X86InstrMMX.td revision 263508
1//===-- X86InstrMMX.td - Describe the MMX Instruction Set --*- tablegen -*-===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This file describes 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. 13// 14// All instructions that use MMX should be in this file, even if they also use 15// SSE. 16// 17//===----------------------------------------------------------------------===// 18 19//===----------------------------------------------------------------------===// 20// MMX Multiclasses 21//===----------------------------------------------------------------------===// 22 23let Sched = WriteVecALU in { 24def MMX_INTALU_ITINS : OpndItins< 25 IIC_MMX_ALU_RR, IIC_MMX_ALU_RM 26>; 27 28def MMX_INTALUQ_ITINS : OpndItins< 29 IIC_MMX_ALUQ_RR, IIC_MMX_ALUQ_RM 30>; 31 32def MMX_PHADDSUBW : OpndItins< 33 IIC_MMX_PHADDSUBW_RR, IIC_MMX_PHADDSUBW_RM 34>; 35 36def MMX_PHADDSUBD : OpndItins< 37 IIC_MMX_PHADDSUBD_RR, IIC_MMX_PHADDSUBD_RM 38>; 39} 40 41let Sched = WriteVecIMul in 42def MMX_PMUL_ITINS : OpndItins< 43 IIC_MMX_PMUL, IIC_MMX_PMUL 44>; 45 46let Sched = WriteVecALU in { 47def MMX_PSADBW_ITINS : OpndItins< 48 IIC_MMX_PSADBW, IIC_MMX_PSADBW 49>; 50 51def MMX_MISC_FUNC_ITINS : OpndItins< 52 IIC_MMX_MISC_FUNC_MEM, IIC_MMX_MISC_FUNC_REG 53>; 54} 55 56def MMX_SHIFT_ITINS : ShiftOpndItins< 57 IIC_MMX_SHIFT_RR, IIC_MMX_SHIFT_RM, IIC_MMX_SHIFT_RI 58>; 59 60let Sched = WriteShuffle in { 61def MMX_UNPCK_H_ITINS : OpndItins< 62 IIC_MMX_UNPCK_H_RR, IIC_MMX_UNPCK_H_RM 63>; 64 65def MMX_UNPCK_L_ITINS : OpndItins< 66 IIC_MMX_UNPCK_L, IIC_MMX_UNPCK_L 67>; 68 69def MMX_PCK_ITINS : OpndItins< 70 IIC_MMX_PCK_RR, IIC_MMX_PCK_RM 71>; 72 73def MMX_PSHUF_ITINS : OpndItins< 74 IIC_MMX_PSHUF, IIC_MMX_PSHUF 75>; 76} // Sched 77 78let Sched = WriteCvtF2I in { 79def MMX_CVT_PD_ITINS : OpndItins< 80 IIC_MMX_CVT_PD_RR, IIC_MMX_CVT_PD_RM 81>; 82 83def MMX_CVT_PS_ITINS : OpndItins< 84 IIC_MMX_CVT_PS_RR, IIC_MMX_CVT_PS_RM 85>; 86} 87 88let Constraints = "$src1 = $dst" in { 89 // MMXI_binop_rm_int - Simple MMX binary operator based on intrinsic. 90 // When this is cleaned up, remove the FIXME from X86RecognizableInstr.cpp. 91 multiclass MMXI_binop_rm_int<bits<8> opc, string OpcodeStr, Intrinsic IntId, 92 OpndItins itins, bit Commutable = 0> { 93 def irr : MMXI<opc, MRMSrcReg, (outs VR64:$dst), 94 (ins VR64:$src1, VR64:$src2), 95 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"), 96 [(set VR64:$dst, (IntId VR64:$src1, VR64:$src2))], itins.rr>, 97 Sched<[itins.Sched]> { 98 let isCommutable = Commutable; 99 } 100 def irm : MMXI<opc, MRMSrcMem, (outs VR64:$dst), 101 (ins VR64:$src1, i64mem:$src2), 102 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"), 103 [(set VR64:$dst, (IntId VR64:$src1, 104 (bitconvert (load_mmx addr:$src2))))], 105 itins.rm>, Sched<[itins.Sched.Folded, ReadAfterLd]>; 106 } 107 108 multiclass MMXI_binop_rmi_int<bits<8> opc, bits<8> opc2, Format ImmForm, 109 string OpcodeStr, Intrinsic IntId, 110 Intrinsic IntId2, ShiftOpndItins itins> { 111 def rr : MMXI<opc, MRMSrcReg, (outs VR64:$dst), 112 (ins VR64:$src1, VR64:$src2), 113 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"), 114 [(set VR64:$dst, (IntId VR64:$src1, VR64:$src2))], itins.rr>, 115 Sched<[WriteVecShift]>; 116 def rm : MMXI<opc, MRMSrcMem, (outs VR64:$dst), 117 (ins VR64:$src1, i64mem:$src2), 118 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"), 119 [(set VR64:$dst, (IntId VR64:$src1, 120 (bitconvert (load_mmx addr:$src2))))], 121 itins.rm>, Sched<[WriteVecShiftLd, ReadAfterLd]>; 122 def ri : MMXIi8<opc2, ImmForm, (outs VR64:$dst), 123 (ins VR64:$src1, i32i8imm:$src2), 124 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"), 125 [(set VR64:$dst, (IntId2 VR64:$src1, (i32 imm:$src2)))], itins.ri>, 126 Sched<[WriteVecShift]>; 127 } 128} 129 130/// Unary MMX instructions requiring SSSE3. 131multiclass SS3I_unop_rm_int_mm<bits<8> opc, string OpcodeStr, 132 Intrinsic IntId64, OpndItins itins> { 133 def rr64 : MMXSS38I<opc, MRMSrcReg, (outs VR64:$dst), (ins VR64:$src), 134 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"), 135 [(set VR64:$dst, (IntId64 VR64:$src))], itins.rr>, 136 Sched<[itins.Sched]>; 137 138 def rm64 : MMXSS38I<opc, MRMSrcMem, (outs VR64:$dst), (ins i64mem:$src), 139 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"), 140 [(set VR64:$dst, 141 (IntId64 (bitconvert (memopmmx addr:$src))))], 142 itins.rm>, Sched<[itins.Sched.Folded]>; 143} 144 145/// Binary MMX instructions requiring SSSE3. 146let ImmT = NoImm, Constraints = "$src1 = $dst" in { 147multiclass SS3I_binop_rm_int_mm<bits<8> opc, string OpcodeStr, 148 Intrinsic IntId64, OpndItins itins> { 149 let isCommutable = 0 in 150 def rr64 : MMXSS38I<opc, MRMSrcReg, (outs VR64:$dst), 151 (ins VR64:$src1, VR64:$src2), 152 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"), 153 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))], itins.rr>, 154 Sched<[itins.Sched]>; 155 def rm64 : MMXSS38I<opc, MRMSrcMem, (outs VR64:$dst), 156 (ins VR64:$src1, i64mem:$src2), 157 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"), 158 [(set VR64:$dst, 159 (IntId64 VR64:$src1, 160 (bitconvert (memopmmx addr:$src2))))], itins.rm>, 161 Sched<[itins.Sched.Folded, ReadAfterLd]>; 162} 163} 164 165/// PALIGN MMX instructions (require SSSE3). 166multiclass ssse3_palign_mm<string asm, Intrinsic IntId> { 167 def R64irr : MMXSS3AI<0x0F, MRMSrcReg, (outs VR64:$dst), 168 (ins VR64:$src1, VR64:$src2, i8imm:$src3), 169 !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"), 170 [(set VR64:$dst, (IntId VR64:$src1, VR64:$src2, (i8 imm:$src3)))]>; 171 def R64irm : MMXSS3AI<0x0F, MRMSrcMem, (outs VR64:$dst), 172 (ins VR64:$src1, i64mem:$src2, i8imm:$src3), 173 !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"), 174 [(set VR64:$dst, (IntId VR64:$src1, 175 (bitconvert (load_mmx addr:$src2)), (i8 imm:$src3)))]>; 176} 177 178multiclass sse12_cvt_pint<bits<8> opc, RegisterClass SrcRC, RegisterClass DstRC, 179 Intrinsic Int, X86MemOperand x86memop, PatFrag ld_frag, 180 string asm, OpndItins itins, Domain d> { 181 def irr : MMXPI<opc, MRMSrcReg, (outs DstRC:$dst), (ins SrcRC:$src), asm, 182 [(set DstRC:$dst, (Int SrcRC:$src))], itins.rr, d>, 183 Sched<[itins.Sched]>; 184 def irm : MMXPI<opc, MRMSrcMem, (outs DstRC:$dst), (ins x86memop:$src), asm, 185 [(set DstRC:$dst, (Int (ld_frag addr:$src)))], itins.rm, d>, 186 Sched<[itins.Sched.Folded]>; 187} 188 189multiclass sse12_cvt_pint_3addr<bits<8> opc, RegisterClass SrcRC, 190 RegisterClass DstRC, Intrinsic Int, X86MemOperand x86memop, 191 PatFrag ld_frag, string asm, Domain d> { 192 def irr : MMXPI<opc, MRMSrcReg, (outs DstRC:$dst), 193 (ins DstRC:$src1, SrcRC:$src2), asm, 194 [(set DstRC:$dst, (Int DstRC:$src1, SrcRC:$src2))], 195 NoItinerary, d>; 196 def irm : MMXPI<opc, MRMSrcMem, (outs DstRC:$dst), 197 (ins DstRC:$src1, x86memop:$src2), asm, 198 [(set DstRC:$dst, (Int DstRC:$src1, (ld_frag addr:$src2)))], 199 NoItinerary, d>; 200} 201 202//===----------------------------------------------------------------------===// 203// MMX EMMS Instruction 204//===----------------------------------------------------------------------===// 205 206def MMX_EMMS : MMXI<0x77, RawFrm, (outs), (ins), "emms", 207 [(int_x86_mmx_emms)], IIC_MMX_EMMS>; 208 209//===----------------------------------------------------------------------===// 210// MMX Scalar Instructions 211//===----------------------------------------------------------------------===// 212 213// Data Transfer Instructions 214def MMX_MOVD64rr : MMXI<0x6E, MRMSrcReg, (outs VR64:$dst), (ins GR32:$src), 215 "movd\t{$src, $dst|$dst, $src}", 216 [(set VR64:$dst, 217 (x86mmx (scalar_to_vector GR32:$src)))], 218 IIC_MMX_MOV_MM_RM>, Sched<[WriteMove]>; 219let canFoldAsLoad = 1 in 220def MMX_MOVD64rm : MMXI<0x6E, MRMSrcMem, (outs VR64:$dst), (ins i32mem:$src), 221 "movd\t{$src, $dst|$dst, $src}", 222 [(set VR64:$dst, 223 (x86mmx (scalar_to_vector (loadi32 addr:$src))))], 224 IIC_MMX_MOV_MM_RM>, Sched<[WriteLoad]>; 225let mayStore = 1 in 226def MMX_MOVD64mr : MMXI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, VR64:$src), 227 "movd\t{$src, $dst|$dst, $src}", [], IIC_MMX_MOV_MM_RM>, 228 Sched<[WriteStore]>; 229 230// Low word of MMX to GPR. 231def MMX_X86movd2w : SDNode<"X86ISD::MMX_MOVD2W", SDTypeProfile<1, 1, 232 [SDTCisVT<0, i32>, SDTCisVT<1, x86mmx>]>>; 233def MMX_MOVD64grr : MMXI<0x7E, MRMDestReg, (outs GR32:$dst), (ins VR64:$src), 234 "movd\t{$src, $dst|$dst, $src}", 235 [(set GR32:$dst, 236 (MMX_X86movd2w (x86mmx VR64:$src)))], 237 IIC_MMX_MOV_REG_MM>, Sched<[WriteMove]>; 238 239def MMX_MOVD64to64rr : MMXRI<0x6E, MRMSrcReg, (outs VR64:$dst), (ins GR64:$src), 240 "movd\t{$src, $dst|$dst, $src}", 241 [(set VR64:$dst, (bitconvert GR64:$src))], 242 IIC_MMX_MOV_MM_RM>, Sched<[WriteMove]>; 243 244// These are 64 bit moves, but since the OS X assembler doesn't 245// recognize a register-register movq, we write them as 246// movd. 247let SchedRW = [WriteMove] in { 248def MMX_MOVD64from64rr : MMXRI<0x7E, MRMDestReg, 249 (outs GR64:$dst), (ins VR64:$src), 250 "movd\t{$src, $dst|$dst, $src}", 251 [(set GR64:$dst, 252 (bitconvert VR64:$src))], IIC_MMX_MOV_REG_MM>; 253let neverHasSideEffects = 1 in 254def MMX_MOVQ64rr : MMXI<0x6F, MRMSrcReg, (outs VR64:$dst), (ins VR64:$src), 255 "movq\t{$src, $dst|$dst, $src}", [], 256 IIC_MMX_MOVQ_RR>; 257} // SchedRW 258 259let SchedRW = [WriteLoad] in { 260let canFoldAsLoad = 1 in 261def MMX_MOVQ64rm : MMXI<0x6F, MRMSrcMem, (outs VR64:$dst), (ins i64mem:$src), 262 "movq\t{$src, $dst|$dst, $src}", 263 [(set VR64:$dst, (load_mmx addr:$src))], 264 IIC_MMX_MOVQ_RM>; 265def MMX_MOVQ64mr : MMXI<0x7F, MRMDestMem, (outs), (ins i64mem:$dst, VR64:$src), 266 "movq\t{$src, $dst|$dst, $src}", 267 [(store (x86mmx VR64:$src), addr:$dst)], 268 IIC_MMX_MOVQ_RM>; 269} // SchedRW 270 271let SchedRW = [WriteMove] in { 272def MMX_MOVDQ2Qrr : MMXSDIi8<0xD6, MRMSrcReg, (outs VR64:$dst), 273 (ins VR128:$src), "movdq2q\t{$src, $dst|$dst, $src}", 274 [(set VR64:$dst, 275 (x86mmx (bitconvert 276 (i64 (vector_extract (v2i64 VR128:$src), 277 (iPTR 0))))))], 278 IIC_MMX_MOVQ_RR>; 279 280def MMX_MOVQ2DQrr : MMXS2SIi8<0xD6, MRMSrcReg, (outs VR128:$dst), 281 (ins VR64:$src), "movq2dq\t{$src, $dst|$dst, $src}", 282 [(set VR128:$dst, 283 (v2i64 284 (scalar_to_vector 285 (i64 (bitconvert (x86mmx VR64:$src))))))], 286 IIC_MMX_MOVQ_RR>; 287 288let isCodeGenOnly = 1, hasSideEffects = 1 in { 289def MMX_MOVQ2FR64rr: MMXS2SIi8<0xD6, MRMSrcReg, (outs FR64:$dst), 290 (ins VR64:$src), "movq2dq\t{$src, $dst|$dst, $src}", 291 [], IIC_MMX_MOVQ_RR>; 292 293def MMX_MOVFR642Qrr: MMXSDIi8<0xD6, MRMSrcReg, (outs VR64:$dst), 294 (ins FR64:$src), "movdq2q\t{$src, $dst|$dst, $src}", 295 [], IIC_MMX_MOVQ_RR>; 296} 297} // SchedRW 298 299def MMX_MOVNTQmr : MMXI<0xE7, MRMDestMem, (outs), (ins i64mem:$dst, VR64:$src), 300 "movntq\t{$src, $dst|$dst, $src}", 301 [(int_x86_mmx_movnt_dq addr:$dst, VR64:$src)], 302 IIC_MMX_MOVQ_RM>, Sched<[WriteStore]>; 303 304let Predicates = [HasMMX] in { 305 let AddedComplexity = 15 in 306 // movd to MMX register zero-extends 307 def : Pat<(x86mmx (X86vzmovl (x86mmx (scalar_to_vector GR32:$src)))), 308 (MMX_MOVD64rr GR32:$src)>; 309 let AddedComplexity = 20 in 310 def : Pat<(x86mmx (X86vzmovl (x86mmx (scalar_to_vector (loadi32 addr:$src))))), 311 (MMX_MOVD64rm addr:$src)>; 312} 313 314// Arithmetic Instructions 315defm MMX_PABSB : SS3I_unop_rm_int_mm<0x1C, "pabsb", int_x86_ssse3_pabs_b, 316 MMX_INTALU_ITINS>; 317defm MMX_PABSW : SS3I_unop_rm_int_mm<0x1D, "pabsw", int_x86_ssse3_pabs_w, 318 MMX_INTALU_ITINS>; 319defm MMX_PABSD : SS3I_unop_rm_int_mm<0x1E, "pabsd", int_x86_ssse3_pabs_d, 320 MMX_INTALU_ITINS>; 321// -- Addition 322defm MMX_PADDB : MMXI_binop_rm_int<0xFC, "paddb", int_x86_mmx_padd_b, 323 MMX_INTALU_ITINS, 1>; 324defm MMX_PADDW : MMXI_binop_rm_int<0xFD, "paddw", int_x86_mmx_padd_w, 325 MMX_INTALU_ITINS, 1>; 326defm MMX_PADDD : MMXI_binop_rm_int<0xFE, "paddd", int_x86_mmx_padd_d, 327 MMX_INTALU_ITINS, 1>; 328defm MMX_PADDQ : MMXI_binop_rm_int<0xD4, "paddq", int_x86_mmx_padd_q, 329 MMX_INTALUQ_ITINS, 1>; 330defm MMX_PADDSB : MMXI_binop_rm_int<0xEC, "paddsb" , int_x86_mmx_padds_b, 331 MMX_INTALU_ITINS, 1>; 332defm MMX_PADDSW : MMXI_binop_rm_int<0xED, "paddsw" , int_x86_mmx_padds_w, 333 MMX_INTALU_ITINS, 1>; 334 335defm MMX_PADDUSB : MMXI_binop_rm_int<0xDC, "paddusb", int_x86_mmx_paddus_b, 336 MMX_INTALU_ITINS, 1>; 337defm MMX_PADDUSW : MMXI_binop_rm_int<0xDD, "paddusw", int_x86_mmx_paddus_w, 338 MMX_INTALU_ITINS, 1>; 339 340defm MMX_PHADDW : SS3I_binop_rm_int_mm<0x01, "phaddw", int_x86_ssse3_phadd_w, 341 MMX_PHADDSUBW>; 342defm MMX_PHADD : SS3I_binop_rm_int_mm<0x02, "phaddd", int_x86_ssse3_phadd_d, 343 MMX_PHADDSUBD>; 344defm MMX_PHADDSW : SS3I_binop_rm_int_mm<0x03, "phaddsw",int_x86_ssse3_phadd_sw, 345 MMX_PHADDSUBW>; 346 347 348// -- Subtraction 349defm MMX_PSUBB : MMXI_binop_rm_int<0xF8, "psubb", int_x86_mmx_psub_b, 350 MMX_INTALU_ITINS>; 351defm MMX_PSUBW : MMXI_binop_rm_int<0xF9, "psubw", int_x86_mmx_psub_w, 352 MMX_INTALU_ITINS>; 353defm MMX_PSUBD : MMXI_binop_rm_int<0xFA, "psubd", int_x86_mmx_psub_d, 354 MMX_INTALU_ITINS>; 355defm MMX_PSUBQ : MMXI_binop_rm_int<0xFB, "psubq", int_x86_mmx_psub_q, 356 MMX_INTALUQ_ITINS>; 357 358defm MMX_PSUBSB : MMXI_binop_rm_int<0xE8, "psubsb" , int_x86_mmx_psubs_b, 359 MMX_INTALU_ITINS>; 360defm MMX_PSUBSW : MMXI_binop_rm_int<0xE9, "psubsw" , int_x86_mmx_psubs_w, 361 MMX_INTALU_ITINS>; 362 363defm MMX_PSUBUSB : MMXI_binop_rm_int<0xD8, "psubusb", int_x86_mmx_psubus_b, 364 MMX_INTALU_ITINS>; 365defm MMX_PSUBUSW : MMXI_binop_rm_int<0xD9, "psubusw", int_x86_mmx_psubus_w, 366 MMX_INTALU_ITINS>; 367 368defm MMX_PHSUBW : SS3I_binop_rm_int_mm<0x05, "phsubw", int_x86_ssse3_phsub_w, 369 MMX_PHADDSUBW>; 370defm MMX_PHSUBD : SS3I_binop_rm_int_mm<0x06, "phsubd", int_x86_ssse3_phsub_d, 371 MMX_PHADDSUBD>; 372defm MMX_PHSUBSW : SS3I_binop_rm_int_mm<0x07, "phsubsw",int_x86_ssse3_phsub_sw, 373 MMX_PHADDSUBW>; 374 375// -- Multiplication 376defm MMX_PMULLW : MMXI_binop_rm_int<0xD5, "pmullw", int_x86_mmx_pmull_w, 377 MMX_PMUL_ITINS, 1>; 378 379defm MMX_PMULHW : MMXI_binop_rm_int<0xE5, "pmulhw", int_x86_mmx_pmulh_w, 380 MMX_PMUL_ITINS, 1>; 381defm MMX_PMULHUW : MMXI_binop_rm_int<0xE4, "pmulhuw", int_x86_mmx_pmulhu_w, 382 MMX_PMUL_ITINS, 1>; 383defm MMX_PMULUDQ : MMXI_binop_rm_int<0xF4, "pmuludq", int_x86_mmx_pmulu_dq, 384 MMX_PMUL_ITINS, 1>; 385let isCommutable = 1 in 386defm MMX_PMULHRSW : SS3I_binop_rm_int_mm<0x0B, "pmulhrsw", 387 int_x86_ssse3_pmul_hr_sw, MMX_PMUL_ITINS>; 388 389// -- Miscellanea 390defm MMX_PMADDWD : MMXI_binop_rm_int<0xF5, "pmaddwd", int_x86_mmx_pmadd_wd, 391 MMX_PMUL_ITINS, 1>; 392 393defm MMX_PMADDUBSW : SS3I_binop_rm_int_mm<0x04, "pmaddubsw", 394 int_x86_ssse3_pmadd_ub_sw, MMX_PMUL_ITINS>; 395defm MMX_PAVGB : MMXI_binop_rm_int<0xE0, "pavgb", int_x86_mmx_pavg_b, 396 MMX_MISC_FUNC_ITINS, 1>; 397defm MMX_PAVGW : MMXI_binop_rm_int<0xE3, "pavgw", int_x86_mmx_pavg_w, 398 MMX_MISC_FUNC_ITINS, 1>; 399 400defm MMX_PMINUB : MMXI_binop_rm_int<0xDA, "pminub", int_x86_mmx_pminu_b, 401 MMX_MISC_FUNC_ITINS, 1>; 402defm MMX_PMINSW : MMXI_binop_rm_int<0xEA, "pminsw", int_x86_mmx_pmins_w, 403 MMX_MISC_FUNC_ITINS, 1>; 404 405defm MMX_PMAXUB : MMXI_binop_rm_int<0xDE, "pmaxub", int_x86_mmx_pmaxu_b, 406 MMX_MISC_FUNC_ITINS, 1>; 407defm MMX_PMAXSW : MMXI_binop_rm_int<0xEE, "pmaxsw", int_x86_mmx_pmaxs_w, 408 MMX_MISC_FUNC_ITINS, 1>; 409 410defm MMX_PSADBW : MMXI_binop_rm_int<0xF6, "psadbw", int_x86_mmx_psad_bw, 411 MMX_PSADBW_ITINS, 1>; 412 413defm MMX_PSIGNB : SS3I_binop_rm_int_mm<0x08, "psignb", int_x86_ssse3_psign_b, 414 MMX_MISC_FUNC_ITINS>; 415defm MMX_PSIGNW : SS3I_binop_rm_int_mm<0x09, "psignw", int_x86_ssse3_psign_w, 416 MMX_MISC_FUNC_ITINS>; 417defm MMX_PSIGND : SS3I_binop_rm_int_mm<0x0A, "psignd", int_x86_ssse3_psign_d, 418 MMX_MISC_FUNC_ITINS>; 419let Constraints = "$src1 = $dst" in 420 defm MMX_PALIGN : ssse3_palign_mm<"palignr", int_x86_mmx_palignr_b>; 421 422// Logical Instructions 423defm MMX_PAND : MMXI_binop_rm_int<0xDB, "pand", int_x86_mmx_pand, 424 MMX_INTALU_ITINS, 1>; 425defm MMX_POR : MMXI_binop_rm_int<0xEB, "por" , int_x86_mmx_por, 426 MMX_INTALU_ITINS, 1>; 427defm MMX_PXOR : MMXI_binop_rm_int<0xEF, "pxor", int_x86_mmx_pxor, 428 MMX_INTALU_ITINS, 1>; 429defm MMX_PANDN : MMXI_binop_rm_int<0xDF, "pandn", int_x86_mmx_pandn, 430 MMX_INTALU_ITINS>; 431 432// Shift Instructions 433defm MMX_PSRLW : MMXI_binop_rmi_int<0xD1, 0x71, MRM2r, "psrlw", 434 int_x86_mmx_psrl_w, int_x86_mmx_psrli_w, 435 MMX_SHIFT_ITINS>; 436defm MMX_PSRLD : MMXI_binop_rmi_int<0xD2, 0x72, MRM2r, "psrld", 437 int_x86_mmx_psrl_d, int_x86_mmx_psrli_d, 438 MMX_SHIFT_ITINS>; 439defm MMX_PSRLQ : MMXI_binop_rmi_int<0xD3, 0x73, MRM2r, "psrlq", 440 int_x86_mmx_psrl_q, int_x86_mmx_psrli_q, 441 MMX_SHIFT_ITINS>; 442 443defm MMX_PSLLW : MMXI_binop_rmi_int<0xF1, 0x71, MRM6r, "psllw", 444 int_x86_mmx_psll_w, int_x86_mmx_pslli_w, 445 MMX_SHIFT_ITINS>; 446defm MMX_PSLLD : MMXI_binop_rmi_int<0xF2, 0x72, MRM6r, "pslld", 447 int_x86_mmx_psll_d, int_x86_mmx_pslli_d, 448 MMX_SHIFT_ITINS>; 449defm MMX_PSLLQ : MMXI_binop_rmi_int<0xF3, 0x73, MRM6r, "psllq", 450 int_x86_mmx_psll_q, int_x86_mmx_pslli_q, 451 MMX_SHIFT_ITINS>; 452 453defm MMX_PSRAW : MMXI_binop_rmi_int<0xE1, 0x71, MRM4r, "psraw", 454 int_x86_mmx_psra_w, int_x86_mmx_psrai_w, 455 MMX_SHIFT_ITINS>; 456defm MMX_PSRAD : MMXI_binop_rmi_int<0xE2, 0x72, MRM4r, "psrad", 457 int_x86_mmx_psra_d, int_x86_mmx_psrai_d, 458 MMX_SHIFT_ITINS>; 459 460// Comparison Instructions 461defm MMX_PCMPEQB : MMXI_binop_rm_int<0x74, "pcmpeqb", int_x86_mmx_pcmpeq_b, 462 MMX_INTALU_ITINS>; 463defm MMX_PCMPEQW : MMXI_binop_rm_int<0x75, "pcmpeqw", int_x86_mmx_pcmpeq_w, 464 MMX_INTALU_ITINS>; 465defm MMX_PCMPEQD : MMXI_binop_rm_int<0x76, "pcmpeqd", int_x86_mmx_pcmpeq_d, 466 MMX_INTALU_ITINS>; 467 468defm MMX_PCMPGTB : MMXI_binop_rm_int<0x64, "pcmpgtb", int_x86_mmx_pcmpgt_b, 469 MMX_INTALU_ITINS>; 470defm MMX_PCMPGTW : MMXI_binop_rm_int<0x65, "pcmpgtw", int_x86_mmx_pcmpgt_w, 471 MMX_INTALU_ITINS>; 472defm MMX_PCMPGTD : MMXI_binop_rm_int<0x66, "pcmpgtd", int_x86_mmx_pcmpgt_d, 473 MMX_INTALU_ITINS>; 474 475// -- Unpack Instructions 476defm MMX_PUNPCKHBW : MMXI_binop_rm_int<0x68, "punpckhbw", 477 int_x86_mmx_punpckhbw, 478 MMX_UNPCK_H_ITINS>; 479defm MMX_PUNPCKHWD : MMXI_binop_rm_int<0x69, "punpckhwd", 480 int_x86_mmx_punpckhwd, 481 MMX_UNPCK_H_ITINS>; 482defm MMX_PUNPCKHDQ : MMXI_binop_rm_int<0x6A, "punpckhdq", 483 int_x86_mmx_punpckhdq, 484 MMX_UNPCK_H_ITINS>; 485defm MMX_PUNPCKLBW : MMXI_binop_rm_int<0x60, "punpcklbw", 486 int_x86_mmx_punpcklbw, 487 MMX_UNPCK_L_ITINS>; 488defm MMX_PUNPCKLWD : MMXI_binop_rm_int<0x61, "punpcklwd", 489 int_x86_mmx_punpcklwd, 490 MMX_UNPCK_L_ITINS>; 491defm MMX_PUNPCKLDQ : MMXI_binop_rm_int<0x62, "punpckldq", 492 int_x86_mmx_punpckldq, 493 MMX_UNPCK_L_ITINS>; 494 495// -- Pack Instructions 496defm MMX_PACKSSWB : MMXI_binop_rm_int<0x63, "packsswb", int_x86_mmx_packsswb, 497 MMX_PCK_ITINS>; 498defm MMX_PACKSSDW : MMXI_binop_rm_int<0x6B, "packssdw", int_x86_mmx_packssdw, 499 MMX_PCK_ITINS>; 500defm MMX_PACKUSWB : MMXI_binop_rm_int<0x67, "packuswb", int_x86_mmx_packuswb, 501 MMX_PCK_ITINS>; 502 503// -- Shuffle Instructions 504defm MMX_PSHUFB : SS3I_binop_rm_int_mm<0x00, "pshufb", int_x86_ssse3_pshuf_b, 505 MMX_PSHUF_ITINS>; 506 507def MMX_PSHUFWri : MMXIi8<0x70, MRMSrcReg, 508 (outs VR64:$dst), (ins VR64:$src1, i8imm:$src2), 509 "pshufw\t{$src2, $src1, $dst|$dst, $src1, $src2}", 510 [(set VR64:$dst, 511 (int_x86_sse_pshuf_w VR64:$src1, imm:$src2))], 512 IIC_MMX_PSHUF>, Sched<[WriteShuffle]>; 513def MMX_PSHUFWmi : MMXIi8<0x70, MRMSrcMem, 514 (outs VR64:$dst), (ins i64mem:$src1, i8imm:$src2), 515 "pshufw\t{$src2, $src1, $dst|$dst, $src1, $src2}", 516 [(set VR64:$dst, 517 (int_x86_sse_pshuf_w (load_mmx addr:$src1), 518 imm:$src2))], 519 IIC_MMX_PSHUF>, Sched<[WriteShuffleLd]>; 520 521 522 523 524// -- Conversion Instructions 525defm MMX_CVTPS2PI : sse12_cvt_pint<0x2D, VR128, VR64, int_x86_sse_cvtps2pi, 526 f64mem, load, "cvtps2pi\t{$src, $dst|$dst, $src}", 527 MMX_CVT_PS_ITINS, SSEPackedSingle>, TB; 528defm MMX_CVTPD2PI : sse12_cvt_pint<0x2D, VR128, VR64, int_x86_sse_cvtpd2pi, 529 f128mem, memop, "cvtpd2pi\t{$src, $dst|$dst, $src}", 530 MMX_CVT_PD_ITINS, SSEPackedDouble>, TB, OpSize; 531defm MMX_CVTTPS2PI : sse12_cvt_pint<0x2C, VR128, VR64, int_x86_sse_cvttps2pi, 532 f64mem, load, "cvttps2pi\t{$src, $dst|$dst, $src}", 533 MMX_CVT_PS_ITINS, SSEPackedSingle>, TB; 534defm MMX_CVTTPD2PI : sse12_cvt_pint<0x2C, VR128, VR64, int_x86_sse_cvttpd2pi, 535 f128mem, memop, "cvttpd2pi\t{$src, $dst|$dst, $src}", 536 MMX_CVT_PD_ITINS, SSEPackedDouble>, TB, OpSize; 537defm MMX_CVTPI2PD : sse12_cvt_pint<0x2A, VR64, VR128, int_x86_sse_cvtpi2pd, 538 i64mem, load, "cvtpi2pd\t{$src, $dst|$dst, $src}", 539 MMX_CVT_PD_ITINS, SSEPackedDouble>, TB, OpSize; 540let Constraints = "$src1 = $dst" in { 541 defm MMX_CVTPI2PS : sse12_cvt_pint_3addr<0x2A, VR64, VR128, 542 int_x86_sse_cvtpi2ps, 543 i64mem, load, "cvtpi2ps\t{$src2, $dst|$dst, $src2}", 544 SSEPackedSingle>, TB; 545} 546 547// Extract / Insert 548def MMX_PEXTRWirri: MMXIi8<0xC5, MRMSrcReg, 549 (outs GR32orGR64:$dst), (ins VR64:$src1, i32i8imm:$src2), 550 "pextrw\t{$src2, $src1, $dst|$dst, $src1, $src2}", 551 [(set GR32orGR64:$dst, (int_x86_mmx_pextr_w VR64:$src1, 552 (iPTR imm:$src2)))], 553 IIC_MMX_PEXTR>, Sched<[WriteShuffle]>; 554let Constraints = "$src1 = $dst" in { 555 def MMX_PINSRWirri : MMXIi8<0xC4, MRMSrcReg, 556 (outs VR64:$dst), 557 (ins VR64:$src1, GR32orGR64:$src2, i32i8imm:$src3), 558 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}", 559 [(set VR64:$dst, (int_x86_mmx_pinsr_w VR64:$src1, 560 GR32orGR64:$src2, (iPTR imm:$src3)))], 561 IIC_MMX_PINSRW>, Sched<[WriteShuffle]>; 562 563 def MMX_PINSRWirmi : MMXIi8<0xC4, MRMSrcMem, 564 (outs VR64:$dst), 565 (ins VR64:$src1, i16mem:$src2, i32i8imm:$src3), 566 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}", 567 [(set VR64:$dst, (int_x86_mmx_pinsr_w VR64:$src1, 568 (i32 (anyext (loadi16 addr:$src2))), 569 (iPTR imm:$src3)))], 570 IIC_MMX_PINSRW>, Sched<[WriteShuffleLd, ReadAfterLd]>; 571} 572 573// Mask creation 574def MMX_PMOVMSKBrr : MMXI<0xD7, MRMSrcReg, (outs GR32orGR64:$dst), 575 (ins VR64:$src), 576 "pmovmskb\t{$src, $dst|$dst, $src}", 577 [(set GR32orGR64:$dst, 578 (int_x86_mmx_pmovmskb VR64:$src))]>; 579 580 581// Low word of XMM to MMX. 582def MMX_X86movdq2q : SDNode<"X86ISD::MOVDQ2Q", SDTypeProfile<1, 1, 583 [SDTCisVT<0, x86mmx>, SDTCisVT<1, v2i64>]>>; 584 585def : Pat<(x86mmx (MMX_X86movdq2q VR128:$src)), 586 (x86mmx (MMX_MOVDQ2Qrr VR128:$src))>; 587 588def : Pat<(x86mmx (MMX_X86movdq2q (loadv2i64 addr:$src))), 589 (x86mmx (MMX_MOVQ64rm addr:$src))>; 590 591// Misc. 592let SchedRW = [WriteShuffle] in { 593let Uses = [EDI] in 594def MMX_MASKMOVQ : MMXI32<0xF7, MRMSrcReg, (outs), (ins VR64:$src, VR64:$mask), 595 "maskmovq\t{$mask, $src|$src, $mask}", 596 [(int_x86_mmx_maskmovq VR64:$src, VR64:$mask, EDI)], 597 IIC_MMX_MASKMOV>; 598let Uses = [RDI] in 599def MMX_MASKMOVQ64: MMXI64<0xF7, MRMSrcReg, (outs), (ins VR64:$src, VR64:$mask), 600 "maskmovq\t{$mask, $src|$src, $mask}", 601 [(int_x86_mmx_maskmovq VR64:$src, VR64:$mask, RDI)], 602 IIC_MMX_MASKMOV>; 603} 604 605// 64-bit bit convert. 606let Predicates = [HasSSE2] in { 607def : Pat<(x86mmx (bitconvert (i64 GR64:$src))), 608 (MMX_MOVD64to64rr GR64:$src)>; 609def : Pat<(i64 (bitconvert (x86mmx VR64:$src))), 610 (MMX_MOVD64from64rr VR64:$src)>; 611def : Pat<(f64 (bitconvert (x86mmx VR64:$src))), 612 (MMX_MOVQ2FR64rr VR64:$src)>; 613def : Pat<(x86mmx (bitconvert (f64 FR64:$src))), 614 (MMX_MOVFR642Qrr FR64:$src)>; 615} 616 617 618