| 1 | //===-- ARMAddressingModes.h - ARM Addressing Modes -------------*- C++ -*-===// |
| 2 | // |
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 | // See https://llvm.org/LICENSE.txt for license information. |
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 | // |
| 7 | //===----------------------------------------------------------------------===// |
| 8 | // |
| 9 | // This file contains the ARM addressing mode implementation stuff. |
| 10 | // |
| 11 | //===----------------------------------------------------------------------===// |
| 12 | |
| 13 | #ifndef LLVM_LIB_TARGET_ARM_MCTARGETDESC_ARMADDRESSINGMODES_H |
| 14 | #define LLVM_LIB_TARGET_ARM_MCTARGETDESC_ARMADDRESSINGMODES_H |
| 15 | |
| 16 | #include "llvm/ADT/APFloat.h" |
| 17 | #include "llvm/ADT/APInt.h" |
| 18 | #include "llvm/ADT/bit.h" |
| 19 | #include "llvm/Support/ErrorHandling.h" |
| 20 | #include "llvm/Support/MathExtras.h" |
| 21 | #include <cassert> |
| 22 | |
| 23 | namespace llvm { |
| 24 | |
| 25 | /// ARM_AM - ARM Addressing Mode Stuff |
| 26 | namespace ARM_AM { |
| 27 | enum ShiftOpc { |
| 28 | no_shift = 0, |
| 29 | asr, |
| 30 | lsl, |
| 31 | lsr, |
| 32 | ror, |
| 33 | rrx, |
| 34 | uxtw |
| 35 | }; |
| 36 | |
| 37 | enum AddrOpc { |
| 38 | sub = 0, |
| 39 | add |
| 40 | }; |
| 41 | |
| 42 | inline const char *getAddrOpcStr(AddrOpc Op) { return Op == sub ? "-" : "" ; } |
| 43 | |
| 44 | inline StringRef getShiftOpcStr(ShiftOpc Op) { |
| 45 | switch (Op) { |
| 46 | default: llvm_unreachable("Unknown shift opc!" ); |
| 47 | case ARM_AM::asr: return "asr" ; |
| 48 | case ARM_AM::lsl: return "lsl" ; |
| 49 | case ARM_AM::lsr: return "lsr" ; |
| 50 | case ARM_AM::ror: return "ror" ; |
| 51 | case ARM_AM::rrx: return "rrx" ; |
| 52 | case ARM_AM::uxtw: return "uxtw" ; |
| 53 | } |
| 54 | } |
| 55 | |
| 56 | inline unsigned getShiftOpcEncoding(ShiftOpc Op) { |
| 57 | switch (Op) { |
| 58 | default: llvm_unreachable("Unknown shift opc!" ); |
| 59 | case ARM_AM::asr: return 2; |
| 60 | case ARM_AM::lsl: return 0; |
| 61 | case ARM_AM::lsr: return 1; |
| 62 | case ARM_AM::ror: return 3; |
| 63 | } |
| 64 | } |
| 65 | |
| 66 | enum AMSubMode { |
| 67 | bad_am_submode = 0, |
| 68 | ia, |
| 69 | ib, |
| 70 | da, |
| 71 | db |
| 72 | }; |
| 73 | |
| 74 | inline const char *getAMSubModeStr(AMSubMode Mode) { |
| 75 | switch (Mode) { |
| 76 | default: llvm_unreachable("Unknown addressing sub-mode!" ); |
| 77 | case ARM_AM::ia: return "ia" ; |
| 78 | case ARM_AM::ib: return "ib" ; |
| 79 | case ARM_AM::da: return "da" ; |
| 80 | case ARM_AM::db: return "db" ; |
| 81 | } |
| 82 | } |
| 83 | |
| 84 | //===--------------------------------------------------------------------===// |
| 85 | // Addressing Mode #1: shift_operand with registers |
| 86 | //===--------------------------------------------------------------------===// |
| 87 | // |
| 88 | // This 'addressing mode' is used for arithmetic instructions. It can |
| 89 | // represent things like: |
| 90 | // reg |
| 91 | // reg [asr|lsl|lsr|ror|rrx] reg |
| 92 | // reg [asr|lsl|lsr|ror|rrx] imm |
| 93 | // |
| 94 | // This is stored three operands [rega, regb, opc]. The first is the base |
| 95 | // reg, the second is the shift amount (or reg0 if not present or imm). The |
| 96 | // third operand encodes the shift opcode and the imm if a reg isn't present. |
| 97 | // |
| 98 | inline unsigned getSORegOpc(ShiftOpc ShOp, unsigned Imm) { |
| 99 | return ShOp | (Imm << 3); |
| 100 | } |
| 101 | inline unsigned getSORegOffset(unsigned Op) { return Op >> 3; } |
| 102 | inline ShiftOpc getSORegShOp(unsigned Op) { return (ShiftOpc)(Op & 7); } |
| 103 | |
| 104 | /// getSOImmValImm - Given an encoded imm field for the reg/imm form, return |
| 105 | /// the 8-bit imm value. |
| 106 | inline unsigned getSOImmValImm(unsigned Imm) { return Imm & 0xFF; } |
| 107 | /// getSOImmValRot - Given an encoded imm field for the reg/imm form, return |
| 108 | /// the rotate amount. |
| 109 | inline unsigned getSOImmValRot(unsigned Imm) { return (Imm >> 8) * 2; } |
| 110 | |
| 111 | /// getSOImmValRotate - Try to handle Imm with an immediate shifter operand, |
| 112 | /// computing the rotate amount to use. If this immediate value cannot be |
| 113 | /// handled with a single shifter-op, determine a good rotate amount that will |
| 114 | /// take a maximal chunk of bits out of the immediate. |
| 115 | inline unsigned getSOImmValRotate(unsigned Imm) { |
| 116 | // 8-bit (or less) immediates are trivially shifter_operands with a rotate |
| 117 | // of zero. |
| 118 | if ((Imm & ~255U) == 0) return 0; |
| 119 | |
| 120 | // Use CTZ to compute the rotate amount. |
| 121 | unsigned TZ = llvm::countr_zero(Val: Imm); |
| 122 | |
| 123 | // Rotate amount must be even. Something like 0x200 must be rotated 8 bits, |
| 124 | // not 9. |
| 125 | unsigned RotAmt = TZ & ~1; |
| 126 | |
| 127 | // If we can handle this spread, return it. |
| 128 | if ((llvm::rotr<uint32_t>(V: Imm, R: RotAmt) & ~255U) == 0) |
| 129 | return (32-RotAmt)&31; // HW rotates right, not left. |
| 130 | |
| 131 | // For values like 0xF000000F, we should ignore the low 6 bits, then |
| 132 | // retry the hunt. |
| 133 | if (Imm & 63U) { |
| 134 | unsigned TZ2 = llvm::countr_zero(Val: Imm & ~63U); |
| 135 | unsigned RotAmt2 = TZ2 & ~1; |
| 136 | if ((llvm::rotr<uint32_t>(V: Imm, R: RotAmt2) & ~255U) == 0) |
| 137 | return (32-RotAmt2)&31; // HW rotates right, not left. |
| 138 | } |
| 139 | |
| 140 | // Otherwise, we have no way to cover this span of bits with a single |
| 141 | // shifter_op immediate. Return a chunk of bits that will be useful to |
| 142 | // handle. |
| 143 | return (32-RotAmt)&31; // HW rotates right, not left. |
| 144 | } |
| 145 | |
| 146 | /// getSOImmVal - Given a 32-bit immediate, if it is something that can fit |
| 147 | /// into an shifter_operand immediate operand, return the 12-bit encoding for |
| 148 | /// it. If not, return -1. |
| 149 | inline int getSOImmVal(unsigned Arg) { |
| 150 | // 8-bit (or less) immediates are trivially shifter_operands with a rotate |
| 151 | // of zero. |
| 152 | if ((Arg & ~255U) == 0) return Arg; |
| 153 | |
| 154 | unsigned RotAmt = getSOImmValRotate(Imm: Arg); |
| 155 | |
| 156 | // If this cannot be handled with a single shifter_op, bail out. |
| 157 | if (llvm::rotr<uint32_t>(V: ~255U, R: RotAmt) & Arg) |
| 158 | return -1; |
| 159 | |
| 160 | // Encode this correctly. |
| 161 | return llvm::rotl<uint32_t>(V: Arg, R: RotAmt) | ((RotAmt >> 1) << 8); |
| 162 | } |
| 163 | |
| 164 | /// isSOImmTwoPartVal - Return true if the specified value can be obtained by |
| 165 | /// or'ing together two SOImmVal's. |
| 166 | inline bool isSOImmTwoPartVal(unsigned V) { |
| 167 | // If this can be handled with a single shifter_op, bail out. |
| 168 | V = llvm::rotr<uint32_t>(V: ~255U, R: getSOImmValRotate(Imm: V)) & V; |
| 169 | if (V == 0) |
| 170 | return false; |
| 171 | |
| 172 | // If this can be handled with two shifter_op's, accept. |
| 173 | V = llvm::rotr<uint32_t>(V: ~255U, R: getSOImmValRotate(Imm: V)) & V; |
| 174 | return V == 0; |
| 175 | } |
| 176 | |
| 177 | /// getSOImmTwoPartFirst - If V is a value that satisfies isSOImmTwoPartVal, |
| 178 | /// return the first chunk of it. |
| 179 | inline unsigned getSOImmTwoPartFirst(unsigned V) { |
| 180 | return llvm::rotr<uint32_t>(V: 255U, R: getSOImmValRotate(Imm: V)) & V; |
| 181 | } |
| 182 | |
| 183 | /// getSOImmTwoPartSecond - If V is a value that satisfies isSOImmTwoPartVal, |
| 184 | /// return the second chunk of it. |
| 185 | inline unsigned getSOImmTwoPartSecond(unsigned V) { |
| 186 | // Mask out the first hunk. |
| 187 | V = llvm::rotr<uint32_t>(V: ~255U, R: getSOImmValRotate(Imm: V)) & V; |
| 188 | |
| 189 | // Take what's left. |
| 190 | assert(V == (llvm::rotr<uint32_t>(255U, getSOImmValRotate(V)) & V)); |
| 191 | return V; |
| 192 | } |
| 193 | |
| 194 | /// isSOImmTwoPartValNeg - Return true if the specified value can be obtained |
| 195 | /// by two SOImmVal, that -V = First + Second. |
| 196 | /// "R+V" can be optimized to (sub (sub R, First), Second). |
| 197 | /// "R=V" can be optimized to (sub (mvn R, ~(-First)), Second). |
| 198 | inline bool isSOImmTwoPartValNeg(unsigned V) { |
| 199 | unsigned First; |
| 200 | if (!isSOImmTwoPartVal(V: -V)) |
| 201 | return false; |
| 202 | // Return false if ~(-First) is not a SoImmval. |
| 203 | First = getSOImmTwoPartFirst(V: -V); |
| 204 | First = ~(-First); |
| 205 | return !(llvm::rotr<uint32_t>(V: ~255U, R: getSOImmValRotate(Imm: First)) & First); |
| 206 | } |
| 207 | |
| 208 | /// getThumbImmValShift - Try to handle Imm with a 8-bit immediate followed |
| 209 | /// by a left shift. Returns the shift amount to use. |
| 210 | inline unsigned getThumbImmValShift(unsigned Imm) { |
| 211 | // 8-bit (or less) immediates are trivially immediate operand with a shift |
| 212 | // of zero. |
| 213 | if ((Imm & ~255U) == 0) return 0; |
| 214 | |
| 215 | // Use CTZ to compute the shift amount. |
| 216 | return llvm::countr_zero(Val: Imm); |
| 217 | } |
| 218 | |
| 219 | /// isThumbImmShiftedVal - Return true if the specified value can be obtained |
| 220 | /// by left shifting a 8-bit immediate. |
| 221 | inline bool isThumbImmShiftedVal(unsigned V) { |
| 222 | // If this can be handled with |
| 223 | V = (~255U << getThumbImmValShift(Imm: V)) & V; |
| 224 | return V == 0; |
| 225 | } |
| 226 | |
| 227 | /// getThumbImm16ValShift - Try to handle Imm with a 16-bit immediate followed |
| 228 | /// by a left shift. Returns the shift amount to use. |
| 229 | inline unsigned getThumbImm16ValShift(unsigned Imm) { |
| 230 | // 16-bit (or less) immediates are trivially immediate operand with a shift |
| 231 | // of zero. |
| 232 | if ((Imm & ~65535U) == 0) return 0; |
| 233 | |
| 234 | // Use CTZ to compute the shift amount. |
| 235 | return llvm::countr_zero(Val: Imm); |
| 236 | } |
| 237 | |
| 238 | /// isThumbImm16ShiftedVal - Return true if the specified value can be |
| 239 | /// obtained by left shifting a 16-bit immediate. |
| 240 | inline bool isThumbImm16ShiftedVal(unsigned V) { |
| 241 | // If this can be handled with |
| 242 | V = (~65535U << getThumbImm16ValShift(Imm: V)) & V; |
| 243 | return V == 0; |
| 244 | } |
| 245 | |
| 246 | /// getThumbImmNonShiftedVal - If V is a value that satisfies |
| 247 | /// isThumbImmShiftedVal, return the non-shiftd value. |
| 248 | inline unsigned getThumbImmNonShiftedVal(unsigned V) { |
| 249 | return V >> getThumbImmValShift(Imm: V); |
| 250 | } |
| 251 | |
| 252 | |
| 253 | /// getT2SOImmValSplat - Return the 12-bit encoded representation |
| 254 | /// if the specified value can be obtained by splatting the low 8 bits |
| 255 | /// into every other byte or every byte of a 32-bit value. i.e., |
| 256 | /// 00000000 00000000 00000000 abcdefgh control = 0 |
| 257 | /// 00000000 abcdefgh 00000000 abcdefgh control = 1 |
| 258 | /// abcdefgh 00000000 abcdefgh 00000000 control = 2 |
| 259 | /// abcdefgh abcdefgh abcdefgh abcdefgh control = 3 |
| 260 | /// Return -1 if none of the above apply. |
| 261 | /// See ARM Reference Manual A6.3.2. |
| 262 | inline int getT2SOImmValSplatVal(unsigned V) { |
| 263 | unsigned u, Vs, Imm; |
| 264 | // control = 0 |
| 265 | if ((V & 0xffffff00) == 0) |
| 266 | return V; |
| 267 | |
| 268 | // If the value is zeroes in the first byte, just shift those off |
| 269 | Vs = ((V & 0xff) == 0) ? V >> 8 : V; |
| 270 | // Any passing value only has 8 bits of payload, splatted across the word |
| 271 | Imm = Vs & 0xff; |
| 272 | // Likewise, any passing values have the payload splatted into the 3rd byte |
| 273 | u = Imm | (Imm << 16); |
| 274 | |
| 275 | // control = 1 or 2 |
| 276 | if (Vs == u) |
| 277 | return (((Vs == V) ? 1 : 2) << 8) | Imm; |
| 278 | |
| 279 | // control = 3 |
| 280 | if (Vs == (u | (u << 8))) |
| 281 | return (3 << 8) | Imm; |
| 282 | |
| 283 | return -1; |
| 284 | } |
| 285 | |
| 286 | /// getT2SOImmValRotateVal - Return the 12-bit encoded representation if the |
| 287 | /// specified value is a rotated 8-bit value. Return -1 if no rotation |
| 288 | /// encoding is possible. |
| 289 | /// See ARM Reference Manual A6.3.2. |
| 290 | inline int getT2SOImmValRotateVal(unsigned V) { |
| 291 | unsigned RotAmt = llvm::countl_zero(Val: V); |
| 292 | if (RotAmt >= 24) |
| 293 | return -1; |
| 294 | |
| 295 | // If 'Arg' can be handled with a single shifter_op return the value. |
| 296 | if ((llvm::rotr<uint32_t>(V: 0xff000000U, R: RotAmt) & V) == V) |
| 297 | return (llvm::rotr<uint32_t>(V, R: 24 - RotAmt) & 0x7f) | |
| 298 | ((RotAmt + 8) << 7); |
| 299 | |
| 300 | return -1; |
| 301 | } |
| 302 | |
| 303 | /// getT2SOImmVal - Given a 32-bit immediate, if it is something that can fit |
| 304 | /// into a Thumb-2 shifter_operand immediate operand, return the 12-bit |
| 305 | /// encoding for it. If not, return -1. |
| 306 | /// See ARM Reference Manual A6.3.2. |
| 307 | inline int getT2SOImmVal(unsigned Arg) { |
| 308 | // If 'Arg' is an 8-bit splat, then get the encoded value. |
| 309 | int Splat = getT2SOImmValSplatVal(V: Arg); |
| 310 | if (Splat != -1) |
| 311 | return Splat; |
| 312 | |
| 313 | // If 'Arg' can be handled with a single shifter_op return the value. |
| 314 | int Rot = getT2SOImmValRotateVal(V: Arg); |
| 315 | if (Rot != -1) |
| 316 | return Rot; |
| 317 | |
| 318 | return -1; |
| 319 | } |
| 320 | |
| 321 | inline unsigned getT2SOImmValRotate(unsigned V) { |
| 322 | if ((V & ~255U) == 0) return 0; |
| 323 | // Use CTZ to compute the rotate amount. |
| 324 | unsigned RotAmt = llvm::countr_zero(Val: V); |
| 325 | return (32 - RotAmt) & 31; |
| 326 | } |
| 327 | |
| 328 | inline bool isT2SOImmTwoPartVal(unsigned Imm) { |
| 329 | unsigned V = Imm; |
| 330 | // Passing values can be any combination of splat values and shifter |
| 331 | // values. If this can be handled with a single shifter or splat, bail |
| 332 | // out. Those should be handled directly, not with a two-part val. |
| 333 | if (getT2SOImmValSplatVal(V) != -1) |
| 334 | return false; |
| 335 | V = llvm::rotr<uint32_t>(V: ~255U, R: getT2SOImmValRotate(V)) & V; |
| 336 | if (V == 0) |
| 337 | return false; |
| 338 | |
| 339 | // If this can be handled as an immediate, accept. |
| 340 | if (getT2SOImmVal(Arg: V) != -1) return true; |
| 341 | |
| 342 | // Likewise, try masking out a splat value first. |
| 343 | V = Imm; |
| 344 | if (getT2SOImmValSplatVal(V: V & 0xff00ff00U) != -1) |
| 345 | V &= ~0xff00ff00U; |
| 346 | else if (getT2SOImmValSplatVal(V: V & 0x00ff00ffU) != -1) |
| 347 | V &= ~0x00ff00ffU; |
| 348 | // If what's left can be handled as an immediate, accept. |
| 349 | if (getT2SOImmVal(Arg: V) != -1) return true; |
| 350 | |
| 351 | // Otherwise, do not accept. |
| 352 | return false; |
| 353 | } |
| 354 | |
| 355 | inline unsigned getT2SOImmTwoPartFirst(unsigned Imm) { |
| 356 | assert (isT2SOImmTwoPartVal(Imm) && |
| 357 | "Immedate cannot be encoded as two part immediate!" ); |
| 358 | // Try a shifter operand as one part |
| 359 | unsigned V = llvm::rotr<uint32_t>(V: ~255, R: getT2SOImmValRotate(V: Imm)) & Imm; |
| 360 | // If the rest is encodable as an immediate, then return it. |
| 361 | if (getT2SOImmVal(Arg: V) != -1) return V; |
| 362 | |
| 363 | // Try masking out a splat value first. |
| 364 | if (getT2SOImmValSplatVal(V: Imm & 0xff00ff00U) != -1) |
| 365 | return Imm & 0xff00ff00U; |
| 366 | |
| 367 | // The other splat is all that's left as an option. |
| 368 | assert (getT2SOImmValSplatVal(Imm & 0x00ff00ffU) != -1); |
| 369 | return Imm & 0x00ff00ffU; |
| 370 | } |
| 371 | |
| 372 | inline unsigned getT2SOImmTwoPartSecond(unsigned Imm) { |
| 373 | // Mask out the first hunk |
| 374 | Imm ^= getT2SOImmTwoPartFirst(Imm); |
| 375 | // Return what's left |
| 376 | assert (getT2SOImmVal(Imm) != -1 && |
| 377 | "Unable to encode second part of T2 two part SO immediate" ); |
| 378 | return Imm; |
| 379 | } |
| 380 | |
| 381 | |
| 382 | //===--------------------------------------------------------------------===// |
| 383 | // Addressing Mode #2 |
| 384 | //===--------------------------------------------------------------------===// |
| 385 | // |
| 386 | // This is used for most simple load/store instructions. |
| 387 | // |
| 388 | // addrmode2 := reg +/- reg shop imm |
| 389 | // addrmode2 := reg +/- imm12 |
| 390 | // |
| 391 | // The first operand is always a Reg. The second operand is a reg if in |
| 392 | // reg/reg form, otherwise it's reg#0. The third field encodes the operation |
| 393 | // in bit 12, the immediate in bits 0-11, and the shift op in 13-15. The |
| 394 | // fourth operand 16-17 encodes the index mode. |
| 395 | // |
| 396 | // If this addressing mode is a frame index (before prolog/epilog insertion |
| 397 | // and code rewriting), this operand will have the form: FI#, reg0, <offs> |
| 398 | // with no shift amount for the frame offset. |
| 399 | // |
| 400 | inline unsigned getAM2Opc(AddrOpc Opc, unsigned Imm12, ShiftOpc SO, |
| 401 | unsigned IdxMode = 0) { |
| 402 | assert(Imm12 < (1 << 12) && "Imm too large!" ); |
| 403 | bool isSub = Opc == sub; |
| 404 | return Imm12 | ((int)isSub << 12) | (SO << 13) | (IdxMode << 16) ; |
| 405 | } |
| 406 | inline unsigned getAM2Offset(unsigned AM2Opc) { |
| 407 | return AM2Opc & ((1 << 12)-1); |
| 408 | } |
| 409 | inline AddrOpc getAM2Op(unsigned AM2Opc) { |
| 410 | return ((AM2Opc >> 12) & 1) ? sub : add; |
| 411 | } |
| 412 | inline ShiftOpc getAM2ShiftOpc(unsigned AM2Opc) { |
| 413 | return (ShiftOpc)((AM2Opc >> 13) & 7); |
| 414 | } |
| 415 | inline unsigned getAM2IdxMode(unsigned AM2Opc) { return (AM2Opc >> 16); } |
| 416 | |
| 417 | //===--------------------------------------------------------------------===// |
| 418 | // Addressing Mode #3 |
| 419 | //===--------------------------------------------------------------------===// |
| 420 | // |
| 421 | // This is used for sign-extending loads, and load/store-pair instructions. |
| 422 | // |
| 423 | // addrmode3 := reg +/- reg |
| 424 | // addrmode3 := reg +/- imm8 |
| 425 | // |
| 426 | // The first operand is always a Reg. The second operand is a reg if in |
| 427 | // reg/reg form, otherwise it's reg#0. The third field encodes the operation |
| 428 | // in bit 8, the immediate in bits 0-7. The fourth operand 9-10 encodes the |
| 429 | // index mode. |
| 430 | |
| 431 | /// getAM3Opc - This function encodes the addrmode3 opc field. |
| 432 | inline unsigned getAM3Opc(AddrOpc Opc, unsigned char Offset, |
| 433 | unsigned IdxMode = 0) { |
| 434 | bool isSub = Opc == sub; |
| 435 | return ((int)isSub << 8) | Offset | (IdxMode << 9); |
| 436 | } |
| 437 | inline unsigned char getAM3Offset(unsigned AM3Opc) { return AM3Opc & 0xFF; } |
| 438 | inline AddrOpc getAM3Op(unsigned AM3Opc) { |
| 439 | return ((AM3Opc >> 8) & 1) ? sub : add; |
| 440 | } |
| 441 | inline unsigned getAM3IdxMode(unsigned AM3Opc) { return (AM3Opc >> 9); } |
| 442 | |
| 443 | //===--------------------------------------------------------------------===// |
| 444 | // Addressing Mode #4 |
| 445 | //===--------------------------------------------------------------------===// |
| 446 | // |
| 447 | // This is used for load / store multiple instructions. |
| 448 | // |
| 449 | // addrmode4 := reg, <mode> |
| 450 | // |
| 451 | // The four modes are: |
| 452 | // IA - Increment after |
| 453 | // IB - Increment before |
| 454 | // DA - Decrement after |
| 455 | // DB - Decrement before |
| 456 | // For VFP instructions, only the IA and DB modes are valid. |
| 457 | |
| 458 | inline AMSubMode getAM4SubMode(unsigned Mode) { |
| 459 | return (AMSubMode)(Mode & 0x7); |
| 460 | } |
| 461 | |
| 462 | inline unsigned getAM4ModeImm(AMSubMode SubMode) { return (int)SubMode; } |
| 463 | |
| 464 | //===--------------------------------------------------------------------===// |
| 465 | // Addressing Mode #5 |
| 466 | //===--------------------------------------------------------------------===// |
| 467 | // |
| 468 | // This is used for coprocessor instructions, such as FP load/stores. |
| 469 | // |
| 470 | // addrmode5 := reg +/- imm8*4 |
| 471 | // |
| 472 | // The first operand is always a Reg. The second operand encodes the |
| 473 | // operation (add or subtract) in bit 8 and the immediate in bits 0-7. |
| 474 | |
| 475 | /// getAM5Opc - This function encodes the addrmode5 opc field. |
| 476 | inline unsigned getAM5Opc(AddrOpc Opc, unsigned char Offset) { |
| 477 | bool isSub = Opc == sub; |
| 478 | return ((int)isSub << 8) | Offset; |
| 479 | } |
| 480 | inline unsigned char getAM5Offset(unsigned AM5Opc) { return AM5Opc & 0xFF; } |
| 481 | inline AddrOpc getAM5Op(unsigned AM5Opc) { |
| 482 | return ((AM5Opc >> 8) & 1) ? sub : add; |
| 483 | } |
| 484 | |
| 485 | //===--------------------------------------------------------------------===// |
| 486 | // Addressing Mode #5 FP16 |
| 487 | //===--------------------------------------------------------------------===// |
| 488 | // |
| 489 | // This is used for coprocessor instructions, such as 16-bit FP load/stores. |
| 490 | // |
| 491 | // addrmode5fp16 := reg +/- imm8*2 |
| 492 | // |
| 493 | // The first operand is always a Reg. The second operand encodes the |
| 494 | // operation (add or subtract) in bit 8 and the immediate in bits 0-7. |
| 495 | |
| 496 | /// getAM5FP16Opc - This function encodes the addrmode5fp16 opc field. |
| 497 | inline unsigned getAM5FP16Opc(AddrOpc Opc, unsigned char Offset) { |
| 498 | bool isSub = Opc == sub; |
| 499 | return ((int)isSub << 8) | Offset; |
| 500 | } |
| 501 | inline unsigned char getAM5FP16Offset(unsigned AM5Opc) { |
| 502 | return AM5Opc & 0xFF; |
| 503 | } |
| 504 | inline AddrOpc getAM5FP16Op(unsigned AM5Opc) { |
| 505 | return ((AM5Opc >> 8) & 1) ? sub : add; |
| 506 | } |
| 507 | |
| 508 | //===--------------------------------------------------------------------===// |
| 509 | // Addressing Mode #6 |
| 510 | //===--------------------------------------------------------------------===// |
| 511 | // |
| 512 | // This is used for NEON load / store instructions. |
| 513 | // |
| 514 | // addrmode6 := reg with optional alignment |
| 515 | // |
| 516 | // This is stored in two operands [regaddr, align]. The first is the |
| 517 | // address register. The second operand is the value of the alignment |
| 518 | // specifier in bytes or zero if no explicit alignment. |
| 519 | // Valid alignments depend on the specific instruction. |
| 520 | |
| 521 | //===--------------------------------------------------------------------===// |
| 522 | // NEON/MVE Modified Immediates |
| 523 | //===--------------------------------------------------------------------===// |
| 524 | // |
| 525 | // Several NEON and MVE instructions (e.g., VMOV) take a "modified immediate" |
| 526 | // vector operand, where a small immediate encoded in the instruction |
| 527 | // specifies a full NEON vector value. These modified immediates are |
| 528 | // represented here as encoded integers. The low 8 bits hold the immediate |
| 529 | // value; bit 12 holds the "Op" field of the instruction, and bits 11-8 hold |
| 530 | // the "Cmode" field of the instruction. The interfaces below treat the |
| 531 | // Op and Cmode values as a single 5-bit value. |
| 532 | |
| 533 | inline unsigned createVMOVModImm(unsigned OpCmode, unsigned Val) { |
| 534 | return (OpCmode << 8) | Val; |
| 535 | } |
| 536 | inline unsigned getVMOVModImmOpCmode(unsigned ModImm) { |
| 537 | return (ModImm >> 8) & 0x1f; |
| 538 | } |
| 539 | inline unsigned getVMOVModImmVal(unsigned ModImm) { return ModImm & 0xff; } |
| 540 | |
| 541 | /// decodeVMOVModImm - Decode a NEON/MVE modified immediate value into the |
| 542 | /// element value and the element size in bits. (If the element size is |
| 543 | /// smaller than the vector, it is splatted into all the elements.) |
| 544 | inline uint64_t decodeVMOVModImm(unsigned ModImm, unsigned &EltBits) { |
| 545 | unsigned OpCmode = getVMOVModImmOpCmode(ModImm); |
| 546 | unsigned Imm8 = getVMOVModImmVal(ModImm); |
| 547 | uint64_t Val = 0; |
| 548 | |
| 549 | if (OpCmode == 0xe) { |
| 550 | // 8-bit vector elements |
| 551 | Val = Imm8; |
| 552 | EltBits = 8; |
| 553 | } else if ((OpCmode & 0xc) == 0x8) { |
| 554 | // 16-bit vector elements |
| 555 | unsigned ByteNum = (OpCmode & 0x6) >> 1; |
| 556 | Val = Imm8 << (8 * ByteNum); |
| 557 | EltBits = 16; |
| 558 | } else if ((OpCmode & 0x8) == 0) { |
| 559 | // 32-bit vector elements, zero with one byte set |
| 560 | unsigned ByteNum = (OpCmode & 0x6) >> 1; |
| 561 | Val = Imm8 << (8 * ByteNum); |
| 562 | EltBits = 32; |
| 563 | } else if ((OpCmode & 0xe) == 0xc) { |
| 564 | // 32-bit vector elements, one byte with low bits set |
| 565 | unsigned ByteNum = 1 + (OpCmode & 0x1); |
| 566 | Val = (Imm8 << (8 * ByteNum)) | (0xffff >> (8 * (2 - ByteNum))); |
| 567 | EltBits = 32; |
| 568 | } else if (OpCmode == 0x1e) { |
| 569 | // 64-bit vector elements |
| 570 | for (unsigned ByteNum = 0; ByteNum < 8; ++ByteNum) { |
| 571 | if ((ModImm >> ByteNum) & 1) |
| 572 | Val |= (uint64_t)0xff << (8 * ByteNum); |
| 573 | } |
| 574 | EltBits = 64; |
| 575 | } else { |
| 576 | llvm_unreachable("Unsupported VMOV immediate" ); |
| 577 | } |
| 578 | return Val; |
| 579 | } |
| 580 | |
| 581 | // Generic validation for single-byte immediate (0X00, 00X0, etc). |
| 582 | inline bool isNEONBytesplat(unsigned Value, unsigned Size) { |
| 583 | assert(Size >= 1 && Size <= 4 && "Invalid size" ); |
| 584 | unsigned count = 0; |
| 585 | for (unsigned i = 0; i < Size; ++i) { |
| 586 | if (Value & 0xff) count++; |
| 587 | Value >>= 8; |
| 588 | } |
| 589 | return count == 1; |
| 590 | } |
| 591 | |
| 592 | /// Checks if Value is a correct immediate for instructions like VBIC/VORR. |
| 593 | inline bool isNEONi16splat(unsigned Value) { |
| 594 | if (Value > 0xffff) |
| 595 | return false; |
| 596 | // i16 value with set bits only in one byte X0 or 0X. |
| 597 | return Value == 0 || isNEONBytesplat(Value, Size: 2); |
| 598 | } |
| 599 | |
| 600 | // Encode NEON 16 bits Splat immediate for instructions like VBIC/VORR |
| 601 | inline unsigned encodeNEONi16splat(unsigned Value) { |
| 602 | assert(isNEONi16splat(Value) && "Invalid NEON splat value" ); |
| 603 | if (Value >= 0x100) |
| 604 | Value = (Value >> 8) | 0xa00; |
| 605 | else |
| 606 | Value |= 0x800; |
| 607 | return Value; |
| 608 | } |
| 609 | |
| 610 | /// Checks if Value is a correct immediate for instructions like VBIC/VORR. |
| 611 | inline bool isNEONi32splat(unsigned Value) { |
| 612 | // i32 value with set bits only in one byte X000, 0X00, 00X0, or 000X. |
| 613 | return Value == 0 || isNEONBytesplat(Value, Size: 4); |
| 614 | } |
| 615 | |
| 616 | /// Encode NEON 32 bits Splat immediate for instructions like VBIC/VORR. |
| 617 | inline unsigned encodeNEONi32splat(unsigned Value) { |
| 618 | assert(isNEONi32splat(Value) && "Invalid NEON splat value" ); |
| 619 | if (Value >= 0x100 && Value <= 0xff00) |
| 620 | Value = (Value >> 8) | 0x200; |
| 621 | else if (Value > 0xffff && Value <= 0xff0000) |
| 622 | Value = (Value >> 16) | 0x400; |
| 623 | else if (Value > 0xffffff) |
| 624 | Value = (Value >> 24) | 0x600; |
| 625 | return Value; |
| 626 | } |
| 627 | |
| 628 | //===--------------------------------------------------------------------===// |
| 629 | // Floating-point Immediates |
| 630 | // |
| 631 | inline float getFPImmFloat(unsigned Imm) { |
| 632 | // We expect an 8-bit binary encoding of a floating-point number here. |
| 633 | |
| 634 | uint8_t Sign = (Imm >> 7) & 0x1; |
| 635 | uint8_t Exp = (Imm >> 4) & 0x7; |
| 636 | uint8_t Mantissa = Imm & 0xf; |
| 637 | |
| 638 | // 8-bit FP IEEE Float Encoding |
| 639 | // abcd efgh aBbbbbbc defgh000 00000000 00000000 |
| 640 | // |
| 641 | // where B = NOT(b); |
| 642 | uint32_t I = 0; |
| 643 | I |= Sign << 31; |
| 644 | I |= ((Exp & 0x4) != 0 ? 0 : 1) << 30; |
| 645 | I |= ((Exp & 0x4) != 0 ? 0x1f : 0) << 25; |
| 646 | I |= (Exp & 0x3) << 23; |
| 647 | I |= Mantissa << 19; |
| 648 | return bit_cast<float>(from: I); |
| 649 | } |
| 650 | |
| 651 | /// getFP16Imm - Return an 8-bit floating-point version of the 16-bit |
| 652 | /// floating-point value. If the value cannot be represented as an 8-bit |
| 653 | /// floating-point value, then return -1. |
| 654 | inline int getFP16Imm(const APInt &Imm) { |
| 655 | uint32_t Sign = Imm.lshr(shiftAmt: 15).getZExtValue() & 1; |
| 656 | int32_t Exp = (Imm.lshr(shiftAmt: 10).getSExtValue() & 0x1f) - 15; // -14 to 15 |
| 657 | int64_t Mantissa = Imm.getZExtValue() & 0x3ff; // 10 bits |
| 658 | |
| 659 | // We can handle 4 bits of mantissa. |
| 660 | // mantissa = (16+UInt(e:f:g:h))/16. |
| 661 | if (Mantissa & 0x3f) |
| 662 | return -1; |
| 663 | Mantissa >>= 6; |
| 664 | |
| 665 | // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3 |
| 666 | if (Exp < -3 || Exp > 4) |
| 667 | return -1; |
| 668 | Exp = ((Exp+3) & 0x7) ^ 4; |
| 669 | |
| 670 | return ((int)Sign << 7) | (Exp << 4) | Mantissa; |
| 671 | } |
| 672 | |
| 673 | inline int getFP16Imm(const APFloat &FPImm) { |
| 674 | return getFP16Imm(Imm: FPImm.bitcastToAPInt()); |
| 675 | } |
| 676 | |
| 677 | /// If this is a FP16Imm encoded as a fp32 value, return the 8-bit encoding |
| 678 | /// for it. Otherwise return -1 like getFP16Imm. |
| 679 | inline int getFP32FP16Imm(const APInt &Imm) { |
| 680 | if (Imm.getActiveBits() > 16) |
| 681 | return -1; |
| 682 | return ARM_AM::getFP16Imm(Imm: Imm.trunc(width: 16)); |
| 683 | } |
| 684 | |
| 685 | inline int getFP32FP16Imm(const APFloat &FPImm) { |
| 686 | return getFP32FP16Imm(Imm: FPImm.bitcastToAPInt()); |
| 687 | } |
| 688 | |
| 689 | /// getFP32Imm - Return an 8-bit floating-point version of the 32-bit |
| 690 | /// floating-point value. If the value cannot be represented as an 8-bit |
| 691 | /// floating-point value, then return -1. |
| 692 | inline int getFP32Imm(const APInt &Imm) { |
| 693 | uint32_t Sign = Imm.lshr(shiftAmt: 31).getZExtValue() & 1; |
| 694 | int32_t Exp = (Imm.lshr(shiftAmt: 23).getSExtValue() & 0xff) - 127; // -126 to 127 |
| 695 | int64_t Mantissa = Imm.getZExtValue() & 0x7fffff; // 23 bits |
| 696 | |
| 697 | // We can handle 4 bits of mantissa. |
| 698 | // mantissa = (16+UInt(e:f:g:h))/16. |
| 699 | if (Mantissa & 0x7ffff) |
| 700 | return -1; |
| 701 | Mantissa >>= 19; |
| 702 | if ((Mantissa & 0xf) != Mantissa) |
| 703 | return -1; |
| 704 | |
| 705 | // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3 |
| 706 | if (Exp < -3 || Exp > 4) |
| 707 | return -1; |
| 708 | Exp = ((Exp+3) & 0x7) ^ 4; |
| 709 | |
| 710 | return ((int)Sign << 7) | (Exp << 4) | Mantissa; |
| 711 | } |
| 712 | |
| 713 | inline int getFP32Imm(const APFloat &FPImm) { |
| 714 | return getFP32Imm(Imm: FPImm.bitcastToAPInt()); |
| 715 | } |
| 716 | |
| 717 | /// getFP64Imm - Return an 8-bit floating-point version of the 64-bit |
| 718 | /// floating-point value. If the value cannot be represented as an 8-bit |
| 719 | /// floating-point value, then return -1. |
| 720 | inline int getFP64Imm(const APInt &Imm) { |
| 721 | uint64_t Sign = Imm.lshr(shiftAmt: 63).getZExtValue() & 1; |
| 722 | int64_t Exp = (Imm.lshr(shiftAmt: 52).getSExtValue() & 0x7ff) - 1023; // -1022 to 1023 |
| 723 | uint64_t Mantissa = Imm.getZExtValue() & 0xfffffffffffffULL; |
| 724 | |
| 725 | // We can handle 4 bits of mantissa. |
| 726 | // mantissa = (16+UInt(e:f:g:h))/16. |
| 727 | if (Mantissa & 0xffffffffffffULL) |
| 728 | return -1; |
| 729 | Mantissa >>= 48; |
| 730 | if ((Mantissa & 0xf) != Mantissa) |
| 731 | return -1; |
| 732 | |
| 733 | // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3 |
| 734 | if (Exp < -3 || Exp > 4) |
| 735 | return -1; |
| 736 | Exp = ((Exp+3) & 0x7) ^ 4; |
| 737 | |
| 738 | return ((int)Sign << 7) | (Exp << 4) | Mantissa; |
| 739 | } |
| 740 | |
| 741 | inline int getFP64Imm(const APFloat &FPImm) { |
| 742 | return getFP64Imm(Imm: FPImm.bitcastToAPInt()); |
| 743 | } |
| 744 | |
| 745 | } // end namespace ARM_AM |
| 746 | } // end namespace llvm |
| 747 | |
| 748 | #endif |
| 749 | |