src, $dst|$dst, $src}", []>;
3348 // AVX 256-bit register conversion intrinsics
3349 def : Pat<(int_x86_avx_cvtdq2_pd_256 VR128:$src),
3350 (VCVTDQ2PDYrr VR128:$src)>;
3351 def : Pat<(int_x86_avx_cvtdq2_pd_256 (memopv4i32 addr:$src)),
3352 (VCVTDQ2PDYrm addr:$src)>;
3354 def : Pat<(int_x86_avx_cvt_pd2dq_256 VR256:$src),
3355 (VCVTPD2DQYrr VR256:$src)>;
3356 def : Pat<(int_x86_avx_cvt_pd2dq_256 (memopv4f64 addr:$src)),
3357 (VCVTPD2DQYrm addr:$src)>;
3359 def : Pat<(v4f64 (sint_to_fp (v4i32 VR128:$src))),
3360 (VCVTDQ2PDYrr VR128:$src)>;
3361 def : Pat<(v4f64 (sint_to_fp (memopv4i32 addr:$src))),
3362 (VCVTDQ2PDYrm addr:$src)>;
3364 //===---------------------------------------------------------------------===//
3365 // SSE3 - Move Instructions
3366 //===---------------------------------------------------------------------===//
3368 //===---------------------------------------------------------------------===//
3369 // Replicate Single FP - MOVSHDUP and MOVSLDUP
3371 multiclass sse3_replicate_sfp<bits<8> op, SDNode OpNode, string OpcodeStr,
3372 ValueType vt, RegisterClass RC, PatFrag mem_frag,
3373 X86MemOperand x86memop> {
3374 def rr : S3SI<op, MRMSrcReg, (outs RC:$dst), (ins RC:$src),
3375 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3376 [(set RC:$dst, (vt (OpNode RC:$src)))]>;
3377 def rm : S3SI<op, MRMSrcMem, (outs RC:$dst), (ins x86memop:$src),
3378 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3379 [(set RC:$dst, (OpNode (mem_frag addr:$src)))]>;
3382 let Predicates = [HasAVX] in {
3383 defm VMOVSHDUP : sse3_replicate_sfp<0x16, X86Movshdup, "vmovshdup",
3384 v4f32, VR128, memopv4f32, f128mem>, VEX;
3385 defm VMOVSLDUP : sse3_replicate_sfp<0x12, X86Movsldup, "vmovsldup",
3386 v4f32, VR128, memopv4f32, f128mem>, VEX;
3387 defm VMOVSHDUPY : sse3_replicate_sfp<0x16, X86Movshdup, "vmovshdup",
3388 v8f32, VR256, memopv8f32, f256mem>, VEX;
3389 defm VMOVSLDUPY : sse3_replicate_sfp<0x12, X86Movsldup, "vmovsldup",
3390 v8f32, VR256, memopv8f32, f256mem>, VEX;
3392 defm MOVSHDUP : sse3_replicate_sfp<0x16, X86Movshdup, "movshdup", v4f32, VR128,
3393 memopv4f32, f128mem>;
3394 defm MOVSLDUP : sse3_replicate_sfp<0x12, X86Movsldup, "movsldup", v4f32, VR128,
3395 memopv4f32, f128mem>;
3397 let Predicates = [HasSSE3] in {
3398 def : Pat<(v4i32 (X86Movshdup VR128:$src)),
3399 (MOVSHDUPrr VR128:$src)>;
3400 def : Pat<(v4i32 (X86Movshdup (bc_v4i32 (memopv2i64 addr:$src)))),
3401 (MOVSHDUPrm addr:$src)>;
3402 def : Pat<(v4i32 (X86Movsldup VR128:$src)),
3403 (MOVSLDUPrr VR128:$src)>;
3404 def : Pat<(v4i32 (X86Movsldup (bc_v4i32 (memopv2i64 addr:$src)))),
3405 (MOVSLDUPrm addr:$src)>;
3408 let Predicates = [HasAVX] in {
3409 def : Pat<(v4i32 (X86Movshdup VR128:$src)),
3410 (VMOVSHDUPrr VR128:$src)>;
3411 def : Pat<(v4i32 (X86Movshdup (bc_v4i32 (memopv2i64 addr:$src)))),
3412 (VMOVSHDUPrm addr:$src)>;
3413 def : Pat<(v4i32 (X86Movsldup VR128:$src)),
3414 (VMOVSLDUPrr VR128:$src)>;
3415 def : Pat<(v4i32 (X86Movsldup (bc_v4i32 (memopv2i64 addr:$src)))),
3416 (VMOVSLDUPrm addr:$src)>;
3417 def : Pat<(v8i32 (X86Movshdup VR256:$src)),
3418 (VMOVSHDUPYrr VR256:$src)>;
3419 def : Pat<(v8i32 (X86Movshdup (bc_v8i32 (memopv4i64 addr:$src)))),
3420 (VMOVSHDUPYrm addr:$src)>;
3421 def : Pat<(v8i32 (X86Movsldup VR256:$src)),
3422 (VMOVSLDUPYrr VR256:$src)>;
3423 def : Pat<(v8i32 (X86Movsldup (bc_v8i32 (memopv4i64 addr:$src)))),
3424 (VMOVSLDUPYrm addr:$src)>;
3427 //===---------------------------------------------------------------------===//
3428 // Replicate Double FP - MOVDDUP
3430 multiclass sse3_replicate_dfp<string OpcodeStr> {
3431 def rr : S3DI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3432 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3433 [(set VR128:$dst,(v2f64 (movddup VR128:$src, (undef))))]>;
3434 def rm : S3DI<0x12, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
3435 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3437 (v2f64 (movddup (scalar_to_vector (loadf64 addr:$src)),
3441 multiclass sse3_replicate_dfp_y<string OpcodeStr> {
3442 def rr : S3DI<0x12, MRMSrcReg, (outs VR256:$dst), (ins VR256:$src),
3443 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3445 def rm : S3DI<0x12, MRMSrcMem, (outs VR256:$dst), (ins f256mem:$src),
3446 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3450 let Predicates = [HasAVX] in {
3451 // FIXME: Merge above classes when we have patterns for the ymm version
3452 defm VMOVDDUP : sse3_replicate_dfp<"vmovddup">, VEX;
3453 defm VMOVDDUPY : sse3_replicate_dfp_y<"vmovddup">, VEX;
3455 defm MOVDDUP : sse3_replicate_dfp<"movddup">;
3457 // Move Unaligned Integer
3458 let Predicates = [HasAVX] in {
3459 def VLDDQUrm : S3DI<0xF0, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
3460 "vlddqu\t{$src, $dst|$dst, $src}",
3461 [(set VR128:$dst, (int_x86_sse3_ldu_dq addr:$src))]>, VEX;
3462 def VLDDQUYrm : S3DI<0xF0, MRMSrcMem, (outs VR256:$dst), (ins i256mem:$src),
3463 "vlddqu\t{$src, $dst|$dst, $src}",
3464 [(set VR256:$dst, (int_x86_avx_ldu_dq_256 addr:$src))]>, VEX;
3466 def LDDQUrm : S3DI<0xF0, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
3467 "lddqu\t{$src, $dst|$dst, $src}",
3468 [(set VR128:$dst, (int_x86_sse3_ldu_dq addr:$src))]>;
3470 def : Pat<(movddup (bc_v2f64 (v2i64 (scalar_to_vector (loadi64 addr:$src)))),
3472 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
3474 // Several Move patterns
3475 let AddedComplexity = 5 in {
3476 def : Pat<(movddup (memopv2f64 addr:$src), (undef)),
3477 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
3478 def : Pat<(movddup (bc_v4f32 (memopv2f64 addr:$src)), (undef)),
3479 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
3480 def : Pat<(movddup (memopv2i64 addr:$src), (undef)),
3481 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
3482 def : Pat<(movddup (bc_v4i32 (memopv2i64 addr:$src)), (undef)),
3483 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
3486 //===---------------------------------------------------------------------===//
3487 // SSE3 - Arithmetic
3488 //===---------------------------------------------------------------------===//
3490 multiclass sse3_addsub<Intrinsic Int, string OpcodeStr, RegisterClass RC,
3491 X86MemOperand x86memop, bit Is2Addr = 1> {
3492 def rr : I<0xD0, MRMSrcReg,
3493 (outs RC:$dst), (ins RC:$src1, RC:$src2),
3495 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3496 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
3497 [(set RC:$dst, (Int RC:$src1, RC:$src2))]>;
3498 def rm : I<0xD0, MRMSrcMem,
3499 (outs RC:$dst), (ins RC:$src1, x86memop:$src2),
3501 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3502 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
3503 [(set RC:$dst, (Int RC:$src1, (memop addr:$src2)))]>;
3506 let Predicates = [HasAVX],
3507 ExeDomain = SSEPackedDouble in {
3508 defm VADDSUBPS : sse3_addsub<int_x86_sse3_addsub_ps, "vaddsubps", VR128,
3509 f128mem, 0>, TB, XD, VEX_4V;
3510 defm VADDSUBPD : sse3_addsub<int_x86_sse3_addsub_pd, "vaddsubpd", VR128,
3511 f128mem, 0>, TB, OpSize, VEX_4V;
3512 defm VADDSUBPSY : sse3_addsub<int_x86_avx_addsub_ps_256, "vaddsubps", VR256,
3513 f256mem, 0>, TB, XD, VEX_4V;
3514 defm VADDSUBPDY : sse3_addsub<int_x86_avx_addsub_pd_256, "vaddsubpd", VR256,
3515 f256mem, 0>, TB, OpSize, VEX_4V;
3517 let Constraints = "$src1 = $dst", Predicates = [HasSSE3],
3518 ExeDomain = SSEPackedDouble in {
3519 defm ADDSUBPS : sse3_addsub<int_x86_sse3_addsub_ps, "addsubps", VR128,
3521 defm ADDSUBPD : sse3_addsub<int_x86_sse3_addsub_pd, "addsubpd", VR128,
3522 f128mem>, TB, OpSize;
3525 //===---------------------------------------------------------------------===//
3526 // SSE3 Instructions
3527 //===---------------------------------------------------------------------===//
3530 multiclass S3D_Int<bits<8> o, string OpcodeStr, ValueType vt, RegisterClass RC,
3531 X86MemOperand x86memop, Intrinsic IntId, bit Is2Addr = 1> {
3532 def rr : S3DI<o, MRMSrcReg, (outs RC:$dst), (ins RC:$src1, RC:$src2),
3534 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3535 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
3536 [(set RC:$dst, (vt (IntId RC:$src1, RC:$src2)))]>;
3538 def rm : S3DI<o, MRMSrcMem, (outs RC:$dst), (ins RC:$src1, x86memop:$src2),
3540 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3541 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
3542 [(set RC:$dst, (vt (IntId RC:$src1, (memop addr:$src2))))]>;
3544 multiclass S3_Int<bits<8> o, string OpcodeStr, ValueType vt, RegisterClass RC,
3545 X86MemOperand x86memop, Intrinsic IntId, bit Is2Addr = 1> {
3546 def rr : S3I<o, MRMSrcReg, (outs RC:$dst), (ins RC:$src1, RC:$src2),
3548 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3549 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
3550 [(set RC:$dst, (vt (IntId RC:$src1, RC:$src2)))]>;
3552 def rm : S3I<o, MRMSrcMem, (outs RC:$dst), (ins RC:$src1, x86memop:$src2),
3554 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3555 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
3556 [(set RC:$dst, (vt (IntId RC:$src1, (memop addr:$src2))))]>;
3559 let Predicates = [HasAVX] in {
3560 defm VHADDPS : S3D_Int<0x7C, "vhaddps", v4f32, VR128, f128mem,
3561 int_x86_sse3_hadd_ps, 0>, VEX_4V;
3562 defm VHADDPD : S3_Int <0x7C, "vhaddpd", v2f64, VR128, f128mem,
3563 int_x86_sse3_hadd_pd, 0>, VEX_4V;
3564 defm VHSUBPS : S3D_Int<0x7D, "vhsubps", v4f32, VR128, f128mem,
3565 int_x86_sse3_hsub_ps, 0>, VEX_4V;
3566 defm VHSUBPD : S3_Int <0x7D, "vhsubpd", v2f64, VR128, f128mem,
3567 int_x86_sse3_hsub_pd, 0>, VEX_4V;
3568 defm VHADDPSY : S3D_Int<0x7C, "vhaddps", v8f32, VR256, f256mem,
3569 int_x86_avx_hadd_ps_256, 0>, VEX_4V;
3570 defm VHADDPDY : S3_Int <0x7C, "vhaddpd", v4f64, VR256, f256mem,
3571 int_x86_avx_hadd_pd_256, 0>, VEX_4V;
3572 defm VHSUBPSY : S3D_Int<0x7D, "vhsubps", v8f32, VR256, f256mem,
3573 int_x86_avx_hsub_ps_256, 0>, VEX_4V;
3574 defm VHSUBPDY : S3_Int <0x7D, "vhsubpd", v4f64, VR256, f256mem,
3575 int_x86_avx_hsub_pd_256, 0>, VEX_4V;
3578 let Constraints = "$src1 = $dst" in {
3579 defm HADDPS : S3D_Int<0x7C, "haddps", v4f32, VR128, f128mem,
3580 int_x86_sse3_hadd_ps>;
3581 defm HADDPD : S3_Int<0x7C, "haddpd", v2f64, VR128, f128mem,
3582 int_x86_sse3_hadd_pd>;
3583 defm HSUBPS : S3D_Int<0x7D, "hsubps", v4f32, VR128, f128mem,
3584 int_x86_sse3_hsub_ps>;
3585 defm HSUBPD : S3_Int<0x7D, "hsubpd", v2f64, VR128, f128mem,
3586 int_x86_sse3_hsub_pd>;
3589 //===---------------------------------------------------------------------===//
3590 // SSSE3 - Packed Absolute Instructions
3591 //===---------------------------------------------------------------------===//
3594 /// SS3I_unop_rm_int - Simple SSSE3 unary op whose type can be v*{i8,i16,i32}.
3595 multiclass SS3I_unop_rm_int<bits<8> opc, string OpcodeStr,
3596 PatFrag mem_frag128, Intrinsic IntId128> {
3597 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
3599 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3600 [(set VR128:$dst, (IntId128 VR128:$src))]>,
3603 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
3605 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3608 (bitconvert (mem_frag128 addr:$src))))]>, OpSize;
3611 let Predicates = [HasAVX] in {
3612 defm VPABSB : SS3I_unop_rm_int<0x1C, "vpabsb", memopv16i8,
3613 int_x86_ssse3_pabs_b_128>, VEX;
3614 defm VPABSW : SS3I_unop_rm_int<0x1D, "vpabsw", memopv8i16,
3615 int_x86_ssse3_pabs_w_128>, VEX;
3616 defm VPABSD : SS3I_unop_rm_int<0x1E, "vpabsd", memopv4i32,
3617 int_x86_ssse3_pabs_d_128>, VEX;
3620 defm PABSB : SS3I_unop_rm_int<0x1C, "pabsb", memopv16i8,
3621 int_x86_ssse3_pabs_b_128>;
3622 defm PABSW : SS3I_unop_rm_int<0x1D, "pabsw", memopv8i16,
3623 int_x86_ssse3_pabs_w_128>;
3624 defm PABSD : SS3I_unop_rm_int<0x1E, "pabsd", memopv4i32,
3625 int_x86_ssse3_pabs_d_128>;
3627 //===---------------------------------------------------------------------===//
3628 // SSSE3 - Packed Binary Operator Instructions
3629 //===---------------------------------------------------------------------===//
3631 /// SS3I_binop_rm_int - Simple SSSE3 bin op whose type can be v*{i8,i16,i32}.
3632 multiclass SS3I_binop_rm_int<bits<8> opc, string OpcodeStr,
3633 PatFrag mem_frag128, Intrinsic IntId128,
3635 let isCommutable = 1 in
3636 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
3637 (ins VR128:$src1, VR128:$src2),
3639 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3640 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
3641 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3643 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
3644 (ins VR128:$src1, i128mem:$src2),
3646 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3647 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
3649 (IntId128 VR128:$src1,
3650 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
3653 let Predicates = [HasAVX] in {
3654 let isCommutable = 0 in {
3655 defm VPHADDW : SS3I_binop_rm_int<0x01, "vphaddw", memopv8i16,
3656 int_x86_ssse3_phadd_w_128, 0>, VEX_4V;
3657 defm VPHADDD : SS3I_binop_rm_int<0x02, "vphaddd", memopv4i32,
3658 int_x86_ssse3_phadd_d_128, 0>, VEX_4V;
3659 defm VPHADDSW : SS3I_binop_rm_int<0x03, "vphaddsw", memopv8i16,
3660 int_x86_ssse3_phadd_sw_128, 0>, VEX_4V;
3661 defm VPHSUBW : SS3I_binop_rm_int<0x05, "vphsubw", memopv8i16,
3662 int_x86_ssse3_phsub_w_128, 0>, VEX_4V;
3663 defm VPHSUBD : SS3I_binop_rm_int<0x06, "vphsubd", memopv4i32,
3664 int_x86_ssse3_phsub_d_128, 0>, VEX_4V;
3665 defm VPHSUBSW : SS3I_binop_rm_int<0x07, "vphsubsw", memopv8i16,
3666 int_x86_ssse3_phsub_sw_128, 0>, VEX_4V;
3667 defm VPMADDUBSW : SS3I_binop_rm_int<0x04, "vpmaddubsw", memopv16i8,
3668 int_x86_ssse3_pmadd_ub_sw_128, 0>, VEX_4V;
3669 defm VPSHUFB : SS3I_binop_rm_int<0x00, "vpshufb", memopv16i8,
3670 int_x86_ssse3_pshuf_b_128, 0>, VEX_4V;
3671 defm VPSIGNB : SS3I_binop_rm_int<0x08, "vpsignb", memopv16i8,
3672 int_x86_ssse3_psign_b_128, 0>, VEX_4V;
3673 defm VPSIGNW : SS3I_binop_rm_int<0x09, "vpsignw", memopv8i16,
3674 int_x86_ssse3_psign_w_128, 0>, VEX_4V;
3675 defm VPSIGND : SS3I_binop_rm_int<0x0A, "vpsignd", memopv4i32,
3676 int_x86_ssse3_psign_d_128, 0>, VEX_4V;
3678 defm VPMULHRSW : SS3I_binop_rm_int<0x0B, "vpmulhrsw", memopv8i16,
3679 int_x86_ssse3_pmul_hr_sw_128, 0>, VEX_4V;
3682 // None of these have i8 immediate fields.
3683 let ImmT = NoImm, Constraints = "$src1 = $dst" in {
3684 let isCommutable = 0 in {
3685 defm PHADDW : SS3I_binop_rm_int<0x01, "phaddw", memopv8i16,
3686 int_x86_ssse3_phadd_w_128>;
3687 defm PHADDD : SS3I_binop_rm_int<0x02, "phaddd", memopv4i32,
3688 int_x86_ssse3_phadd_d_128>;
3689 defm PHADDSW : SS3I_binop_rm_int<0x03, "phaddsw", memopv8i16,
3690 int_x86_ssse3_phadd_sw_128>;
3691 defm PHSUBW : SS3I_binop_rm_int<0x05, "phsubw", memopv8i16,
3692 int_x86_ssse3_phsub_w_128>;
3693 defm PHSUBD : SS3I_binop_rm_int<0x06, "phsubd", memopv4i32,
3694 int_x86_ssse3_phsub_d_128>;
3695 defm PHSUBSW : SS3I_binop_rm_int<0x07, "phsubsw", memopv8i16,
3696 int_x86_ssse3_phsub_sw_128>;
3697 defm PMADDUBSW : SS3I_binop_rm_int<0x04, "pmaddubsw", memopv16i8,
3698 int_x86_ssse3_pmadd_ub_sw_128>;
3699 defm PSHUFB : SS3I_binop_rm_int<0x00, "pshufb", memopv16i8,
3700 int_x86_ssse3_pshuf_b_128>;
3701 defm PSIGNB : SS3I_binop_rm_int<0x08, "psignb", memopv16i8,
3702 int_x86_ssse3_psign_b_128>;
3703 defm PSIGNW : SS3I_binop_rm_int<0x09, "psignw", memopv8i16,
3704 int_x86_ssse3_psign_w_128>;
3705 defm PSIGND : SS3I_binop_rm_int<0x0A, "psignd", memopv4i32,
3706 int_x86_ssse3_psign_d_128>;
3708 defm PMULHRSW : SS3I_binop_rm_int<0x0B, "pmulhrsw", memopv8i16,
3709 int_x86_ssse3_pmul_hr_sw_128>;
3712 def : Pat<(X86pshufb VR128:$src, VR128:$mask),
3713 (PSHUFBrr128 VR128:$src, VR128:$mask)>, Requires<[HasSSSE3]>;
3714 def : Pat<(X86pshufb VR128:$src, (bc_v16i8 (memopv2i64 addr:$mask))),
3715 (PSHUFBrm128 VR128:$src, addr:$mask)>, Requires<[HasSSSE3]>;
3717 def : Pat<(X86psignb VR128:$src1, VR128:$src2),
3718 (PSIGNBrr128 VR128:$src1, VR128:$src2)>, Requires<[HasSSSE3]>;
3719 def : Pat<(X86psignw VR128:$src1, VR128:$src2),
3720 (PSIGNWrr128 VR128:$src1, VR128:$src2)>, Requires<[HasSSSE3]>;
3721 def : Pat<(X86psignd VR128:$src1, VR128:$src2),
3722 (PSIGNDrr128 VR128:$src1, VR128:$src2)>, Requires<[HasSSSE3]>;
3724 //===---------------------------------------------------------------------===//
3725 // SSSE3 - Packed Align Instruction Patterns
3726 //===---------------------------------------------------------------------===//
3728 multiclass ssse3_palign<string asm, bit Is2Addr = 1> {
3729 def R128rr : SS3AI<0x0F, MRMSrcReg, (outs VR128:$dst),
3730 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
3732 !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3734 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
3736 def R128rm : SS3AI<0x0F, MRMSrcMem, (outs VR128:$dst),
3737 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
3739 !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3741 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
3745 let Predicates = [HasAVX] in
3746 defm VPALIGN : ssse3_palign<"vpalignr", 0>, VEX_4V;
3747 let Constraints = "$src1 = $dst", Predicates = [HasSSSE3] in
3748 defm PALIGN : ssse3_palign<"palignr">;
3750 let Predicates = [HasSSSE3] in {
3751 def : Pat<(v4i32 (X86PAlign VR128:$src1, VR128:$src2, (i8 imm:$imm))),
3752 (PALIGNR128rr VR128:$src2, VR128:$src1, imm:$imm)>;
3753 def : Pat<(v4f32 (X86PAlign VR128:$src1, VR128:$src2, (i8 imm:$imm))),
3754 (PALIGNR128rr VR128:$src2, VR128:$src1, imm:$imm)>;
3755 def : Pat<(v8i16 (X86PAlign VR128:$src1, VR128:$src2, (i8 imm:$imm))),
3756 (PALIGNR128rr VR128:$src2, VR128:$src1, imm:$imm)>;
3757 def : Pat<(v16i8 (X86PAlign VR128:$src1, VR128:$src2, (i8 imm:$imm))),
3758 (PALIGNR128rr VR128:$src2, VR128:$src1, imm:$imm)>;
3761 let Predicates = [HasAVX] in {
3762 def : Pat<(v4i32 (X86PAlign VR128:$src1, VR128:$src2, (i8 imm:$imm))),
3763 (VPALIGNR128rr VR128:$src2, VR128:$src1, imm:$imm)>;
3764 def : Pat<(v4f32 (X86PAlign VR128:$src1, VR128:$src2, (i8 imm:$imm))),
3765 (VPALIGNR128rr VR128:$src2, VR128:$src1, imm:$imm)>;
3766 def : Pat<(v8i16 (X86PAlign VR128:$src1, VR128:$src2, (i8 imm:$imm))),
3767 (VPALIGNR128rr VR128:$src2, VR128:$src1, imm:$imm)>;
3768 def : Pat<(v16i8 (X86PAlign VR128:$src1, VR128:$src2, (i8 imm:$imm))),
3769 (VPALIGNR128rr VR128:$src2, VR128:$src1, imm:$imm)>;
3772 //===---------------------------------------------------------------------===//
3773 // SSSE3 Misc Instructions
3774 //===---------------------------------------------------------------------===//
3776 // Thread synchronization
3777 let usesCustomInserter = 1 in {
3778 def MONITOR : PseudoI<(outs), (ins i32mem:$src1, GR32:$src2, GR32:$src3),
3779 [(int_x86_sse3_monitor addr:$src1, GR32:$src2, GR32:$src3)]>;
3780 def MWAIT : PseudoI<(outs), (ins GR32:$src1, GR32:$src2),
3781 [(int_x86_sse3_mwait GR32:$src1, GR32:$src2)]>;
3784 let Uses = [EAX, ECX, EDX] in
3785 def MONITORrrr : I<0x01, MRM_C8, (outs), (ins), "monitor", []>, TB,
3786 Requires<[HasSSE3]>;
3787 let Uses = [ECX, EAX] in
3788 def MWAITrr : I<0x01, MRM_C9, (outs), (ins), "mwait", []>, TB,
3789 Requires<[HasSSE3]>;
3791 def : InstAlias<"mwait %eax, %ecx", (MWAITrr)>, Requires<[In32BitMode]>;
3792 def : InstAlias<"mwait %rax, %rcx", (MWAITrr)>, Requires<[In64BitMode]>;
3794 def : InstAlias<"monitor %eax, %ecx, %edx", (MONITORrrr)>,
3795 Requires<[In32BitMode]>;
3796 def : InstAlias<"monitor %rax, %rcx, %rdx", (MONITORrrr)>,
3797 Requires<[In64BitMode]>;
3799 //===---------------------------------------------------------------------===//
3800 // Non-Instruction Patterns
3801 //===---------------------------------------------------------------------===//
3803 // extload f32 -> f64. This matches load+fextend because we have a hack in
3804 // the isel (PreprocessForFPConvert) that can introduce loads after dag
3806 // Since these loads aren't folded into the fextend, we have to match it
3808 let Predicates = [HasSSE2] in
3809 def : Pat<(fextend (loadf32 addr:$src)),
3810 (CVTSS2SDrm addr:$src)>;
3812 // Bitcasts between 128-bit vector types. Return the original type since
3813 // no instruction is needed for the conversion
3814 let Predicates = [HasXMMInt] in {
3815 def : Pat<(v2i64 (bitconvert (v4i32 VR128:$src))), (v2i64 VR128:$src)>;
3816 def : Pat<(v2i64 (bitconvert (v8i16 VR128:$src))), (v2i64 VR128:$src)>;
3817 def : Pat<(v2i64 (bitconvert (v16i8 VR128:$src))), (v2i64 VR128:$src)>;
3818 def : Pat<(v2i64 (bitconvert (v2f64 VR128:$src))), (v2i64 VR128:$src)>;
3819 def : Pat<(v2i64 (bitconvert (v4f32 VR128:$src))), (v2i64 VR128:$src)>;
3820 def : Pat<(v4i32 (bitconvert (v2i64 VR128:$src))), (v4i32 VR128:$src)>;
3821 def : Pat<(v4i32 (bitconvert (v8i16 VR128:$src))), (v4i32 VR128:$src)>;
3822 def : Pat<(v4i32 (bitconvert (v16i8 VR128:$src))), (v4i32 VR128:$src)>;
3823 def : Pat<(v4i32 (bitconvert (v2f64 VR128:$src))), (v4i32 VR128:$src)>;
3824 def : Pat<(v4i32 (bitconvert (v4f32 VR128:$src))), (v4i32 VR128:$src)>;
3825 def : Pat<(v8i16 (bitconvert (v2i64 VR128:$src))), (v8i16 VR128:$src)>;
3826 def : Pat<(v8i16 (bitconvert (v4i32 VR128:$src))), (v8i16 VR128:$src)>;
3827 def : Pat<(v8i16 (bitconvert (v16i8 VR128:$src))), (v8i16 VR128:$src)>;
3828 def : Pat<(v8i16 (bitconvert (v2f64 VR128:$src))), (v8i16 VR128:$src)>;
3829 def : Pat<(v8i16 (bitconvert (v4f32 VR128:$src))), (v8i16 VR128:$src)>;
3830 def : Pat<(v16i8 (bitconvert (v2i64 VR128:$src))), (v16i8 VR128:$src)>;
3831 def : Pat<(v16i8 (bitconvert (v4i32 VR128:$src))), (v16i8 VR128:$src)>;
3832 def : Pat<(v16i8 (bitconvert (v8i16 VR128:$src))), (v16i8 VR128:$src)>;
3833 def : Pat<(v16i8 (bitconvert (v2f64 VR128:$src))), (v16i8 VR128:$src)>;
3834 def : Pat<(v16i8 (bitconvert (v4f32 VR128:$src))), (v16i8 VR128:$src)>;
3835 def : Pat<(v4f32 (bitconvert (v2i64 VR128:$src))), (v4f32 VR128:$src)>;
3836 def : Pat<(v4f32 (bitconvert (v4i32 VR128:$src))), (v4f32 VR128:$src)>;
3837 def : Pat<(v4f32 (bitconvert (v8i16 VR128:$src))), (v4f32 VR128:$src)>;
3838 def : Pat<(v4f32 (bitconvert (v16i8 VR128:$src))), (v4f32 VR128:$src)>;
3839 def : Pat<(v4f32 (bitconvert (v2f64 VR128:$src))), (v4f32 VR128:$src)>;
3840 def : Pat<(v2f64 (bitconvert (v2i64 VR128:$src))), (v2f64 VR128:$src)>;
3841 def : Pat<(v2f64 (bitconvert (v4i32 VR128:$src))), (v2f64 VR128:$src)>;
3842 def : Pat<(v2f64 (bitconvert (v8i16 VR128:$src))), (v2f64 VR128:$src)>;
3843 def : Pat<(v2f64 (bitconvert (v16i8 VR128:$src))), (v2f64 VR128:$src)>;
3844 def : Pat<(v2f64 (bitconvert (v4f32 VR128:$src))), (v2f64 VR128:$src)>;
3847 // Bitcasts between 256-bit vector types. Return the original type since
3848 // no instruction is needed for the conversion
3849 let Predicates = [HasAVX] in {
3850 def : Pat<(v4f64 (bitconvert (v8f32 VR256:$src))), (v4f64 VR256:$src)>;
3851 def : Pat<(v4f64 (bitconvert (v8i32 VR256:$src))), (v4f64 VR256:$src)>;
3852 def : Pat<(v4f64 (bitconvert (v4i64 VR256:$src))), (v4f64 VR256:$src)>;
3853 def : Pat<(v4f64 (bitconvert (v16i16 VR256:$src))), (v4f64 VR256:$src)>;
3854 def : Pat<(v4f64 (bitconvert (v32i8 VR256:$src))), (v4f64 VR256:$src)>;
3855 def : Pat<(v8f32 (bitconvert (v8i32 VR256:$src))), (v8f32 VR256:$src)>;
3856 def : Pat<(v8f32 (bitconvert (v4i64 VR256:$src))), (v8f32 VR256:$src)>;
3857 def : Pat<(v8f32 (bitconvert (v4f64 VR256:$src))), (v8f32 VR256:$src)>;
3858 def : Pat<(v8f32 (bitconvert (v32i8 VR256:$src))), (v8f32 VR256:$src)>;
3859 def : Pat<(v8f32 (bitconvert (v16i16 VR256:$src))), (v8f32 VR256:$src)>;
3860 def : Pat<(v4i64 (bitconvert (v8f32 VR256:$src))), (v4i64 VR256:$src)>;
3861 def : Pat<(v4i64 (bitconvert (v8i32 VR256:$src))), (v4i64 VR256:$src)>;
3862 def : Pat<(v4i64 (bitconvert (v4f64 VR256:$src))), (v4i64 VR256:$src)>;
3863 def : Pat<(v4i64 (bitconvert (v32i8 VR256:$src))), (v4i64 VR256:$src)>;
3864 def : Pat<(v4i64 (bitconvert (v16i16 VR256:$src))), (v4i64 VR256:$src)>;
3865 def : Pat<(v32i8 (bitconvert (v4f64 VR256:$src))), (v32i8 VR256:$src)>;
3866 def : Pat<(v32i8 (bitconvert (v4i64 VR256:$src))), (v32i8 VR256:$src)>;
3867 def : Pat<(v32i8 (bitconvert (v8f32 VR256:$src))), (v32i8 VR256:$src)>;
3868 def : Pat<(v32i8 (bitconvert (v8i32 VR256:$src))), (v32i8 VR256:$src)>;
3869 def : Pat<(v32i8 (bitconvert (v16i16 VR256:$src))), (v32i8 VR256:$src)>;
3870 def : Pat<(v8i32 (bitconvert (v32i8 VR256:$src))), (v8i32 VR256:$src)>;
3871 def : Pat<(v8i32 (bitconvert (v16i16 VR256:$src))), (v8i32 VR256:$src)>;
3872 def : Pat<(v8i32 (bitconvert (v8f32 VR256:$src))), (v8i32 VR256:$src)>;
3873 def : Pat<(v8i32 (bitconvert (v4i64 VR256:$src))), (v8i32 VR256:$src)>;
3874 def : Pat<(v8i32 (bitconvert (v4f64 VR256:$src))), (v8i32 VR256:$src)>;
3875 def : Pat<(v16i16 (bitconvert (v8f32 VR256:$src))), (v16i16 VR256:$src)>;
3876 def : Pat<(v16i16 (bitconvert (v8i32 VR256:$src))), (v16i16 VR256:$src)>;
3877 def : Pat<(v16i16 (bitconvert (v4i64 VR256:$src))), (v16i16 VR256:$src)>;
3878 def : Pat<(v16i16 (bitconvert (v4f64 VR256:$src))), (v16i16 VR256:$src)>;
3879 def : Pat<(v16i16 (bitconvert (v32i8 VR256:$src))), (v16i16 VR256:$src)>;
3882 // Move scalar to XMM zero-extended
3883 // movd to XMM register zero-extends
3884 let AddedComplexity = 15 in {
3885 // Zeroing a VR128 then do a MOVS{S|D} to the lower bits.
3886 def : Pat<(v2f64 (X86vzmovl (v2f64 (scalar_to_vector FR64:$src)))),
3887 (MOVSDrr (v2f64 (V_SET0PS)), FR64:$src)>;
3888 def : Pat<(v4f32 (X86vzmovl (v4f32 (scalar_to_vector FR32:$src)))),
3889 (MOVSSrr (v4f32 (V_SET0PS)), FR32:$src)>;
3890 def : Pat<(v4f32 (X86vzmovl (v4f32 VR128:$src))),
3891 (MOVSSrr (v4f32 (V_SET0PS)),
3892 (f32 (EXTRACT_SUBREG (v4f32 VR128:$src), sub_ss)))>;
3893 def : Pat<(v4i32 (X86vzmovl (v4i32 VR128:$src))),
3894 (MOVSSrr (v4i32 (V_SET0PI)),
3895 (EXTRACT_SUBREG (v4i32 VR128:$src), sub_ss))>;
3898 // Splat v2f64 / v2i64
3899 let AddedComplexity = 10 in {
3900 def : Pat<(splat_lo (v2f64 VR128:$src), (undef)),
3901 (UNPCKLPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
3902 def : Pat<(splat_lo (v2i64 VR128:$src), (undef)),
3903 (PUNPCKLQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
3906 // Special unary SHUFPSrri case.
3907 def : Pat<(v4f32 (pshufd:$src3 VR128:$src1, (undef))),
3908 (SHUFPSrri VR128:$src1, VR128:$src1,
3909 (SHUFFLE_get_shuf_imm VR128:$src3))>;
3910 let AddedComplexity = 5 in
3911 def : Pat<(v4f32 (pshufd:$src2 VR128:$src1, (undef))),
3912 (PSHUFDri VR128:$src1, (SHUFFLE_get_shuf_imm VR128:$src2))>,
3913 Requires<[HasSSE2]>;
3914 // Special unary SHUFPDrri case.
3915 def : Pat<(v2i64 (pshufd:$src3 VR128:$src1, (undef))),
3916 (SHUFPDrri VR128:$src1, VR128:$src1,
3917 (SHUFFLE_get_shuf_imm VR128:$src3))>,
3918 Requires<[HasSSE2]>;
3919 // Special unary SHUFPDrri case.
3920 def : Pat<(v2f64 (pshufd:$src3 VR128:$src1, (undef))),
3921 (SHUFPDrri VR128:$src1, VR128:$src1,
3922 (SHUFFLE_get_shuf_imm VR128:$src3))>,
3923 Requires<[HasSSE2]>;
3924 // Unary v4f32 shuffle with PSHUF* in order to fold a load.
3925 def : Pat<(pshufd:$src2 (bc_v4i32 (memopv4f32 addr:$src1)), (undef)),
3926 (PSHUFDmi addr:$src1, (SHUFFLE_get_shuf_imm VR128:$src2))>,
3927 Requires<[HasSSE2]>;
3929 // Special binary v4i32 shuffle cases with SHUFPS.
3930 def : Pat<(v4i32 (shufp:$src3 VR128:$src1, (v4i32 VR128:$src2))),
3931 (SHUFPSrri VR128:$src1, VR128:$src2,
3932 (SHUFFLE_get_shuf_imm VR128:$src3))>,
3933 Requires<[HasSSE2]>;
3934 def : Pat<(v4i32 (shufp:$src3 VR128:$src1, (bc_v4i32 (memopv2i64 addr:$src2)))),
3935 (SHUFPSrmi VR128:$src1, addr:$src2,
3936 (SHUFFLE_get_shuf_imm VR128:$src3))>,
3937 Requires<[HasSSE2]>;
3938 // Special binary v2i64 shuffle cases using SHUFPDrri.
3939 def : Pat<(v2i64 (shufp:$src3 VR128:$src1, VR128:$src2)),
3940 (SHUFPDrri VR128:$src1, VR128:$src2,
3941 (SHUFFLE_get_shuf_imm VR128:$src3))>,
3942 Requires<[HasSSE2]>;
3944 let AddedComplexity = 20 in {
3945 // vector_shuffle v1, v2 <0, 1, 4, 5> using MOVLHPS
3946 def : Pat<(v4i32 (movlhps VR128:$src1, VR128:$src2)),
3947 (MOVLHPSrr VR128:$src1, VR128:$src2)>;
3949 // vector_shuffle v1, v2 <6, 7, 2, 3> using MOVHLPS
3950 def : Pat<(v4i32 (movhlps VR128:$src1, VR128:$src2)),
3951 (MOVHLPSrr VR128:$src1, VR128:$src2)>;
3953 // vector_shuffle v1, undef <2, ?, ?, ?> using MOVHLPS
3954 def : Pat<(v4f32 (movhlps_undef VR128:$src1, (undef))),
3955 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
3956 def : Pat<(v4i32 (movhlps_undef VR128:$src1, (undef))),
3957 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
3960 let AddedComplexity = 20 in {
3961 // vector_shuffle v1, (load v2) <4, 5, 2, 3> using MOVLPS
3962 def : Pat<(v4f32 (movlp VR128:$src1, (load addr:$src2))),
3963 (MOVLPSrm VR128:$src1, addr:$src2)>;
3964 def : Pat<(v2f64 (movlp VR128:$src1, (load addr:$src2))),
3965 (MOVLPDrm VR128:$src1, addr:$src2)>;
3966 def : Pat<(v4i32 (movlp VR128:$src1, (load addr:$src2))),
3967 (MOVLPSrm VR128:$src1, addr:$src2)>;
3968 def : Pat<(v2i64 (movlp VR128:$src1, (load addr:$src2))),
3969 (MOVLPDrm VR128:$src1, addr:$src2)>;
3972 // (store (vector_shuffle (load addr), v2, <4, 5, 2, 3>), addr) using MOVLPS
3973 def : Pat<(store (v4f32 (movlp (load addr:$src1), VR128:$src2)), addr:$src1),
3974 (MOVLPSmr addr:$src1, VR128:$src2)>;
3975 def : Pat<(store (v2f64 (movlp (load addr:$src1), VR128:$src2)), addr:$src1),
3976 (MOVLPDmr addr:$src1, VR128:$src2)>;
3977 def : Pat<(store (v4i32 (movlp (bc_v4i32 (loadv2i64 addr:$src1)), VR128:$src2)),
3979 (MOVLPSmr addr:$src1, VR128:$src2)>;
3980 def : Pat<(store (v2i64 (movlp (load addr:$src1), VR128:$src2)), addr:$src1),
3981 (MOVLPDmr addr:$src1, VR128:$src2)>;
3983 let AddedComplexity = 15 in {
3984 // Setting the lowest element in the vector.
3985 def : Pat<(v4i32 (movl VR128:$src1, VR128:$src2)),
3986 (MOVSSrr (v4i32 VR128:$src1),
3987 (EXTRACT_SUBREG (v4i32 VR128:$src2), sub_ss))>;
3988 def : Pat<(v2i64 (movl VR128:$src1, VR128:$src2)),
3989 (MOVSDrr (v2i64 VR128:$src1),
3990 (EXTRACT_SUBREG (v2i64 VR128:$src2), sub_sd))>;
3992 // vector_shuffle v1, v2 <4, 5, 2, 3> using movsd
3993 def : Pat<(v4f32 (movlp VR128:$src1, VR128:$src2)),
3994 (MOVSDrr VR128:$src1, (EXTRACT_SUBREG VR128:$src2, sub_sd))>,
3995 Requires<[HasSSE2]>;
3996 def : Pat<(v4i32 (movlp VR128:$src1, VR128:$src2)),
3997 (MOVSDrr VR128:$src1, (EXTRACT_SUBREG VR128:$src2, sub_sd))>,
3998 Requires<[HasSSE2]>;
4001 // vector_shuffle v1, v2 <4, 5, 2, 3> using SHUFPSrri (we prefer movsd, but
4002 // fall back to this for SSE1)
4003 def : Pat<(v4f32 (movlp:$src3 VR128:$src1, (v4f32 VR128:$src2))),
4004 (SHUFPSrri VR128:$src2, VR128:$src1,
4005 (SHUFFLE_get_shuf_imm VR128:$src3))>;
4007 // Set lowest element and zero upper elements.
4008 def : Pat<(v2f64 (X86vzmovl (v2f64 VR128:$src))),
4009 (MOVZPQILo2PQIrr VR128:$src)>, Requires<[HasSSE2]>;
4011 // Use movaps / movups for SSE integer load / store (one byte shorter).
4012 // The instructions selected below are then converted to MOVDQA/MOVDQU
4013 // during the SSE domain pass.
4014 let Predicates = [HasSSE1] in {
4015 def : Pat<(alignedloadv4i32 addr:$src),
4016 (MOVAPSrm addr:$src)>;
4017 def : Pat<(loadv4i32 addr:$src),
4018 (MOVUPSrm addr:$src)>;
4019 def : Pat<(alignedloadv2i64 addr:$src),
4020 (MOVAPSrm addr:$src)>;
4021 def : Pat<(loadv2i64 addr:$src),
4022 (MOVUPSrm addr:$src)>;
4024 def : Pat<(alignedstore (v2i64 VR128:$src), addr:$dst),
4025 (MOVAPSmr addr:$dst, VR128:$src)>;
4026 def : Pat<(alignedstore (v4i32 VR128:$src), addr:$dst),
4027 (MOVAPSmr addr:$dst, VR128:$src)>;
4028 def : Pat<(alignedstore (v8i16 VR128:$src), addr:$dst),
4029 (MOVAPSmr addr:$dst, VR128:$src)>;
4030 def : Pat<(alignedstore (v16i8 VR128:$src), addr:$dst),
4031 (MOVAPSmr addr:$dst, VR128:$src)>;
4032 def : Pat<(store (v2i64 VR128:$src), addr:$dst),
4033 (MOVUPSmr addr:$dst, VR128:$src)>;
4034 def : Pat<(store (v4i32 VR128:$src), addr:$dst),
4035 (MOVUPSmr addr:$dst, VR128:$src)>;
4036 def : Pat<(store (v8i16 VR128:$src), addr:$dst),
4037 (MOVUPSmr addr:$dst, VR128:$src)>;
4038 def : Pat<(store (v16i8 VR128:$src), addr:$dst),
4039 (MOVUPSmr addr:$dst, VR128:$src)>;
4042 // Use vmovaps/vmovups for AVX integer load/store.
4043 let Predicates = [HasAVX] in {
4044 // 128-bit load/store
4045 def : Pat<(alignedloadv4i32 addr:$src),
4046 (VMOVAPSrm addr:$src)>;
4047 def : Pat<(loadv4i32 addr:$src),
4048 (VMOVUPSrm addr:$src)>;
4049 def : Pat<(alignedloadv2i64 addr:$src),
4050 (VMOVAPSrm addr:$src)>;
4051 def : Pat<(loadv2i64 addr:$src),
4052 (VMOVUPSrm addr:$src)>;
4054 def : Pat<(alignedstore (v2i64 VR128:$src), addr:$dst),
4055 (VMOVAPSmr addr:$dst, VR128:$src)>;
4056 def : Pat<(alignedstore (v4i32 VR128:$src), addr:$dst),
4057 (VMOVAPSmr addr:$dst, VR128:$src)>;
4058 def : Pat<(alignedstore (v8i16 VR128:$src), addr:$dst),
4059 (VMOVAPSmr addr:$dst, VR128:$src)>;
4060 def : Pat<(alignedstore (v16i8 VR128:$src), addr:$dst),
4061 (VMOVAPSmr addr:$dst, VR128:$src)>;
4062 def : Pat<(store (v2i64 VR128:$src), addr:$dst),
4063 (VMOVUPSmr addr:$dst, VR128:$src)>;
4064 def : Pat<(store (v4i32 VR128:$src), addr:$dst),
4065 (VMOVUPSmr addr:$dst, VR128:$src)>;
4066 def : Pat<(store (v8i16 VR128:$src), addr:$dst),
4067 (VMOVUPSmr addr:$dst, VR128:$src)>;
4068 def : Pat<(store (v16i8 VR128:$src), addr:$dst),
4069 (VMOVUPSmr addr:$dst, VR128:$src)>;
4071 // 256-bit load/store
4072 def : Pat<(alignedloadv4i64 addr:$src),
4073 (VMOVAPSYrm addr:$src)>;
4074 def : Pat<(loadv4i64 addr:$src),
4075 (VMOVUPSYrm addr:$src)>;
4076 def : Pat<(alignedloadv8i32 addr:$src),
4077 (VMOVAPSYrm addr:$src)>;
4078 def : Pat<(loadv8i32 addr:$src),
4079 (VMOVUPSYrm addr:$src)>;
4080 def : Pat<(alignedstore (v4i64 VR256:$src), addr:$dst),
4081 (VMOVAPSYmr addr:$dst, VR256:$src)>;
4082 def : Pat<(alignedstore (v8i32 VR256:$src), addr:$dst),
4083 (VMOVAPSYmr addr:$dst, VR256:$src)>;
4084 def : Pat<(alignedstore (v16i16 VR256:$src), addr:$dst),
4085 (VMOVAPSYmr addr:$dst, VR256:$src)>;
4086 def : Pat<(alignedstore (v32i8 VR256:$src), addr:$dst),
4087 (VMOVAPSYmr addr:$dst, VR256:$src)>;
4088 def : Pat<(store (v4i64 VR256:$src), addr:$dst),
4089 (VMOVUPSYmr addr:$dst, VR256:$src)>;
4090 def : Pat<(store (v8i32 VR256:$src), addr:$dst),
4091 (VMOVUPSYmr addr:$dst, VR256:$src)>;
4092 def : Pat<(store (v16i16 VR256:$src), addr:$dst),
4093 (VMOVUPSYmr addr:$dst, VR256:$src)>;
4094 def : Pat<(store (v32i8 VR256:$src), addr:$dst),
4095 (VMOVUPSYmr addr:$dst, VR256:$src)>;
4098 //===----------------------------------------------------------------------===//
4099 // SSE4.1 - Packed Move with Sign/Zero Extend
4100 //===----------------------------------------------------------------------===//
4102 multiclass SS41I_binop_rm_int8<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
4103 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
4104 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
4105 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
4107 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
4108 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
4110 (IntId (bitconvert (v2i64 (scalar_to_vector (loadi64 addr:$src))))))]>,
4114 let Predicates = [HasAVX] in {
4115 defm VPMOVSXBW : SS41I_binop_rm_int8<0x20, "vpmovsxbw", int_x86_sse41_pmovsxbw>,
4117 defm VPMOVSXWD : SS41I_binop_rm_int8<0x23, "vpmovsxwd", int_x86_sse41_pmovsxwd>,
4119 defm VPMOVSXDQ : SS41I_binop_rm_int8<0x25, "vpmovsxdq", int_x86_sse41_pmovsxdq>,
4121 defm VPMOVZXBW : SS41I_binop_rm_int8<0x30, "vpmovzxbw", int_x86_sse41_pmovzxbw>,
4123 defm VPMOVZXWD : SS41I_binop_rm_int8<0x33, "vpmovzxwd", int_x86_sse41_pmovzxwd>,
4125 defm VPMOVZXDQ : SS41I_binop_rm_int8<0x35, "vpmovzxdq", int_x86_sse41_pmovzxdq>,
4129 defm PMOVSXBW : SS41I_binop_rm_int8<0x20, "pmovsxbw", int_x86_sse41_pmovsxbw>;
4130 defm PMOVSXWD : SS41I_binop_rm_int8<0x23, "pmovsxwd", int_x86_sse41_pmovsxwd>;
4131 defm PMOVSXDQ : SS41I_binop_rm_int8<0x25, "pmovsxdq", int_x86_sse41_pmovsxdq>;
4132 defm PMOVZXBW : SS41I_binop_rm_int8<0x30, "pmovzxbw", int_x86_sse41_pmovzxbw>;
4133 defm PMOVZXWD : SS41I_binop_rm_int8<0x33, "pmovzxwd", int_x86_sse41_pmovzxwd>;
4134 defm PMOVZXDQ : SS41I_binop_rm_int8<0x35, "pmovzxdq", int_x86_sse41_pmovzxdq>;
4136 // Common patterns involving scalar load.
4137 def : Pat<(int_x86_sse41_pmovsxbw (vzmovl_v2i64 addr:$src)),
4138 (PMOVSXBWrm addr:$src)>, Requires<[HasSSE41]>;
4139 def : Pat<(int_x86_sse41_pmovsxbw (vzload_v2i64 addr:$src)),
4140 (PMOVSXBWrm addr:$src)>, Requires<[HasSSE41]>;
4142 def : Pat<(int_x86_sse41_pmovsxwd (vzmovl_v2i64 addr:$src)),
4143 (PMOVSXWDrm addr:$src)>, Requires<[HasSSE41]>;
4144 def : Pat<(int_x86_sse41_pmovsxwd (vzload_v2i64 addr:$src)),
4145 (PMOVSXWDrm addr:$src)>, Requires<[HasSSE41]>;
4147 def : Pat<(int_x86_sse41_pmovsxdq (vzmovl_v2i64 addr:$src)),
4148 (PMOVSXDQrm addr:$src)>, Requires<[HasSSE41]>;
4149 def : Pat<(int_x86_sse41_pmovsxdq (vzload_v2i64 addr:$src)),
4150 (PMOVSXDQrm addr:$src)>, Requires<[HasSSE41]>;
4152 def : Pat<(int_x86_sse41_pmovzxbw (vzmovl_v2i64 addr:$src)),
4153 (PMOVZXBWrm addr:$src)>, Requires<[HasSSE41]>;
4154 def : Pat<(int_x86_sse41_pmovzxbw (vzload_v2i64 addr:$src)),
4155 (PMOVZXBWrm addr:$src)>, Requires<[HasSSE41]>;
4157 def : Pat<(int_x86_sse41_pmovzxwd (vzmovl_v2i64 addr:$src)),
4158 (PMOVZXWDrm addr:$src)>, Requires<[HasSSE41]>;
4159 def : Pat<(int_x86_sse41_pmovzxwd (vzload_v2i64 addr:$src)),
4160 (PMOVZXWDrm addr:$src)>, Requires<[HasSSE41]>;
4162 def : Pat<(int_x86_sse41_pmovzxdq (vzmovl_v2i64 addr:$src)),
4163 (PMOVZXDQrm addr:$src)>, Requires<[HasSSE41]>;
4164 def : Pat<(int_x86_sse41_pmovzxdq (vzload_v2i64 addr:$src)),
4165 (PMOVZXDQrm addr:$src)>, Requires<[HasSSE41]>;
4168 multiclass SS41I_binop_rm_int4<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
4169 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
4170 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
4171 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
4173 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
4174 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
4176 (IntId (bitconvert (v4i32 (scalar_to_vector (loadi32 addr:$src))))))]>,
4180 let Predicates = [HasAVX] in {
4181 defm VPMOVSXBD : SS41I_binop_rm_int4<0x21, "vpmovsxbd", int_x86_sse41_pmovsxbd>,
4183 defm VPMOVSXWQ : SS41I_binop_rm_int4<0x24, "vpmovsxwq", int_x86_sse41_pmovsxwq>,
4185 defm VPMOVZXBD : SS41I_binop_rm_int4<0x31, "vpmovzxbd", int_x86_sse41_pmovzxbd>,
4187 defm VPMOVZXWQ : SS41I_binop_rm_int4<0x34, "vpmovzxwq", int_x86_sse41_pmovzxwq>,
4191 defm PMOVSXBD : SS41I_binop_rm_int4<0x21, "pmovsxbd", int_x86_sse41_pmovsxbd>;
4192 defm PMOVSXWQ : SS41I_binop_rm_int4<0x24, "pmovsxwq", int_x86_sse41_pmovsxwq>;
4193 defm PMOVZXBD : SS41I_binop_rm_int4<0x31, "pmovzxbd", int_x86_sse41_pmovzxbd>;
4194 defm PMOVZXWQ : SS41I_binop_rm_int4<0x34, "pmovzxwq", int_x86_sse41_pmovzxwq>;
4196 // Common patterns involving scalar load
4197 def : Pat<(int_x86_sse41_pmovsxbd (vzmovl_v4i32 addr:$src)),
4198 (PMOVSXBDrm addr:$src)>, Requires<[HasSSE41]>;
4199 def : Pat<(int_x86_sse41_pmovsxwq (vzmovl_v4i32 addr:$src)),
4200 (PMOVSXWQrm addr:$src)>, Requires<[HasSSE41]>;
4202 def : Pat<(int_x86_sse41_pmovzxbd (vzmovl_v4i32 addr:$src)),
4203 (PMOVZXBDrm addr:$src)>, Requires<[HasSSE41]>;
4204 def : Pat<(int_x86_sse41_pmovzxwq (vzmovl_v4i32 addr:$src)),
4205 (PMOVZXWQrm addr:$src)>, Requires<[HasSSE41]>;
4208 multiclass SS41I_binop_rm_int2<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
4209 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
4210 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
4211 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
4213 // Expecting a i16 load any extended to i32 value.
4214 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i16mem:$src),
4215 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
4216 [(set VR128:$dst, (IntId (bitconvert
4217 (v4i32 (scalar_to_vector (loadi16_anyext addr:$src))))))]>,
4221 let Predicates = [HasAVX] in {
4222 defm VPMOVSXBQ : SS41I_binop_rm_int2<0x22, "vpmovsxbq", int_x86_sse41_pmovsxbq>,
4224 defm VPMOVZXBQ : SS41I_binop_rm_int2<0x32, "vpmovzxbq", int_x86_sse41_pmovzxbq>,
4227 defm PMOVSXBQ : SS41I_binop_rm_int2<0x22, "pmovsxbq", int_x86_sse41_pmovsxbq>;
4228 defm PMOVZXBQ : SS41I_binop_rm_int2<0x32, "pmovzxbq", int_x86_sse41_pmovzxbq>;
4230 // Common patterns involving scalar load
4231 def : Pat<(int_x86_sse41_pmovsxbq
4232 (bitconvert (v4i32 (X86vzmovl
4233 (v4i32 (scalar_to_vector (loadi32 addr:$src))))))),
4234 (PMOVSXBQrm addr:$src)>, Requires<[HasSSE41]>;
4236 def : Pat<(int_x86_sse41_pmovzxbq
4237 (bitconvert (v4i32 (X86vzmovl
4238 (v4i32 (scalar_to_vector (loadi32 addr:$src))))))),
4239 (PMOVZXBQrm addr:$src)>, Requires<[HasSSE41]>;
4241 //===----------------------------------------------------------------------===//
4242 // SSE4.1 - Extract Instructions
4243 //===----------------------------------------------------------------------===//
4245 /// SS41I_binop_ext8 - SSE 4.1 extract 8 bits to 32 bit reg or 8 bit mem
4246 multiclass SS41I_extract8<bits<8> opc, string OpcodeStr> {
4247 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
4248 (ins VR128:$src1, i32i8imm:$src2),
4249 !strconcat(OpcodeStr,
4250 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4251 [(set GR32:$dst, (X86pextrb (v16i8 VR128:$src1), imm:$src2))]>,
4253 def mr : SS4AIi8<opc, MRMDestMem, (outs),
4254 (ins i8mem:$dst, VR128:$src1, i32i8imm:$src2),
4255 !strconcat(OpcodeStr,
4256 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4259 // There's an AssertZext in the way of writing the store pattern
4260 // (store (i8 (trunc (X86pextrb (v16i8 VR128:$src1), imm:$src2))), addr:$dst)
4263 let Predicates = [HasAVX] in {
4264 defm VPEXTRB : SS41I_extract8<0x14, "vpextrb">, VEX;
4265 def VPEXTRBrr64 : SS4AIi8<0x14, MRMDestReg, (outs GR64:$dst),
4266 (ins VR128:$src1, i32i8imm:$src2),
4267 "vpextrb\t{$src2, $src1, $dst|$dst, $src1, $src2}", []>, OpSize, VEX;
4270 defm PEXTRB : SS41I_extract8<0x14, "pextrb">;
4273 /// SS41I_extract16 - SSE 4.1 extract 16 bits to memory destination
4274 multiclass SS41I_extract16<bits<8> opc, string OpcodeStr> {
4275 def mr : SS4AIi8<opc, MRMDestMem, (outs),
4276 (ins i16mem:$dst, VR128:$src1, i32i8imm:$src2),
4277 !strconcat(OpcodeStr,
4278 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4281 // There's an AssertZext in the way of writing the store pattern
4282 // (store (i16 (trunc (X86pextrw (v16i8 VR128:$src1), imm:$src2))), addr:$dst)
4285 let Predicates = [HasAVX] in
4286 defm VPEXTRW : SS41I_extract16<0x15, "vpextrw">, VEX;
4288 defm PEXTRW : SS41I_extract16<0x15, "pextrw">;
4291 /// SS41I_extract32 - SSE 4.1 extract 32 bits to int reg or memory destination
4292 multiclass SS41I_extract32<bits<8> opc, string OpcodeStr> {
4293 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
4294 (ins VR128:$src1, i32i8imm:$src2),
4295 !strconcat(OpcodeStr,
4296 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4298 (extractelt (v4i32 VR128:$src1), imm:$src2))]>, OpSize;
4299 def mr : SS4AIi8<opc, MRMDestMem, (outs),
4300 (ins i32mem:$dst, VR128:$src1, i32i8imm:$src2),
4301 !strconcat(OpcodeStr,
4302 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4303 [(store (extractelt (v4i32 VR128:$src1), imm:$src2),
4304 addr:$dst)]>, OpSize;
4307 let Predicates = [HasAVX] in
4308 defm VPEXTRD : SS41I_extract32<0x16, "vpextrd">, VEX;
4310 defm PEXTRD : SS41I_extract32<0x16, "pextrd">;
4312 /// SS41I_extract32 - SSE 4.1 extract 32 bits to int reg or memory destination
4313 multiclass SS41I_extract64<bits<8> opc, string OpcodeStr> {
4314 def rr : SS4AIi8<opc, MRMDestReg, (outs GR64:$dst),
4315 (ins VR128:$src1, i32i8imm:$src2),
4316 !strconcat(OpcodeStr,
4317 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4319 (extractelt (v2i64 VR128:$src1), imm:$src2))]>, OpSize, REX_W;
4320 def mr : SS4AIi8<opc, MRMDestMem, (outs),
4321 (ins i64mem:$dst, VR128:$src1, i32i8imm:$src2),
4322 !strconcat(OpcodeStr,
4323 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4324 [(store (extractelt (v2i64 VR128:$src1), imm:$src2),
4325 addr:$dst)]>, OpSize, REX_W;
4328 let Predicates = [HasAVX] in
4329 defm VPEXTRQ : SS41I_extract64<0x16, "vpextrq">, VEX, VEX_W;
4331 defm PEXTRQ : SS41I_extract64<0x16, "pextrq">;
4333 /// SS41I_extractf32 - SSE 4.1 extract 32 bits fp value to int reg or memory
4335 multiclass SS41I_extractf32<bits<8> opc, string OpcodeStr> {
4336 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
4337 (ins VR128:$src1, i32i8imm:$src2),
4338 !strconcat(OpcodeStr,
4339 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4341 (extractelt (bc_v4i32 (v4f32 VR128:$src1)), imm:$src2))]>,
4343 def mr : SS4AIi8<opc, MRMDestMem, (outs),
4344 (ins f32mem:$dst, VR128:$src1, i32i8imm:$src2),
4345 !strconcat(OpcodeStr,
4346 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4347 [(store (extractelt (bc_v4i32 (v4f32 VR128:$src1)), imm:$src2),
4348 addr:$dst)]>, OpSize;
4351 let Predicates = [HasAVX] in {
4352 defm VEXTRACTPS : SS41I_extractf32<0x17, "vextractps">, VEX;
4353 def VEXTRACTPSrr64 : SS4AIi8<0x17, MRMDestReg, (outs GR64:$dst),
4354 (ins VR128:$src1, i32i8imm:$src2),
4355 "vextractps \t{$src2, $src1, $dst|$dst, $src1, $src2}",
4358 defm EXTRACTPS : SS41I_extractf32<0x17, "extractps">;
4360 // Also match an EXTRACTPS store when the store is done as f32 instead of i32.
4361 def : Pat<(store (f32 (bitconvert (extractelt (bc_v4i32 (v4f32 VR128:$src1)),
4364 (EXTRACTPSmr addr:$dst, VR128:$src1, imm:$src2)>,
4365 Requires<[HasSSE41]>;
4367 //===----------------------------------------------------------------------===//
4368 // SSE4.1 - Insert Instructions
4369 //===----------------------------------------------------------------------===//
4371 multiclass SS41I_insert8<bits<8> opc, string asm, bit Is2Addr = 1> {
4372 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
4373 (ins VR128:$src1, GR32:$src2, i32i8imm:$src3),
4375 !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
4377 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
4379 (X86pinsrb VR128:$src1, GR32:$src2, imm:$src3))]>, OpSize;
4380 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
4381 (ins VR128:$src1, i8mem:$src2, i32i8imm:$src3),
4383 !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
4385 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
4387 (X86pinsrb VR128:$src1, (extloadi8 addr:$src2),
4388 imm:$src3))]>, OpSize;
4391 let Predicates = [HasAVX] in
4392 defm VPINSRB : SS41I_insert8<0x20, "vpinsrb", 0>, VEX_4V;
4393 let Constraints = "$src1 = $dst" in
4394 defm PINSRB : SS41I_insert8<0x20, "pinsrb">;
4396 multiclass SS41I_insert32<bits<8> opc, string asm, bit Is2Addr = 1> {
4397 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
4398 (ins VR128:$src1, GR32:$src2, i32i8imm:$src3),
4400 !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
4402 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
4404 (v4i32 (insertelt VR128:$src1, GR32:$src2, imm:$src3)))]>,
4406 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
4407 (ins VR128:$src1, i32mem:$src2, i32i8imm:$src3),
4409 !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
4411 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
4413 (v4i32 (insertelt VR128:$src1, (loadi32 addr:$src2),
4414 imm:$src3)))]>, OpSize;
4417 let Predicates = [HasAVX] in
4418 defm VPINSRD : SS41I_insert32<0x22, "vpinsrd", 0>, VEX_4V;
4419 let Constraints = "$src1 = $dst" in
4420 defm PINSRD : SS41I_insert32<0x22, "pinsrd">;
4422 multiclass SS41I_insert64<bits<8> opc, string asm, bit Is2Addr = 1> {
4423 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
4424 (ins VR128:$src1, GR64:$src2, i32i8imm:$src3),
4426 !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
4428 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
4430 (v2i64 (insertelt VR128:$src1, GR64:$src2, imm:$src3)))]>,
4432 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
4433 (ins VR128:$src1, i64mem:$src2, i32i8imm:$src3),
4435 !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
4437 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
4439 (v2i64 (insertelt VR128:$src1, (loadi64 addr:$src2),
4440 imm:$src3)))]>, OpSize;
4443 let Predicates = [HasAVX] in
4444 defm VPINSRQ : SS41I_insert64<0x22, "vpinsrq", 0>, VEX_4V, VEX_W;
4445 let Constraints = "$src1 = $dst" in
4446 defm PINSRQ : SS41I_insert64<0x22, "pinsrq">, REX_W;
4448 // insertps has a few different modes, there's the first two here below which
4449 // are optimized inserts that won't zero arbitrary elements in the destination
4450 // vector. The next one matches the intrinsic and could zero arbitrary elements
4451 // in the target vector.
4452 multiclass SS41I_insertf32<bits<8> opc, string asm, bit Is2Addr = 1> {
4453 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
4454 (ins VR128:$src1, VR128:$src2, u32u8imm:$src3),
4456 !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
4458 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
4460 (X86insrtps VR128:$src1, VR128:$src2, imm:$src3))]>,
4462 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
4463 (ins VR128:$src1, f32mem:$src2, u32u8imm:$src3),
4465 !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
4467 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
4469 (X86insrtps VR128:$src1,
4470 (v4f32 (scalar_to_vector (loadf32 addr:$src2))),
4471 imm:$src3))]>, OpSize;
4474 let Constraints = "$src1 = $dst" in
4475 defm INSERTPS : SS41I_insertf32<0x21, "insertps">;
4476 let Predicates = [HasAVX] in
4477 defm VINSERTPS : SS41I_insertf32<0x21, "vinsertps", 0>, VEX_4V;
4479 def : Pat<(int_x86_sse41_insertps VR128:$src1, VR128:$src2, imm:$src3),
4480 (VINSERTPSrr VR128:$src1, VR128:$src2, imm:$src3)>,
4482 def : Pat<(int_x86_sse41_insertps VR128:$src1, VR128:$src2, imm:$src3),
4483 (INSERTPSrr VR128:$src1, VR128:$src2, imm:$src3)>,
4484 Requires<[HasSSE41]>;
4486 //===----------------------------------------------------------------------===//
4487 // SSE4.1 - Round Instructions
4488 //===----------------------------------------------------------------------===//
4490 multiclass sse41_fp_unop_rm<bits<8> opcps, bits<8> opcpd, string OpcodeStr,
4491 X86MemOperand x86memop, RegisterClass RC,
4492 PatFrag mem_frag32, PatFrag mem_frag64,
4493 Intrinsic V4F32Int, Intrinsic V2F64Int> {
4494 // Intrinsic operation, reg.
4495 // Vector intrinsic operation, reg
4496 def PSr : SS4AIi8<opcps, MRMSrcReg,
4497 (outs RC:$dst), (ins RC:$src1, i32i8imm:$src2),
4498 !strconcat(OpcodeStr,
4499 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4500 [(set RC:$dst, (V4F32Int RC:$src1, imm:$src2))]>,
4503 // Vector intrinsic operation, mem
4504 def PSm : Ii8<opcps, MRMSrcMem,
4505 (outs RC:$dst), (ins f256mem:$src1, i32i8imm:$src2),
4506 !strconcat(OpcodeStr,
4507 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4509 (V4F32Int (mem_frag32 addr:$src1),imm:$src2))]>,
4511 Requires<[HasSSE41]>;
4513 // Vector intrinsic operation, reg
4514 def PDr : SS4AIi8<opcpd, MRMSrcReg,
4515 (outs RC:$dst), (ins RC:$src1, i32i8imm:$src2),
4516 !strconcat(OpcodeStr,
4517 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4518 [(set RC:$dst, (V2F64Int RC:$src1, imm:$src2))]>,
4521 // Vector intrinsic operation, mem
4522 def PDm : SS4AIi8<opcpd, MRMSrcMem,
4523 (outs RC:$dst), (ins f256mem:$src1, i32i8imm:$src2),
4524 !strconcat(OpcodeStr,
4525 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4527 (V2F64Int (mem_frag64 addr:$src1),imm:$src2))]>,
4531 multiclass sse41_fp_unop_rm_avx_p<bits<8> opcps, bits<8> opcpd,
4532 RegisterClass RC, X86MemOperand x86memop, string OpcodeStr> {
4533 // Intrinsic operation, reg.
4534 // Vector intrinsic operation, reg
4535 def PSr_AVX : SS4AIi8<opcps, MRMSrcReg,
4536 (outs RC:$dst), (ins RC:$src1, i32i8imm:$src2),
4537 !strconcat(OpcodeStr,
4538 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4541 // Vector intrinsic operation, mem
4542 def PSm_AVX : Ii8<opcps, MRMSrcMem,
4543 (outs RC:$dst), (ins x86memop:$src1, i32i8imm:$src2),
4544 !strconcat(OpcodeStr,
4545 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4546 []>, TA, OpSize, Requires<[HasSSE41]>;
4548 // Vector intrinsic operation, reg
4549 def PDr_AVX : SS4AIi8<opcpd, MRMSrcReg,
4550 (outs RC:$dst), (ins RC:$src1, i32i8imm:$src2),
4551 !strconcat(OpcodeStr,
4552 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4555 // Vector intrinsic operation, mem
4556 def PDm_AVX : SS4AIi8<opcpd, MRMSrcMem,
4557 (outs RC:$dst), (ins x86memop:$src1, i32i8imm:$src2),
4558 !strconcat(OpcodeStr,
4559 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
4563 multiclass sse41_fp_binop_rm<bits<8> opcss, bits<8> opcsd,
4566 Intrinsic F64Int, bit Is2Addr = 1> {
4567 // Intrinsic operation, reg.
4568 def SSr : SS4AIi8<opcss, MRMSrcReg,
4569 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
4571 !strconcat(OpcodeStr,
4572 "ss\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
4573 !strconcat(OpcodeStr,
4574 "ss\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
4575 [(set VR128:$dst, (F32Int VR128:$src1, VR128:$src2, imm:$src3))]>,
4578 // Intrinsic operation, mem.
4579 def SSm : SS4AIi8<opcss, MRMSrcMem,
4580 (outs VR128:$dst), (ins VR128:$src1, ssmem:$src2, i32i8imm:$src3),
4582 !strconcat(OpcodeStr,
4583 "ss\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
4584 !strconcat(OpcodeStr,
4585 "ss\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
4587 (F32Int VR128:$src1, sse_load_f32:$src2, imm:$src3))]>,
4590 // Intrinsic operation, reg.
4591 def SDr : SS4AIi8<opcsd, MRMSrcReg,
4592 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
4594 !strconcat(OpcodeStr,
4595 "sd\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
4596 !strconcat(OpcodeStr,
4597 "sd\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
4598 [(set VR128:$dst, (F64Int VR128:$src1, VR128:$src2, imm:$src3))]>,
4601 // Intrinsic operation, mem.
4602 def SDm : SS4AIi8<opcsd, MRMSrcMem,
4603 (outs VR128:$dst), (ins VR128:$src1, sdmem:$src2, i32i8imm:$src3),
4605 !strconcat(OpcodeStr,
4606 "sd\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
4607 !strconcat(OpcodeStr,
4608 "sd\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
4610 (F64Int VR128:$src1, sse_load_f64:$src2, imm:$src3))]>,
4614 multiclass sse41_fp_binop_rm_avx_s<bits<8> opcss, bits<8> opcsd,
4616 // Intrinsic operation, reg.
4617 def SSr_AVX : SS4AIi8<opcss, MRMSrcReg,
4618 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
4619 !strconcat(OpcodeStr,
4620 "ss\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
4623 // Intrinsic operation, mem.
4624 def SSm_AVX : SS4AIi8<opcss, MRMSrcMem,
4625 (outs VR128:$dst), (ins VR128:$src1, ssmem:$src2, i32i8imm:$src3),
4626 !strconcat(OpcodeStr,
4627 "ss\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
4630 // Intrinsic operation, reg.
4631 def SDr_AVX : SS4AIi8<opcsd, MRMSrcReg,
4632 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
4633 !strconcat(OpcodeStr,
4634 "sd\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
4637 // Intrinsic operation, mem.
4638 def SDm_AVX : SS4AIi8<opcsd, MRMSrcMem,
4639 (outs VR128:$dst), (ins VR128:$src1, sdmem:$src2, i32i8imm:$src3),
4640 !strconcat(OpcodeStr,
4641 "sd\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
4645 // FP round - roundss, roundps, roundsd, roundpd
4646 let Predicates = [HasAVX] in {
4648 defm VROUND : sse41_fp_unop_rm<0x08, 0x09, "vround", f128mem, VR128,
4649 memopv4f32, memopv2f64,
4650 int_x86_sse41_round_ps,
4651 int_x86_sse41_round_pd>, VEX;
4652 defm VROUNDY : sse41_fp_unop_rm<0x08, 0x09, "vround", f256mem, VR256,
4653 memopv8f32, memopv4f64,
4654 int_x86_avx_round_ps_256,
4655 int_x86_avx_round_pd_256>, VEX;
4656 defm VROUND : sse41_fp_binop_rm<0x0A, 0x0B, "vround",
4657 int_x86_sse41_round_ss,
4658 int_x86_sse41_round_sd, 0>, VEX_4V;
4660 // Instructions for the assembler
4661 defm VROUND : sse41_fp_unop_rm_avx_p<0x08, 0x09, VR128, f128mem, "vround">,
4663 defm VROUNDY : sse41_fp_unop_rm_avx_p<0x08, 0x09, VR256, f256mem, "vround">,
4665 defm VROUND : sse41_fp_binop_rm_avx_s<0x0A, 0x0B, "vround">, VEX_4V;
4668 defm ROUND : sse41_fp_unop_rm<0x08, 0x09, "round", f128mem, VR128,
4669 memopv4f32, memopv2f64,
4670 int_x86_sse41_round_ps, int_x86_sse41_round_pd>;
4671 let Constraints = "$src1 = $dst" in
4672 defm ROUND : sse41_fp_binop_rm<0x0A, 0x0B, "round",
4673 int_x86_sse41_round_ss, int_x86_sse41_round_sd>;
4675 //===----------------------------------------------------------------------===//
4676 // SSE4.1 - Packed Bit Test
4677 //===----------------------------------------------------------------------===//
4679 // ptest instruction we'll lower to this in X86ISelLowering primarily from
4680 // the intel intrinsic that corresponds to this.
4681 let Defs = [EFLAGS], Predicates = [HasAVX] in {
4682 def VPTESTrr : SS48I<0x17, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
4683 "vptest\t{$src2, $src1|$src1, $src2}",
4684 [(set EFLAGS, (X86ptest VR128:$src1, (v4f32 VR128:$src2)))]>,
4686 def VPTESTrm : SS48I<0x17, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
4687 "vptest\t{$src2, $src1|$src1, $src2}",
4688 [(set EFLAGS,(X86ptest VR128:$src1, (memopv4f32 addr:$src2)))]>,
4691 def VPTESTYrr : SS48I<0x17, MRMSrcReg, (outs), (ins VR256:$src1, VR256:$src2),
4692 "vptest\t{$src2, $src1|$src1, $src2}",
4693 [(set EFLAGS, (X86ptest VR256:$src1, (v4i64 VR256:$src2)))]>,
4695 def VPTESTYrm : SS48I<0x17, MRMSrcMem, (outs), (ins VR256:$src1, i256mem:$src2),
4696 "vptest\t{$src2, $src1|$src1, $src2}",
4697 [(set EFLAGS,(X86ptest VR256:$src1, (memopv4i64 addr:$src2)))]>,
4701 let Defs = [EFLAGS] in {
4702 def PTESTrr : SS48I<0x17, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
4703 "ptest \t{$src2, $src1|$src1, $src2}",
4704 [(set EFLAGS, (X86ptest VR128:$src1, (v4f32 VR128:$src2)))]>,
4706 def PTESTrm : SS48I<0x17, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
4707 "ptest \t{$src2, $src1|$src1, $src2}",
4708 [(set EFLAGS, (X86ptest VR128:$src1, (memopv4f32 addr:$src2)))]>,
4712 // The bit test instructions below are AVX only
4713 multiclass avx_bittest<bits<8> opc, string OpcodeStr, RegisterClass RC,
4714 X86MemOperand x86memop, PatFrag mem_frag, ValueType vt> {
4715 def rr : SS48I<opc, MRMSrcReg, (outs), (ins RC:$src1, RC:$src2),
4716 !strconcat(OpcodeStr, "\t{$src2, $src1|$src1, $src2}"),
4717 [(set EFLAGS, (X86testp RC:$src1, (vt RC:$src2)))]>, OpSize, VEX;
4718 def rm : SS48I<opc, MRMSrcMem, (outs), (ins RC:$src1, x86memop:$src2),
4719 !strconcat(OpcodeStr, "\t{$src2, $src1|$src1, $src2}"),
4720 [(set EFLAGS, (X86testp RC:$src1, (mem_frag addr:$src2)))]>,
4724 let Defs = [EFLAGS], Predicates = [HasAVX] in {
4725 defm VTESTPS : avx_bittest<0x0E, "vtestps", VR128, f128mem, memopv4f32, v4f32>;
4726 defm VTESTPSY : avx_bittest<0x0E, "vtestps", VR256, f256mem, memopv8f32, v8f32>;
4727 defm VTESTPD : avx_bittest<0x0F, "vtestpd", VR128, f128mem, memopv2f64, v2f64>;
4728 defm VTESTPDY : avx_bittest<0x0F, "vtestpd", VR256, f256mem, memopv4f64, v4f64>;
4731 //===----------------------------------------------------------------------===//
4732 // SSE4.1 - Misc Instructions
4733 //===----------------------------------------------------------------------===//
4735 def POPCNT16rr : I<0xB8, MRMSrcReg, (outs GR16:$dst), (ins GR16:$src),
4736 "popcnt{w}\t{$src, $dst|$dst, $src}",
4737 [(set GR16:$dst, (ctpop GR16:$src))]>, OpSize, XS;
4738 def POPCNT16rm : I<0xB8, MRMSrcMem, (outs GR16:$dst), (ins i16mem:$src),
4739 "popcnt{w}\t{$src, $dst|$dst, $src}",
4740 [(set GR16:$dst, (ctpop (loadi16 addr:$src)))]>, OpSize, XS;
4742 def POPCNT32rr : I<0xB8, MRMSrcReg, (outs GR32:$dst), (ins GR32:$src),
4743 "popcnt{l}\t{$src, $dst|$dst, $src}",
4744 [(set GR32:$dst, (ctpop GR32:$src))]>, XS;
4745 def POPCNT32rm : I<0xB8, MRMSrcMem, (outs GR32:$dst), (ins i32mem:$src),
4746 "popcnt{l}\t{$src, $dst|$dst, $src}",
4747 [(set GR32:$dst, (ctpop (loadi32 addr:$src)))]>, XS;
4749 def POPCNT64rr : RI<0xB8, MRMSrcReg, (outs GR64:$dst), (ins GR64:$src),
4750 "popcnt{q}\t{$src, $dst|$dst, $src}",
4751 [(set GR64:$dst, (ctpop GR64:$src))]>, XS;
4752 def POPCNT64rm : RI<0xB8, MRMSrcMem, (outs GR64:$dst), (ins i64mem:$src),
4753 "popcnt{q}\t{$src, $dst|$dst, $src}",
4754 [(set GR64:$dst, (ctpop (loadi64 addr:$src)))]>, XS;
4758 // SS41I_unop_rm_int_v16 - SSE 4.1 unary operator whose type is v8i16.
4759 multiclass SS41I_unop_rm_int_v16<bits<8> opc, string OpcodeStr,
4760 Intrinsic IntId128> {
4761 def rr128 : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
4763 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
4764 [(set VR128:$dst, (IntId128 VR128:$src))]>, OpSize;
4765 def rm128 : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
4767 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
4770 (bitconvert (memopv8i16 addr:$src))))]>, OpSize;
4773 let Predicates = [HasAVX] in
4774 defm VPHMINPOSUW : SS41I_unop_rm_int_v16 <0x41, "vphminposuw",
4775 int_x86_sse41_phminposuw>, VEX;
4776 defm PHMINPOSUW : SS41I_unop_rm_int_v16 <0x41, "phminposuw",
4777 int_x86_sse41_phminposuw>;
4779 /// SS41I_binop_rm_int - Simple SSE 4.1 binary operator
4780 multiclass SS41I_binop_rm_int<bits<8> opc, string OpcodeStr,
4781 Intrinsic IntId128, bit Is2Addr = 1> {
4782 let isCommutable = 1 in
4783 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
4784 (ins VR128:$src1, VR128:$src2),
4786 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
4787 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
4788 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>, OpSize;
4789 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
4790 (ins VR128:$src1, i128mem:$src2),
4792 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
4793 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
4795 (IntId128 VR128:$src1,
4796 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
4799 let Predicates = [HasAVX] in {
4800 let isCommutable = 0 in
4801 defm VPACKUSDW : SS41I_binop_rm_int<0x2B, "vpackusdw", int_x86_sse41_packusdw,
4803 defm VPCMPEQQ : SS41I_binop_rm_int<0x29, "vpcmpeqq", int_x86_sse41_pcmpeqq,
4805 defm VPMINSB : SS41I_binop_rm_int<0x38, "vpminsb", int_x86_sse41_pminsb,
4807 defm VPMINSD : SS41I_binop_rm_int<0x39, "vpminsd", int_x86_sse41_pminsd,
4809 defm VPMINUD : SS41I_binop_rm_int<0x3B, "vpminud", int_x86_sse41_pminud,
4811 defm VPMINUW : SS41I_binop_rm_int<0x3A, "vpminuw", int_x86_sse41_pminuw,
4813 defm VPMAXSB : SS41I_binop_rm_int<0x3C, "vpmaxsb", int_x86_sse41_pmaxsb,
4815 defm VPMAXSD : SS41I_binop_rm_int<0x3D, "vpmaxsd", int_x86_sse41_pmaxsd,
4817 defm VPMAXUD : SS41I_binop_rm_int<0x3F, "vpmaxud", int_x86_sse41_pmaxud,
4819 defm VPMAXUW : SS41I_binop_rm_int<0x3E, "vpmaxuw", int_x86_sse41_pmaxuw,
4821 defm VPMULDQ : SS41I_binop_rm_int<0x28, "vpmuldq", int_x86_sse41_pmuldq,
4825 let Constraints = "$src1 = $dst" in {
4826 let isCommutable = 0 in
4827 defm PACKUSDW : SS41I_binop_rm_int<0x2B, "packusdw", int_x86_sse41_packusdw>;
4828 defm PCMPEQQ : SS41I_binop_rm_int<0x29, "pcmpeqq", int_x86_sse41_pcmpeqq>;
4829 defm PMINSB : SS41I_binop_rm_int<0x38, "pminsb", int_x86_sse41_pminsb>;
4830 defm PMINSD : SS41I_binop_rm_int<0x39, "pminsd", int_x86_sse41_pminsd>;
4831 defm PMINUD : SS41I_binop_rm_int<0x3B, "pminud", int_x86_sse41_pminud>;
4832 defm PMINUW : SS41I_binop_rm_int<0x3A, "pminuw", int_x86_sse41_pminuw>;
4833 defm PMAXSB : SS41I_binop_rm_int<0x3C, "pmaxsb", int_x86_sse41_pmaxsb>;
4834 defm PMAXSD : SS41I_binop_rm_int<0x3D, "pmaxsd", int_x86_sse41_pmaxsd>;
4835 defm PMAXUD : SS41I_binop_rm_int<0x3F, "pmaxud", int_x86_sse41_pmaxud>;
4836 defm PMAXUW : SS41I_binop_rm_int<0x3E, "pmaxuw", int_x86_sse41_pmaxuw>;
4837 defm PMULDQ : SS41I_binop_rm_int<0x28, "pmuldq", int_x86_sse41_pmuldq>;
4840 def : Pat<(v2i64 (X86pcmpeqq VR128:$src1, VR128:$src2)),
4841 (PCMPEQQrr VR128:$src1, VR128:$src2)>;
4842 def : Pat<(v2i64 (X86pcmpeqq VR128:$src1, (memop addr:$src2))),
4843 (PCMPEQQrm VR128:$src1, addr:$src2)>;
4845 /// SS48I_binop_rm - Simple SSE41 binary operator.
4846 multiclass SS48I_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
4847 ValueType OpVT, bit Is2Addr = 1> {
4848 let isCommutable = 1 in
4849 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
4850 (ins VR128:$src1, VR128:$src2),
4852 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
4853 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
4854 [(set VR128:$dst, (OpVT (OpNode VR128:$src1, VR128:$src2)))]>,
4856 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
4857 (ins VR128:$src1, i128mem:$src2),
4859 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
4860 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
4861 [(set VR128:$dst, (OpNode VR128:$src1,
4862 (bc_v4i32 (memopv2i64 addr:$src2))))]>,
4866 let Predicates = [HasAVX] in
4867 defm VPMULLD : SS48I_binop_rm<0x40, "vpmulld", mul, v4i32, 0>, VEX_4V;
4868 let Constraints = "$src1 = $dst" in
4869 defm PMULLD : SS48I_binop_rm<0x40, "pmulld", mul, v4i32>;
4871 /// SS41I_binop_rmi_int - SSE 4.1 binary operator with 8-bit immediate
4872 multiclass SS41I_binop_rmi_int<bits<8> opc, string OpcodeStr,
4873 Intrinsic IntId, RegisterClass RC, PatFrag memop_frag,
4874 X86MemOperand x86memop, bit Is2Addr = 1> {
4875 let isCommutable = 1 in
4876 def rri : SS4AIi8<opc, MRMSrcReg, (outs RC:$dst),
4877 (ins RC:$src1, RC:$src2, u32u8imm:$src3),
4879 !strconcat(OpcodeStr,
4880 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
4881 !strconcat(OpcodeStr,
4882 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
4883 [(set RC:$dst, (IntId RC:$src1, RC:$src2, imm:$src3))]>,
4885 def rmi : SS4AIi8<opc, MRMSrcMem, (outs RC:$dst),
4886 (ins RC:$src1, x86memop:$src2, u32u8imm:$src3),
4888 !strconcat(OpcodeStr,
4889 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
4890 !strconcat(OpcodeStr,
4891 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
4894 (bitconvert (memop_frag addr:$src2)), imm:$src3))]>,
4898 let Predicates = [HasAVX] in {
4899 let isCommutable = 0 in {
4900 defm VBLENDPS : SS41I_binop_rmi_int<0x0C, "vblendps", int_x86_sse41_blendps,
4901 VR128, memopv16i8, i128mem, 0>, VEX_4V;
4902 defm VBLENDPD : SS41I_binop_rmi_int<0x0D, "vblendpd", int_x86_sse41_blendpd,
4903 VR128, memopv16i8, i128mem, 0>, VEX_4V;
4904 defm VBLENDPSY : SS41I_binop_rmi_int<0x0C, "vblendps",
4905 int_x86_avx_blend_ps_256, VR256, memopv32i8, i256mem, 0>, VEX_4V;
4906 defm VBLENDPDY : SS41I_binop_rmi_int<0x0D, "vblendpd",
4907 int_x86_avx_blend_pd_256, VR256, memopv32i8, i256mem, 0>, VEX_4V;
4908 defm VPBLENDW : SS41I_binop_rmi_int<0x0E, "vpblendw", int_x86_sse41_pblendw,
4909 VR128, memopv16i8, i128mem, 0>, VEX_4V;
4910 defm VMPSADBW : SS41I_binop_rmi_int<0x42, "vmpsadbw", int_x86_sse41_mpsadbw,
4911 VR128, memopv16i8, i128mem, 0>, VEX_4V;
4913 defm VDPPS : SS41I_binop_rmi_int<0x40, "vdpps", int_x86_sse41_dpps,
4914 VR128, memopv16i8, i128mem, 0>, VEX_4V;
4915 defm VDPPD : SS41I_binop_rmi_int<0x41, "vdppd", int_x86_sse41_dppd,
4916 VR128, memopv16i8, i128mem, 0>, VEX_4V;
4917 defm VDPPSY : SS41I_binop_rmi_int<0x40, "vdpps", int_x86_avx_dp_ps_256,
4918 VR256, memopv32i8, i256mem, 0>, VEX_4V;
4921 let Constraints = "$src1 = $dst" in {
4922 let isCommutable = 0 in {
4923 defm BLENDPS : SS41I_binop_rmi_int<0x0C, "blendps", int_x86_sse41_blendps,
4924 VR128, memopv16i8, i128mem>;
4925 defm BLENDPD : SS41I_binop_rmi_int<0x0D, "blendpd", int_x86_sse41_blendpd,
4926 VR128, memopv16i8, i128mem>;
4927 defm PBLENDW : SS41I_binop_rmi_int<0x0E, "pblendw", int_x86_sse41_pblendw,
4928 VR128, memopv16i8, i128mem>;
4929 defm MPSADBW : SS41I_binop_rmi_int<0x42, "mpsadbw", int_x86_sse41_mpsadbw,
4930 VR128, memopv16i8, i128mem>;
4932 defm DPPS : SS41I_binop_rmi_int<0x40, "dpps", int_x86_sse41_dpps,
4933 VR128, memopv16i8, i128mem>;
4934 defm DPPD : SS41I_binop_rmi_int<0x41, "dppd", int_x86_sse41_dppd,
4935 VR128, memopv16i8, i128mem>;
4938 /// SS41I_quaternary_int_avx - AVX SSE 4.1 with 4 operators
4939 let Predicates = [HasAVX] in {
4940 multiclass SS41I_quaternary_int_avx<bits<8> opc, string OpcodeStr,
4941 RegisterClass RC, X86MemOperand x86memop,
4942 PatFrag mem_frag, Intrinsic IntId> {
4943 def rr : I<opc, MRMSrcReg, (outs RC:$dst),
4944 (ins RC:$src1, RC:$src2, RC:$src3),
4945 !strconcat(OpcodeStr,
4946 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
4947 [(set RC:$dst, (IntId RC:$src1, RC:$src2, RC:$src3))],
4948 SSEPackedInt>, OpSize, TA, VEX_4V, VEX_I8IMM;
4950 def rm : I<opc, MRMSrcMem, (outs RC:$dst),
4951 (ins RC:$src1, x86memop:$src2, RC:$src3),
4952 !strconcat(OpcodeStr,
4953 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
4955 (IntId RC:$src1, (bitconvert (mem_frag addr:$src2)),
4957 SSEPackedInt>, OpSize, TA, VEX_4V, VEX_I8IMM;
4961 defm VBLENDVPD : SS41I_quaternary_int_avx<0x4B, "vblendvpd", VR128, i128mem,
4962 memopv16i8, int_x86_sse41_blendvpd>;
4963 defm VBLENDVPS : SS41I_quaternary_int_avx<0x4A, "vblendvps", VR128, i128mem,
4964 memopv16i8, int_x86_sse41_blendvps>;
4965 defm VPBLENDVB : SS41I_quaternary_int_avx<0x4C, "vpblendvb", VR128, i128mem,
4966 memopv16i8, int_x86_sse41_pblendvb>;
4967 defm VBLENDVPDY : SS41I_quaternary_int_avx<0x4B, "vblendvpd", VR256, i256mem,
4968 memopv32i8, int_x86_avx_blendv_pd_256>;
4969 defm VBLENDVPSY : SS41I_quaternary_int_avx<0x4A, "vblendvps", VR256, i256mem,
4970 memopv32i8, int_x86_avx_blendv_ps_256>;
4972 /// SS41I_ternary_int - SSE 4.1 ternary operator
4973 let Uses = [XMM0], Constraints = "$src1 = $dst" in {
4974 multiclass SS41I_ternary_int<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
4975 def rr0 : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
4976 (ins VR128:$src1, VR128:$src2),
4977 !strconcat(OpcodeStr,
4978 "\t{$src2, $dst|$dst, $src2}"),
4979 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2, XMM0))]>,
4982 def rm0 : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
4983 (ins VR128:$src1, i128mem:$src2),
4984 !strconcat(OpcodeStr,
4985 "\t{$src2, $dst|$dst, $src2}"),
4988 (bitconvert (memopv16i8 addr:$src2)), XMM0))]>, OpSize;
4992 defm BLENDVPD : SS41I_ternary_int<0x15, "blendvpd", int_x86_sse41_blendvpd>;
4993 defm BLENDVPS : SS41I_ternary_int<0x14, "blendvps", int_x86_sse41_blendvps>;
4994 defm PBLENDVB : SS41I_ternary_int<0x10, "pblendvb", int_x86_sse41_pblendvb>;
4996 def : Pat<(X86pblendv VR128:$src1, VR128:$src2, XMM0),
4997 (PBLENDVBrr0 VR128:$src1, VR128:$src2)>;
4999 let Predicates = [HasAVX] in
5000 def VMOVNTDQArm : SS48I<0x2A, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
5001 "vmovntdqa\t{$src, $dst|$dst, $src}",
5002 [(set VR128:$dst, (int_x86_sse41_movntdqa addr:$src))]>,
5004 def MOVNTDQArm : SS48I<0x2A, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
5005 "movntdqa\t{$src, $dst|$dst, $src}",
5006 [(set VR128:$dst, (int_x86_sse41_movntdqa addr:$src))]>,
5009 //===----------------------------------------------------------------------===//
5010 // SSE4.2 - Compare Instructions
5011 //===----------------------------------------------------------------------===//
5013 /// SS42I_binop_rm_int - Simple SSE 4.2 binary operator
5014 multiclass SS42I_binop_rm_int<bits<8> opc, string OpcodeStr,
5015 Intrinsic IntId128, bit Is2Addr = 1> {
5016 def rr : SS428I<opc, MRMSrcReg, (outs VR128:$dst),
5017 (ins VR128:$src1, VR128:$src2),
5019 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
5020 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
5021 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
5023 def rm : SS428I<opc, MRMSrcMem, (outs VR128:$dst),
5024 (ins VR128:$src1, i128mem:$src2),
5026 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
5027 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
5029 (IntId128 VR128:$src1,
5030 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
5033 let Predicates = [HasAVX] in
5034 defm VPCMPGTQ : SS42I_binop_rm_int<0x37, "vpcmpgtq", int_x86_sse42_pcmpgtq,
5036 let Constraints = "$src1 = $dst" in
5037 defm PCMPGTQ : SS42I_binop_rm_int<0x37, "pcmpgtq", int_x86_sse42_pcmpgtq>;
5039 def : Pat<(v2i64 (X86pcmpgtq VR128:$src1, VR128:$src2)),
5040 (PCMPGTQrr VR128:$src1, VR128:$src2)>;
5041 def : Pat<(v2i64 (X86pcmpgtq VR128:$src1, (memop addr:$src2))),
5042 (PCMPGTQrm VR128:$src1, addr:$src2)>;
5044 //===----------------------------------------------------------------------===//
5045 // SSE4.2 - String/text Processing Instructions
5046 //===----------------------------------------------------------------------===//
5048 // Packed Compare Implicit Length Strings, Return Mask
5049 multiclass pseudo_pcmpistrm<string asm> {
5050 def REG : PseudoI<(outs VR128:$dst),
5051 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
5052 [(set VR128:$dst, (int_x86_sse42_pcmpistrm128 VR128:$src1, VR128:$src2,
5054 def MEM : PseudoI<(outs VR128:$dst),
5055 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
5056 [(set VR128:$dst, (int_x86_sse42_pcmpistrm128
5057 VR128:$src1, (load addr:$src2), imm:$src3))]>;
5060 let Defs = [EFLAGS], usesCustomInserter = 1 in {
5061 defm PCMPISTRM128 : pseudo_pcmpistrm<"#PCMPISTRM128">, Requires<[HasSSE42]>;
5062 defm VPCMPISTRM128 : pseudo_pcmpistrm<"#VPCMPISTRM128">, Requires<[HasAVX]>;
5065 let Defs = [XMM0, EFLAGS], Predicates = [HasAVX] in {
5066 def VPCMPISTRM128rr : SS42AI<0x62, MRMSrcReg, (outs),
5067 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
5068 "vpcmpistrm\t{$src3, $src2, $src1|$src1, $src2, $src3}", []>, OpSize, VEX;
5069 def VPCMPISTRM128rm : SS42AI<0x62, MRMSrcMem, (outs),
5070 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
5071 "vpcmpistrm\t{$src3, $src2, $src1|$src1, $src2, $src3}", []>, OpSize, VEX;
5074 let Defs = [XMM0, EFLAGS] in {
5075 def PCMPISTRM128rr : SS42AI<0x62, MRMSrcReg, (outs),
5076 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
5077 "pcmpistrm\t{$src3, $src2, $src1|$src1, $src2, $src3}", []>, OpSize;
5078 def PCMPISTRM128rm : SS42AI<0x62, MRMSrcMem, (outs),
5079 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
5080 "pcmpistrm\t{$src3, $src2, $src1|$src1, $src2, $src3}", []>, OpSize;
5083 // Packed Compare Explicit Length Strings, Return Mask
5084 multiclass pseudo_pcmpestrm<string asm> {
5085 def REG : PseudoI<(outs VR128:$dst),
5086 (ins VR128:$src1, VR128:$src3, i8imm:$src5),
5087 [(set VR128:$dst, (int_x86_sse42_pcmpestrm128
5088 VR128:$src1, EAX, VR128:$src3, EDX, imm:$src5))]>;
5089 def MEM : PseudoI<(outs VR128:$dst),
5090 (ins VR128:$src1, i128mem:$src3, i8imm:$src5),
5091 [(set VR128:$dst, (int_x86_sse42_pcmpestrm128
5092 VR128:$src1, EAX, (load addr:$src3), EDX, imm:$src5))]>;
5095 let Defs = [EFLAGS], Uses = [EAX, EDX], usesCustomInserter = 1 in {
5096 defm PCMPESTRM128 : pseudo_pcmpestrm<"#PCMPESTRM128">, Requires<[HasSSE42]>;
5097 defm VPCMPESTRM128 : pseudo_pcmpestrm<"#VPCMPESTRM128">, Requires<[HasAVX]>;
5100 let Predicates = [HasAVX],
5101 Defs = [XMM0, EFLAGS], Uses = [EAX, EDX] in {
5102 def VPCMPESTRM128rr : SS42AI<0x60, MRMSrcReg, (outs),
5103 (ins VR128:$src1, VR128:$src3, i8imm:$src5),
5104 "vpcmpestrm\t{$src5, $src3, $src1|$src1, $src3, $src5}", []>, OpSize, VEX;
5105 def VPCMPESTRM128rm : SS42AI<0x60, MRMSrcMem, (outs),
5106 (ins VR128:$src1, i128mem:$src3, i8imm:$src5),
5107 "vpcmpestrm\t{$src5, $src3, $src1|$src1, $src3, $src5}", []>, OpSize, VEX;
5110 let Defs = [XMM0, EFLAGS], Uses = [EAX, EDX] in {
5111 def PCMPESTRM128rr : SS42AI<0x60, MRMSrcReg, (outs),
5112 (ins VR128:$src1, VR128:$src3, i8imm:$src5),
5113 "pcmpestrm\t{$src5, $src3, $src1|$src1, $src3, $src5}", []>, OpSize;
5114 def PCMPESTRM128rm : SS42AI<0x60, MRMSrcMem, (outs),
5115 (ins VR128:$src1, i128mem:$src3, i8imm:$src5),
5116 "pcmpestrm\t{$src5, $src3, $src1|$src1, $src3, $src5}", []>, OpSize;
5119 // Packed Compare Implicit Length Strings, Return Index
5120 let Defs = [ECX, EFLAGS] in {
5121 multiclass SS42AI_pcmpistri<Intrinsic IntId128, string asm = "pcmpistri"> {
5122 def rr : SS42AI<0x63, MRMSrcReg, (outs),
5123 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
5124 !strconcat(asm, "\t{$src3, $src2, $src1|$src1, $src2, $src3}"),
5125 [(set ECX, (IntId128 VR128:$src1, VR128:$src2, imm:$src3)),
5126 (implicit EFLAGS)]>, OpSize;
5127 def rm : SS42AI<0x63, MRMSrcMem, (outs),
5128 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
5129 !strconcat(asm, "\t{$src3, $src2, $src1|$src1, $src2, $src3}"),
5130 [(set ECX, (IntId128 VR128:$src1, (load addr:$src2), imm:$src3)),
5131 (implicit EFLAGS)]>, OpSize;
5135 let Predicates = [HasAVX] in {
5136 defm VPCMPISTRI : SS42AI_pcmpistri<int_x86_sse42_pcmpistri128, "vpcmpistri">,
5138 defm VPCMPISTRIA : SS42AI_pcmpistri<int_x86_sse42_pcmpistria128, "vpcmpistri">,
5140 defm VPCMPISTRIC : SS42AI_pcmpistri<int_x86_sse42_pcmpistric128, "vpcmpistri">,
5142 defm VPCMPISTRIO : SS42AI_pcmpistri<int_x86_sse42_pcmpistrio128, "vpcmpistri">,
5144 defm VPCMPISTRIS : SS42AI_pcmpistri<int_x86_sse42_pcmpistris128, "vpcmpistri">,
5146 defm VPCMPISTRIZ : SS42AI_pcmpistri<int_x86_sse42_pcmpistriz128, "vpcmpistri">,
5150 defm PCMPISTRI : SS42AI_pcmpistri<int_x86_sse42_pcmpistri128>;
5151 defm PCMPISTRIA : SS42AI_pcmpistri<int_x86_sse42_pcmpistria128>;
5152 defm PCMPISTRIC : SS42AI_pcmpistri<int_x86_sse42_pcmpistric128>;
5153 defm PCMPISTRIO : SS42AI_pcmpistri<int_x86_sse42_pcmpistrio128>;
5154 defm PCMPISTRIS : SS42AI_pcmpistri<int_x86_sse42_pcmpistris128>;
5155 defm PCMPISTRIZ : SS42AI_pcmpistri<int_x86_sse42_pcmpistriz128>;
5157 // Packed Compare Explicit Length Strings, Return Index
5158 let Defs = [ECX, EFLAGS], Uses = [EAX, EDX] in {
5159 multiclass SS42AI_pcmpestri<Intrinsic IntId128, string asm = "pcmpestri"> {
5160 def rr : SS42AI<0x61, MRMSrcReg, (outs),
5161 (ins VR128:$src1, VR128:$src3, i8imm:$src5),
5162 !strconcat(asm, "\t{$src5, $src3, $src1|$src1, $src3, $src5}"),
5163 [(set ECX, (IntId128 VR128:$src1, EAX, VR128:$src3, EDX, imm:$src5)),
5164 (implicit EFLAGS)]>, OpSize;
5165 def rm : SS42AI<0x61, MRMSrcMem, (outs),
5166 (ins VR128:$src1, i128mem:$src3, i8imm:$src5),
5167 !strconcat(asm, "\t{$src5, $src3, $src1|$src1, $src3, $src5}"),
5169 (IntId128 VR128:$src1, EAX, (load addr:$src3), EDX, imm:$src5)),
5170 (implicit EFLAGS)]>, OpSize;
5174 let Predicates = [HasAVX] in {
5175 defm VPCMPESTRI : SS42AI_pcmpestri<int_x86_sse42_pcmpestri128, "vpcmpestri">,
5177 defm VPCMPESTRIA : SS42AI_pcmpestri<int_x86_sse42_pcmpestria128, "vpcmpestri">,
5179 defm VPCMPESTRIC : SS42AI_pcmpestri<int_x86_sse42_pcmpestric128, "vpcmpestri">,
5181 defm VPCMPESTRIO : SS42AI_pcmpestri<int_x86_sse42_pcmpestrio128, "vpcmpestri">,
5183 defm VPCMPESTRIS : SS42AI_pcmpestri<int_x86_sse42_pcmpestris128, "vpcmpestri">,
5185 defm VPCMPESTRIZ : SS42AI_pcmpestri<int_x86_sse42_pcmpestriz128, "vpcmpestri">,
5189 defm PCMPESTRI : SS42AI_pcmpestri<int_x86_sse42_pcmpestri128>;
5190 defm PCMPESTRIA : SS42AI_pcmpestri<int_x86_sse42_pcmpestria128>;
5191 defm PCMPESTRIC : SS42AI_pcmpestri<int_x86_sse42_pcmpestric128>;
5192 defm PCMPESTRIO : SS42AI_pcmpestri<int_x86_sse42_pcmpestrio128>;
5193 defm PCMPESTRIS : SS42AI_pcmpestri<int_x86_sse42_pcmpestris128>;
5194 defm PCMPESTRIZ : SS42AI_pcmpestri<int_x86_sse42_pcmpestriz128>;
5196 //===----------------------------------------------------------------------===//
5197 // SSE4.2 - CRC Instructions
5198 //===----------------------------------------------------------------------===//
5200 // No CRC instructions have AVX equivalents
5202 // crc intrinsic instruction
5203 // This set of instructions are only rm, the only difference is the size
5205 let Constraints = "$src1 = $dst" in {
5206 def CRC32r32m8 : SS42FI<0xF0, MRMSrcMem, (outs GR32:$dst),
5207 (ins GR32:$src1, i8mem:$src2),
5208 "crc32{b} \t{$src2, $src1|$src1, $src2}",
5210 (int_x86_sse42_crc32_32_8 GR32:$src1,
5211 (load addr:$src2)))]>;
5212 def CRC32r32r8 : SS42FI<0xF0, MRMSrcReg, (outs GR32:$dst),
5213 (ins GR32:$src1, GR8:$src2),
5214 "crc32{b} \t{$src2, $src1|$src1, $src2}",
5216 (int_x86_sse42_crc32_32_8 GR32:$src1, GR8:$src2))]>;
5217 def CRC32r32m16 : SS42FI<0xF1, MRMSrcMem, (outs GR32:$dst),
5218 (ins GR32:$src1, i16mem:$src2),
5219 "crc32{w} \t{$src2, $src1|$src1, $src2}",
5221 (int_x86_sse42_crc32_32_16 GR32:$src1,
5222 (load addr:$src2)))]>,
5224 def CRC32r32r16 : SS42FI<0xF1, MRMSrcReg, (outs GR32:$dst),
5225 (ins GR32:$src1, GR16:$src2),
5226 "crc32{w} \t{$src2, $src1|$src1, $src2}",
5228 (int_x86_sse42_crc32_32_16 GR32:$src1, GR16:$src2))]>,
5230 def CRC32r32m32 : SS42FI<0xF1, MRMSrcMem, (outs GR32:$dst),
5231 (ins GR32:$src1, i32mem:$src2),
5232 "crc32{l} \t{$src2, $src1|$src1, $src2}",
5234 (int_x86_sse42_crc32_32_32 GR32:$src1,
5235 (load addr:$src2)))]>;
5236 def CRC32r32r32 : SS42FI<0xF1, MRMSrcReg, (outs GR32:$dst),
5237 (ins GR32:$src1, GR32:$src2),
5238 "crc32{l} \t{$src2, $src1|$src1, $src2}",
5240 (int_x86_sse42_crc32_32_32 GR32:$src1, GR32:$src2))]>;
5241 def CRC32r64m8 : SS42FI<0xF0, MRMSrcMem, (outs GR64:$dst),
5242 (ins GR64:$src1, i8mem:$src2),
5243 "crc32{b} \t{$src2, $src1|$src1, $src2}",
5245 (int_x86_sse42_crc32_64_8 GR64:$src1,
5246 (load addr:$src2)))]>,
5248 def CRC32r64r8 : SS42FI<0xF0, MRMSrcReg, (outs GR64:$dst),
5249 (ins GR64:$src1, GR8:$src2),
5250 "crc32{b} \t{$src2, $src1|$src1, $src2}",
5252 (int_x86_sse42_crc32_64_8 GR64:$src1, GR8:$src2))]>,
5254 def CRC32r64m64 : SS42FI<0xF1, MRMSrcMem, (outs GR64:$dst),
5255 (ins GR64:$src1, i64mem:$src2),
5256 "crc32{q} \t{$src2, $src1|$src1, $src2}",
5258 (int_x86_sse42_crc32_64_64 GR64:$src1,
5259 (load addr:$src2)))]>,
5261 def CRC32r64r64 : SS42FI<0xF1, MRMSrcReg, (outs GR64:$dst),
5262 (ins GR64:$src1, GR64:$src2),
5263 "crc32{q} \t{$src2, $src1|$src1, $src2}",
5265 (int_x86_sse42_crc32_64_64 GR64:$src1, GR64:$src2))]>,
5269 //===----------------------------------------------------------------------===//
5270 // AES-NI Instructions
5271 //===----------------------------------------------------------------------===//
5273 multiclass AESI_binop_rm_int<bits<8> opc, string OpcodeStr,
5274 Intrinsic IntId128, bit Is2Addr = 1> {
5275 def rr : AES8I<opc, MRMSrcReg, (outs VR128:$dst),
5276 (ins VR128:$src1, VR128:$src2),
5278 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
5279 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
5280 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
5282 def rm : AES8I<opc, MRMSrcMem, (outs VR128:$dst),
5283 (ins VR128:$src1, i128mem:$src2),
5285 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
5286 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
5288 (IntId128 VR128:$src1,
5289 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
5292 // Perform One Round of an AES Encryption/Decryption Flow
5293 let Predicates = [HasAVX, HasAES] in {
5294 defm VAESENC : AESI_binop_rm_int<0xDC, "vaesenc",
5295 int_x86_aesni_aesenc, 0>, VEX_4V;
5296 defm VAESENCLAST : AESI_binop_rm_int<0xDD, "vaesenclast",
5297 int_x86_aesni_aesenclast, 0>, VEX_4V;
5298 defm VAESDEC : AESI_binop_rm_int<0xDE, "vaesdec",
5299 int_x86_aesni_aesdec, 0>, VEX_4V;
5300 defm VAESDECLAST : AESI_binop_rm_int<0xDF, "vaesdeclast",
5301 int_x86_aesni_aesdeclast, 0>, VEX_4V;
5304 let Constraints = "$src1 = $dst" in {
5305 defm AESENC : AESI_binop_rm_int<0xDC, "aesenc",
5306 int_x86_aesni_aesenc>;
5307 defm AESENCLAST : AESI_binop_rm_int<0xDD, "aesenclast",
5308 int_x86_aesni_aesenclast>;
5309 defm AESDEC : AESI_binop_rm_int<0xDE, "aesdec",
5310 int_x86_aesni_aesdec>;
5311 defm AESDECLAST : AESI_binop_rm_int<0xDF, "aesdeclast",
5312 int_x86_aesni_aesdeclast>;
5315 def : Pat<(v2i64 (int_x86_aesni_aesenc VR128:$src1, VR128:$src2)),
5316 (AESENCrr VR128:$src1, VR128:$src2)>;
5317 def : Pat<(v2i64 (int_x86_aesni_aesenc VR128:$src1, (memop addr:$src2))),
5318 (AESENCrm VR128:$src1, addr:$src2)>;
5319 def : Pat<(v2i64 (int_x86_aesni_aesenclast VR128:$src1, VR128:$src2)),
5320 (AESENCLASTrr VR128:$src1, VR128:$src2)>;
5321 def : Pat<(v2i64 (int_x86_aesni_aesenclast VR128:$src1, (memop addr:$src2))),
5322 (AESENCLASTrm VR128:$src1, addr:$src2)>;
5323 def : Pat<(v2i64 (int_x86_aesni_aesdec VR128:$src1, VR128:$src2)),
5324 (AESDECrr VR128:$src1, VR128:$src2)>;
5325 def : Pat<(v2i64 (int_x86_aesni_aesdec VR128:$src1, (memop addr:$src2))),
5326 (AESDECrm VR128:$src1, addr:$src2)>;
5327 def : Pat<(v2i64 (int_x86_aesni_aesdeclast VR128:$src1, VR128:$src2)),
5328 (AESDECLASTrr VR128:$src1, VR128:$src2)>;
5329 def : Pat<(v2i64 (int_x86_aesni_aesdeclast VR128:$src1, (memop addr:$src2))),
5330 (AESDECLASTrm VR128:$src1, addr:$src2)>;
5332 // Perform the AES InvMixColumn Transformation
5333 let Predicates = [HasAVX, HasAES] in {
5334 def VAESIMCrr : AES8I<0xDB, MRMSrcReg, (outs VR128:$dst),
5336 "vaesimc\t{$src1, $dst|$dst, $src1}",
5338 (int_x86_aesni_aesimc VR128:$src1))]>,
5340 def VAESIMCrm : AES8I<0xDB, MRMSrcMem, (outs VR128:$dst),
5341 (ins i128mem:$src1),
5342 "vaesimc\t{$src1, $dst|$dst, $src1}",
5344 (int_x86_aesni_aesimc (bitconvert (memopv2i64 addr:$src1))))]>,
5347 def AESIMCrr : AES8I<0xDB, MRMSrcReg, (outs VR128:$dst),
5349 "aesimc\t{$src1, $dst|$dst, $src1}",
5351 (int_x86_aesni_aesimc VR128:$src1))]>,
5353 def AESIMCrm : AES8I<0xDB, MRMSrcMem, (outs VR128:$dst),
5354 (ins i128mem:$src1),
5355 "aesimc\t{$src1, $dst|$dst, $src1}",
5357 (int_x86_aesni_aesimc (bitconvert (memopv2i64 addr:$src1))))]>,
5360 // AES Round Key Generation Assist
5361 let Predicates = [HasAVX, HasAES] in {
5362 def VAESKEYGENASSIST128rr : AESAI<0xDF, MRMSrcReg, (outs VR128:$dst),
5363 (ins VR128:$src1, i8imm:$src2),
5364 "vaeskeygenassist\t{$src2, $src1, $dst|$dst, $src1, $src2}",
5366 (int_x86_aesni_aeskeygenassist VR128:$src1, imm:$src2))]>,
5368 def VAESKEYGENASSIST128rm : AESAI<0xDF, MRMSrcMem, (outs VR128:$dst),
5369 (ins i128mem:$src1, i8imm:$src2),
5370 "vaeskeygenassist\t{$src2, $src1, $dst|$dst, $src1, $src2}",
5372 (int_x86_aesni_aeskeygenassist (bitconvert (memopv2i64 addr:$src1)),
5376 def AESKEYGENASSIST128rr : AESAI<0xDF, MRMSrcReg, (outs VR128:$dst),
5377 (ins VR128:$src1, i8imm:$src2),
5378 "aeskeygenassist\t{$src2, $src1, $dst|$dst, $src1, $src2}",
5380 (int_x86_aesni_aeskeygenassist VR128:$src1, imm:$src2))]>,
5382 def AESKEYGENASSIST128rm : AESAI<0xDF, MRMSrcMem, (outs VR128:$dst),
5383 (ins i128mem:$src1, i8imm:$src2),
5384 "aeskeygenassist\t{$src2, $src1, $dst|$dst, $src1, $src2}",
5386 (int_x86_aesni_aeskeygenassist (bitconvert (memopv2i64 addr:$src1)),
5390 //===----------------------------------------------------------------------===//
5391 // CLMUL Instructions
5392 //===----------------------------------------------------------------------===//
5394 // Carry-less Multiplication instructions
5395 let Constraints = "$src1 = $dst" in {
5396 def PCLMULQDQrr : CLMULIi8<0x44, MRMSrcReg, (outs VR128:$dst),
5397 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
5398 "pclmulqdq\t{$src3, $src2, $dst|$dst, $src2, $src3}",
5401 def PCLMULQDQrm : CLMULIi8<0x44, MRMSrcMem, (outs VR128:$dst),
5402 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
5403 "pclmulqdq\t{$src3, $src2, $dst|$dst, $src2, $src3}",
5407 // AVX carry-less Multiplication instructions
5408 def VPCLMULQDQrr : AVXCLMULIi8<0x44, MRMSrcReg, (outs VR128:$dst),
5409 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
5410 "vpclmulqdq\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
5413 def VPCLMULQDQrm : AVXCLMULIi8<0x44, MRMSrcMem, (outs VR128:$dst),
5414 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
5415 "vpclmulqdq\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
5419 multiclass pclmul_alias<string asm, int immop> {
5420 def : InstAlias<!strconcat("pclmul", asm,
5421 "dq {$src, $dst|$dst, $src}"),
5422 (PCLMULQDQrr VR128:$dst, VR128:$src, immop)>;
5424 def : InstAlias<!strconcat("pclmul", asm,
5425 "dq {$src, $dst|$dst, $src}"),
5426 (PCLMULQDQrm VR128:$dst, i128mem:$src, immop)>;
5428 def : InstAlias<!strconcat("vpclmul", asm,
5429 "dq {$src2, $src1, $dst|$dst, $src1, $src2}"),
5430 (VPCLMULQDQrr VR128:$dst, VR128:$src1, VR128:$src2, immop)>;
5432 def : InstAlias<!strconcat("vpclmul", asm,
5433 "dq {$src2, $src1, $dst|$dst, $src1, $src2}"),
5434 (VPCLMULQDQrm VR128:$dst, VR128:$src1, i128mem:$src2, immop)>;
5436 defm : pclmul_alias<"hqhq", 0x11>;
5437 defm : pclmul_alias<"hqlq", 0x01>;
5438 defm : pclmul_alias<"lqhq", 0x10>;
5439 defm : pclmul_alias<"lqlq", 0x00>;
5441 //===----------------------------------------------------------------------===//
5443 //===----------------------------------------------------------------------===//
5445 //===----------------------------------------------------------------------===//
5446 // VBROADCAST - Load from memory and broadcast to all elements of the
5447 // destination operand
5449 class avx_broadcast<bits<8> opc, string OpcodeStr, RegisterClass RC,
5450 X86MemOperand x86memop, Intrinsic Int> :
5451 AVX8I<opc, MRMSrcMem, (outs RC:$dst), (ins x86memop:$src),
5452 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
5453 [(set RC:$dst, (Int addr:$src))]>, VEX;
5455 def VBROADCASTSS : avx_broadcast<0x18, "vbroadcastss", VR128, f32mem,
5456 int_x86_avx_vbroadcastss>;
5457 def VBROADCASTSSY : avx_broadcast<0x18, "vbroadcastss", VR256, f32mem,
5458 int_x86_avx_vbroadcastss_256>;
5459 def VBROADCASTSD : avx_broadcast<0x19, "vbroadcastsd", VR256, f64mem,
5460 int_x86_avx_vbroadcast_sd_256>;
5461 def VBROADCASTF128 : avx_broadcast<0x1A, "vbroadcastf128", VR256, f128mem,
5462 int_x86_avx_vbroadcastf128_pd_256>;
5464 def : Pat<(int_x86_avx_vbroadcastf128_ps_256 addr:$src),
5465 (VBROADCASTF128 addr:$src)>;
5467 //===----------------------------------------------------------------------===//
5468 // VINSERTF128 - Insert packed floating-point values
5470 def VINSERTF128rr : AVXAIi8<0x18, MRMSrcReg, (outs VR256:$dst),
5471 (ins VR256:$src1, VR128:$src2, i8imm:$src3),
5472 "vinsertf128\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
5474 def VINSERTF128rm : AVXAIi8<0x18, MRMSrcMem, (outs VR256:$dst),
5475 (ins VR256:$src1, f128mem:$src2, i8imm:$src3),
5476 "vinsertf128\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
5479 def : Pat<(int_x86_avx_vinsertf128_pd_256 VR256:$src1, VR128:$src2, imm:$src3),
5480 (VINSERTF128rr VR256:$src1, VR128:$src2, imm:$src3)>;
5481 def : Pat<(int_x86_avx_vinsertf128_ps_256 VR256:$src1, VR128:$src2, imm:$src3),
5482 (VINSERTF128rr VR256:$src1, VR128:$src2, imm:$src3)>;
5483 def : Pat<(int_x86_avx_vinsertf128_si_256 VR256:$src1, VR128:$src2, imm:$src3),
5484 (VINSERTF128rr VR256:$src1, VR128:$src2, imm:$src3)>;
5486 def : Pat<(vinsertf128_insert:$ins (v8f32 VR256:$src1), (v4f32 VR128:$src2),
5488 (VINSERTF128rr VR256:$src1, VR128:$src2,
5489 (INSERT_get_vinsertf128_imm VR256:$ins))>;
5490 def : Pat<(vinsertf128_insert:$ins (v4f64 VR256:$src1), (v2f64 VR128:$src2),
5492 (VINSERTF128rr VR256:$src1, VR128:$src2,
5493 (INSERT_get_vinsertf128_imm VR256:$ins))>;
5494 def : Pat<(vinsertf128_insert:$ins (v8i32 VR256:$src1), (v4i32 VR128:$src2),
5496 (VINSERTF128rr VR256:$src1, VR128:$src2,
5497 (INSERT_get_vinsertf128_imm VR256:$ins))>;
5498 def : Pat<(vinsertf128_insert:$ins (v4i64 VR256:$src1), (v2i64 VR128:$src2),
5500 (VINSERTF128rr VR256:$src1, VR128:$src2,
5501 (INSERT_get_vinsertf128_imm VR256:$ins))>;
5502 def : Pat<(vinsertf128_insert:$ins (v32i8 VR256:$src1), (v16i8 VR128:$src2),
5504 (VINSERTF128rr VR256:$src1, VR128:$src2,
5505 (INSERT_get_vinsertf128_imm VR256:$ins))>;
5506 def : Pat<(vinsertf128_insert:$ins (v16i16 VR256:$src1), (v8i16 VR128:$src2),
5508 (VINSERTF128rr VR256:$src1, VR128:$src2,
5509 (INSERT_get_vinsertf128_imm VR256:$ins))>;
5511 // Special COPY patterns
5512 def : Pat<(insert_subvector undef, (v2i64 VR128:$src), (i32 0)),
5513 (INSERT_SUBREG (v4i64 (IMPLICIT_DEF)), VR128:$src, sub_xmm)>;
5514 def : Pat<(insert_subvector undef, (v2f64 VR128:$src), (i32 0)),
5515 (INSERT_SUBREG (v4f64 (IMPLICIT_DEF)), VR128:$src, sub_xmm)>;
5516 def : Pat<(insert_subvector undef, (v4i32 VR128:$src), (i32 0)),
5517 (INSERT_SUBREG (v8i32 (IMPLICIT_DEF)), VR128:$src, sub_xmm)>;
5518 def : Pat<(insert_subvector undef, (v4f32 VR128:$src), (i32 0)),
5519 (INSERT_SUBREG (v8f32 (IMPLICIT_DEF)), VR128:$src, sub_xmm)>;
5520 def : Pat<(insert_subvector undef, (v8i16 VR128:$src), (i32 0)),
5521 (INSERT_SUBREG (v16i16 (IMPLICIT_DEF)), VR128:$src, sub_xmm)>;
5522 def : Pat<(insert_subvector undef, (v16i8 VR128:$src), (i32 0)),
5523 (INSERT_SUBREG (v32i8 (IMPLICIT_DEF)), VR128:$src, sub_xmm)>;
5525 //===----------------------------------------------------------------------===//
5526 // VEXTRACTF128 - Extract packed floating-point values
5528 def VEXTRACTF128rr : AVXAIi8<0x19, MRMDestReg, (outs VR128:$dst),
5529 (ins VR256:$src1, i8imm:$src2),
5530 "vextractf128\t{$src2, $src1, $dst|$dst, $src1, $src2}",
5532 def VEXTRACTF128mr : AVXAIi8<0x19, MRMDestMem, (outs),
5533 (ins f128mem:$dst, VR256:$src1, i8imm:$src2),
5534 "vextractf128\t{$src2, $src1, $dst|$dst, $src1, $src2}",
5537 def : Pat<(int_x86_avx_vextractf128_pd_256 VR256:$src1, imm:$src2),
5538 (VEXTRACTF128rr VR256:$src1, imm:$src2)>;
5539 def : Pat<(int_x86_avx_vextractf128_ps_256 VR256:$src1, imm:$src2),
5540 (VEXTRACTF128rr VR256:$src1, imm:$src2)>;
5541 def : Pat<(int_x86_avx_vextractf128_si_256 VR256:$src1, imm:$src2),
5542 (VEXTRACTF128rr VR256:$src1, imm:$src2)>;
5544 def : Pat<(vextractf128_extract:$ext VR256:$src1, (i32 imm)),
5545 (v4f32 (VEXTRACTF128rr
5546 (v8f32 VR256:$src1),
5547 (EXTRACT_get_vextractf128_imm VR128:$ext)))>;
5548 def : Pat<(vextractf128_extract:$ext VR256:$src1, (i32 imm)),
5549 (v2f64 (VEXTRACTF128rr
5550 (v4f64 VR256:$src1),
5551 (EXTRACT_get_vextractf128_imm VR128:$ext)))>;
5552 def : Pat<(vextractf128_extract:$ext VR256:$src1, (i32 imm)),
5553 (v4i32 (VEXTRACTF128rr
5554 (v8i32 VR256:$src1),
5555 (EXTRACT_get_vextractf128_imm VR128:$ext)))>;
5556 def : Pat<(vextractf128_extract:$ext VR256:$src1, (i32 imm)),
5557 (v2i64 (VEXTRACTF128rr
5558 (v4i64 VR256:$src1),
5559 (EXTRACT_get_vextractf128_imm VR128:$ext)))>;
5560 def : Pat<(vextractf128_extract:$ext VR256:$src1, (i32 imm)),
5561 (v8i16 (VEXTRACTF128rr
5562 (v16i16 VR256:$src1),
5563 (EXTRACT_get_vextractf128_imm VR128:$ext)))>;
5564 def : Pat<(vextractf128_extract:$ext VR256:$src1, (i32 imm)),
5565 (v16i8 (VEXTRACTF128rr
5566 (v32i8 VR256:$src1),
5567 (EXTRACT_get_vextractf128_imm VR128:$ext)))>;
5569 // Special COPY patterns
5570 def : Pat<(v4i32 (extract_subvector (v8i32 VR256:$src), (i32 0))),
5571 (v4i32 (EXTRACT_SUBREG (v8i32 VR256:$src), sub_xmm))>;
5572 def : Pat<(v4f32 (extract_subvector (v8f32 VR256:$src), (i32 0))),
5573 (v4f32 (EXTRACT_SUBREG (v8f32 VR256:$src), sub_xmm))>;
5575 def : Pat<(v2i64 (extract_subvector (v4i64 VR256:$src), (i32 0))),
5576 (v2i64 (EXTRACT_SUBREG (v4i64 VR256:$src), sub_xmm))>;
5577 def : Pat<(v2f64 (extract_subvector (v4f64 VR256:$src), (i32 0))),
5578 (v2f64 (EXTRACT_SUBREG (v4f64 VR256:$src), sub_xmm))>;
5581 //===----------------------------------------------------------------------===//
5582 // VMASKMOV - Conditional SIMD Packed Loads and Stores
5584 multiclass avx_movmask_rm<bits<8> opc_rm, bits<8> opc_mr, string OpcodeStr,
5585 Intrinsic IntLd, Intrinsic IntLd256,
5586 Intrinsic IntSt, Intrinsic IntSt256,
5587 PatFrag pf128, PatFrag pf256> {
5588 def rm : AVX8I<opc_rm, MRMSrcMem, (outs VR128:$dst),
5589 (ins VR128:$src1, f128mem:$src2),
5590 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
5591 [(set VR128:$dst, (IntLd addr:$src2, VR128:$src1))]>,
5593 def Yrm : AVX8I<opc_rm, MRMSrcMem, (outs VR256:$dst),
5594 (ins VR256:$src1, f256mem:$src2),
5595 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
5596 [(set VR256:$dst, (IntLd256 addr:$src2, VR256:$src1))]>,
5598 def mr : AVX8I<opc_mr, MRMDestMem, (outs),
5599 (ins f128mem:$dst, VR128:$src1, VR128:$src2),
5600 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
5601 [(IntSt addr:$dst, VR128:$src1, VR128:$src2)]>, VEX_4V;
5602 def Ymr : AVX8I<opc_mr, MRMDestMem, (outs),
5603 (ins f256mem:$dst, VR256:$src1, VR256:$src2),
5604 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
5605 [(IntSt256 addr:$dst, VR256:$src1, VR256:$src2)]>, VEX_4V;
5608 defm VMASKMOVPS : avx_movmask_rm<0x2C, 0x2E, "vmaskmovps",
5609 int_x86_avx_maskload_ps,
5610 int_x86_avx_maskload_ps_256,
5611 int_x86_avx_maskstore_ps,
5612 int_x86_avx_maskstore_ps_256,
5613 memopv4f32, memopv8f32>;
5614 defm VMASKMOVPD : avx_movmask_rm<0x2D, 0x2F, "vmaskmovpd",
5615 int_x86_avx_maskload_pd,
5616 int_x86_avx_maskload_pd_256,
5617 int_x86_avx_maskstore_pd,
5618 int_x86_avx_maskstore_pd_256,
5619 memopv2f64, memopv4f64>;
5621 //===----------------------------------------------------------------------===//
5622 // VPERMIL - Permute Single and Double Floating-Point Values
5624 multiclass avx_permil<bits<8> opc_rm, bits<8> opc_rmi, string OpcodeStr,
5625 RegisterClass RC, X86MemOperand x86memop_f,
5626 X86MemOperand x86memop_i, PatFrag f_frag, PatFrag i_frag,
5627 Intrinsic IntVar, Intrinsic IntImm> {
5628 def rr : AVX8I<opc_rm, MRMSrcReg, (outs RC:$dst),
5629 (ins RC:$src1, RC:$src2),
5630 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
5631 [(set RC:$dst, (IntVar RC:$src1, RC:$src2))]>, VEX_4V;
5632 def rm : AVX8I<opc_rm, MRMSrcMem, (outs RC:$dst),
5633 (ins RC:$src1, x86memop_i:$src2),
5634 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
5635 [(set RC:$dst, (IntVar RC:$src1, (i_frag addr:$src2)))]>, VEX_4V;
5637 def ri : AVXAIi8<opc_rmi, MRMSrcReg, (outs RC:$dst),
5638 (ins RC:$src1, i8imm:$src2),
5639 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
5640 [(set RC:$dst, (IntImm RC:$src1, imm:$src2))]>, VEX;
5641 def mi : AVXAIi8<opc_rmi, MRMSrcMem, (outs RC:$dst),
5642 (ins x86memop_f:$src1, i8imm:$src2),
5643 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
5644 [(set RC:$dst, (IntImm (f_frag addr:$src1), imm:$src2))]>, VEX;
5647 defm VPERMILPS : avx_permil<0x0C, 0x04, "vpermilps", VR128, f128mem, i128mem,
5648 memopv4f32, memopv4i32,
5649 int_x86_avx_vpermilvar_ps,
5650 int_x86_avx_vpermil_ps>;
5651 defm VPERMILPSY : avx_permil<0x0C, 0x04, "vpermilps", VR256, f256mem, i256mem,
5652 memopv8f32, memopv8i32,
5653 int_x86_avx_vpermilvar_ps_256,
5654 int_x86_avx_vpermil_ps_256>;
5655 defm VPERMILPD : avx_permil<0x0D, 0x05, "vpermilpd", VR128, f128mem, i128mem,
5656 memopv2f64, memopv2i64,
5657 int_x86_avx_vpermilvar_pd,
5658 int_x86_avx_vpermil_pd>;
5659 defm VPERMILPDY : avx_permil<0x0D, 0x05, "vpermilpd", VR256, f256mem, i256mem,
5660 memopv4f64, memopv4i64,
5661 int_x86_avx_vpermilvar_pd_256,
5662 int_x86_avx_vpermil_pd_256>;
5664 def : Pat<(v8f32 (X86VPermilpsy VR256:$src1, (i8 imm:$imm))),
5665 (VPERMILPSYri VR256:$src1, imm:$imm)>;
5666 def : Pat<(v4f64 (X86VPermilpdy VR256:$src1, (i8 imm:$imm))),
5667 (VPERMILPDYri VR256:$src1, imm:$imm)>;
5668 def : Pat<(v8i32 (X86VPermilpsy VR256:$src1, (i8 imm:$imm))),
5669 (VPERMILPSYri VR256:$src1, imm:$imm)>;
5670 def : Pat<(v4i64 (X86VPermilpdy VR256:$src1, (i8 imm:$imm))),
5671 (VPERMILPDYri VR256:$src1, imm:$imm)>;
5673 //===----------------------------------------------------------------------===//
5674 // VPERM2F128 - Permute Floating-Point Values in 128-bit chunks
5676 def VPERM2F128rr : AVXAIi8<0x06, MRMSrcReg, (outs VR256:$dst),
5677 (ins VR256:$src1, VR256:$src2, i8imm:$src3),
5678 "vperm2f128\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
5680 def VPERM2F128rm : AVXAIi8<0x06, MRMSrcMem, (outs VR256:$dst),
5681 (ins VR256:$src1, f256mem:$src2, i8imm:$src3),
5682 "vperm2f128\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
5685 def : Pat<(int_x86_avx_vperm2f128_ps_256 VR256:$src1, VR256:$src2, imm:$src3),
5686 (VPERM2F128rr VR256:$src1, VR256:$src2, imm:$src3)>;
5687 def : Pat<(int_x86_avx_vperm2f128_pd_256 VR256:$src1, VR256:$src2, imm:$src3),
5688 (VPERM2F128rr VR256:$src1, VR256:$src2, imm:$src3)>;
5689 def : Pat<(int_x86_avx_vperm2f128_si_256 VR256:$src1, VR256:$src2, imm:$src3),
5690 (VPERM2F128rr VR256:$src1, VR256:$src2, imm:$src3)>;
5692 def : Pat<(int_x86_avx_vperm2f128_ps_256
5693 VR256:$src1, (memopv8f32 addr:$src2), imm:$src3),
5694 (VPERM2F128rm VR256:$src1, addr:$src2, imm:$src3)>;
5695 def : Pat<(int_x86_avx_vperm2f128_pd_256
5696 VR256:$src1, (memopv4f64 addr:$src2), imm:$src3),
5697 (VPERM2F128rm VR256:$src1, addr:$src2, imm:$src3)>;
5698 def : Pat<(int_x86_avx_vperm2f128_si_256
5699 VR256:$src1, (memopv8i32 addr:$src2), imm:$src3),
5700 (VPERM2F128rm VR256:$src1, addr:$src2, imm:$src3)>;
5702 def : Pat<(v8f32 (X86VPerm2f128 VR256:$src1, VR256:$src2, (i8 imm:$imm))),
5703 (VPERM2F128rr VR256:$src1, VR256:$src2, imm:$imm)>;
5704 def : Pat<(v8i32 (X86VPerm2f128 VR256:$src1, VR256:$src2, (i8 imm:$imm))),
5705 (VPERM2F128rr VR256:$src1, VR256:$src2, imm:$imm)>;
5706 def : Pat<(v4i64 (X86VPerm2f128 VR256:$src1, VR256:$src2, (i8 imm:$imm))),
5707 (VPERM2F128rr VR256:$src1, VR256:$src2, imm:$imm)>;
5708 def : Pat<(v4f64 (X86VPerm2f128 VR256:$src1, VR256:$src2, (i8 imm:$imm))),
5709 (VPERM2F128rr VR256:$src1, VR256:$src2, imm:$imm)>;
5710 def : Pat<(v32i8 (X86VPerm2f128 VR256:$src1, VR256:$src2, (i8 imm:$imm))),
5711 (VPERM2F128rr VR256:$src1, VR256:$src2, imm:$imm)>;
5712 def : Pat<(v16i16 (X86VPerm2f128 VR256:$src1, VR256:$src2, (i8 imm:$imm))),
5713 (VPERM2F128rr VR256:$src1, VR256:$src2, imm:$imm)>;
5715 //===----------------------------------------------------------------------===//
5716 // VZERO - Zero YMM registers
5718 // Zero All YMM registers
5719 def VZEROALL : I<0x77, RawFrm, (outs), (ins), "vzeroall",
5720 [(int_x86_avx_vzeroall)]>, VEX, VEX_L, Requires<[HasAVX]>;
5722 // Zero Upper bits of YMM registers
5723 def VZEROUPPER : I<0x77, RawFrm, (outs), (ins), "vzeroupper",
5724 [(int_x86_avx_vzeroupper)]>, VEX, Requires<[HasAVX]>;
5726 //===----------------------------------------------------------------------===//
5727 // SSE Shuffle pattern fragments
5728 //===----------------------------------------------------------------------===//
5730 // This is part of a "work in progress" refactoring. The idea is that all
5731 // vector shuffles are going to be translated into target specific nodes and
5732 // directly matched by the patterns below (which can be changed along the way)
5733 // The AVX version of some but not all of them are described here, and more
5734 // should come in a near future.
5736 // Shuffle with PSHUFD instruction folding loads. The first two patterns match
5737 // SSE2 loads, which are always promoted to v2i64. The last one should match
5738 // the SSE1 case, where the only legal load is v4f32, but there is no PSHUFD
5739 // in SSE2, how does it ever worked? Anyway, the pattern will remain here until
5740 // we investigate further.
5741 def : Pat<(v4i32 (X86PShufd (bc_v4i32 (memopv2i64 addr:$src1)),
5743 (VPSHUFDmi addr:$src1, imm:$imm)>, Requires<[HasAVX]>;
5744 def : Pat<(v4i32 (X86PShufd (bc_v4i32 (memopv2i64 addr:$src1)),
5746 (PSHUFDmi addr:$src1, imm:$imm)>;
5747 def : Pat<(v4i32 (X86PShufd (bc_v4i32 (memopv4f32 addr:$src1)),
5749 (PSHUFDmi addr:$src1, imm:$imm)>; // FIXME: has this ever worked?
5751 // Shuffle with PSHUFD instruction.
5752 def : Pat<(v4f32 (X86PShufd VR128:$src1, (i8 imm:$imm))),
5753 (VPSHUFDri VR128:$src1, imm:$imm)>, Requires<[HasAVX]>;
5754 def : Pat<(v4f32 (X86PShufd VR128:$src1, (i8 imm:$imm))),
5755 (PSHUFDri VR128:$src1, imm:$imm)>;
5757 def : Pat<(v4i32 (X86PShufd VR128:$src1, (i8 imm:$imm))),
5758 (VPSHUFDri VR128:$src1, imm:$imm)>, Requires<[HasAVX]>;
5759 def : Pat<(v4i32 (X86PShufd VR128:$src1, (i8 imm:$imm))),
5760 (PSHUFDri VR128:$src1, imm:$imm)>;
5762 // Shuffle with SHUFPD instruction.
5763 def : Pat<(v2f64 (X86Shufps VR128:$src1,
5764 (memopv2f64 addr:$src2), (i8 imm:$imm))),
5765 (VSHUFPDrmi VR128:$src1, addr:$src2, imm:$imm)>, Requires<[HasAVX]>;
5766 def : Pat<(v2f64 (X86Shufps VR128:$src1,
5767 (memopv2f64 addr:$src2), (i8 imm:$imm))),
5768 (SHUFPDrmi VR128:$src1, addr:$src2, imm:$imm)>;
5770 def : Pat<(v2i64 (X86Shufpd VR128:$src1, VR128:$src2, (i8 imm:$imm))),
5771 (VSHUFPDrri VR128:$src1, VR128:$src2, imm:$imm)>, Requires<[HasAVX]>;
5772 def : Pat<(v2i64 (X86Shufpd VR128:$src1, VR128:$src2, (i8 imm:$imm))),
5773 (SHUFPDrri VR128:$src1, VR128:$src2, imm:$imm)>;
5775 def : Pat<(v2f64 (X86Shufpd VR128:$src1, VR128:$src2, (i8 imm:$imm))),
5776 (VSHUFPDrri VR128:$src1, VR128:$src2, imm:$imm)>, Requires<[HasAVX]>;
5777 def : Pat<(v2f64 (X86Shufpd VR128:$src1, VR128:$src2, (i8 imm:$imm))),
5778 (SHUFPDrri VR128:$src1, VR128:$src2, imm:$imm)>;
5780 // Shuffle with SHUFPS instruction.
5781 def : Pat<(v4f32 (X86Shufps VR128:$src1,
5782 (memopv4f32 addr:$src2), (i8 imm:$imm))),
5783 (VSHUFPSrmi VR128:$src1, addr:$src2, imm:$imm)>, Requires<[HasAVX]>;
5784 def : Pat<(v4f32 (X86Shufps VR128:$src1,
5785 (memopv4f32 addr:$src2), (i8 imm:$imm))),
5786 (SHUFPSrmi VR128:$src1, addr:$src2, imm:$imm)>;
5788 def : Pat<(v4f32 (X86Shufps VR128:$src1, VR128:$src2, (i8 imm:$imm))),
5789 (VSHUFPSrri VR128:$src1, VR128:$src2, imm:$imm)>, Requires<[HasAVX]>;
5790 def : Pat<(v4f32 (X86Shufps VR128:$src1, VR128:$src2, (i8 imm:$imm))),
5791 (SHUFPSrri VR128:$src1, VR128:$src2, imm:$imm)>;
5793 def : Pat<(v4i32 (X86Shufps VR128:$src1,
5794 (bc_v4i32 (memopv2i64 addr:$src2)), (i8 imm:$imm))),
5795 (VSHUFPSrmi VR128:$src1, addr:$src2, imm:$imm)>, Requires<[HasAVX]>;
5796 def : Pat<(v4i32 (X86Shufps VR128:$src1,
5797 (bc_v4i32 (memopv2i64 addr:$src2)), (i8 imm:$imm))),
5798 (SHUFPSrmi VR128:$src1, addr:$src2, imm:$imm)>;
5800 def : Pat<(v4i32 (X86Shufps VR128:$src1, VR128:$src2, (i8 imm:$imm))),
5801 (VSHUFPSrri VR128:$src1, VR128:$src2, imm:$imm)>, Requires<[HasAVX]>;
5802 def : Pat<(v4i32 (X86Shufps VR128:$src1, VR128:$src2, (i8 imm:$imm))),
5803 (SHUFPSrri VR128:$src1, VR128:$src2, imm:$imm)>;
5805 // Shuffle with MOVHLPS instruction
5806 def : Pat<(v4f32 (X86Movhlps VR128:$src1, VR128:$src2)),
5807 (MOVHLPSrr VR128:$src1, VR128:$src2)>;
5808 def : Pat<(v4i32 (X86Movhlps VR128:$src1, VR128:$src2)),
5809 (MOVHLPSrr VR128:$src1, VR128:$src2)>;
5811 // Shuffle with MOVDDUP instruction
5812 def : Pat<(X86Movddup (memopv2f64 addr:$src)),
5813 (VMOVDDUPrm addr:$src)>, Requires<[HasAVX]>;
5814 def : Pat<(X86Movddup (memopv2f64 addr:$src)),
5815 (MOVDDUPrm addr:$src)>;
5817 def : Pat<(X86Movddup (bc_v2f64 (memopv4f32 addr:$src))),
5818 (VMOVDDUPrm addr:$src)>, Requires<[HasAVX]>;
5819 def : Pat<(X86Movddup (bc_v2f64 (memopv4f32 addr:$src))),
5820 (MOVDDUPrm addr:$src)>;
5822 def : Pat<(X86Movddup (bc_v2f64 (memopv2i64 addr:$src))),
5823 (VMOVDDUPrm addr:$src)>, Requires<[HasAVX]>;
5824 def : Pat<(X86Movddup (bc_v2f64 (memopv2i64 addr:$src))),
5825 (MOVDDUPrm addr:$src)>;
5827 def : Pat<(X86Movddup (v2f64 (scalar_to_vector (loadf64 addr:$src)))),
5828 (VMOVDDUPrm addr:$src)>, Requires<[HasAVX]>;
5829 def : Pat<(X86Movddup (v2f64 (scalar_to_vector (loadf64 addr:$src)))),
5830 (MOVDDUPrm addr:$src)>;
5832 def : Pat<(X86Movddup (bc_v2f64
5833 (v2i64 (scalar_to_vector (loadi64 addr:$src))))),
5834 (VMOVDDUPrm addr:$src)>, Requires<[HasAVX]>;
5835 def : Pat<(X86Movddup (bc_v2f64
5836 (v2i64 (scalar_to_vector (loadi64 addr:$src))))),
5837 (MOVDDUPrm addr:$src)>;
5840 // Shuffle with UNPCKLPS
5841 def : Pat<(v4f32 (X86Unpcklps VR128:$src1, (memopv4f32 addr:$src2))),
5842 (VUNPCKLPSrm VR128:$src1, addr:$src2)>, Requires<[HasAVX]>;
5843 def : Pat<(v4f32 (X86Unpcklps VR128:$src1, (memopv4f32 addr:$src2))),
5844 (UNPCKLPSrm VR128:$src1, addr:$src2)>;
5846 def : Pat<(v4f32 (X86Unpcklps VR128:$src1, VR128:$src2)),
5847 (VUNPCKLPSrr VR128:$src1, VR128:$src2)>, Requires<[HasAVX]>;
5848 def : Pat<(v4f32 (X86Unpcklps VR128:$src1, VR128:$src2)),
5849 (UNPCKLPSrr VR128:$src1, VR128:$src2)>;
5851 // Shuffle with VUNPCKHPSY
5852 def : Pat<(v8f32 (X86Unpcklpsy VR256:$src1, (memopv8f32 addr:$src2))),
5853 (VUNPCKLPSYrm VR256:$src1, addr:$src2)>, Requires<[HasAVX]>;
5854 def : Pat<(v8f32 (X86Unpcklpsy VR256:$src1, VR256:$src2)),
5855 (VUNPCKLPSYrr VR256:$src1, VR256:$src2)>, Requires<[HasAVX]>;
5856 def : Pat<(v8i32 (X86Unpcklpsy VR256:$src1, VR256:$src2)),
5857 (VUNPCKLPSYrr VR256:$src1, VR256:$src2)>, Requires<[HasAVX]>;
5858 def : Pat<(v8i32 (X86Unpcklpsy VR256:$src1, (memopv8i32 addr:$src2))),
5859 (VUNPCKLPSYrm VR256:$src1, addr:$src2)>, Requires<[HasAVX]>;
5861 // Shuffle with UNPCKHPS
5862 def : Pat<(v4f32 (X86Unpckhps VR128:$src1, (memopv4f32 addr:$src2))),
5863 (VUNPCKHPSrm VR128:$src1, addr:$src2)>, Requires<[HasAVX]>;
5864 def : Pat<(v4f32 (X86Unpckhps VR128:$src1, (memopv4f32 addr:$src2))),
5865 (UNPCKHPSrm VR128:$src1, addr:$src2)>;
5867 def : Pat<(v4f32 (X86Unpckhps VR128:$src1, VR128:$src2)),
5868 (VUNPCKHPSrr VR128:$src1, VR128:$src2)>, Requires<[HasAVX]>;
5869 def : Pat<(v4f32 (X86Unpckhps VR128:$src1, VR128:$src2)),
5870 (UNPCKHPSrr VR128:$src1, VR128:$src2)>;
5872 // Shuffle with VUNPCKHPSY
5873 def : Pat<(v8f32 (X86Unpckhpsy VR256:$src1, (memopv8f32 addr:$src2))),
5874 (VUNPCKHPSYrm VR256:$src1, addr:$src2)>, Requires<[HasAVX]>;
5875 def : Pat<(v8f32 (X86Unpckhpsy VR256:$src1, VR256:$src2)),
5876 (VUNPCKHPSYrr VR256:$src1, VR256:$src2)>, Requires<[HasAVX]>;
5878 def : Pat<(v8i32 (X86Unpckhpsy VR256:$src1, (memopv8i32 addr:$src2))),
5879 (VUNPCKHPSYrm VR256:$src1, addr:$src2)>, Requires<[HasAVX]>;
5880 def : Pat<(v8i32 (X86Unpckhpsy VR256:$src1, VR256:$src2)),
5881 (VUNPCKHPSYrr VR256:$src1, VR256:$src2)>, Requires<[HasAVX]>;
5883 // Shuffle with UNPCKLPD
5884 def : Pat<(v2f64 (X86Unpcklpd VR128:$src1, (memopv2f64 addr:$src2))),
5885 (VUNPCKLPDrm VR128:$src1, addr:$src2)>, Requires<[HasAVX]>;
5886 def : Pat<(v2f64 (X86Unpcklpd VR128:$src1, (memopv2f64 addr:$src2))),
5887 (UNPCKLPDrm VR128:$src1, addr:$src2)>;
5889 def : Pat<(v2f64 (X86Unpcklpd VR128:$src1, VR128:$src2)),
5890 (VUNPCKLPDrr VR128:$src1, VR128:$src2)>, Requires<[HasAVX]>;
5891 def : Pat<(v2f64 (X86Unpcklpd VR128:$src1, VR128:$src2)),
5892 (UNPCKLPDrr VR128:$src1, VR128:$src2)>;
5894 // Shuffle with VUNPCKLPDY
5895 def : Pat<(v4f64 (X86Unpcklpdy VR256:$src1, (memopv4f64 addr:$src2))),
5896 (VUNPCKLPDYrm VR256:$src1, addr:$src2)>, Requires<[HasAVX]>;
5897 def : Pat<(v4f64 (X86Unpcklpdy VR256:$src1, VR256:$src2)),
5898 (VUNPCKLPDYrr VR256:$src1, VR256:$src2)>, Requires<[HasAVX]>;
5900 def : Pat<(v4i64 (X86Unpcklpdy VR256:$src1, (memopv4i64 addr:$src2))),
5901 (VUNPCKLPDYrm VR256:$src1, addr:$src2)>, Requires<[HasAVX]>;
5902 def : Pat<(v4i64 (X86Unpcklpdy VR256:$src1, VR256:$src2)),
5903 (VUNPCKLPDYrr VR256:$src1, VR256:$src2)>, Requires<[HasAVX]>;
5905 // Shuffle with UNPCKHPD
5906 def : Pat<(v2f64 (X86Unpckhpd VR128:$src1, (memopv2f64 addr:$src2))),
5907 (VUNPCKHPDrm VR128:$src1, addr:$src2)>, Requires<[HasAVX]>;
5908 def : Pat<(v2f64 (X86Unpckhpd VR128:$src1, (memopv2f64 addr:$src2))),
5909 (UNPCKHPDrm VR128:$src1, addr:$src2)>;
5911 def : Pat<(v2f64 (X86Unpckhpd VR128:$src1, VR128:$src2)),
5912 (VUNPCKHPDrr VR128:$src1, VR128:$src2)>, Requires<[HasAVX]>;
5913 def : Pat<(v2f64 (X86Unpckhpd VR128:$src1, VR128:$src2)),
5914 (UNPCKHPDrr VR128:$src1, VR128:$src2)>;
5916 // Shuffle with VUNPCKHPDY
5917 def : Pat<(v4f64 (X86Unpckhpdy VR256:$src1, (memopv4f64 addr:$src2))),
5918 (VUNPCKHPDYrm VR256:$src1, addr:$src2)>, Requires<[HasAVX]>;
5919 def : Pat<(v4f64 (X86Unpckhpdy VR256:$src1, VR256:$src2)),
5920 (VUNPCKHPDYrr VR256:$src1, VR256:$src2)>, Requires<[HasAVX]>;
5921 def : Pat<(v4i64 (X86Unpckhpdy VR256:$src1, (memopv4i64 addr:$src2))),
5922 (VUNPCKHPDYrm VR256:$src1, addr:$src2)>, Requires<[HasAVX]>;
5923 def : Pat<(v4i64 (X86Unpckhpdy VR256:$src1, VR256:$src2)),
5924 (VUNPCKHPDYrr VR256:$src1, VR256:$src2)>, Requires<[HasAVX]>;
5926 // Shuffle with MOVLHPS
5927 def : Pat<(X86Movlhps VR128:$src1,
5928 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2))))),
5929 (MOVHPSrm VR128:$src1, addr:$src2)>;
5930 def : Pat<(X86Movlhps VR128:$src1,
5931 (bc_v4i32 (v2i64 (X86vzload addr:$src2)))),
5932 (MOVHPSrm VR128:$src1, addr:$src2)>;
5933 def : Pat<(v4f32 (X86Movlhps VR128:$src1, VR128:$src2)),
5934 (MOVLHPSrr VR128:$src1, VR128:$src2)>;
5935 def : Pat<(v4i32 (X86Movlhps VR128:$src1, VR128:$src2)),
5936 (MOVLHPSrr VR128:$src1, VR128:$src2)>;
5937 def : Pat<(v2i64 (X86Movlhps VR128:$src1, VR128:$src2)),
5938 (MOVLHPSrr (v2i64 VR128:$src1), VR128:$src2)>;
5940 // FIXME: Instead of X86Movddup, there should be a X86Unpcklpd here, the problem
5941 // is during lowering, where it's not possible to recognize the load fold cause
5942 // it has two uses through a bitcast. One use disappears at isel time and the
5943 // fold opportunity reappears.
5944 def : Pat<(v2f64 (X86Movddup VR128:$src)),
5945 (UNPCKLPDrr VR128:$src, VR128:$src)>;
5947 // Shuffle with MOVLHPD
5948 def : Pat<(v2f64 (X86Movlhpd VR128:$src1,
5949 (scalar_to_vector (loadf64 addr:$src2)))),
5950 (MOVHPDrm VR128:$src1, addr:$src2)>;
5952 // FIXME: Instead of X86Unpcklpd, there should be a X86Movlhpd here, the problem
5953 // is during lowering, where it's not possible to recognize the load fold cause
5954 // it has two uses through a bitcast. One use disappears at isel time and the
5955 // fold opportunity reappears.
5956 def : Pat<(v2f64 (X86Unpcklpd VR128:$src1,
5957 (scalar_to_vector (loadf64 addr:$src2)))),
5958 (MOVHPDrm VR128:$src1, addr:$src2)>;
5960 // Shuffle with MOVSS
5961 def : Pat<(v4f32 (X86Movss VR128:$src1, (scalar_to_vector FR32:$src2))),
5962 (MOVSSrr VR128:$src1, FR32:$src2)>;
5963 def : Pat<(v4i32 (X86Movss VR128:$src1, VR128:$src2)),
5964 (MOVSSrr (v4i32 VR128:$src1),
5965 (EXTRACT_SUBREG (v4i32 VR128:$src2), sub_ss))>;
5966 def : Pat<(v4f32 (X86Movss VR128:$src1, VR128:$src2)),
5967 (MOVSSrr (v4f32 VR128:$src1),
5968 (EXTRACT_SUBREG (v4f32 VR128:$src2), sub_ss))>;
5970 // Shuffle with MOVSD
5971 def : Pat<(v2f64 (X86Movsd VR128:$src1, (scalar_to_vector FR64:$src2))),
5972 (MOVSDrr VR128:$src1, FR64:$src2)>;
5973 def : Pat<(v2i64 (X86Movsd VR128:$src1, VR128:$src2)),
5974 (MOVSDrr (v2i64 VR128:$src1),
5975 (EXTRACT_SUBREG (v2i64 VR128:$src2), sub_sd))>;
5976 def : Pat<(v2f64 (X86Movsd VR128:$src1, VR128:$src2)),
5977 (MOVSDrr (v2f64 VR128:$src1),
5978 (EXTRACT_SUBREG (v2f64 VR128:$src2), sub_sd))>;
5979 def : Pat<(v4f32 (X86Movsd VR128:$src1, VR128:$src2)),
5980 (MOVSDrr VR128:$src1, (EXTRACT_SUBREG (v4f32 VR128:$src2), sub_sd))>;
5981 def : Pat<(v4i32 (X86Movsd VR128:$src1, VR128:$src2)),
5982 (MOVSDrr VR128:$src1, (EXTRACT_SUBREG (v4i32 VR128:$src2), sub_sd))>;
5984 // Shuffle with PSHUFHW
5985 def : Pat<(v8i16 (X86PShufhw VR128:$src, (i8 imm:$imm))),
5986 (PSHUFHWri VR128:$src, imm:$imm)>;
5987 def : Pat<(v8i16 (X86PShufhw (bc_v8i16 (memopv2i64 addr:$src)), (i8 imm:$imm))),
5988 (PSHUFHWmi addr:$src, imm:$imm)>;
5990 // Shuffle with PSHUFLW
5991 def : Pat<(v8i16 (X86PShuflw VR128:$src, (i8 imm:$imm))),
5992 (PSHUFLWri VR128:$src, imm:$imm)>;
5993 def : Pat<(v8i16 (X86PShuflw (bc_v8i16 (memopv2i64 addr:$src)), (i8 imm:$imm))),
5994 (PSHUFLWmi addr:$src, imm:$imm)>;
5996 // Shuffle with MOVLPS
5997 def : Pat<(v4f32 (X86Movlps VR128:$src1, (load addr:$src2))),
5998 (MOVLPSrm VR128:$src1, addr:$src2)>;
5999 def : Pat<(v4i32 (X86Movlps VR128:$src1, (load addr:$src2))),
6000 (MOVLPSrm VR128:$src1, addr:$src2)>;
6001 def : Pat<(X86Movlps VR128:$src1,
6002 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2))))),
6003 (MOVLPSrm VR128:$src1, addr:$src2)>;
6004 // FIXME: Instead of a X86Movlps there should be a X86Movsd here, the problem
6005 // is during lowering, where it's not possible to recognize the load fold cause
6006 // it has two uses through a bitcast. One use disappears at isel time and the
6007 // fold opportunity reappears.
6008 def : Pat<(v4f32 (X86Movlps VR128:$src1, VR128:$src2)),
6009 (MOVSDrr VR128:$src1, (EXTRACT_SUBREG (v4f32 VR128:$src2), sub_sd))>;
6011 def : Pat<(v4i32 (X86Movlps VR128:$src1, VR128:$src2)),
6012 (MOVSDrr VR128:$src1, (EXTRACT_SUBREG (v4i32 VR128:$src2), sub_sd))>;
6014 // Shuffle with MOVLPD
6015 def : Pat<(v2f64 (X86Movlpd VR128:$src1, (load addr:$src2))),
6016 (MOVLPDrm VR128:$src1, addr:$src2)>;
6017 def : Pat<(v2i64 (X86Movlpd VR128:$src1, (load addr:$src2))),
6018 (MOVLPDrm VR128:$src1, addr:$src2)>;
6019 def : Pat<(v2f64 (X86Movlpd VR128:$src1,
6020 (scalar_to_vector (loadf64 addr:$src2)))),
6021 (MOVLPDrm VR128:$src1, addr:$src2)>;
6023 // Extra patterns to match stores with MOVHPS/PD and MOVLPS/PD
6024 def : Pat<(store (f64 (vector_extract
6025 (v2f64 (X86Unpckhps VR128:$src, (undef))), (iPTR 0))),addr:$dst),
6026 (MOVHPSmr addr:$dst, VR128:$src)>;
6027 def : Pat<(store (f64 (vector_extract
6028 (v2f64 (X86Unpckhpd VR128:$src, (undef))), (iPTR 0))),addr:$dst),
6029 (MOVHPDmr addr:$dst, VR128:$src)>;
6031 def : Pat<(store (v4f32 (X86Movlps (load addr:$src1), VR128:$src2)),addr:$src1),
6032 (MOVLPSmr addr:$src1, VR128:$src2)>;
6033 def : Pat<(store (v4i32 (X86Movlps
6034 (bc_v4i32 (loadv2i64 addr:$src1)), VR128:$src2)), addr:$src1),
6035 (MOVLPSmr addr:$src1, VR128:$src2)>;
6037 def : Pat<(store (v2f64 (X86Movlpd (load addr:$src1), VR128:$src2)),addr:$src1),
6038 (MOVLPDmr addr:$src1, VR128:$src2)>;
6039 def : Pat<(store (v2i64 (X86Movlpd (load addr:$src1), VR128:$src2)),addr:$src1),
6040 (MOVLPDmr addr:$src1, VR128:$src2)>;