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 Constraints = "$src1 = $dst" in { 24 // MMXI_binop_rm_int - Simple MMX binary operator based on intrinsic. 25 // When this is cleaned up, remove the FIXME from X86RecognizableInstr.cpp. 26 multiclass MMXI_binop_rm_int<bits<8> opc, string OpcodeStr, Intrinsic IntId, 27 bit Commutable = 0> { 28 def irr : MMXI<opc, MRMSrcReg, (outs VR64:$dst), 29 (ins VR64:$src1, VR64:$src2), 30 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"), 31 [(set VR64:$dst, (IntId VR64:$src1, VR64:$src2))]> { 32 let isCommutable = Commutable; 33 } 34 def irm : MMXI<opc, MRMSrcMem, (outs VR64:$dst), 35 (ins VR64:$src1, i64mem:$src2), 36 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"), 37 [(set VR64:$dst, (IntId VR64:$src1, 38 (bitconvert (load_mmx addr:$src2))))]>; 39 } 40 41 multiclass MMXI_binop_rmi_int<bits<8> opc, bits<8> opc2, Format ImmForm, 42 string OpcodeStr, Intrinsic IntId, 43 Intrinsic IntId2> { 44 def rr : MMXI<opc, MRMSrcReg, (outs VR64:$dst), 45 (ins VR64:$src1, VR64:$src2), 46 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"), 47 [(set VR64:$dst, (IntId VR64:$src1, VR64:$src2))]>; 48 def rm : MMXI<opc, MRMSrcMem, (outs VR64:$dst), 49 (ins VR64:$src1, i64mem:$src2), 50 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"), 51 [(set VR64:$dst, (IntId VR64:$src1, 52 (bitconvert (load_mmx addr:$src2))))]>; 53 def ri : MMXIi8<opc2, ImmForm, (outs VR64:$dst), 54 (ins VR64:$src1, i32i8imm:$src2), 55 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"), 56 [(set VR64:$dst, (IntId2 VR64:$src1, (i32 imm:$src2)))]>; 57 } 58} 59 60/// Unary MMX instructions requiring SSSE3. 61multiclass SS3I_unop_rm_int_mm<bits<8> opc, string OpcodeStr, 62 Intrinsic IntId64> { 63 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst), (ins VR64:$src), 64 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"), 65 [(set VR64:$dst, (IntId64 VR64:$src))]>; 66 67 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst), (ins i64mem:$src), 68 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"), 69 [(set VR64:$dst, 70 (IntId64 (bitconvert (memopmmx addr:$src))))]>; 71} 72 73/// Binary MMX instructions requiring SSSE3. 74let ImmT = NoImm, Constraints = "$src1 = $dst" in { 75multiclass SS3I_binop_rm_int_mm<bits<8> opc, string OpcodeStr, 76 Intrinsic IntId64> { 77 let isCommutable = 0 in 78 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst), 79 (ins VR64:$src1, VR64:$src2), 80 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"), 81 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]>; 82 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst), 83 (ins VR64:$src1, i64mem:$src2), 84 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"), 85 [(set VR64:$dst, 86 (IntId64 VR64:$src1, 87 (bitconvert (memopmmx addr:$src2))))]>; 88} 89} 90 91/// PALIGN MMX instructions (require SSSE3). 92multiclass ssse3_palign_mm<string asm, Intrinsic IntId> { 93 def R64irr : SS3AI<0x0F, MRMSrcReg, (outs VR64:$dst), 94 (ins VR64:$src1, VR64:$src2, i8imm:$src3), 95 !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"), 96 [(set VR64:$dst, (IntId VR64:$src1, VR64:$src2, (i8 imm:$src3)))]>; 97 def R64irm : SS3AI<0x0F, MRMSrcMem, (outs VR64:$dst), 98 (ins VR64:$src1, i64mem:$src2, i8imm:$src3), 99 !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"), 100 [(set VR64:$dst, (IntId VR64:$src1, 101 (bitconvert (load_mmx addr:$src2)), (i8 imm:$src3)))]>; 102} 103 104multiclass sse12_cvt_pint<bits<8> opc, RegisterClass SrcRC, RegisterClass DstRC, 105 Intrinsic Int, X86MemOperand x86memop, PatFrag ld_frag, 106 string asm, Domain d> { 107 def irr : PI<opc, MRMSrcReg, (outs DstRC:$dst), (ins SrcRC:$src), asm, 108 [(set DstRC:$dst, (Int SrcRC:$src))], d>; 109 def irm : PI<opc, MRMSrcMem, (outs DstRC:$dst), (ins x86memop:$src), asm, 110 [(set DstRC:$dst, (Int (ld_frag addr:$src)))], d>; 111} 112 113multiclass sse12_cvt_pint_3addr<bits<8> opc, RegisterClass SrcRC, 114 RegisterClass DstRC, Intrinsic Int, X86MemOperand x86memop, 115 PatFrag ld_frag, string asm, Domain d> { 116 def irr : PI<opc, MRMSrcReg, (outs DstRC:$dst),(ins DstRC:$src1, SrcRC:$src2), 117 asm, [(set DstRC:$dst, (Int DstRC:$src1, SrcRC:$src2))], d>; 118 def irm : PI<opc, MRMSrcMem, (outs DstRC:$dst), 119 (ins DstRC:$src1, x86memop:$src2), asm, 120 [(set DstRC:$dst, (Int DstRC:$src1, (ld_frag addr:$src2)))], d>; 121} 122 123//===----------------------------------------------------------------------===// 124// MMX EMMS Instruction 125//===----------------------------------------------------------------------===// 126 127def MMX_EMMS : MMXI<0x77, RawFrm, (outs), (ins), "emms", 128 [(int_x86_mmx_emms)]>; 129 130//===----------------------------------------------------------------------===// 131// MMX Scalar Instructions 132//===----------------------------------------------------------------------===// 133 134// Data Transfer Instructions 135def MMX_MOVD64rr : MMXI<0x6E, MRMSrcReg, (outs VR64:$dst), (ins GR32:$src), 136 "movd\t{$src, $dst|$dst, $src}", 137 [(set VR64:$dst, 138 (x86mmx (scalar_to_vector GR32:$src)))]>; 139let canFoldAsLoad = 1 in 140def MMX_MOVD64rm : MMXI<0x6E, MRMSrcMem, (outs VR64:$dst), (ins i32mem:$src), 141 "movd\t{$src, $dst|$dst, $src}", 142 [(set VR64:$dst, 143 (x86mmx (scalar_to_vector (loadi32 addr:$src))))]>; 144 145let Predicates = [HasMMX] in { 146 let AddedComplexity = 15 in 147 def : Pat<(x86mmx (MMX_X86movw2d GR32:$src)), 148 (MMX_MOVD64rr GR32:$src)>; 149 let AddedComplexity = 20 in 150 def : Pat<(x86mmx (MMX_X86movw2d (loadi32 addr:$src))), 151 (MMX_MOVD64rm addr:$src)>; 152} 153 154let mayStore = 1 in 155def MMX_MOVD64mr : MMXI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, VR64:$src), 156 "movd\t{$src, $dst|$dst, $src}", []>; 157 158def MMX_MOVD64grr : MMXI<0x7E, MRMDestReg, (outs GR32:$dst), (ins VR64:$src), 159 "movd\t{$src, $dst|$dst, $src}", 160 [(set GR32:$dst, 161 (MMX_X86movd2w (x86mmx VR64:$src)))]>; 162 163let neverHasSideEffects = 1 in 164def MMX_MOVD64to64rr : MMXRI<0x6E, MRMSrcReg, (outs VR64:$dst), (ins GR64:$src), 165 "movd\t{$src, $dst|$dst, $src}", 166 []>; 167 168// These are 64 bit moves, but since the OS X assembler doesn't 169// recognize a register-register movq, we write them as 170// movd. 171def MMX_MOVD64from64rr : MMXRI<0x7E, MRMDestReg, 172 (outs GR64:$dst), (ins VR64:$src), 173 "movd\t{$src, $dst|$dst, $src}", 174 [(set GR64:$dst, 175 (bitconvert VR64:$src))]>; 176def MMX_MOVD64rrv164 : MMXRI<0x6E, MRMSrcReg, (outs VR64:$dst), (ins GR64:$src), 177 "movd\t{$src, $dst|$dst, $src}", 178 [(set VR64:$dst, 179 (bitconvert GR64:$src))]>; 180let neverHasSideEffects = 1 in 181def MMX_MOVQ64rr : MMXI<0x6F, MRMSrcReg, (outs VR64:$dst), (ins VR64:$src), 182 "movq\t{$src, $dst|$dst, $src}", []>; 183let canFoldAsLoad = 1 in 184def MMX_MOVQ64rm : MMXI<0x6F, MRMSrcMem, (outs VR64:$dst), (ins i64mem:$src), 185 "movq\t{$src, $dst|$dst, $src}", 186 [(set VR64:$dst, (load_mmx addr:$src))]>; 187def MMX_MOVQ64mr : MMXI<0x7F, MRMDestMem, (outs), (ins i64mem:$dst, VR64:$src), 188 "movq\t{$src, $dst|$dst, $src}", 189 [(store (x86mmx VR64:$src), addr:$dst)]>; 190 191def MMX_MOVDQ2Qrr : SDIi8<0xD6, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src), 192 "movdq2q\t{$src, $dst|$dst, $src}", 193 [(set VR64:$dst, 194 (x86mmx (bitconvert 195 (i64 (vector_extract (v2i64 VR128:$src), 196 (iPTR 0))))))]>; 197 198def MMX_MOVQ2DQrr : SSDIi8<0xD6, MRMSrcReg, (outs VR128:$dst), (ins VR64:$src), 199 "movq2dq\t{$src, $dst|$dst, $src}", 200 [(set VR128:$dst, 201 (v2i64 (scalar_to_vector 202 (i64 (bitconvert (x86mmx VR64:$src))))))]>; 203 204let neverHasSideEffects = 1 in 205def MMX_MOVQ2FR64rr: SSDIi8<0xD6, MRMSrcReg, (outs FR64:$dst), (ins VR64:$src), 206 "movq2dq\t{$src, $dst|$dst, $src}", []>; 207 208def MMX_MOVFR642Qrr: SDIi8<0xD6, MRMSrcReg, (outs VR64:$dst), (ins FR64:$src), 209 "movdq2q\t{$src, $dst|$dst, $src}", []>; 210 211def MMX_MOVNTQmr : MMXI<0xE7, MRMDestMem, (outs), (ins i64mem:$dst, VR64:$src), 212 "movntq\t{$src, $dst|$dst, $src}", 213 [(int_x86_mmx_movnt_dq addr:$dst, VR64:$src)]>; 214 215let AddedComplexity = 15 in 216// movd to MMX register zero-extends 217def MMX_MOVZDI2PDIrr : MMXI<0x6E, MRMSrcReg, (outs VR64:$dst), (ins GR32:$src), 218 "movd\t{$src, $dst|$dst, $src}", 219 [(set VR64:$dst, 220 (x86mmx (X86vzmovl (x86mmx (scalar_to_vector GR32:$src)))))]>; 221let AddedComplexity = 20 in 222def MMX_MOVZDI2PDIrm : MMXI<0x6E, MRMSrcMem, (outs VR64:$dst), 223 (ins i32mem:$src), 224 "movd\t{$src, $dst|$dst, $src}", 225 [(set VR64:$dst, 226 (x86mmx (X86vzmovl (x86mmx 227 (scalar_to_vector (loadi32 addr:$src))))))]>; 228 229// Arithmetic Instructions 230defm MMX_PABSB : SS3I_unop_rm_int_mm<0x1C, "pabsb", int_x86_ssse3_pabs_b>; 231defm MMX_PABSW : SS3I_unop_rm_int_mm<0x1D, "pabsw", int_x86_ssse3_pabs_w>; 232defm MMX_PABSD : SS3I_unop_rm_int_mm<0x1E, "pabsd", int_x86_ssse3_pabs_d>; 233// -- Addition 234defm MMX_PADDB : MMXI_binop_rm_int<0xFC, "paddb", int_x86_mmx_padd_b, 1>; 235defm MMX_PADDW : MMXI_binop_rm_int<0xFD, "paddw", int_x86_mmx_padd_w, 1>; 236defm MMX_PADDD : MMXI_binop_rm_int<0xFE, "paddd", int_x86_mmx_padd_d, 1>; 237defm MMX_PADDQ : MMXI_binop_rm_int<0xD4, "paddq", int_x86_mmx_padd_q, 1>; 238defm MMX_PADDSB : MMXI_binop_rm_int<0xEC, "paddsb" , int_x86_mmx_padds_b, 1>; 239defm MMX_PADDSW : MMXI_binop_rm_int<0xED, "paddsw" , int_x86_mmx_padds_w, 1>; 240 241defm MMX_PADDUSB : MMXI_binop_rm_int<0xDC, "paddusb", int_x86_mmx_paddus_b, 1>; 242defm MMX_PADDUSW : MMXI_binop_rm_int<0xDD, "paddusw", int_x86_mmx_paddus_w, 1>; 243 244defm MMX_PHADDW : SS3I_binop_rm_int_mm<0x01, "phaddw", int_x86_ssse3_phadd_w>; 245defm MMX_PHADD : SS3I_binop_rm_int_mm<0x02, "phaddd", int_x86_ssse3_phadd_d>; 246defm MMX_PHADDSW : SS3I_binop_rm_int_mm<0x03, "phaddsw",int_x86_ssse3_phadd_sw>; 247 248 249// -- Subtraction 250defm MMX_PSUBB : MMXI_binop_rm_int<0xF8, "psubb", int_x86_mmx_psub_b>; 251defm MMX_PSUBW : MMXI_binop_rm_int<0xF9, "psubw", int_x86_mmx_psub_w>; 252defm MMX_PSUBD : MMXI_binop_rm_int<0xFA, "psubd", int_x86_mmx_psub_d>; 253defm MMX_PSUBQ : MMXI_binop_rm_int<0xFB, "psubq", int_x86_mmx_psub_q>; 254 255defm MMX_PSUBSB : MMXI_binop_rm_int<0xE8, "psubsb" , int_x86_mmx_psubs_b>; 256defm MMX_PSUBSW : MMXI_binop_rm_int<0xE9, "psubsw" , int_x86_mmx_psubs_w>; 257 258defm MMX_PSUBUSB : MMXI_binop_rm_int<0xD8, "psubusb", int_x86_mmx_psubus_b>; 259defm MMX_PSUBUSW : MMXI_binop_rm_int<0xD9, "psubusw", int_x86_mmx_psubus_w>; 260 261defm MMX_PHSUBW : SS3I_binop_rm_int_mm<0x05, "phsubw", int_x86_ssse3_phsub_w>; 262defm MMX_PHSUBD : SS3I_binop_rm_int_mm<0x06, "phsubd", int_x86_ssse3_phsub_d>; 263defm MMX_PHSUBSW : SS3I_binop_rm_int_mm<0x07, "phsubsw",int_x86_ssse3_phsub_sw>; 264 265// -- Multiplication 266defm MMX_PMULLW : MMXI_binop_rm_int<0xD5, "pmullw", int_x86_mmx_pmull_w, 1>; 267 268defm MMX_PMULHW : MMXI_binop_rm_int<0xE5, "pmulhw", int_x86_mmx_pmulh_w, 1>; 269defm MMX_PMULHUW : MMXI_binop_rm_int<0xE4, "pmulhuw", int_x86_mmx_pmulhu_w, 1>; 270defm MMX_PMULUDQ : MMXI_binop_rm_int<0xF4, "pmuludq", int_x86_mmx_pmulu_dq, 1>; 271let isCommutable = 1 in 272defm MMX_PMULHRSW : SS3I_binop_rm_int_mm<0x0B, "pmulhrsw", 273 int_x86_ssse3_pmul_hr_sw>; 274 275// -- Miscellanea 276defm MMX_PMADDWD : MMXI_binop_rm_int<0xF5, "pmaddwd", int_x86_mmx_pmadd_wd, 1>; 277 278defm MMX_PMADDUBSW : SS3I_binop_rm_int_mm<0x04, "pmaddubsw", 279 int_x86_ssse3_pmadd_ub_sw>; 280defm MMX_PAVGB : MMXI_binop_rm_int<0xE0, "pavgb", int_x86_mmx_pavg_b, 1>; 281defm MMX_PAVGW : MMXI_binop_rm_int<0xE3, "pavgw", int_x86_mmx_pavg_w, 1>; 282 283defm MMX_PMINUB : MMXI_binop_rm_int<0xDA, "pminub", int_x86_mmx_pminu_b, 1>; 284defm MMX_PMINSW : MMXI_binop_rm_int<0xEA, "pminsw", int_x86_mmx_pmins_w, 1>; 285 286defm MMX_PMAXUB : MMXI_binop_rm_int<0xDE, "pmaxub", int_x86_mmx_pmaxu_b, 1>; 287defm MMX_PMAXSW : MMXI_binop_rm_int<0xEE, "pmaxsw", int_x86_mmx_pmaxs_w, 1>; 288 289defm MMX_PSADBW : MMXI_binop_rm_int<0xF6, "psadbw", int_x86_mmx_psad_bw, 1>; 290 291defm MMX_PSIGNB : SS3I_binop_rm_int_mm<0x08, "psignb", int_x86_ssse3_psign_b>; 292defm MMX_PSIGNW : SS3I_binop_rm_int_mm<0x09, "psignw", int_x86_ssse3_psign_w>; 293defm MMX_PSIGND : SS3I_binop_rm_int_mm<0x0A, "psignd", int_x86_ssse3_psign_d>; 294let Constraints = "$src1 = $dst" in 295 defm MMX_PALIGN : ssse3_palign_mm<"palignr", int_x86_mmx_palignr_b>; 296 297// Logical Instructions 298defm MMX_PAND : MMXI_binop_rm_int<0xDB, "pand", int_x86_mmx_pand, 1>; 299defm MMX_POR : MMXI_binop_rm_int<0xEB, "por" , int_x86_mmx_por, 1>; 300defm MMX_PXOR : MMXI_binop_rm_int<0xEF, "pxor", int_x86_mmx_pxor, 1>; 301defm MMX_PANDN : MMXI_binop_rm_int<0xDF, "pandn", int_x86_mmx_pandn>; 302 303// Shift Instructions 304defm MMX_PSRLW : MMXI_binop_rmi_int<0xD1, 0x71, MRM2r, "psrlw", 305 int_x86_mmx_psrl_w, int_x86_mmx_psrli_w>; 306defm MMX_PSRLD : MMXI_binop_rmi_int<0xD2, 0x72, MRM2r, "psrld", 307 int_x86_mmx_psrl_d, int_x86_mmx_psrli_d>; 308defm MMX_PSRLQ : MMXI_binop_rmi_int<0xD3, 0x73, MRM2r, "psrlq", 309 int_x86_mmx_psrl_q, int_x86_mmx_psrli_q>; 310 311defm MMX_PSLLW : MMXI_binop_rmi_int<0xF1, 0x71, MRM6r, "psllw", 312 int_x86_mmx_psll_w, int_x86_mmx_pslli_w>; 313defm MMX_PSLLD : MMXI_binop_rmi_int<0xF2, 0x72, MRM6r, "pslld", 314 int_x86_mmx_psll_d, int_x86_mmx_pslli_d>; 315defm MMX_PSLLQ : MMXI_binop_rmi_int<0xF3, 0x73, MRM6r, "psllq", 316 int_x86_mmx_psll_q, int_x86_mmx_pslli_q>; 317 318defm MMX_PSRAW : MMXI_binop_rmi_int<0xE1, 0x71, MRM4r, "psraw", 319 int_x86_mmx_psra_w, int_x86_mmx_psrai_w>; 320defm MMX_PSRAD : MMXI_binop_rmi_int<0xE2, 0x72, MRM4r, "psrad", 321 int_x86_mmx_psra_d, int_x86_mmx_psrai_d>; 322 323// Comparison Instructions 324defm MMX_PCMPEQB : MMXI_binop_rm_int<0x74, "pcmpeqb", int_x86_mmx_pcmpeq_b>; 325defm MMX_PCMPEQW : MMXI_binop_rm_int<0x75, "pcmpeqw", int_x86_mmx_pcmpeq_w>; 326defm MMX_PCMPEQD : MMXI_binop_rm_int<0x76, "pcmpeqd", int_x86_mmx_pcmpeq_d>; 327 328defm MMX_PCMPGTB : MMXI_binop_rm_int<0x64, "pcmpgtb", int_x86_mmx_pcmpgt_b>; 329defm MMX_PCMPGTW : MMXI_binop_rm_int<0x65, "pcmpgtw", int_x86_mmx_pcmpgt_w>; 330defm MMX_PCMPGTD : MMXI_binop_rm_int<0x66, "pcmpgtd", int_x86_mmx_pcmpgt_d>; 331 332// -- Unpack Instructions 333defm MMX_PUNPCKHBW : MMXI_binop_rm_int<0x68, "punpckhbw", 334 int_x86_mmx_punpckhbw>; 335defm MMX_PUNPCKHWD : MMXI_binop_rm_int<0x69, "punpckhwd", 336 int_x86_mmx_punpckhwd>; 337defm MMX_PUNPCKHDQ : MMXI_binop_rm_int<0x6A, "punpckhdq", 338 int_x86_mmx_punpckhdq>; 339defm MMX_PUNPCKLBW : MMXI_binop_rm_int<0x60, "punpcklbw", 340 int_x86_mmx_punpcklbw>; 341defm MMX_PUNPCKLWD : MMXI_binop_rm_int<0x61, "punpcklwd", 342 int_x86_mmx_punpcklwd>; 343defm MMX_PUNPCKLDQ : MMXI_binop_rm_int<0x62, "punpckldq", 344 int_x86_mmx_punpckldq>; 345 346// -- Pack Instructions 347defm MMX_PACKSSWB : MMXI_binop_rm_int<0x63, "packsswb", int_x86_mmx_packsswb>; 348defm MMX_PACKSSDW : MMXI_binop_rm_int<0x6B, "packssdw", int_x86_mmx_packssdw>; 349defm MMX_PACKUSWB : MMXI_binop_rm_int<0x67, "packuswb", int_x86_mmx_packuswb>; 350 351// -- Shuffle Instructions 352defm MMX_PSHUFB : SS3I_binop_rm_int_mm<0x00, "pshufb", int_x86_ssse3_pshuf_b>; 353 354def MMX_PSHUFWri : MMXIi8<0x70, MRMSrcReg, 355 (outs VR64:$dst), (ins VR64:$src1, i8imm:$src2), 356 "pshufw\t{$src2, $src1, $dst|$dst, $src1, $src2}", 357 [(set VR64:$dst, 358 (int_x86_sse_pshuf_w VR64:$src1, imm:$src2))]>; 359def MMX_PSHUFWmi : MMXIi8<0x70, MRMSrcMem, 360 (outs VR64:$dst), (ins i64mem:$src1, i8imm:$src2), 361 "pshufw\t{$src2, $src1, $dst|$dst, $src1, $src2}", 362 [(set VR64:$dst, 363 (int_x86_sse_pshuf_w (load_mmx addr:$src1), 364 imm:$src2))]>; 365 366 367 368 369 370// -- Conversion Instructions 371defm MMX_CVTPS2PI : sse12_cvt_pint<0x2D, VR128, VR64, int_x86_sse_cvtps2pi, 372 f64mem, load, "cvtps2pi\t{$src, $dst|$dst, $src}", 373 SSEPackedSingle>, TB; 374defm MMX_CVTPD2PI : sse12_cvt_pint<0x2D, VR128, VR64, int_x86_sse_cvtpd2pi, 375 f128mem, memop, "cvtpd2pi\t{$src, $dst|$dst, $src}", 376 SSEPackedDouble>, TB, OpSize; 377defm MMX_CVTTPS2PI : sse12_cvt_pint<0x2C, VR128, VR64, int_x86_sse_cvttps2pi, 378 f64mem, load, "cvttps2pi\t{$src, $dst|$dst, $src}", 379 SSEPackedSingle>, TB; 380defm MMX_CVTTPD2PI : sse12_cvt_pint<0x2C, VR128, VR64, int_x86_sse_cvttpd2pi, 381 f128mem, memop, "cvttpd2pi\t{$src, $dst|$dst, $src}", 382 SSEPackedDouble>, TB, OpSize; 383defm MMX_CVTPI2PD : sse12_cvt_pint<0x2A, VR64, VR128, int_x86_sse_cvtpi2pd, 384 i64mem, load, "cvtpi2pd\t{$src, $dst|$dst, $src}", 385 SSEPackedDouble>, TB, OpSize; 386let Constraints = "$src1 = $dst" in { 387 defm MMX_CVTPI2PS : sse12_cvt_pint_3addr<0x2A, VR64, VR128, 388 int_x86_sse_cvtpi2ps, 389 i64mem, load, "cvtpi2ps\t{$src2, $dst|$dst, $src2}", 390 SSEPackedSingle>, TB; 391} 392 393// Extract / Insert 394def MMX_PEXTRWirri: MMXIi8<0xC5, MRMSrcReg, 395 (outs GR32:$dst), (ins VR64:$src1, i32i8imm:$src2), 396 "pextrw\t{$src2, $src1, $dst|$dst, $src1, $src2}", 397 [(set GR32:$dst, (int_x86_mmx_pextr_w VR64:$src1, 398 (iPTR imm:$src2)))]>; 399let Constraints = "$src1 = $dst" in { 400 def MMX_PINSRWirri : MMXIi8<0xC4, MRMSrcReg, 401 (outs VR64:$dst), 402 (ins VR64:$src1, GR32:$src2, i32i8imm:$src3), 403 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}", 404 [(set VR64:$dst, (int_x86_mmx_pinsr_w VR64:$src1, 405 GR32:$src2, (iPTR imm:$src3)))]>; 406 407 def MMX_PINSRWirmi : MMXIi8<0xC4, MRMSrcMem, 408 (outs VR64:$dst), 409 (ins VR64:$src1, i16mem:$src2, i32i8imm:$src3), 410 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}", 411 [(set VR64:$dst, (int_x86_mmx_pinsr_w VR64:$src1, 412 (i32 (anyext (loadi16 addr:$src2))), 413 (iPTR imm:$src3)))]>; 414} 415 416// Mask creation 417def MMX_PMOVMSKBrr : MMXI<0xD7, MRMSrcReg, (outs GR32:$dst), (ins VR64:$src), 418 "pmovmskb\t{$src, $dst|$dst, $src}", 419 [(set GR32:$dst, 420 (int_x86_mmx_pmovmskb VR64:$src))]>; 421 422 423// MMX to XMM for vector types 424def MMX_X86movq2dq : SDNode<"X86ISD::MOVQ2DQ", SDTypeProfile<1, 1, 425 [SDTCisVT<0, v2i64>, SDTCisVT<1, x86mmx>]>>; 426 427def : Pat<(v2i64 (MMX_X86movq2dq VR64:$src)), 428 (v2i64 (MMX_MOVQ2DQrr VR64:$src))>; 429 430def : Pat<(v2i64 (MMX_X86movq2dq (load_mmx addr:$src))), 431 (v2i64 (MOVQI2PQIrm addr:$src))>; 432 433def : Pat<(v2i64 (MMX_X86movq2dq 434 (x86mmx (scalar_to_vector (loadi32 addr:$src))))), 435 (v2i64 (MOVDI2PDIrm addr:$src))>; 436 437// Low word of XMM to MMX. 438def MMX_X86movdq2q : SDNode<"X86ISD::MOVDQ2Q", SDTypeProfile<1, 1, 439 [SDTCisVT<0, x86mmx>, SDTCisVT<1, v2i64>]>>; 440 441def : Pat<(x86mmx (MMX_X86movdq2q VR128:$src)), 442 (x86mmx (MMX_MOVDQ2Qrr VR128:$src))>; 443 444def : Pat<(x86mmx (MMX_X86movdq2q (loadv2i64 addr:$src))), 445 (x86mmx (MMX_MOVQ64rm addr:$src))>; 446 447// Misc. 448let Uses = [EDI] in 449def MMX_MASKMOVQ : MMXI<0xF7, MRMSrcReg, (outs), (ins VR64:$src, VR64:$mask), 450 "maskmovq\t{$mask, $src|$src, $mask}", 451 [(int_x86_mmx_maskmovq VR64:$src, VR64:$mask, EDI)]>; 452let Uses = [RDI] in 453def MMX_MASKMOVQ64: MMXI64<0xF7, MRMSrcReg, (outs), (ins VR64:$src, VR64:$mask), 454 "maskmovq\t{$mask, $src|$src, $mask}", 455 [(int_x86_mmx_maskmovq VR64:$src, VR64:$mask, RDI)]>; 456 457// 64-bit bit convert. 458def : Pat<(x86mmx (bitconvert (i64 GR64:$src))), 459 (MMX_MOVD64to64rr GR64:$src)>; 460def : Pat<(i64 (bitconvert (x86mmx VR64:$src))), 461 (MMX_MOVD64from64rr VR64:$src)>; 462def : Pat<(f64 (bitconvert (x86mmx VR64:$src))), 463 (MMX_MOVQ2FR64rr VR64:$src)>; 464def : Pat<(x86mmx (bitconvert (f64 FR64:$src))), 465 (MMX_MOVFR642Qrr FR64:$src)>; 466 467 468