1 //====- X86InstrMMX.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 MMX instruction set, defining the instructions,
11 // and properties of the instructions which are needed for code generation,
12 // machine code emission, and analysis.
14 //===----------------------------------------------------------------------===//
16 //===----------------------------------------------------------------------===//
17 // MMX Pattern Fragments
18 //===----------------------------------------------------------------------===//
20 def load_mmx : PatFrag<(ops node:$ptr), (v1i64 (load node:$ptr))>;
22 def bc_v8i8 : PatFrag<(ops node:$in), (v8i8 (bitconvert node:$in))>;
23 def bc_v4i16 : PatFrag<(ops node:$in), (v4i16 (bitconvert node:$in))>;
24 def bc_v2i32 : PatFrag<(ops node:$in), (v2i32 (bitconvert node:$in))>;
25 def bc_v1i64 : PatFrag<(ops node:$in), (v1i64 (bitconvert node:$in))>;
27 //===----------------------------------------------------------------------===//
29 //===----------------------------------------------------------------------===//
31 // MMX_SHUFFLE_get_shuf_imm xform function: convert vector_shuffle mask to
33 def MMX_SHUFFLE_get_shuf_imm : SDNodeXForm<vector_shuffle, [{
34 return getI8Imm(X86::getShuffleSHUFImmediate(N));
37 // Patterns for: vector_shuffle v1, v2, <2, 6, 3, 7, ...>
38 def mmx_unpckh : PatFrag<(ops node:$lhs, node:$rhs),
39 (vector_shuffle node:$lhs, node:$rhs), [{
40 return X86::isUNPCKHMask(cast<ShuffleVectorSDNode>(N));
43 // Patterns for: vector_shuffle v1, v2, <0, 4, 2, 5, ...>
44 def mmx_unpckl : PatFrag<(ops node:$lhs, node:$rhs),
45 (vector_shuffle node:$lhs, node:$rhs), [{
46 return X86::isUNPCKLMask(cast<ShuffleVectorSDNode>(N));
49 // Patterns for: vector_shuffle v1, <undef>, <0, 0, 1, 1, ...>
50 def mmx_unpckh_undef : PatFrag<(ops node:$lhs, node:$rhs),
51 (vector_shuffle node:$lhs, node:$rhs), [{
52 return X86::isUNPCKH_v_undef_Mask(cast<ShuffleVectorSDNode>(N));
55 // Patterns for: vector_shuffle v1, <undef>, <2, 2, 3, 3, ...>
56 def mmx_unpckl_undef : PatFrag<(ops node:$lhs, node:$rhs),
57 (vector_shuffle node:$lhs, node:$rhs), [{
58 return X86::isUNPCKL_v_undef_Mask(cast<ShuffleVectorSDNode>(N));
61 def mmx_pshufw : PatFrag<(ops node:$lhs, node:$rhs),
62 (vector_shuffle node:$lhs, node:$rhs), [{
63 return X86::isPSHUFDMask(cast<ShuffleVectorSDNode>(N));
64 }], MMX_SHUFFLE_get_shuf_imm>;
66 //===----------------------------------------------------------------------===//
68 //===----------------------------------------------------------------------===//
70 let Constraints = "$src1 = $dst" in {
71 // MMXI_binop_rm - Simple MMX binary operator.
72 multiclass MMXI_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
73 ValueType OpVT, bit Commutable = 0> {
74 def rr : MMXI<opc, MRMSrcReg, (outs VR64:$dst),
75 (ins VR64:$src1, VR64:$src2),
76 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
77 [(set VR64:$dst, (OpVT (OpNode VR64:$src1, VR64:$src2)))]> {
78 let isCommutable = Commutable;
80 def rm : MMXI<opc, MRMSrcMem, (outs VR64:$dst),
81 (ins VR64:$src1, i64mem:$src2),
82 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
83 [(set VR64:$dst, (OpVT (OpNode VR64:$src1,
85 (load_mmx addr:$src2)))))]>;
88 multiclass MMXI_binop_rm_int<bits<8> opc, string OpcodeStr, Intrinsic IntId,
90 def rr : MMXI<opc, MRMSrcReg, (outs VR64:$dst),
91 (ins VR64:$src1, VR64:$src2),
92 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
93 [(set VR64:$dst, (IntId VR64:$src1, VR64:$src2))]> {
94 let isCommutable = Commutable;
96 def rm : MMXI<opc, MRMSrcMem, (outs VR64:$dst),
97 (ins VR64:$src1, i64mem:$src2),
98 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
99 [(set VR64:$dst, (IntId VR64:$src1,
100 (bitconvert (load_mmx addr:$src2))))]>;
103 // MMXI_binop_rm_v1i64 - Simple MMX binary operator whose type is v1i64.
105 // FIXME: we could eliminate this and use MMXI_binop_rm instead if tblgen knew
106 // to collapse (bitconvert VT to VT) into its operand.
108 multiclass MMXI_binop_rm_v1i64<bits<8> opc, string OpcodeStr, SDNode OpNode,
109 bit Commutable = 0> {
110 def rr : MMXI<opc, MRMSrcReg, (outs VR64:$dst),
111 (ins VR64:$src1, VR64:$src2),
112 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
113 [(set VR64:$dst, (v1i64 (OpNode VR64:$src1, VR64:$src2)))]> {
114 let isCommutable = Commutable;
116 def rm : MMXI<opc, MRMSrcMem, (outs VR64:$dst),
117 (ins VR64:$src1, i64mem:$src2),
118 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
120 (OpNode VR64:$src1,(load_mmx addr:$src2)))]>;
123 multiclass MMXI_binop_rmi_int<bits<8> opc, bits<8> opc2, Format ImmForm,
124 string OpcodeStr, Intrinsic IntId,
126 def rr : MMXI<opc, MRMSrcReg, (outs VR64:$dst),
127 (ins VR64:$src1, VR64:$src2),
128 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
129 [(set VR64:$dst, (IntId VR64:$src1, VR64:$src2))]>;
130 def rm : MMXI<opc, MRMSrcMem, (outs VR64:$dst),
131 (ins VR64:$src1, i64mem:$src2),
132 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
133 [(set VR64:$dst, (IntId VR64:$src1,
134 (bitconvert (load_mmx addr:$src2))))]>;
135 def ri : MMXIi8<opc2, ImmForm, (outs VR64:$dst),
136 (ins VR64:$src1, i32i8imm:$src2),
137 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
138 [(set VR64:$dst, (IntId2 VR64:$src1, (i32 imm:$src2)))]>;
142 //===----------------------------------------------------------------------===//
143 // MMX EMMS & FEMMS Instructions
144 //===----------------------------------------------------------------------===//
146 def MMX_EMMS : MMXI<0x77, RawFrm, (outs), (ins), "emms",
147 [(int_x86_mmx_emms)]>;
148 def MMX_FEMMS : MMXI<0x0E, RawFrm, (outs), (ins), "femms",
149 [(int_x86_mmx_femms)]>;
151 //===----------------------------------------------------------------------===//
152 // MMX Scalar Instructions
153 //===----------------------------------------------------------------------===//
155 // Data Transfer Instructions
156 def MMX_MOVD64rr : MMXI<0x6E, MRMSrcReg, (outs VR64:$dst), (ins GR32:$src),
157 "movd\t{$src, $dst|$dst, $src}",
159 (v2i32 (scalar_to_vector GR32:$src)))]>;
160 let canFoldAsLoad = 1, isReMaterializable = 1 in
161 def MMX_MOVD64rm : MMXI<0x6E, MRMSrcMem, (outs VR64:$dst), (ins i32mem:$src),
162 "movd\t{$src, $dst|$dst, $src}",
164 (v2i32 (scalar_to_vector (loadi32 addr:$src))))]>;
166 def MMX_MOVD64mr : MMXI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, VR64:$src),
167 "movd\t{$src, $dst|$dst, $src}", []>;
169 let neverHasSideEffects = 1 in
170 def MMX_MOVD64to64rr : MMXRI<0x6E, MRMSrcReg, (outs VR64:$dst), (ins GR64:$src),
171 "movd\t{$src, $dst|$dst, $src}",
174 let neverHasSideEffects = 1 in
175 // These are 64 bit moves, but since the OS X assembler doesn't
176 // recognize a register-register movq, we write them as
178 def MMX_MOVD64from64rr : MMXRI<0x7E, MRMDestReg,
179 (outs GR64:$dst), (ins VR64:$src),
180 "movd\t{$src, $dst|$dst, $src}", []>;
181 def MMX_MOVD64rrv164 : MMXI<0x6E, MRMSrcReg, (outs VR64:$dst), (ins GR64:$src),
182 "movd\t{$src, $dst|$dst, $src}",
184 (v1i64 (scalar_to_vector GR64:$src)))]>;
186 let neverHasSideEffects = 1 in
187 def MMX_MOVQ64rr : MMXI<0x6F, MRMSrcReg, (outs VR64:$dst), (ins VR64:$src),
188 "movq\t{$src, $dst|$dst, $src}", []>;
189 let canFoldAsLoad = 1, isReMaterializable = 1, mayHaveSideEffects = 1 in
190 def MMX_MOVQ64rm : MMXI<0x6F, MRMSrcMem, (outs VR64:$dst), (ins i64mem:$src),
191 "movq\t{$src, $dst|$dst, $src}",
192 [(set VR64:$dst, (load_mmx addr:$src))]>;
193 def MMX_MOVQ64mr : MMXI<0x7F, MRMDestMem, (outs), (ins i64mem:$dst, VR64:$src),
194 "movq\t{$src, $dst|$dst, $src}",
195 [(store (v1i64 VR64:$src), addr:$dst)]>;
197 def MMX_MOVDQ2Qrr : SDIi8<0xD6, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
198 "movdq2q\t{$src, $dst|$dst, $src}",
201 (i64 (vector_extract (v2i64 VR128:$src),
204 def MMX_MOVQ2DQrr : SSDIi8<0xD6, MRMSrcReg, (outs VR128:$dst), (ins VR64:$src),
205 "movq2dq\t{$src, $dst|$dst, $src}",
208 (v2i64 (scalar_to_vector (i64 (bitconvert VR64:$src))))))]>;
210 let neverHasSideEffects = 1 in
211 def MMX_MOVQ2FR64rr: SSDIi8<0xD6, MRMSrcReg, (outs FR64:$dst), (ins VR64:$src),
212 "movq2dq\t{$src, $dst|$dst, $src}", []>;
214 def MMX_MOVNTQmr : MMXI<0xE7, MRMDestMem, (outs), (ins i64mem:$dst, VR64:$src),
215 "movntq\t{$src, $dst|$dst, $src}",
216 [(int_x86_mmx_movnt_dq addr:$dst, VR64:$src)]>;
218 let AddedComplexity = 15 in
219 // movd to MMX register zero-extends
220 def MMX_MOVZDI2PDIrr : MMXI<0x6E, MRMSrcReg, (outs VR64:$dst), (ins GR32:$src),
221 "movd\t{$src, $dst|$dst, $src}",
223 (v2i32 (X86vzmovl (v2i32 (scalar_to_vector GR32:$src)))))]>;
224 let AddedComplexity = 20 in
225 def MMX_MOVZDI2PDIrm : MMXI<0x6E, MRMSrcMem, (outs VR64:$dst),
227 "movd\t{$src, $dst|$dst, $src}",
229 (v2i32 (X86vzmovl (v2i32
230 (scalar_to_vector (loadi32 addr:$src))))))]>;
232 // Arithmetic Instructions
235 defm MMX_PADDB : MMXI_binop_rm<0xFC, "paddb", add, v8i8, 1>;
236 defm MMX_PADDW : MMXI_binop_rm<0xFD, "paddw", add, v4i16, 1>;
237 defm MMX_PADDD : MMXI_binop_rm<0xFE, "paddd", add, v2i32, 1>;
238 defm MMX_PADDQ : MMXI_binop_rm<0xD4, "paddq", add, v1i64, 1>;
240 defm MMX_PADDSB : MMXI_binop_rm_int<0xEC, "paddsb" , int_x86_mmx_padds_b, 1>;
241 defm MMX_PADDSW : MMXI_binop_rm_int<0xED, "paddsw" , int_x86_mmx_padds_w, 1>;
243 defm MMX_PADDUSB : MMXI_binop_rm_int<0xDC, "paddusb", int_x86_mmx_paddus_b, 1>;
244 defm MMX_PADDUSW : MMXI_binop_rm_int<0xDD, "paddusw", int_x86_mmx_paddus_w, 1>;
247 defm MMX_PSUBB : MMXI_binop_rm<0xF8, "psubb", sub, v8i8>;
248 defm MMX_PSUBW : MMXI_binop_rm<0xF9, "psubw", sub, v4i16>;
249 defm MMX_PSUBD : MMXI_binop_rm<0xFA, "psubd", sub, v2i32>;
250 defm MMX_PSUBQ : MMXI_binop_rm<0xFB, "psubq", sub, v1i64>;
252 defm MMX_PSUBSB : MMXI_binop_rm_int<0xE8, "psubsb" , int_x86_mmx_psubs_b>;
253 defm MMX_PSUBSW : MMXI_binop_rm_int<0xE9, "psubsw" , int_x86_mmx_psubs_w>;
255 defm MMX_PSUBUSB : MMXI_binop_rm_int<0xD8, "psubusb", int_x86_mmx_psubus_b>;
256 defm MMX_PSUBUSW : MMXI_binop_rm_int<0xD9, "psubusw", int_x86_mmx_psubus_w>;
259 defm MMX_PMULLW : MMXI_binop_rm<0xD5, "pmullw", mul, v4i16, 1>;
261 defm MMX_PMULHW : MMXI_binop_rm_int<0xE5, "pmulhw", int_x86_mmx_pmulh_w, 1>;
262 defm MMX_PMULHUW : MMXI_binop_rm_int<0xE4, "pmulhuw", int_x86_mmx_pmulhu_w, 1>;
263 defm MMX_PMULUDQ : MMXI_binop_rm_int<0xF4, "pmuludq", int_x86_mmx_pmulu_dq, 1>;
266 defm MMX_PMADDWD : MMXI_binop_rm_int<0xF5, "pmaddwd", int_x86_mmx_pmadd_wd, 1>;
268 defm MMX_PAVGB : MMXI_binop_rm_int<0xE0, "pavgb", int_x86_mmx_pavg_b, 1>;
269 defm MMX_PAVGW : MMXI_binop_rm_int<0xE3, "pavgw", int_x86_mmx_pavg_w, 1>;
271 defm MMX_PMINUB : MMXI_binop_rm_int<0xDA, "pminub", int_x86_mmx_pminu_b, 1>;
272 defm MMX_PMINSW : MMXI_binop_rm_int<0xEA, "pminsw", int_x86_mmx_pmins_w, 1>;
274 defm MMX_PMAXUB : MMXI_binop_rm_int<0xDE, "pmaxub", int_x86_mmx_pmaxu_b, 1>;
275 defm MMX_PMAXSW : MMXI_binop_rm_int<0xEE, "pmaxsw", int_x86_mmx_pmaxs_w, 1>;
277 defm MMX_PSADBW : MMXI_binop_rm_int<0xF6, "psadbw", int_x86_mmx_psad_bw, 1>;
279 // Logical Instructions
280 defm MMX_PAND : MMXI_binop_rm_v1i64<0xDB, "pand", and, 1>;
281 defm MMX_POR : MMXI_binop_rm_v1i64<0xEB, "por" , or, 1>;
282 defm MMX_PXOR : MMXI_binop_rm_v1i64<0xEF, "pxor", xor, 1>;
284 let Constraints = "$src1 = $dst" in {
285 def MMX_PANDNrr : MMXI<0xDF, MRMSrcReg,
286 (outs VR64:$dst), (ins VR64:$src1, VR64:$src2),
287 "pandn\t{$src2, $dst|$dst, $src2}",
288 [(set VR64:$dst, (v1i64 (and (vnot VR64:$src1),
290 def MMX_PANDNrm : MMXI<0xDF, MRMSrcMem,
291 (outs VR64:$dst), (ins VR64:$src1, i64mem:$src2),
292 "pandn\t{$src2, $dst|$dst, $src2}",
293 [(set VR64:$dst, (v1i64 (and (vnot VR64:$src1),
294 (load addr:$src2))))]>;
297 // Shift Instructions
298 defm MMX_PSRLW : MMXI_binop_rmi_int<0xD1, 0x71, MRM2r, "psrlw",
299 int_x86_mmx_psrl_w, int_x86_mmx_psrli_w>;
300 defm MMX_PSRLD : MMXI_binop_rmi_int<0xD2, 0x72, MRM2r, "psrld",
301 int_x86_mmx_psrl_d, int_x86_mmx_psrli_d>;
302 defm MMX_PSRLQ : MMXI_binop_rmi_int<0xD3, 0x73, MRM2r, "psrlq",
303 int_x86_mmx_psrl_q, int_x86_mmx_psrli_q>;
305 defm MMX_PSLLW : MMXI_binop_rmi_int<0xF1, 0x71, MRM6r, "psllw",
306 int_x86_mmx_psll_w, int_x86_mmx_pslli_w>;
307 defm MMX_PSLLD : MMXI_binop_rmi_int<0xF2, 0x72, MRM6r, "pslld",
308 int_x86_mmx_psll_d, int_x86_mmx_pslli_d>;
309 defm MMX_PSLLQ : MMXI_binop_rmi_int<0xF3, 0x73, MRM6r, "psllq",
310 int_x86_mmx_psll_q, int_x86_mmx_pslli_q>;
312 defm MMX_PSRAW : MMXI_binop_rmi_int<0xE1, 0x71, MRM4r, "psraw",
313 int_x86_mmx_psra_w, int_x86_mmx_psrai_w>;
314 defm MMX_PSRAD : MMXI_binop_rmi_int<0xE2, 0x72, MRM4r, "psrad",
315 int_x86_mmx_psra_d, int_x86_mmx_psrai_d>;
317 // Shift up / down and insert zero's.
318 def : Pat<(v1i64 (X86vshl VR64:$src, (i8 imm:$amt))),
319 (v1i64 (MMX_PSLLQri VR64:$src, imm:$amt))>;
320 def : Pat<(v1i64 (X86vshr VR64:$src, (i8 imm:$amt))),
321 (v1i64 (MMX_PSRLQri VR64:$src, imm:$amt))>;
323 // Comparison Instructions
324 defm MMX_PCMPEQB : MMXI_binop_rm_int<0x74, "pcmpeqb", int_x86_mmx_pcmpeq_b>;
325 defm MMX_PCMPEQW : MMXI_binop_rm_int<0x75, "pcmpeqw", int_x86_mmx_pcmpeq_w>;
326 defm MMX_PCMPEQD : MMXI_binop_rm_int<0x76, "pcmpeqd", int_x86_mmx_pcmpeq_d>;
328 defm MMX_PCMPGTB : MMXI_binop_rm_int<0x64, "pcmpgtb", int_x86_mmx_pcmpgt_b>;
329 defm MMX_PCMPGTW : MMXI_binop_rm_int<0x65, "pcmpgtw", int_x86_mmx_pcmpgt_w>;
330 defm MMX_PCMPGTD : MMXI_binop_rm_int<0x66, "pcmpgtd", int_x86_mmx_pcmpgt_d>;
332 // Conversion Instructions
334 // -- Unpack Instructions
335 let Constraints = "$src1 = $dst" in {
336 // Unpack High Packed Data Instructions
337 def MMX_PUNPCKHBWrr : MMXI<0x68, MRMSrcReg,
338 (outs VR64:$dst), (ins VR64:$src1, VR64:$src2),
339 "punpckhbw\t{$src2, $dst|$dst, $src2}",
341 (v8i8 (mmx_unpckh VR64:$src1, VR64:$src2)))]>;
342 def MMX_PUNPCKHBWrm : MMXI<0x68, MRMSrcMem,
343 (outs VR64:$dst), (ins VR64:$src1, i64mem:$src2),
344 "punpckhbw\t{$src2, $dst|$dst, $src2}",
346 (v8i8 (mmx_unpckh VR64:$src1,
347 (bc_v8i8 (load_mmx addr:$src2)))))]>;
349 def MMX_PUNPCKHWDrr : MMXI<0x69, MRMSrcReg,
350 (outs VR64:$dst), (ins VR64:$src1, VR64:$src2),
351 "punpckhwd\t{$src2, $dst|$dst, $src2}",
353 (v4i16 (mmx_unpckh VR64:$src1, VR64:$src2)))]>;
354 def MMX_PUNPCKHWDrm : MMXI<0x69, MRMSrcMem,
355 (outs VR64:$dst), (ins VR64:$src1, i64mem:$src2),
356 "punpckhwd\t{$src2, $dst|$dst, $src2}",
358 (v4i16 (mmx_unpckh VR64:$src1,
359 (bc_v4i16 (load_mmx addr:$src2)))))]>;
361 def MMX_PUNPCKHDQrr : MMXI<0x6A, MRMSrcReg,
362 (outs VR64:$dst), (ins VR64:$src1, VR64:$src2),
363 "punpckhdq\t{$src2, $dst|$dst, $src2}",
365 (v2i32 (mmx_unpckh VR64:$src1, VR64:$src2)))]>;
366 def MMX_PUNPCKHDQrm : MMXI<0x6A, MRMSrcMem,
367 (outs VR64:$dst), (ins VR64:$src1, i64mem:$src2),
368 "punpckhdq\t{$src2, $dst|$dst, $src2}",
370 (v2i32 (mmx_unpckh VR64:$src1,
371 (bc_v2i32 (load_mmx addr:$src2)))))]>;
373 // Unpack Low Packed Data Instructions
374 def MMX_PUNPCKLBWrr : MMXI<0x60, MRMSrcReg,
375 (outs VR64:$dst), (ins VR64:$src1, VR64:$src2),
376 "punpcklbw\t{$src2, $dst|$dst, $src2}",
378 (v8i8 (mmx_unpckl VR64:$src1, VR64:$src2)))]>;
379 def MMX_PUNPCKLBWrm : MMXI<0x60, MRMSrcMem,
380 (outs VR64:$dst), (ins VR64:$src1, i64mem:$src2),
381 "punpcklbw\t{$src2, $dst|$dst, $src2}",
383 (v8i8 (mmx_unpckl VR64:$src1,
384 (bc_v8i8 (load_mmx addr:$src2)))))]>;
386 def MMX_PUNPCKLWDrr : MMXI<0x61, MRMSrcReg,
387 (outs VR64:$dst), (ins VR64:$src1, VR64:$src2),
388 "punpcklwd\t{$src2, $dst|$dst, $src2}",
390 (v4i16 (mmx_unpckl VR64:$src1, VR64:$src2)))]>;
391 def MMX_PUNPCKLWDrm : MMXI<0x61, MRMSrcMem,
392 (outs VR64:$dst), (ins VR64:$src1, i64mem:$src2),
393 "punpcklwd\t{$src2, $dst|$dst, $src2}",
395 (v4i16 (mmx_unpckl VR64:$src1,
396 (bc_v4i16 (load_mmx addr:$src2)))))]>;
398 def MMX_PUNPCKLDQrr : MMXI<0x62, MRMSrcReg,
399 (outs VR64:$dst), (ins VR64:$src1, VR64:$src2),
400 "punpckldq\t{$src2, $dst|$dst, $src2}",
402 (v2i32 (mmx_unpckl VR64:$src1, VR64:$src2)))]>;
403 def MMX_PUNPCKLDQrm : MMXI<0x62, MRMSrcMem,
404 (outs VR64:$dst), (ins VR64:$src1, i64mem:$src2),
405 "punpckldq\t{$src2, $dst|$dst, $src2}",
407 (v2i32 (mmx_unpckl VR64:$src1,
408 (bc_v2i32 (load_mmx addr:$src2)))))]>;
411 // -- Pack Instructions
412 defm MMX_PACKSSWB : MMXI_binop_rm_int<0x63, "packsswb", int_x86_mmx_packsswb>;
413 defm MMX_PACKSSDW : MMXI_binop_rm_int<0x6B, "packssdw", int_x86_mmx_packssdw>;
414 defm MMX_PACKUSWB : MMXI_binop_rm_int<0x67, "packuswb", int_x86_mmx_packuswb>;
416 // -- Shuffle Instructions
417 def MMX_PSHUFWri : MMXIi8<0x70, MRMSrcReg,
418 (outs VR64:$dst), (ins VR64:$src1, i8imm:$src2),
419 "pshufw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
421 (v4i16 (mmx_pshufw:$src2 VR64:$src1, (undef))))]>;
422 def MMX_PSHUFWmi : MMXIi8<0x70, MRMSrcMem,
423 (outs VR64:$dst), (ins i64mem:$src1, i8imm:$src2),
424 "pshufw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
426 (mmx_pshufw:$src2 (bc_v4i16 (load_mmx addr:$src1)),
429 // -- Conversion Instructions
430 let neverHasSideEffects = 1 in {
431 def MMX_CVTPD2PIrr : MMX2I<0x2D, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
432 "cvtpd2pi\t{$src, $dst|$dst, $src}", []>;
434 def MMX_CVTPD2PIrm : MMX2I<0x2D, MRMSrcMem, (outs VR64:$dst),
436 "cvtpd2pi\t{$src, $dst|$dst, $src}", []>;
438 def MMX_CVTPI2PDrr : MMX2I<0x2A, MRMSrcReg, (outs VR128:$dst), (ins VR64:$src),
439 "cvtpi2pd\t{$src, $dst|$dst, $src}", []>;
441 def MMX_CVTPI2PDrm : MMX2I<0x2A, MRMSrcMem, (outs VR128:$dst),
443 "cvtpi2pd\t{$src, $dst|$dst, $src}", []>;
445 def MMX_CVTPI2PSrr : MMXI<0x2A, MRMSrcReg, (outs VR128:$dst), (ins VR64:$src),
446 "cvtpi2ps\t{$src, $dst|$dst, $src}", []>;
448 def MMX_CVTPI2PSrm : MMXI<0x2A, MRMSrcMem, (outs VR128:$dst),
450 "cvtpi2ps\t{$src, $dst|$dst, $src}", []>;
452 def MMX_CVTPS2PIrr : MMXI<0x2D, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
453 "cvtps2pi\t{$src, $dst|$dst, $src}", []>;
455 def MMX_CVTPS2PIrm : MMXI<0x2D, MRMSrcMem, (outs VR64:$dst), (ins f64mem:$src),
456 "cvtps2pi\t{$src, $dst|$dst, $src}", []>;
458 def MMX_CVTTPD2PIrr : MMX2I<0x2C, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
459 "cvttpd2pi\t{$src, $dst|$dst, $src}", []>;
461 def MMX_CVTTPD2PIrm : MMX2I<0x2C, MRMSrcMem, (outs VR64:$dst),
463 "cvttpd2pi\t{$src, $dst|$dst, $src}", []>;
465 def MMX_CVTTPS2PIrr : MMXI<0x2C, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
466 "cvttps2pi\t{$src, $dst|$dst, $src}", []>;
468 def MMX_CVTTPS2PIrm : MMXI<0x2C, MRMSrcMem, (outs VR64:$dst), (ins f64mem:$src),
469 "cvttps2pi\t{$src, $dst|$dst, $src}", []>;
470 } // end neverHasSideEffects
474 def MMX_X86pextrw : SDNode<"X86ISD::PEXTRW", SDTypeProfile<1, 2, []>, []>;
475 def MMX_X86pinsrw : SDNode<"X86ISD::PINSRW", SDTypeProfile<1, 3, []>, []>;
477 def MMX_PEXTRWri : MMXIi8<0xC5, MRMSrcReg,
478 (outs GR32:$dst), (ins VR64:$src1, i16i8imm:$src2),
479 "pextrw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
480 [(set GR32:$dst, (MMX_X86pextrw (v4i16 VR64:$src1),
481 (iPTR imm:$src2)))]>;
482 let Constraints = "$src1 = $dst" in {
483 def MMX_PINSRWrri : MMXIi8<0xC4, MRMSrcReg,
484 (outs VR64:$dst), (ins VR64:$src1, GR32:$src2,
486 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
487 [(set VR64:$dst, (v4i16 (MMX_X86pinsrw (v4i16 VR64:$src1),
488 GR32:$src2,(iPTR imm:$src3))))]>;
489 def MMX_PINSRWrmi : MMXIi8<0xC4, MRMSrcMem,
490 (outs VR64:$dst), (ins VR64:$src1, i16mem:$src2,
492 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
494 (v4i16 (MMX_X86pinsrw (v4i16 VR64:$src1),
495 (i32 (anyext (loadi16 addr:$src2))),
496 (iPTR imm:$src3))))]>;
500 def MMX_PMOVMSKBrr : MMXI<0xD7, MRMSrcReg, (outs GR32:$dst), (ins VR64:$src),
501 "pmovmskb\t{$src, $dst|$dst, $src}",
502 [(set GR32:$dst, (int_x86_mmx_pmovmskb VR64:$src))]>;
506 def MMX_MASKMOVQ : MMXI<0xF7, MRMSrcReg, (outs), (ins VR64:$src, VR64:$mask),
507 "maskmovq\t{$mask, $src|$src, $mask}",
508 [(int_x86_mmx_maskmovq VR64:$src, VR64:$mask, EDI)]>;
510 def MMX_MASKMOVQ64: MMXI64<0xF7, MRMSrcReg, (outs), (ins VR64:$src, VR64:$mask),
511 "maskmovq\t{$mask, $src|$src, $mask}",
512 [(int_x86_mmx_maskmovq VR64:$src, VR64:$mask, RDI)]>;
514 //===----------------------------------------------------------------------===//
515 // Alias Instructions
516 //===----------------------------------------------------------------------===//
518 // Alias instructions that map zero vector to pxor.
519 let isReMaterializable = 1, isCodeGenOnly = 1 in {
520 def MMX_V_SET0 : MMXI<0xEF, MRMInitReg, (outs VR64:$dst), (ins),
522 [(set VR64:$dst, (v2i32 immAllZerosV))]>;
523 def MMX_V_SETALLONES : MMXI<0x76, MRMInitReg, (outs VR64:$dst), (ins),
524 "pcmpeqd\t$dst, $dst",
525 [(set VR64:$dst, (v2i32 immAllOnesV))]>;
528 let Predicates = [HasMMX] in {
529 def : Pat<(v1i64 immAllZerosV), (MMX_V_SET0)>;
530 def : Pat<(v4i16 immAllZerosV), (MMX_V_SET0)>;
531 def : Pat<(v8i8 immAllZerosV), (MMX_V_SET0)>;
534 //===----------------------------------------------------------------------===//
535 // Non-Instruction Patterns
536 //===----------------------------------------------------------------------===//
538 // Store 64-bit integer vector values.
539 def : Pat<(store (v8i8 VR64:$src), addr:$dst),
540 (MMX_MOVQ64mr addr:$dst, VR64:$src)>;
541 def : Pat<(store (v4i16 VR64:$src), addr:$dst),
542 (MMX_MOVQ64mr addr:$dst, VR64:$src)>;
543 def : Pat<(store (v2i32 VR64:$src), addr:$dst),
544 (MMX_MOVQ64mr addr:$dst, VR64:$src)>;
545 def : Pat<(store (v2f32 VR64:$src), addr:$dst),
546 (MMX_MOVQ64mr addr:$dst, VR64:$src)>;
547 def : Pat<(store (v1i64 VR64:$src), addr:$dst),
548 (MMX_MOVQ64mr addr:$dst, VR64:$src)>;
551 def : Pat<(v8i8 (bitconvert (v1i64 VR64:$src))), (v8i8 VR64:$src)>;
552 def : Pat<(v8i8 (bitconvert (v2i32 VR64:$src))), (v8i8 VR64:$src)>;
553 def : Pat<(v8i8 (bitconvert (v2f32 VR64:$src))), (v8i8 VR64:$src)>;
554 def : Pat<(v8i8 (bitconvert (v4i16 VR64:$src))), (v8i8 VR64:$src)>;
555 def : Pat<(v4i16 (bitconvert (v1i64 VR64:$src))), (v4i16 VR64:$src)>;
556 def : Pat<(v4i16 (bitconvert (v2i32 VR64:$src))), (v4i16 VR64:$src)>;
557 def : Pat<(v4i16 (bitconvert (v2f32 VR64:$src))), (v4i16 VR64:$src)>;
558 def : Pat<(v4i16 (bitconvert (v8i8 VR64:$src))), (v4i16 VR64:$src)>;
559 def : Pat<(v2i32 (bitconvert (v1i64 VR64:$src))), (v2i32 VR64:$src)>;
560 def : Pat<(v2i32 (bitconvert (v2f32 VR64:$src))), (v2i32 VR64:$src)>;
561 def : Pat<(v2i32 (bitconvert (v4i16 VR64:$src))), (v2i32 VR64:$src)>;
562 def : Pat<(v2i32 (bitconvert (v8i8 VR64:$src))), (v2i32 VR64:$src)>;
563 def : Pat<(v2f32 (bitconvert (v1i64 VR64:$src))), (v2f32 VR64:$src)>;
564 def : Pat<(v2f32 (bitconvert (v2i32 VR64:$src))), (v2f32 VR64:$src)>;
565 def : Pat<(v2f32 (bitconvert (v4i16 VR64:$src))), (v2f32 VR64:$src)>;
566 def : Pat<(v2f32 (bitconvert (v8i8 VR64:$src))), (v2f32 VR64:$src)>;
567 def : Pat<(v1i64 (bitconvert (v2i32 VR64:$src))), (v1i64 VR64:$src)>;
568 def : Pat<(v1i64 (bitconvert (v2f32 VR64:$src))), (v1i64 VR64:$src)>;
569 def : Pat<(v1i64 (bitconvert (v4i16 VR64:$src))), (v1i64 VR64:$src)>;
570 def : Pat<(v1i64 (bitconvert (v8i8 VR64:$src))), (v1i64 VR64:$src)>;
572 // 64-bit bit convert.
573 def : Pat<(v1i64 (bitconvert (i64 GR64:$src))),
574 (MMX_MOVD64to64rr GR64:$src)>;
575 def : Pat<(v2i32 (bitconvert (i64 GR64:$src))),
576 (MMX_MOVD64to64rr GR64:$src)>;
577 def : Pat<(v2f32 (bitconvert (i64 GR64:$src))),
578 (MMX_MOVD64to64rr GR64:$src)>;
579 def : Pat<(v4i16 (bitconvert (i64 GR64:$src))),
580 (MMX_MOVD64to64rr GR64:$src)>;
581 def : Pat<(v8i8 (bitconvert (i64 GR64:$src))),
582 (MMX_MOVD64to64rr GR64:$src)>;
583 def : Pat<(i64 (bitconvert (v1i64 VR64:$src))),
584 (MMX_MOVD64from64rr VR64:$src)>;
585 def : Pat<(i64 (bitconvert (v2i32 VR64:$src))),
586 (MMX_MOVD64from64rr VR64:$src)>;
587 def : Pat<(i64 (bitconvert (v2f32 VR64:$src))),
588 (MMX_MOVD64from64rr VR64:$src)>;
589 def : Pat<(i64 (bitconvert (v4i16 VR64:$src))),
590 (MMX_MOVD64from64rr VR64:$src)>;
591 def : Pat<(i64 (bitconvert (v8i8 VR64:$src))),
592 (MMX_MOVD64from64rr VR64:$src)>;
593 def : Pat<(f64 (bitconvert (v1i64 VR64:$src))),
594 (MMX_MOVQ2FR64rr VR64:$src)>;
595 def : Pat<(f64 (bitconvert (v2i32 VR64:$src))),
596 (MMX_MOVQ2FR64rr VR64:$src)>;
597 def : Pat<(f64 (bitconvert (v4i16 VR64:$src))),
598 (MMX_MOVQ2FR64rr VR64:$src)>;
599 def : Pat<(f64 (bitconvert (v8i8 VR64:$src))),
600 (MMX_MOVQ2FR64rr VR64:$src)>;
602 let AddedComplexity = 20 in {
603 def : Pat<(v2i32 (X86vzmovl (bc_v2i32 (load_mmx addr:$src)))),
604 (MMX_MOVZDI2PDIrm addr:$src)>;
608 let AddedComplexity = 15 in {
609 def : Pat<(v2i32 (X86vzmovl VR64:$src)),
610 (MMX_PUNPCKLDQrr VR64:$src, (MMX_V_SET0))>;
613 // Patterns to perform canonical versions of vector shuffling.
614 let AddedComplexity = 10 in {
615 def : Pat<(v8i8 (mmx_unpckl_undef VR64:$src, (undef))),
616 (MMX_PUNPCKLBWrr VR64:$src, VR64:$src)>;
617 def : Pat<(v4i16 (mmx_unpckl_undef VR64:$src, (undef))),
618 (MMX_PUNPCKLWDrr VR64:$src, VR64:$src)>;
619 def : Pat<(v2i32 (mmx_unpckl_undef VR64:$src, (undef))),
620 (MMX_PUNPCKLDQrr VR64:$src, VR64:$src)>;
623 let AddedComplexity = 10 in {
624 def : Pat<(v8i8 (mmx_unpckh_undef VR64:$src, (undef))),
625 (MMX_PUNPCKHBWrr VR64:$src, VR64:$src)>;
626 def : Pat<(v4i16 (mmx_unpckh_undef VR64:$src, (undef))),
627 (MMX_PUNPCKHWDrr VR64:$src, VR64:$src)>;
628 def : Pat<(v2i32 (mmx_unpckh_undef VR64:$src, (undef))),
629 (MMX_PUNPCKHDQrr VR64:$src, VR64:$src)>;
632 // Patterns to perform vector shuffling with a zeroed out vector.
633 let AddedComplexity = 20 in {
634 def : Pat<(bc_v2i32 (mmx_unpckl immAllZerosV,
635 (v2i32 (scalar_to_vector (load_mmx addr:$src))))),
636 (MMX_PUNPCKLDQrm VR64:$src, VR64:$src)>;
639 // Some special case PANDN patterns.
640 // FIXME: Get rid of these.
641 def : Pat<(v1i64 (and (xor VR64:$src1, (bc_v1i64 (v2i32 immAllOnesV))),
643 (MMX_PANDNrr VR64:$src1, VR64:$src2)>;
644 def : Pat<(v1i64 (and (xor VR64:$src1, (bc_v1i64 (v4i16 immAllOnesV_bc))),
646 (MMX_PANDNrr VR64:$src1, VR64:$src2)>;
647 def : Pat<(v1i64 (and (xor VR64:$src1, (bc_v1i64 (v8i8 immAllOnesV_bc))),
649 (MMX_PANDNrr VR64:$src1, VR64:$src2)>;
651 def : Pat<(v1i64 (and (xor VR64:$src1, (bc_v1i64 (v2i32 immAllOnesV))),
653 (MMX_PANDNrm VR64:$src1, addr:$src2)>;
654 def : Pat<(v1i64 (and (xor VR64:$src1, (bc_v1i64 (v4i16 immAllOnesV_bc))),
656 (MMX_PANDNrm VR64:$src1, addr:$src2)>;
657 def : Pat<(v1i64 (and (xor VR64:$src1, (bc_v1i64 (v8i8 immAllOnesV_bc))),
659 (MMX_PANDNrm VR64:$src1, addr:$src2)>;
661 // Move MMX to lower 64-bit of XMM
662 def : Pat<(v2i64 (scalar_to_vector (i64 (bitconvert (v8i8 VR64:$src))))),
663 (v2i64 (MMX_MOVQ2DQrr VR64:$src))>;
664 def : Pat<(v2i64 (scalar_to_vector (i64 (bitconvert (v4i16 VR64:$src))))),
665 (v2i64 (MMX_MOVQ2DQrr VR64:$src))>;
666 def : Pat<(v2i64 (scalar_to_vector (i64 (bitconvert (v2i32 VR64:$src))))),
667 (v2i64 (MMX_MOVQ2DQrr VR64:$src))>;
668 def : Pat<(v2i64 (scalar_to_vector (i64 (bitconvert (v1i64 VR64:$src))))),
669 (v2i64 (MMX_MOVQ2DQrr VR64:$src))>;
671 // Move lower 64-bit of XMM to MMX.
672 def : Pat<(v2i32 (bitconvert (i64 (vector_extract (v2i64 VR128:$src),
674 (v2i32 (MMX_MOVDQ2Qrr VR128:$src))>;
675 def : Pat<(v4i16 (bitconvert (i64 (vector_extract (v2i64 VR128:$src),
677 (v4i16 (MMX_MOVDQ2Qrr VR128:$src))>;
678 def : Pat<(v8i8 (bitconvert (i64 (vector_extract (v2i64 VR128:$src),
680 (v8i8 (MMX_MOVDQ2Qrr VR128:$src))>;
682 // Patterns for vector comparisons
683 def : Pat<(v8i8 (X86pcmpeqb VR64:$src1, VR64:$src2)),
684 (MMX_PCMPEQBrr VR64:$src1, VR64:$src2)>;
685 def : Pat<(v8i8 (X86pcmpeqb VR64:$src1, (bitconvert (load_mmx addr:$src2)))),
686 (MMX_PCMPEQBrm VR64:$src1, addr:$src2)>;
687 def : Pat<(v4i16 (X86pcmpeqw VR64:$src1, VR64:$src2)),
688 (MMX_PCMPEQWrr VR64:$src1, VR64:$src2)>;
689 def : Pat<(v4i16 (X86pcmpeqw VR64:$src1, (bitconvert (load_mmx addr:$src2)))),
690 (MMX_PCMPEQWrm VR64:$src1, addr:$src2)>;
691 def : Pat<(v2i32 (X86pcmpeqd VR64:$src1, VR64:$src2)),
692 (MMX_PCMPEQDrr VR64:$src1, VR64:$src2)>;
693 def : Pat<(v2i32 (X86pcmpeqd VR64:$src1, (bitconvert (load_mmx addr:$src2)))),
694 (MMX_PCMPEQDrm VR64:$src1, addr:$src2)>;
696 def : Pat<(v8i8 (X86pcmpgtb VR64:$src1, VR64:$src2)),
697 (MMX_PCMPGTBrr VR64:$src1, VR64:$src2)>;
698 def : Pat<(v8i8 (X86pcmpgtb VR64:$src1, (bitconvert (load_mmx addr:$src2)))),
699 (MMX_PCMPGTBrm VR64:$src1, addr:$src2)>;
700 def : Pat<(v4i16 (X86pcmpgtw VR64:$src1, VR64:$src2)),
701 (MMX_PCMPGTWrr VR64:$src1, VR64:$src2)>;
702 def : Pat<(v4i16 (X86pcmpgtw VR64:$src1, (bitconvert (load_mmx addr:$src2)))),
703 (MMX_PCMPGTWrm VR64:$src1, addr:$src2)>;
704 def : Pat<(v2i32 (X86pcmpgtd VR64:$src1, VR64:$src2)),
705 (MMX_PCMPGTDrr VR64:$src1, VR64:$src2)>;
706 def : Pat<(v2i32 (X86pcmpgtd VR64:$src1, (bitconvert (load_mmx addr:$src2)))),
707 (MMX_PCMPGTDrm VR64:$src1, addr:$src2)>;
709 // CMOV* - Used to implement the SELECT DAG operation. Expanded by the
710 // scheduler into a branch sequence.
711 // These are expanded by the scheduler.
712 let Uses = [EFLAGS], usesCustomDAGSchedInserter = 1 in {
713 def CMOV_V1I64 : I<0, Pseudo,
714 (outs VR64:$dst), (ins VR64:$t, VR64:$f, i8imm:$cond),
715 "#CMOV_V1I64 PSEUDO!",
717 (v1i64 (X86cmov VR64:$t, VR64:$f, imm:$cond,