1//===-- IR/VPIntrinsics.def - Describes llvm.vp.* Intrinsics -*- 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 descriptions of the various Vector Predication intrinsics. 10// This is used as a central place for enumerating the different instructions 11// and should eventually be the place to put comments about the instructions. 12// 13//===----------------------------------------------------------------------===// 14 15// NOTE: NO INCLUDE GUARD DESIRED! 16 17// Provide definitions of macros so that users of this file do not have to 18// define everything to use it... 19// 20// Register a VP intrinsic and begin its property scope. 21// All VP intrinsic scopes are top level, ie it is illegal to place a 22// BEGIN_REGISTER_VP_INTRINSIC within a VP intrinsic scope. 23// \p VPID The VP intrinsic id. 24// \p MASKPOS The mask operand position. 25// \p EVLPOS The explicit vector length operand position. 26#ifndef BEGIN_REGISTER_VP_INTRINSIC 27#define BEGIN_REGISTER_VP_INTRINSIC(VPID, MASKPOS, EVLPOS) 28#endif 29 30// End the property scope of a VP intrinsic. 31#ifndef END_REGISTER_VP_INTRINSIC 32#define END_REGISTER_VP_INTRINSIC(VPID) 33#endif 34 35// Register a new VP SDNode and begin its property scope. 36// When the SDNode scope is nested within a VP intrinsic scope, it is 37// implicitly registered as the canonical SDNode for this VP intrinsic. There 38// is one VP intrinsic that maps directly to one SDNode that goes by the 39// same name. Since the operands are also the same, we open the property 40// scopes for both the VPIntrinsic and the SDNode at once. 41// \p VPSD The SelectionDAG Node id (eg VP_ADD). 42// \p LEGALPOS The operand position of the SDNode that is used for legalizing. 43// If LEGALPOS < 0, then the return type given by 44// TheNode->getValueType(-1-LEGALPOS) is used. 45// \p TDNAME The name of the TableGen definition of this SDNode. 46// \p MASKPOS The mask operand position. 47// \p EVLPOS The explicit vector length operand position. 48#ifndef BEGIN_REGISTER_VP_SDNODE 49#define BEGIN_REGISTER_VP_SDNODE(VPSD, LEGALPOS, TDNAME, MASKPOS, EVLPOS) 50#endif 51 52// End the property scope of a new VP SDNode. 53#ifndef END_REGISTER_VP_SDNODE 54#define END_REGISTER_VP_SDNODE(VPSD) 55#endif 56 57// Helper macro to set up the mapping from VP intrinsic to ISD opcode. 58// Note: More than one VP intrinsic may map to one ISD opcode. 59#ifndef HELPER_MAP_VPID_TO_VPSD 60#define HELPER_MAP_VPID_TO_VPSD(VPID, VPSD) 61#endif 62 63// Helper macros for the common "1:1 - Intrinsic : SDNode" case. 64// 65// There is one VP intrinsic that maps directly to one SDNode that goes by the 66// same name. Since the operands are also the same, we open the property 67// scopes for both the VPIntrinsic and the SDNode at once. 68// 69// \p VPID The canonical name (eg `vp_add`, which at the same time is the 70// name of the intrinsic and the TableGen def of the SDNode). 71// \p MASKPOS The mask operand position. 72// \p EVLPOS The explicit vector length operand position. 73// \p VPSD The SelectionDAG Node id (eg VP_ADD). 74// \p LEGALPOS The operand position of the SDNode that is used for legalizing 75// this SDNode. This can be `-1`, in which case the return type of 76// the SDNode is used. 77#define BEGIN_REGISTER_VP(VPID, MASKPOS, EVLPOS, VPSD, LEGALPOS) \ 78 BEGIN_REGISTER_VP_INTRINSIC(VPID, MASKPOS, EVLPOS) \ 79 BEGIN_REGISTER_VP_SDNODE(VPSD, LEGALPOS, VPID, MASKPOS, EVLPOS) \ 80 HELPER_MAP_VPID_TO_VPSD(VPID, VPSD) 81 82#define END_REGISTER_VP(VPID, VPSD) \ 83 END_REGISTER_VP_INTRINSIC(VPID) \ 84 END_REGISTER_VP_SDNODE(VPSD) 85 86// The following macros attach properties to the scope they are placed in. This 87// assigns the property to the VP Intrinsic and/or SDNode that belongs to the 88// scope. 89// 90// Property Macros { 91 92// The intrinsic and/or SDNode has the same function as this LLVM IR Opcode. 93// \p OPC The opcode of the instruction with the same function. 94#ifndef VP_PROPERTY_FUNCTIONAL_OPC 95#define VP_PROPERTY_FUNCTIONAL_OPC(OPC) 96#endif 97 98// Whether the intrinsic may have a rounding mode or exception behavior operand 99// bundle. 100// \p HASROUND '1' if the intrinsic can have a rounding mode operand bundle, 101// '0' otherwise. 102// \p HASEXCEPT '1' if the intrinsic can have an exception behavior operand 103// bundle, '0' otherwise. 104// \p INTRINID The constrained fp intrinsic this VP intrinsic corresponds to. 105#ifndef VP_PROPERTY_CONSTRAINEDFP 106#define VP_PROPERTY_CONSTRAINEDFP(HASROUND, HASEXCEPT, INTRINID) 107#endif 108 109// The intrinsic and/or SDNode has the same function as this ISD Opcode. 110// \p SDOPC The opcode of the instruction with the same function. 111#ifndef VP_PROPERTY_FUNCTIONAL_SDOPC 112#define VP_PROPERTY_FUNCTIONAL_SDOPC(SDOPC) 113#endif 114 115// Map this VP intrinsic to its canonical functional intrinsic. 116// \p INTRIN The non-VP intrinsics with the same function. 117#ifndef VP_PROPERTY_FUNCTIONAL_INTRINSIC 118#define VP_PROPERTY_FUNCTIONAL_INTRINSIC(INTRIN) 119#endif 120 121// This VP Intrinsic has no functionally-equivalent non-VP opcode or intrinsic. 122#ifndef VP_PROPERTY_NO_FUNCTIONAL 123#define VP_PROPERTY_NO_FUNCTIONAL 124#endif 125 126// This VP Intrinsic is a memory operation 127// The pointer arg is at POINTERPOS and the data arg is at DATAPOS. 128#ifndef VP_PROPERTY_MEMOP 129#define VP_PROPERTY_MEMOP(POINTERPOS, DATAPOS) 130#endif 131 132// Map this VP reduction intrinsic to its reduction operand positions. 133#ifndef VP_PROPERTY_REDUCTION 134#define VP_PROPERTY_REDUCTION(STARTPOS, VECTORPOS) 135#endif 136 137// A property to infer VP binary-op SDNode opcodes automatically. 138#ifndef VP_PROPERTY_BINARYOP 139#define VP_PROPERTY_BINARYOP 140#endif 141 142// A property to infer VP type casts automatically. 143#ifndef VP_PROPERTY_CASTOP 144#define VP_PROPERTY_CASTOP 145#endif 146 147// This VP Intrinsic is a comparison operation 148// The condition code arg is at CCPOS and accepts floating-point condition 149// codes if ISFP is set, else it accepts integer condition codes. 150#ifndef VP_PROPERTY_CMP 151#define VP_PROPERTY_CMP(CCPOS, ISFP) 152#endif 153 154/// } Property Macros 155 156///// Integer Arithmetic { 157 158// Specialized helper macro for integer binary operators (%x, %y, %mask, %evl). 159#ifdef HELPER_REGISTER_BINARY_INT_VP 160#error \ 161 "The internal helper macro HELPER_REGISTER_BINARY_INT_VP is already defined!" 162#endif 163#define HELPER_REGISTER_BINARY_INT_VP(VPID, VPSD, IROPC, SDOPC) \ 164 BEGIN_REGISTER_VP(VPID, 2, 3, VPSD, -1) \ 165 VP_PROPERTY_FUNCTIONAL_OPC(IROPC) \ 166 VP_PROPERTY_FUNCTIONAL_SDOPC(SDOPC) \ 167 VP_PROPERTY_BINARYOP \ 168 END_REGISTER_VP(VPID, VPSD) 169 170// llvm.vp.add(x,y,mask,vlen) 171HELPER_REGISTER_BINARY_INT_VP(vp_add, VP_ADD, Add, ADD) 172 173// llvm.vp.and(x,y,mask,vlen) 174HELPER_REGISTER_BINARY_INT_VP(vp_and, VP_AND, And, AND) 175 176// llvm.vp.ashr(x,y,mask,vlen) 177HELPER_REGISTER_BINARY_INT_VP(vp_ashr, VP_ASHR, AShr, SRA) 178 179// llvm.vp.lshr(x,y,mask,vlen) 180HELPER_REGISTER_BINARY_INT_VP(vp_lshr, VP_LSHR, LShr, SRL) 181 182// llvm.vp.mul(x,y,mask,vlen) 183HELPER_REGISTER_BINARY_INT_VP(vp_mul, VP_MUL, Mul, MUL) 184 185// llvm.vp.or(x,y,mask,vlen) 186HELPER_REGISTER_BINARY_INT_VP(vp_or, VP_OR, Or, OR) 187 188// llvm.vp.sdiv(x,y,mask,vlen) 189HELPER_REGISTER_BINARY_INT_VP(vp_sdiv, VP_SDIV, SDiv, SDIV) 190 191// llvm.vp.shl(x,y,mask,vlen) 192HELPER_REGISTER_BINARY_INT_VP(vp_shl, VP_SHL, Shl, SHL) 193 194// llvm.vp.srem(x,y,mask,vlen) 195HELPER_REGISTER_BINARY_INT_VP(vp_srem, VP_SREM, SRem, SREM) 196 197// llvm.vp.sub(x,y,mask,vlen) 198HELPER_REGISTER_BINARY_INT_VP(vp_sub, VP_SUB, Sub, SUB) 199 200// llvm.vp.udiv(x,y,mask,vlen) 201HELPER_REGISTER_BINARY_INT_VP(vp_udiv, VP_UDIV, UDiv, UDIV) 202 203// llvm.vp.urem(x,y,mask,vlen) 204HELPER_REGISTER_BINARY_INT_VP(vp_urem, VP_UREM, URem, UREM) 205 206// llvm.vp.xor(x,y,mask,vlen) 207HELPER_REGISTER_BINARY_INT_VP(vp_xor, VP_XOR, Xor, XOR) 208 209#undef HELPER_REGISTER_BINARY_INT_VP 210 211// llvm.vp.smin(x,y,mask,vlen) 212BEGIN_REGISTER_VP(vp_smin, 2, 3, VP_SMIN, -1) 213VP_PROPERTY_BINARYOP 214VP_PROPERTY_FUNCTIONAL_SDOPC(SMIN) 215VP_PROPERTY_FUNCTIONAL_INTRINSIC(smin) 216END_REGISTER_VP(vp_smin, VP_SMIN) 217 218// llvm.vp.smax(x,y,mask,vlen) 219BEGIN_REGISTER_VP(vp_smax, 2, 3, VP_SMAX, -1) 220VP_PROPERTY_BINARYOP 221VP_PROPERTY_FUNCTIONAL_SDOPC(SMAX) 222VP_PROPERTY_FUNCTIONAL_INTRINSIC(smax) 223END_REGISTER_VP(vp_smax, VP_SMAX) 224 225// llvm.vp.umin(x,y,mask,vlen) 226BEGIN_REGISTER_VP(vp_umin, 2, 3, VP_UMIN, -1) 227VP_PROPERTY_BINARYOP 228VP_PROPERTY_FUNCTIONAL_SDOPC(UMIN) 229VP_PROPERTY_FUNCTIONAL_INTRINSIC(umin) 230END_REGISTER_VP(vp_umin, VP_UMIN) 231 232// llvm.vp.umax(x,y,mask,vlen) 233BEGIN_REGISTER_VP(vp_umax, 2, 3, VP_UMAX, -1) 234VP_PROPERTY_BINARYOP 235VP_PROPERTY_FUNCTIONAL_SDOPC(UMAX) 236VP_PROPERTY_FUNCTIONAL_INTRINSIC(umax) 237END_REGISTER_VP(vp_umax, VP_UMAX) 238 239// llvm.vp.abs(x,is_int_min_poison,mask,vlen) 240BEGIN_REGISTER_VP_INTRINSIC(vp_abs, 2, 3) 241BEGIN_REGISTER_VP_SDNODE(VP_ABS, -1, vp_abs, 1, 2) 242HELPER_MAP_VPID_TO_VPSD(vp_abs, VP_ABS) 243VP_PROPERTY_FUNCTIONAL_INTRINSIC(abs) 244VP_PROPERTY_FUNCTIONAL_SDOPC(ABS) 245END_REGISTER_VP(vp_abs, VP_ABS) 246 247// llvm.vp.bswap(x,mask,vlen) 248BEGIN_REGISTER_VP(vp_bswap, 1, 2, VP_BSWAP, -1) 249VP_PROPERTY_FUNCTIONAL_INTRINSIC(bswap) 250VP_PROPERTY_FUNCTIONAL_SDOPC(BSWAP) 251END_REGISTER_VP(vp_bswap, VP_BSWAP) 252 253// llvm.vp.bitreverse(x,mask,vlen) 254BEGIN_REGISTER_VP(vp_bitreverse, 1, 2, VP_BITREVERSE, -1) 255VP_PROPERTY_FUNCTIONAL_INTRINSIC(bitreverse) 256VP_PROPERTY_FUNCTIONAL_SDOPC(BITREVERSE) 257END_REGISTER_VP(vp_bitreverse, VP_BITREVERSE) 258 259// llvm.vp.ctpop(x,mask,vlen) 260BEGIN_REGISTER_VP(vp_ctpop, 1, 2, VP_CTPOP, -1) 261VP_PROPERTY_FUNCTIONAL_INTRINSIC(ctpop) 262VP_PROPERTY_FUNCTIONAL_SDOPC(CTPOP) 263END_REGISTER_VP(vp_ctpop, VP_CTPOP) 264 265// llvm.vp.ctlz(x,is_zero_poison,mask,vlen) 266BEGIN_REGISTER_VP_INTRINSIC(vp_ctlz, 2, 3) 267BEGIN_REGISTER_VP_SDNODE(VP_CTLZ, -1, vp_ctlz, 1, 2) 268VP_PROPERTY_FUNCTIONAL_INTRINSIC(ctlz) 269VP_PROPERTY_FUNCTIONAL_SDOPC(CTLZ) 270END_REGISTER_VP_SDNODE(VP_CTLZ) 271BEGIN_REGISTER_VP_SDNODE(VP_CTLZ_ZERO_UNDEF, -1, vp_ctlz_zero_undef, 1, 2) 272END_REGISTER_VP_SDNODE(VP_CTLZ_ZERO_UNDEF) 273END_REGISTER_VP_INTRINSIC(vp_ctlz) 274 275// llvm.vp.cttz(x,is_zero_poison,mask,vlen) 276BEGIN_REGISTER_VP_INTRINSIC(vp_cttz, 2, 3) 277BEGIN_REGISTER_VP_SDNODE(VP_CTTZ, -1, vp_cttz, 1, 2) 278VP_PROPERTY_FUNCTIONAL_INTRINSIC(cttz) 279VP_PROPERTY_FUNCTIONAL_SDOPC(CTTZ) 280END_REGISTER_VP_SDNODE(VP_CTTZ) 281BEGIN_REGISTER_VP_SDNODE(VP_CTTZ_ZERO_UNDEF, -1, vp_cttz_zero_undef, 1, 2) 282END_REGISTER_VP_SDNODE(VP_CTTZ_ZERO_UNDEF) 283END_REGISTER_VP_INTRINSIC(vp_cttz) 284 285// llvm.vp.fshl(x,y,z,mask,vlen) 286BEGIN_REGISTER_VP(vp_fshl, 3, 4, VP_FSHL, -1) 287VP_PROPERTY_FUNCTIONAL_INTRINSIC(fshl) 288VP_PROPERTY_FUNCTIONAL_SDOPC(FSHL) 289END_REGISTER_VP(vp_fshl, VP_FSHL) 290 291// llvm.vp.fshr(x,y,z,mask,vlen) 292BEGIN_REGISTER_VP(vp_fshr, 3, 4, VP_FSHR, -1) 293VP_PROPERTY_FUNCTIONAL_INTRINSIC(fshr) 294VP_PROPERTY_FUNCTIONAL_SDOPC(FSHR) 295END_REGISTER_VP(vp_fshr, VP_FSHR) 296///// } Integer Arithmetic 297 298///// Floating-Point Arithmetic { 299 300// Specialized helper macro for floating-point binary operators 301// <operation>(%x, %y, %mask, %evl). 302#ifdef HELPER_REGISTER_BINARY_FP_VP 303#error \ 304 "The internal helper macro HELPER_REGISTER_BINARY_FP_VP is already defined!" 305#endif 306#define HELPER_REGISTER_BINARY_FP_VP(OPSUFFIX, VPSD, IROPC, SDOPC) \ 307 BEGIN_REGISTER_VP(vp_##OPSUFFIX, 2, 3, VPSD, -1) \ 308 VP_PROPERTY_FUNCTIONAL_OPC(IROPC) \ 309 VP_PROPERTY_CONSTRAINEDFP(1, 1, experimental_constrained_##OPSUFFIX) \ 310 VP_PROPERTY_FUNCTIONAL_SDOPC(SDOPC) \ 311 VP_PROPERTY_BINARYOP \ 312 END_REGISTER_VP(vp_##OPSUFFIX, VPSD) 313 314// llvm.vp.fadd(x,y,mask,vlen) 315HELPER_REGISTER_BINARY_FP_VP(fadd, VP_FADD, FAdd, FADD) 316 317// llvm.vp.fsub(x,y,mask,vlen) 318HELPER_REGISTER_BINARY_FP_VP(fsub, VP_FSUB, FSub, FSUB) 319 320// llvm.vp.fmul(x,y,mask,vlen) 321HELPER_REGISTER_BINARY_FP_VP(fmul, VP_FMUL, FMul, FMUL) 322 323// llvm.vp.fdiv(x,y,mask,vlen) 324HELPER_REGISTER_BINARY_FP_VP(fdiv, VP_FDIV, FDiv, FDIV) 325 326// llvm.vp.frem(x,y,mask,vlen) 327HELPER_REGISTER_BINARY_FP_VP(frem, VP_FREM, FRem, FREM) 328 329#undef HELPER_REGISTER_BINARY_FP_VP 330 331// llvm.vp.fneg(x,mask,vlen) 332BEGIN_REGISTER_VP(vp_fneg, 1, 2, VP_FNEG, -1) 333VP_PROPERTY_FUNCTIONAL_OPC(FNeg) 334VP_PROPERTY_FUNCTIONAL_SDOPC(FNEG) 335END_REGISTER_VP(vp_fneg, VP_FNEG) 336 337// llvm.vp.fabs(x,mask,vlen) 338BEGIN_REGISTER_VP(vp_fabs, 1, 2, VP_FABS, -1) 339VP_PROPERTY_FUNCTIONAL_INTRINSIC(fabs) 340VP_PROPERTY_FUNCTIONAL_SDOPC(FABS) 341END_REGISTER_VP(vp_fabs, VP_FABS) 342 343// llvm.vp.sqrt(x,mask,vlen) 344BEGIN_REGISTER_VP(vp_sqrt, 1, 2, VP_SQRT, -1) 345VP_PROPERTY_FUNCTIONAL_INTRINSIC(sqrt) 346VP_PROPERTY_FUNCTIONAL_SDOPC(FSQRT) 347END_REGISTER_VP(vp_sqrt, VP_SQRT) 348 349// llvm.vp.fma(x,y,z,mask,vlen) 350BEGIN_REGISTER_VP(vp_fma, 3, 4, VP_FMA, -1) 351VP_PROPERTY_CONSTRAINEDFP(1, 1, experimental_constrained_fma) 352VP_PROPERTY_FUNCTIONAL_INTRINSIC(fma) 353VP_PROPERTY_FUNCTIONAL_SDOPC(FMA) 354END_REGISTER_VP(vp_fma, VP_FMA) 355 356// llvm.vp.fmuladd(x,y,z,mask,vlen) 357BEGIN_REGISTER_VP(vp_fmuladd, 3, 4, VP_FMULADD, -1) 358VP_PROPERTY_CONSTRAINEDFP(1, 1, experimental_constrained_fmuladd) 359VP_PROPERTY_FUNCTIONAL_INTRINSIC(fmuladd) 360VP_PROPERTY_FUNCTIONAL_SDOPC(FMAD) 361END_REGISTER_VP(vp_fmuladd, VP_FMULADD) 362 363// llvm.vp.copysign(x,y,mask,vlen) 364BEGIN_REGISTER_VP(vp_copysign, 2, 3, VP_FCOPYSIGN, -1) 365VP_PROPERTY_BINARYOP 366VP_PROPERTY_FUNCTIONAL_SDOPC(FCOPYSIGN) 367VP_PROPERTY_FUNCTIONAL_INTRINSIC(copysign) 368END_REGISTER_VP(vp_copysign, VP_FCOPYSIGN) 369 370// llvm.vp.minnum(x,y,mask,vlen) 371BEGIN_REGISTER_VP(vp_minnum, 2, 3, VP_FMINNUM, -1) 372VP_PROPERTY_BINARYOP 373VP_PROPERTY_FUNCTIONAL_SDOPC(FMINNUM) 374VP_PROPERTY_FUNCTIONAL_INTRINSIC(minnum) 375END_REGISTER_VP(vp_minnum, VP_FMINNUM) 376 377// llvm.vp.maxnum(x,y,mask,vlen) 378BEGIN_REGISTER_VP(vp_maxnum, 2, 3, VP_FMAXNUM, -1) 379VP_PROPERTY_BINARYOP 380VP_PROPERTY_FUNCTIONAL_SDOPC(FMAXNUM) 381VP_PROPERTY_FUNCTIONAL_INTRINSIC(maxnum) 382END_REGISTER_VP(vp_maxnum, VP_FMAXNUM) 383 384// llvm.vp.minimum(x,y,mask,vlen) 385BEGIN_REGISTER_VP(vp_minimum, 2, 3, VP_FMINIMUM, -1) 386VP_PROPERTY_BINARYOP 387VP_PROPERTY_FUNCTIONAL_SDOPC(FMINIMUM) 388VP_PROPERTY_FUNCTIONAL_INTRINSIC(minimum) 389END_REGISTER_VP(vp_minimum, VP_FMINIMUM) 390 391// llvm.vp.maximum(x,y,mask,vlen) 392BEGIN_REGISTER_VP(vp_maximum, 2, 3, VP_FMAXIMUM, -1) 393VP_PROPERTY_BINARYOP 394VP_PROPERTY_FUNCTIONAL_SDOPC(FMAXIMUM) 395VP_PROPERTY_FUNCTIONAL_INTRINSIC(maximum) 396END_REGISTER_VP(vp_maximum, VP_FMAXIMUM) 397 398// llvm.vp.ceil(x,mask,vlen) 399BEGIN_REGISTER_VP(vp_ceil, 1, 2, VP_FCEIL, -1) 400VP_PROPERTY_FUNCTIONAL_INTRINSIC(ceil) 401VP_PROPERTY_FUNCTIONAL_SDOPC(FCEIL) 402END_REGISTER_VP(vp_ceil, VP_FCEIL) 403 404// llvm.vp.floor(x,mask,vlen) 405BEGIN_REGISTER_VP(vp_floor, 1, 2, VP_FFLOOR, -1) 406VP_PROPERTY_FUNCTIONAL_INTRINSIC(floor) 407VP_PROPERTY_FUNCTIONAL_SDOPC(FFLOOR) 408END_REGISTER_VP(vp_floor, VP_FFLOOR) 409 410// llvm.vp.round(x,mask,vlen) 411BEGIN_REGISTER_VP(vp_round, 1, 2, VP_FROUND, -1) 412VP_PROPERTY_FUNCTIONAL_INTRINSIC(round) 413VP_PROPERTY_FUNCTIONAL_SDOPC(FROUND) 414END_REGISTER_VP(vp_round, VP_FROUND) 415 416// llvm.vp.roundeven(x,mask,vlen) 417BEGIN_REGISTER_VP(vp_roundeven, 1, 2, VP_FROUNDEVEN, -1) 418VP_PROPERTY_FUNCTIONAL_INTRINSIC(roundeven) 419VP_PROPERTY_FUNCTIONAL_SDOPC(FROUNDEVEN) 420END_REGISTER_VP(vp_roundeven, VP_FROUNDEVEN) 421 422// llvm.vp.roundtozero(x,mask,vlen) 423BEGIN_REGISTER_VP(vp_roundtozero, 1, 2, VP_FROUNDTOZERO, -1) 424VP_PROPERTY_FUNCTIONAL_INTRINSIC(trunc) 425VP_PROPERTY_FUNCTIONAL_SDOPC(FTRUNC) 426END_REGISTER_VP(vp_roundtozero, VP_FROUNDTOZERO) 427 428// llvm.vp.rint(x,mask,vlen) 429BEGIN_REGISTER_VP(vp_rint, 1, 2, VP_FRINT, -1) 430VP_PROPERTY_FUNCTIONAL_INTRINSIC(rint) 431VP_PROPERTY_FUNCTIONAL_SDOPC(FRINT) 432END_REGISTER_VP(vp_rint, VP_FRINT) 433 434// llvm.vp.nearbyint(x,mask,vlen) 435BEGIN_REGISTER_VP(vp_nearbyint, 1, 2, VP_FNEARBYINT, -1) 436VP_PROPERTY_FUNCTIONAL_INTRINSIC(nearbyint) 437VP_PROPERTY_FUNCTIONAL_SDOPC(FNEARBYINT) 438END_REGISTER_VP(vp_nearbyint, VP_FNEARBYINT) 439 440///// } Floating-Point Arithmetic 441 442///// Type Casts { 443// Specialized helper macro for type conversions. 444// <operation>(%x, %mask, %evl). 445#ifdef HELPER_REGISTER_FP_CAST_VP 446#error \ 447 "The internal helper macro HELPER_REGISTER_FP_CAST_VP is already defined!" 448#endif 449#define HELPER_REGISTER_FP_CAST_VP(OPSUFFIX, VPSD, IROPC, SDOPC, HASROUND) \ 450 BEGIN_REGISTER_VP(vp_##OPSUFFIX, 1, 2, VPSD, -1) \ 451 VP_PROPERTY_FUNCTIONAL_OPC(IROPC) \ 452 VP_PROPERTY_FUNCTIONAL_SDOPC(SDOPC) \ 453 VP_PROPERTY_CONSTRAINEDFP(HASROUND, 1, experimental_constrained_##OPSUFFIX) \ 454 VP_PROPERTY_CASTOP \ 455 END_REGISTER_VP(vp_##OPSUFFIX, VPSD) 456 457// llvm.vp.fptoui(x,mask,vlen) 458HELPER_REGISTER_FP_CAST_VP(fptoui, VP_FP_TO_UINT, FPToUI, FP_TO_UINT, 0) 459 460// llvm.vp.fptosi(x,mask,vlen) 461HELPER_REGISTER_FP_CAST_VP(fptosi, VP_FP_TO_SINT, FPToSI, FP_TO_SINT, 0) 462 463// llvm.vp.uitofp(x,mask,vlen) 464HELPER_REGISTER_FP_CAST_VP(uitofp, VP_UINT_TO_FP, UIToFP, UINT_TO_FP, 1) 465 466// llvm.vp.sitofp(x,mask,vlen) 467HELPER_REGISTER_FP_CAST_VP(sitofp, VP_SINT_TO_FP, SIToFP, SINT_TO_FP, 1) 468 469// llvm.vp.fptrunc(x,mask,vlen) 470HELPER_REGISTER_FP_CAST_VP(fptrunc, VP_FP_ROUND, FPTrunc, FP_ROUND, 1) 471 472// llvm.vp.fpext(x,mask,vlen) 473HELPER_REGISTER_FP_CAST_VP(fpext, VP_FP_EXTEND, FPExt, FP_EXTEND, 0) 474 475#undef HELPER_REGISTER_FP_CAST_VP 476 477// Specialized helper macro for integer type conversions. 478// <operation>(%x, %mask, %evl). 479#ifdef HELPER_REGISTER_INT_CAST_VP 480#error \ 481 "The internal helper macro HELPER_REGISTER_INT_CAST_VP is already defined!" 482#endif 483#define HELPER_REGISTER_INT_CAST_VP(OPSUFFIX, VPSD, IROPC, SDOPC) \ 484 BEGIN_REGISTER_VP(vp_##OPSUFFIX, 1, 2, VPSD, -1) \ 485 VP_PROPERTY_FUNCTIONAL_OPC(IROPC) \ 486 VP_PROPERTY_FUNCTIONAL_SDOPC(SDOPC) \ 487 VP_PROPERTY_CASTOP \ 488 END_REGISTER_VP(vp_##OPSUFFIX, VPSD) 489 490// llvm.vp.trunc(x,mask,vlen) 491HELPER_REGISTER_INT_CAST_VP(trunc, VP_TRUNCATE, Trunc, TRUNCATE) 492 493// llvm.vp.zext(x,mask,vlen) 494HELPER_REGISTER_INT_CAST_VP(zext, VP_ZERO_EXTEND, ZExt, ZERO_EXTEND) 495 496// llvm.vp.sext(x,mask,vlen) 497HELPER_REGISTER_INT_CAST_VP(sext, VP_SIGN_EXTEND, SExt, SIGN_EXTEND) 498 499// llvm.vp.ptrtoint(x,mask,vlen) 500BEGIN_REGISTER_VP(vp_ptrtoint, 1, 2, VP_PTRTOINT, -1) 501VP_PROPERTY_FUNCTIONAL_OPC(PtrToInt) 502VP_PROPERTY_CASTOP 503END_REGISTER_VP(vp_ptrtoint, VP_PTRTOINT) 504 505// llvm.vp.inttoptr(x,mask,vlen) 506BEGIN_REGISTER_VP(vp_inttoptr, 1, 2, VP_INTTOPTR, -1) 507VP_PROPERTY_FUNCTIONAL_OPC(IntToPtr) 508VP_PROPERTY_CASTOP 509END_REGISTER_VP(vp_inttoptr, VP_INTTOPTR) 510 511#undef HELPER_REGISTER_INT_CAST_VP 512 513///// } Type Casts 514 515///// Comparisons { 516 517// VP_SETCC (ISel only) 518BEGIN_REGISTER_VP_SDNODE(VP_SETCC, 0, vp_setcc, 3, 4) 519END_REGISTER_VP_SDNODE(VP_SETCC) 520 521// llvm.vp.fcmp(x,y,cc,mask,vlen) 522BEGIN_REGISTER_VP_INTRINSIC(vp_fcmp, 3, 4) 523HELPER_MAP_VPID_TO_VPSD(vp_fcmp, VP_SETCC) 524VP_PROPERTY_FUNCTIONAL_OPC(FCmp) 525VP_PROPERTY_CMP(2, true) 526VP_PROPERTY_CONSTRAINEDFP(0, 1, experimental_constrained_fcmp) 527END_REGISTER_VP_INTRINSIC(vp_fcmp) 528 529// llvm.vp.icmp(x,y,cc,mask,vlen) 530BEGIN_REGISTER_VP_INTRINSIC(vp_icmp, 3, 4) 531HELPER_MAP_VPID_TO_VPSD(vp_icmp, VP_SETCC) 532VP_PROPERTY_FUNCTIONAL_OPC(ICmp) 533VP_PROPERTY_CMP(2, false) 534END_REGISTER_VP_INTRINSIC(vp_icmp) 535 536///// } Comparisons 537 538// llvm.vp.is.fpclass(on_true,on_false,mask,vlen) 539BEGIN_REGISTER_VP(vp_is_fpclass, 2, 3, VP_IS_FPCLASS, 0) 540VP_PROPERTY_FUNCTIONAL_INTRINSIC(is_fpclass) 541END_REGISTER_VP(vp_is_fpclass, VP_IS_FPCLASS) 542 543///// Memory Operations { 544// llvm.vp.store(val,ptr,mask,vlen) 545BEGIN_REGISTER_VP_INTRINSIC(vp_store, 2, 3) 546// chain = VP_STORE chain,val,base,offset,mask,evl 547BEGIN_REGISTER_VP_SDNODE(VP_STORE, 1, vp_store, 4, 5) 548HELPER_MAP_VPID_TO_VPSD(vp_store, VP_STORE) 549VP_PROPERTY_FUNCTIONAL_OPC(Store) 550VP_PROPERTY_FUNCTIONAL_INTRINSIC(masked_store) 551VP_PROPERTY_MEMOP(1, 0) 552END_REGISTER_VP(vp_store, VP_STORE) 553 554// llvm.experimental.vp.strided.store(val,ptr,stride,mask,vlen) 555BEGIN_REGISTER_VP_INTRINSIC(experimental_vp_strided_store, 3, 4) 556// chain = EXPERIMENTAL_VP_STRIDED_STORE chain,val,base,offset,stride,mask,evl 557VP_PROPERTY_NO_FUNCTIONAL 558BEGIN_REGISTER_VP_SDNODE(EXPERIMENTAL_VP_STRIDED_STORE, 1, experimental_vp_strided_store, 5, 6) 559HELPER_MAP_VPID_TO_VPSD(experimental_vp_strided_store, EXPERIMENTAL_VP_STRIDED_STORE) 560VP_PROPERTY_MEMOP(1, 0) 561END_REGISTER_VP(experimental_vp_strided_store, EXPERIMENTAL_VP_STRIDED_STORE) 562 563// llvm.vp.scatter(ptr,val,mask,vlen) 564BEGIN_REGISTER_VP_INTRINSIC(vp_scatter, 2, 3) 565// chain = VP_SCATTER chain,val,base,indices,scale,mask,evl 566BEGIN_REGISTER_VP_SDNODE(VP_SCATTER, 1, vp_scatter, 5, 6) 567HELPER_MAP_VPID_TO_VPSD(vp_scatter, VP_SCATTER) 568VP_PROPERTY_FUNCTIONAL_INTRINSIC(masked_scatter) 569VP_PROPERTY_MEMOP(1, 0) 570END_REGISTER_VP(vp_scatter, VP_SCATTER) 571 572// llvm.vp.load(ptr,mask,vlen) 573BEGIN_REGISTER_VP_INTRINSIC(vp_load, 1, 2) 574// val,chain = VP_LOAD chain,base,offset,mask,evl 575BEGIN_REGISTER_VP_SDNODE(VP_LOAD, -1, vp_load, 3, 4) 576HELPER_MAP_VPID_TO_VPSD(vp_load, VP_LOAD) 577VP_PROPERTY_FUNCTIONAL_OPC(Load) 578VP_PROPERTY_FUNCTIONAL_INTRINSIC(masked_load) 579VP_PROPERTY_MEMOP(0, std::nullopt) 580END_REGISTER_VP(vp_load, VP_LOAD) 581 582// llvm.experimental.vp.strided.load(ptr,stride,mask,vlen) 583BEGIN_REGISTER_VP_INTRINSIC(experimental_vp_strided_load, 2, 3) 584// chain = EXPERIMENTAL_VP_STRIDED_LOAD chain,base,offset,stride,mask,evl 585VP_PROPERTY_NO_FUNCTIONAL 586BEGIN_REGISTER_VP_SDNODE(EXPERIMENTAL_VP_STRIDED_LOAD, -1, experimental_vp_strided_load, 4, 5) 587HELPER_MAP_VPID_TO_VPSD(experimental_vp_strided_load, EXPERIMENTAL_VP_STRIDED_LOAD) 588VP_PROPERTY_MEMOP(0, std::nullopt) 589END_REGISTER_VP(experimental_vp_strided_load, EXPERIMENTAL_VP_STRIDED_LOAD) 590 591// llvm.vp.gather(ptr,mask,vlen) 592BEGIN_REGISTER_VP_INTRINSIC(vp_gather, 1, 2) 593// val,chain = VP_GATHER chain,base,indices,scale,mask,evl 594BEGIN_REGISTER_VP_SDNODE(VP_GATHER, -1, vp_gather, 4, 5) 595HELPER_MAP_VPID_TO_VPSD(vp_gather, VP_GATHER) 596VP_PROPERTY_FUNCTIONAL_INTRINSIC(masked_gather) 597VP_PROPERTY_MEMOP(0, std::nullopt) 598END_REGISTER_VP(vp_gather, VP_GATHER) 599 600///// } Memory Operations 601 602///// Reductions { 603 604// Specialized helper macro for VP reductions (%start, %x, %mask, %evl). 605#ifdef HELPER_REGISTER_REDUCTION_VP 606#error \ 607 "The internal helper macro HELPER_REGISTER_REDUCTION_VP is already defined!" 608#endif 609#define HELPER_REGISTER_REDUCTION_VP(VPID, VPSD, INTRIN) \ 610 BEGIN_REGISTER_VP(VPID, 2, 3, VPSD, 1) \ 611 VP_PROPERTY_FUNCTIONAL_INTRINSIC(INTRIN) \ 612 VP_PROPERTY_REDUCTION(0, 1) \ 613 END_REGISTER_VP(VPID, VPSD) 614 615// llvm.vp.reduce.add(start,x,mask,vlen) 616HELPER_REGISTER_REDUCTION_VP(vp_reduce_add, VP_REDUCE_ADD, 617 vector_reduce_add) 618 619// llvm.vp.reduce.mul(start,x,mask,vlen) 620HELPER_REGISTER_REDUCTION_VP(vp_reduce_mul, VP_REDUCE_MUL, 621 vector_reduce_mul) 622 623// llvm.vp.reduce.and(start,x,mask,vlen) 624HELPER_REGISTER_REDUCTION_VP(vp_reduce_and, VP_REDUCE_AND, 625 vector_reduce_and) 626 627// llvm.vp.reduce.or(start,x,mask,vlen) 628HELPER_REGISTER_REDUCTION_VP(vp_reduce_or, VP_REDUCE_OR, 629 vector_reduce_or) 630 631// llvm.vp.reduce.xor(start,x,mask,vlen) 632HELPER_REGISTER_REDUCTION_VP(vp_reduce_xor, VP_REDUCE_XOR, 633 vector_reduce_xor) 634 635// llvm.vp.reduce.smax(start,x,mask,vlen) 636HELPER_REGISTER_REDUCTION_VP(vp_reduce_smax, VP_REDUCE_SMAX, 637 vector_reduce_smax) 638 639// llvm.vp.reduce.smin(start,x,mask,vlen) 640HELPER_REGISTER_REDUCTION_VP(vp_reduce_smin, VP_REDUCE_SMIN, 641 vector_reduce_smin) 642 643// llvm.vp.reduce.umax(start,x,mask,vlen) 644HELPER_REGISTER_REDUCTION_VP(vp_reduce_umax, VP_REDUCE_UMAX, 645 vector_reduce_umax) 646 647// llvm.vp.reduce.umin(start,x,mask,vlen) 648HELPER_REGISTER_REDUCTION_VP(vp_reduce_umin, VP_REDUCE_UMIN, 649 vector_reduce_umin) 650 651// llvm.vp.reduce.fmax(start,x,mask,vlen) 652HELPER_REGISTER_REDUCTION_VP(vp_reduce_fmax, VP_REDUCE_FMAX, 653 vector_reduce_fmax) 654 655// llvm.vp.reduce.fmin(start,x,mask,vlen) 656HELPER_REGISTER_REDUCTION_VP(vp_reduce_fmin, VP_REDUCE_FMIN, 657 vector_reduce_fmin) 658 659#undef HELPER_REGISTER_REDUCTION_VP 660 661// Specialized helper macro for VP reductions as above but with two forms: 662// sequential and reassociative. These manifest as the presence of 'reassoc' 663// fast-math flags in the IR and as two distinct ISD opcodes in the 664// SelectionDAG. 665// Note we by default map from the VP intrinsic to the SEQ ISD opcode, which 666// can then be relaxed to the non-SEQ ISD opcode if the 'reassoc' flag is set. 667#ifdef HELPER_REGISTER_REDUCTION_SEQ_VP 668#error \ 669 "The internal helper macro HELPER_REGISTER_REDUCTION_SEQ_VP is already defined!" 670#endif 671#define HELPER_REGISTER_REDUCTION_SEQ_VP(VPID, VPSD, SEQ_VPSD, INTRIN) \ 672 BEGIN_REGISTER_VP_INTRINSIC(VPID, 2, 3) \ 673 BEGIN_REGISTER_VP_SDNODE(VPSD, 1, VPID, 2, 3) \ 674 VP_PROPERTY_REDUCTION(0, 1) \ 675 END_REGISTER_VP_SDNODE(VPSD) \ 676 BEGIN_REGISTER_VP_SDNODE(SEQ_VPSD, 1, VPID, 2, 3) \ 677 HELPER_MAP_VPID_TO_VPSD(VPID, SEQ_VPSD) \ 678 VP_PROPERTY_REDUCTION(0, 1) \ 679 END_REGISTER_VP_SDNODE(SEQ_VPSD) \ 680 VP_PROPERTY_FUNCTIONAL_INTRINSIC(INTRIN) \ 681 END_REGISTER_VP_INTRINSIC(VPID) 682 683// llvm.vp.reduce.fadd(start,x,mask,vlen) 684HELPER_REGISTER_REDUCTION_SEQ_VP(vp_reduce_fadd, VP_REDUCE_FADD, 685 VP_REDUCE_SEQ_FADD, 686 vector_reduce_fadd) 687 688// llvm.vp.reduce.fmul(start,x,mask,vlen) 689HELPER_REGISTER_REDUCTION_SEQ_VP(vp_reduce_fmul, VP_REDUCE_FMUL, 690 VP_REDUCE_SEQ_FMUL, 691 vector_reduce_fmul) 692 693#undef HELPER_REGISTER_REDUCTION_SEQ_VP 694 695///// } Reduction 696 697///// Shuffles { 698 699// The mask 'cond' operand of llvm.vp.select and llvm.vp.merge are not reported 700// as masks with the BEGIN_REGISTER_VP_* macros. This is because, unlike other 701// VP intrinsics, these two have a defined result on lanes where the mask is 702// false. 703// 704// llvm.vp.select(cond,on_true,on_false,vlen) 705BEGIN_REGISTER_VP(vp_select, std::nullopt, 3, VP_SELECT, -1) 706VP_PROPERTY_FUNCTIONAL_OPC(Select) 707VP_PROPERTY_FUNCTIONAL_SDOPC(VSELECT) 708END_REGISTER_VP(vp_select, VP_SELECT) 709 710// llvm.vp.merge(cond,on_true,on_false,pivot) 711BEGIN_REGISTER_VP(vp_merge, std::nullopt, 3, VP_MERGE, -1) 712VP_PROPERTY_NO_FUNCTIONAL 713END_REGISTER_VP(vp_merge, VP_MERGE) 714 715BEGIN_REGISTER_VP(experimental_vp_splice, 3, 5, EXPERIMENTAL_VP_SPLICE, -1) 716VP_PROPERTY_NO_FUNCTIONAL 717END_REGISTER_VP(experimental_vp_splice, EXPERIMENTAL_VP_SPLICE) 718 719// llvm.experimental.vp.reverse(x,mask,vlen) 720BEGIN_REGISTER_VP(experimental_vp_reverse, 1, 2, 721 EXPERIMENTAL_VP_REVERSE, -1) 722VP_PROPERTY_NO_FUNCTIONAL 723END_REGISTER_VP(experimental_vp_reverse, EXPERIMENTAL_VP_REVERSE) 724 725///// } Shuffles 726 727#undef BEGIN_REGISTER_VP 728#undef BEGIN_REGISTER_VP_INTRINSIC 729#undef BEGIN_REGISTER_VP_SDNODE 730#undef END_REGISTER_VP 731#undef END_REGISTER_VP_INTRINSIC 732#undef END_REGISTER_VP_SDNODE 733#undef HELPER_MAP_VPID_TO_VPSD 734#undef VP_PROPERTY_BINARYOP 735#undef VP_PROPERTY_CASTOP 736#undef VP_PROPERTY_CMP 737#undef VP_PROPERTY_CONSTRAINEDFP 738#undef VP_PROPERTY_FUNCTIONAL_INTRINSIC 739#undef VP_PROPERTY_FUNCTIONAL_OPC 740#undef VP_PROPERTY_FUNCTIONAL_SDOPC 741#undef VP_PROPERTY_NO_FUNCTIONAL 742#undef VP_PROPERTY_MEMOP 743#undef VP_PROPERTY_REDUCTION 744