| 1 | //===-- AutoUpgrade.cpp - Implement auto-upgrade helper functions ---------===// |
| 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 implements the auto-upgrade helper functions. |
| 10 | // This is where deprecated IR intrinsics and other IR features are updated to |
| 11 | // current specifications. |
| 12 | // |
| 13 | //===----------------------------------------------------------------------===// |
| 14 | |
| 15 | #include "llvm/IR/AutoUpgrade.h" |
| 16 | #include "llvm/ADT/ArrayRef.h" |
| 17 | #include "llvm/ADT/StringExtras.h" |
| 18 | #include "llvm/ADT/StringRef.h" |
| 19 | #include "llvm/ADT/StringSwitch.h" |
| 20 | #include "llvm/BinaryFormat/Dwarf.h" |
| 21 | #include "llvm/IR/AttributeMask.h" |
| 22 | #include "llvm/IR/Attributes.h" |
| 23 | #include "llvm/IR/CallingConv.h" |
| 24 | #include "llvm/IR/Constants.h" |
| 25 | #include "llvm/IR/DebugInfo.h" |
| 26 | #include "llvm/IR/DebugInfoMetadata.h" |
| 27 | #include "llvm/IR/DiagnosticInfo.h" |
| 28 | #include "llvm/IR/Function.h" |
| 29 | #include "llvm/IR/IRBuilder.h" |
| 30 | #include "llvm/IR/InstVisitor.h" |
| 31 | #include "llvm/IR/Instruction.h" |
| 32 | #include "llvm/IR/IntrinsicInst.h" |
| 33 | #include "llvm/IR/Intrinsics.h" |
| 34 | #include "llvm/IR/IntrinsicsAArch64.h" |
| 35 | #include "llvm/IR/IntrinsicsAMDGPU.h" |
| 36 | #include "llvm/IR/IntrinsicsARM.h" |
| 37 | #include "llvm/IR/IntrinsicsNVPTX.h" |
| 38 | #include "llvm/IR/IntrinsicsRISCV.h" |
| 39 | #include "llvm/IR/IntrinsicsWebAssembly.h" |
| 40 | #include "llvm/IR/IntrinsicsX86.h" |
| 41 | #include "llvm/IR/LLVMContext.h" |
| 42 | #include "llvm/IR/MDBuilder.h" |
| 43 | #include "llvm/IR/Metadata.h" |
| 44 | #include "llvm/IR/Module.h" |
| 45 | #include "llvm/IR/Value.h" |
| 46 | #include "llvm/IR/Verifier.h" |
| 47 | #include "llvm/Support/AMDGPUAddrSpace.h" |
| 48 | #include "llvm/Support/CommandLine.h" |
| 49 | #include "llvm/Support/ErrorHandling.h" |
| 50 | #include "llvm/Support/NVPTXAddrSpace.h" |
| 51 | #include "llvm/Support/Regex.h" |
| 52 | #include "llvm/Support/TimeProfiler.h" |
| 53 | #include "llvm/TargetParser/Triple.h" |
| 54 | #include <cstdint> |
| 55 | #include <cstring> |
| 56 | #include <numeric> |
| 57 | |
| 58 | using namespace llvm; |
| 59 | |
| 60 | static cl::opt<bool> |
| 61 | DisableAutoUpgradeDebugInfo("disable-auto-upgrade-debug-info" , |
| 62 | cl::desc("Disable autoupgrade of debug info" )); |
| 63 | |
| 64 | static void rename(GlobalValue *GV) { GV->setName(GV->getName() + ".old" ); } |
| 65 | |
| 66 | // Upgrade the declarations of the SSE4.1 ptest intrinsics whose arguments have |
| 67 | // changed their type from v4f32 to v2i64. |
| 68 | static bool upgradePTESTIntrinsic(Function *F, Intrinsic::ID IID, |
| 69 | Function *&NewFn) { |
| 70 | // Check whether this is an old version of the function, which received |
| 71 | // v4f32 arguments. |
| 72 | Type *Arg0Type = F->getFunctionType()->getParamType(i: 0); |
| 73 | if (Arg0Type != FixedVectorType::get(ElementType: Type::getFloatTy(C&: F->getContext()), NumElts: 4)) |
| 74 | return false; |
| 75 | |
| 76 | // Yes, it's old, replace it with new version. |
| 77 | rename(GV: F); |
| 78 | NewFn = Intrinsic::getOrInsertDeclaration(M: F->getParent(), id: IID); |
| 79 | return true; |
| 80 | } |
| 81 | |
| 82 | // Upgrade the declarations of intrinsic functions whose 8-bit immediate mask |
| 83 | // arguments have changed their type from i32 to i8. |
| 84 | static bool upgradeX86IntrinsicsWith8BitMask(Function *F, Intrinsic::ID IID, |
| 85 | Function *&NewFn) { |
| 86 | // Check that the last argument is an i32. |
| 87 | Type *LastArgType = F->getFunctionType()->getParamType( |
| 88 | i: F->getFunctionType()->getNumParams() - 1); |
| 89 | if (!LastArgType->isIntegerTy(Bitwidth: 32)) |
| 90 | return false; |
| 91 | |
| 92 | // Move this function aside and map down. |
| 93 | rename(GV: F); |
| 94 | NewFn = Intrinsic::getOrInsertDeclaration(M: F->getParent(), id: IID); |
| 95 | return true; |
| 96 | } |
| 97 | |
| 98 | // Upgrade the declaration of fp compare intrinsics that change return type |
| 99 | // from scalar to vXi1 mask. |
| 100 | static bool upgradeX86MaskedFPCompare(Function *F, Intrinsic::ID IID, |
| 101 | Function *&NewFn) { |
| 102 | // Check if the return type is a vector. |
| 103 | if (F->getReturnType()->isVectorTy()) |
| 104 | return false; |
| 105 | |
| 106 | rename(GV: F); |
| 107 | NewFn = Intrinsic::getOrInsertDeclaration(M: F->getParent(), id: IID); |
| 108 | return true; |
| 109 | } |
| 110 | |
| 111 | // Upgrade the declaration of multiply and add bytes intrinsics whose input |
| 112 | // arguments' types have changed from vectors of i32 to vectors of i8 |
| 113 | static bool upgradeX86MultiplyAddBytes(Function *F, Intrinsic::ID IID, |
| 114 | Function *&NewFn) { |
| 115 | // check if input argument type is a vector of i8 |
| 116 | Type *Arg1Type = F->getFunctionType()->getParamType(i: 1); |
| 117 | Type *Arg2Type = F->getFunctionType()->getParamType(i: 2); |
| 118 | if (Arg1Type->isVectorTy() && |
| 119 | cast<VectorType>(Val: Arg1Type)->getElementType()->isIntegerTy(Bitwidth: 8) && |
| 120 | Arg2Type->isVectorTy() && |
| 121 | cast<VectorType>(Val: Arg2Type)->getElementType()->isIntegerTy(Bitwidth: 8)) |
| 122 | return false; |
| 123 | |
| 124 | rename(GV: F); |
| 125 | NewFn = Intrinsic::getOrInsertDeclaration(M: F->getParent(), id: IID); |
| 126 | return true; |
| 127 | } |
| 128 | |
| 129 | // Upgrade the declaration of multipy and add words intrinsics whose input |
| 130 | // arguments' types have changed to vectors of i32 to vectors of i16 |
| 131 | static bool upgradeX86MultiplyAddWords(Function *F, Intrinsic::ID IID, |
| 132 | Function *&NewFn) { |
| 133 | // check if input argument type is a vector of i16 |
| 134 | Type *Arg1Type = F->getFunctionType()->getParamType(i: 1); |
| 135 | Type *Arg2Type = F->getFunctionType()->getParamType(i: 2); |
| 136 | if (Arg1Type->isVectorTy() && |
| 137 | cast<VectorType>(Val: Arg1Type)->getElementType()->isIntegerTy(Bitwidth: 16) && |
| 138 | Arg2Type->isVectorTy() && |
| 139 | cast<VectorType>(Val: Arg2Type)->getElementType()->isIntegerTy(Bitwidth: 16)) |
| 140 | return false; |
| 141 | |
| 142 | rename(GV: F); |
| 143 | NewFn = Intrinsic::getOrInsertDeclaration(M: F->getParent(), id: IID); |
| 144 | return true; |
| 145 | } |
| 146 | |
| 147 | static bool upgradeX86BF16Intrinsic(Function *F, Intrinsic::ID IID, |
| 148 | Function *&NewFn) { |
| 149 | if (F->getReturnType()->getScalarType()->isBFloatTy()) |
| 150 | return false; |
| 151 | |
| 152 | rename(GV: F); |
| 153 | NewFn = Intrinsic::getOrInsertDeclaration(M: F->getParent(), id: IID); |
| 154 | return true; |
| 155 | } |
| 156 | |
| 157 | static bool upgradeX86BF16DPIntrinsic(Function *F, Intrinsic::ID IID, |
| 158 | Function *&NewFn) { |
| 159 | if (F->getFunctionType()->getParamType(i: 1)->getScalarType()->isBFloatTy()) |
| 160 | return false; |
| 161 | |
| 162 | rename(GV: F); |
| 163 | NewFn = Intrinsic::getOrInsertDeclaration(M: F->getParent(), id: IID); |
| 164 | return true; |
| 165 | } |
| 166 | |
| 167 | static bool shouldUpgradeX86Intrinsic(Function *F, StringRef Name) { |
| 168 | // All of the intrinsics matches below should be marked with which llvm |
| 169 | // version started autoupgrading them. At some point in the future we would |
| 170 | // like to use this information to remove upgrade code for some older |
| 171 | // intrinsics. It is currently undecided how we will determine that future |
| 172 | // point. |
| 173 | if (Name.consume_front(Prefix: "avx." )) |
| 174 | return (Name.starts_with(Prefix: "blend.p" ) || // Added in 3.7 |
| 175 | Name == "cvt.ps2.pd.256" || // Added in 3.9 |
| 176 | Name == "cvtdq2.pd.256" || // Added in 3.9 |
| 177 | Name == "cvtdq2.ps.256" || // Added in 7.0 |
| 178 | Name.starts_with(Prefix: "movnt." ) || // Added in 3.2 |
| 179 | Name.starts_with(Prefix: "sqrt.p" ) || // Added in 7.0 |
| 180 | Name.starts_with(Prefix: "storeu." ) || // Added in 3.9 |
| 181 | Name.starts_with(Prefix: "vbroadcast.s" ) || // Added in 3.5 |
| 182 | Name.starts_with(Prefix: "vbroadcastf128" ) || // Added in 4.0 |
| 183 | Name.starts_with(Prefix: "vextractf128." ) || // Added in 3.7 |
| 184 | Name.starts_with(Prefix: "vinsertf128." ) || // Added in 3.7 |
| 185 | Name.starts_with(Prefix: "vperm2f128." ) || // Added in 6.0 |
| 186 | Name.starts_with(Prefix: "vpermil." )); // Added in 3.1 |
| 187 | |
| 188 | if (Name.consume_front(Prefix: "avx2." )) |
| 189 | return (Name == "movntdqa" || // Added in 5.0 |
| 190 | Name.starts_with(Prefix: "pabs." ) || // Added in 6.0 |
| 191 | Name.starts_with(Prefix: "padds." ) || // Added in 8.0 |
| 192 | Name.starts_with(Prefix: "paddus." ) || // Added in 8.0 |
| 193 | Name.starts_with(Prefix: "pblendd." ) || // Added in 3.7 |
| 194 | Name == "pblendw" || // Added in 3.7 |
| 195 | Name.starts_with(Prefix: "pbroadcast" ) || // Added in 3.8 |
| 196 | Name.starts_with(Prefix: "pcmpeq." ) || // Added in 3.1 |
| 197 | Name.starts_with(Prefix: "pcmpgt." ) || // Added in 3.1 |
| 198 | Name.starts_with(Prefix: "pmax" ) || // Added in 3.9 |
| 199 | Name.starts_with(Prefix: "pmin" ) || // Added in 3.9 |
| 200 | Name.starts_with(Prefix: "pmovsx" ) || // Added in 3.9 |
| 201 | Name.starts_with(Prefix: "pmovzx" ) || // Added in 3.9 |
| 202 | Name == "pmul.dq" || // Added in 7.0 |
| 203 | Name == "pmulu.dq" || // Added in 7.0 |
| 204 | Name.starts_with(Prefix: "psll.dq" ) || // Added in 3.7 |
| 205 | Name.starts_with(Prefix: "psrl.dq" ) || // Added in 3.7 |
| 206 | Name.starts_with(Prefix: "psubs." ) || // Added in 8.0 |
| 207 | Name.starts_with(Prefix: "psubus." ) || // Added in 8.0 |
| 208 | Name.starts_with(Prefix: "vbroadcast" ) || // Added in 3.8 |
| 209 | Name == "vbroadcasti128" || // Added in 3.7 |
| 210 | Name == "vextracti128" || // Added in 3.7 |
| 211 | Name == "vinserti128" || // Added in 3.7 |
| 212 | Name == "vperm2i128" ); // Added in 6.0 |
| 213 | |
| 214 | if (Name.consume_front(Prefix: "avx512." )) { |
| 215 | if (Name.consume_front(Prefix: "mask." )) |
| 216 | // 'avx512.mask.*' |
| 217 | return (Name.starts_with(Prefix: "add.p" ) || // Added in 7.0. 128/256 in 4.0 |
| 218 | Name.starts_with(Prefix: "and." ) || // Added in 3.9 |
| 219 | Name.starts_with(Prefix: "andn." ) || // Added in 3.9 |
| 220 | Name.starts_with(Prefix: "broadcast.s" ) || // Added in 3.9 |
| 221 | Name.starts_with(Prefix: "broadcastf32x4." ) || // Added in 6.0 |
| 222 | Name.starts_with(Prefix: "broadcastf32x8." ) || // Added in 6.0 |
| 223 | Name.starts_with(Prefix: "broadcastf64x2." ) || // Added in 6.0 |
| 224 | Name.starts_with(Prefix: "broadcastf64x4." ) || // Added in 6.0 |
| 225 | Name.starts_with(Prefix: "broadcasti32x4." ) || // Added in 6.0 |
| 226 | Name.starts_with(Prefix: "broadcasti32x8." ) || // Added in 6.0 |
| 227 | Name.starts_with(Prefix: "broadcasti64x2." ) || // Added in 6.0 |
| 228 | Name.starts_with(Prefix: "broadcasti64x4." ) || // Added in 6.0 |
| 229 | Name.starts_with(Prefix: "cmp.b" ) || // Added in 5.0 |
| 230 | Name.starts_with(Prefix: "cmp.d" ) || // Added in 5.0 |
| 231 | Name.starts_with(Prefix: "cmp.q" ) || // Added in 5.0 |
| 232 | Name.starts_with(Prefix: "cmp.w" ) || // Added in 5.0 |
| 233 | Name.starts_with(Prefix: "compress.b" ) || // Added in 9.0 |
| 234 | Name.starts_with(Prefix: "compress.d" ) || // Added in 9.0 |
| 235 | Name.starts_with(Prefix: "compress.p" ) || // Added in 9.0 |
| 236 | Name.starts_with(Prefix: "compress.q" ) || // Added in 9.0 |
| 237 | Name.starts_with(Prefix: "compress.store." ) || // Added in 7.0 |
| 238 | Name.starts_with(Prefix: "compress.w" ) || // Added in 9.0 |
| 239 | Name.starts_with(Prefix: "conflict." ) || // Added in 9.0 |
| 240 | Name.starts_with(Prefix: "cvtdq2pd." ) || // Added in 4.0 |
| 241 | Name.starts_with(Prefix: "cvtdq2ps." ) || // Added in 7.0 updated 9.0 |
| 242 | Name == "cvtpd2dq.256" || // Added in 7.0 |
| 243 | Name == "cvtpd2ps.256" || // Added in 7.0 |
| 244 | Name == "cvtps2pd.128" || // Added in 7.0 |
| 245 | Name == "cvtps2pd.256" || // Added in 7.0 |
| 246 | Name.starts_with(Prefix: "cvtqq2pd." ) || // Added in 7.0 updated 9.0 |
| 247 | Name == "cvtqq2ps.256" || // Added in 9.0 |
| 248 | Name == "cvtqq2ps.512" || // Added in 9.0 |
| 249 | Name == "cvttpd2dq.256" || // Added in 7.0 |
| 250 | Name == "cvttps2dq.128" || // Added in 7.0 |
| 251 | Name == "cvttps2dq.256" || // Added in 7.0 |
| 252 | Name.starts_with(Prefix: "cvtudq2pd." ) || // Added in 4.0 |
| 253 | Name.starts_with(Prefix: "cvtudq2ps." ) || // Added in 7.0 updated 9.0 |
| 254 | Name.starts_with(Prefix: "cvtuqq2pd." ) || // Added in 7.0 updated 9.0 |
| 255 | Name == "cvtuqq2ps.256" || // Added in 9.0 |
| 256 | Name == "cvtuqq2ps.512" || // Added in 9.0 |
| 257 | Name.starts_with(Prefix: "dbpsadbw." ) || // Added in 7.0 |
| 258 | Name.starts_with(Prefix: "div.p" ) || // Added in 7.0. 128/256 in 4.0 |
| 259 | Name.starts_with(Prefix: "expand.b" ) || // Added in 9.0 |
| 260 | Name.starts_with(Prefix: "expand.d" ) || // Added in 9.0 |
| 261 | Name.starts_with(Prefix: "expand.load." ) || // Added in 7.0 |
| 262 | Name.starts_with(Prefix: "expand.p" ) || // Added in 9.0 |
| 263 | Name.starts_with(Prefix: "expand.q" ) || // Added in 9.0 |
| 264 | Name.starts_with(Prefix: "expand.w" ) || // Added in 9.0 |
| 265 | Name.starts_with(Prefix: "fpclass.p" ) || // Added in 7.0 |
| 266 | Name.starts_with(Prefix: "insert" ) || // Added in 4.0 |
| 267 | Name.starts_with(Prefix: "load." ) || // Added in 3.9 |
| 268 | Name.starts_with(Prefix: "loadu." ) || // Added in 3.9 |
| 269 | Name.starts_with(Prefix: "lzcnt." ) || // Added in 5.0 |
| 270 | Name.starts_with(Prefix: "max.p" ) || // Added in 7.0. 128/256 in 5.0 |
| 271 | Name.starts_with(Prefix: "min.p" ) || // Added in 7.0. 128/256 in 5.0 |
| 272 | Name.starts_with(Prefix: "movddup" ) || // Added in 3.9 |
| 273 | Name.starts_with(Prefix: "move.s" ) || // Added in 4.0 |
| 274 | Name.starts_with(Prefix: "movshdup" ) || // Added in 3.9 |
| 275 | Name.starts_with(Prefix: "movsldup" ) || // Added in 3.9 |
| 276 | Name.starts_with(Prefix: "mul.p" ) || // Added in 7.0. 128/256 in 4.0 |
| 277 | Name.starts_with(Prefix: "or." ) || // Added in 3.9 |
| 278 | Name.starts_with(Prefix: "pabs." ) || // Added in 6.0 |
| 279 | Name.starts_with(Prefix: "packssdw." ) || // Added in 5.0 |
| 280 | Name.starts_with(Prefix: "packsswb." ) || // Added in 5.0 |
| 281 | Name.starts_with(Prefix: "packusdw." ) || // Added in 5.0 |
| 282 | Name.starts_with(Prefix: "packuswb." ) || // Added in 5.0 |
| 283 | Name.starts_with(Prefix: "padd." ) || // Added in 4.0 |
| 284 | Name.starts_with(Prefix: "padds." ) || // Added in 8.0 |
| 285 | Name.starts_with(Prefix: "paddus." ) || // Added in 8.0 |
| 286 | Name.starts_with(Prefix: "palignr." ) || // Added in 3.9 |
| 287 | Name.starts_with(Prefix: "pand." ) || // Added in 3.9 |
| 288 | Name.starts_with(Prefix: "pandn." ) || // Added in 3.9 |
| 289 | Name.starts_with(Prefix: "pavg" ) || // Added in 6.0 |
| 290 | Name.starts_with(Prefix: "pbroadcast" ) || // Added in 6.0 |
| 291 | Name.starts_with(Prefix: "pcmpeq." ) || // Added in 3.9 |
| 292 | Name.starts_with(Prefix: "pcmpgt." ) || // Added in 3.9 |
| 293 | Name.starts_with(Prefix: "perm.df." ) || // Added in 3.9 |
| 294 | Name.starts_with(Prefix: "perm.di." ) || // Added in 3.9 |
| 295 | Name.starts_with(Prefix: "permvar." ) || // Added in 7.0 |
| 296 | Name.starts_with(Prefix: "pmaddubs.w." ) || // Added in 7.0 |
| 297 | Name.starts_with(Prefix: "pmaddw.d." ) || // Added in 7.0 |
| 298 | Name.starts_with(Prefix: "pmax" ) || // Added in 4.0 |
| 299 | Name.starts_with(Prefix: "pmin" ) || // Added in 4.0 |
| 300 | Name == "pmov.qd.256" || // Added in 9.0 |
| 301 | Name == "pmov.qd.512" || // Added in 9.0 |
| 302 | Name == "pmov.wb.256" || // Added in 9.0 |
| 303 | Name == "pmov.wb.512" || // Added in 9.0 |
| 304 | Name.starts_with(Prefix: "pmovsx" ) || // Added in 4.0 |
| 305 | Name.starts_with(Prefix: "pmovzx" ) || // Added in 4.0 |
| 306 | Name.starts_with(Prefix: "pmul.dq." ) || // Added in 4.0 |
| 307 | Name.starts_with(Prefix: "pmul.hr.sw." ) || // Added in 7.0 |
| 308 | Name.starts_with(Prefix: "pmulh.w." ) || // Added in 7.0 |
| 309 | Name.starts_with(Prefix: "pmulhu.w." ) || // Added in 7.0 |
| 310 | Name.starts_with(Prefix: "pmull." ) || // Added in 4.0 |
| 311 | Name.starts_with(Prefix: "pmultishift.qb." ) || // Added in 8.0 |
| 312 | Name.starts_with(Prefix: "pmulu.dq." ) || // Added in 4.0 |
| 313 | Name.starts_with(Prefix: "por." ) || // Added in 3.9 |
| 314 | Name.starts_with(Prefix: "prol." ) || // Added in 8.0 |
| 315 | Name.starts_with(Prefix: "prolv." ) || // Added in 8.0 |
| 316 | Name.starts_with(Prefix: "pror." ) || // Added in 8.0 |
| 317 | Name.starts_with(Prefix: "prorv." ) || // Added in 8.0 |
| 318 | Name.starts_with(Prefix: "pshuf.b." ) || // Added in 4.0 |
| 319 | Name.starts_with(Prefix: "pshuf.d." ) || // Added in 3.9 |
| 320 | Name.starts_with(Prefix: "pshufh.w." ) || // Added in 3.9 |
| 321 | Name.starts_with(Prefix: "pshufl.w." ) || // Added in 3.9 |
| 322 | Name.starts_with(Prefix: "psll.d" ) || // Added in 4.0 |
| 323 | Name.starts_with(Prefix: "psll.q" ) || // Added in 4.0 |
| 324 | Name.starts_with(Prefix: "psll.w" ) || // Added in 4.0 |
| 325 | Name.starts_with(Prefix: "pslli" ) || // Added in 4.0 |
| 326 | Name.starts_with(Prefix: "psllv" ) || // Added in 4.0 |
| 327 | Name.starts_with(Prefix: "psra.d" ) || // Added in 4.0 |
| 328 | Name.starts_with(Prefix: "psra.q" ) || // Added in 4.0 |
| 329 | Name.starts_with(Prefix: "psra.w" ) || // Added in 4.0 |
| 330 | Name.starts_with(Prefix: "psrai" ) || // Added in 4.0 |
| 331 | Name.starts_with(Prefix: "psrav" ) || // Added in 4.0 |
| 332 | Name.starts_with(Prefix: "psrl.d" ) || // Added in 4.0 |
| 333 | Name.starts_with(Prefix: "psrl.q" ) || // Added in 4.0 |
| 334 | Name.starts_with(Prefix: "psrl.w" ) || // Added in 4.0 |
| 335 | Name.starts_with(Prefix: "psrli" ) || // Added in 4.0 |
| 336 | Name.starts_with(Prefix: "psrlv" ) || // Added in 4.0 |
| 337 | Name.starts_with(Prefix: "psub." ) || // Added in 4.0 |
| 338 | Name.starts_with(Prefix: "psubs." ) || // Added in 8.0 |
| 339 | Name.starts_with(Prefix: "psubus." ) || // Added in 8.0 |
| 340 | Name.starts_with(Prefix: "pternlog." ) || // Added in 7.0 |
| 341 | Name.starts_with(Prefix: "punpckh" ) || // Added in 3.9 |
| 342 | Name.starts_with(Prefix: "punpckl" ) || // Added in 3.9 |
| 343 | Name.starts_with(Prefix: "pxor." ) || // Added in 3.9 |
| 344 | Name.starts_with(Prefix: "shuf.f" ) || // Added in 6.0 |
| 345 | Name.starts_with(Prefix: "shuf.i" ) || // Added in 6.0 |
| 346 | Name.starts_with(Prefix: "shuf.p" ) || // Added in 4.0 |
| 347 | Name.starts_with(Prefix: "sqrt.p" ) || // Added in 7.0 |
| 348 | Name.starts_with(Prefix: "store.b." ) || // Added in 3.9 |
| 349 | Name.starts_with(Prefix: "store.d." ) || // Added in 3.9 |
| 350 | Name.starts_with(Prefix: "store.p" ) || // Added in 3.9 |
| 351 | Name.starts_with(Prefix: "store.q." ) || // Added in 3.9 |
| 352 | Name.starts_with(Prefix: "store.w." ) || // Added in 3.9 |
| 353 | Name == "store.ss" || // Added in 7.0 |
| 354 | Name.starts_with(Prefix: "storeu." ) || // Added in 3.9 |
| 355 | Name.starts_with(Prefix: "sub.p" ) || // Added in 7.0. 128/256 in 4.0 |
| 356 | Name.starts_with(Prefix: "ucmp." ) || // Added in 5.0 |
| 357 | Name.starts_with(Prefix: "unpckh." ) || // Added in 3.9 |
| 358 | Name.starts_with(Prefix: "unpckl." ) || // Added in 3.9 |
| 359 | Name.starts_with(Prefix: "valign." ) || // Added in 4.0 |
| 360 | Name == "vcvtph2ps.128" || // Added in 11.0 |
| 361 | Name == "vcvtph2ps.256" || // Added in 11.0 |
| 362 | Name.starts_with(Prefix: "vextract" ) || // Added in 4.0 |
| 363 | Name.starts_with(Prefix: "vfmadd." ) || // Added in 7.0 |
| 364 | Name.starts_with(Prefix: "vfmaddsub." ) || // Added in 7.0 |
| 365 | Name.starts_with(Prefix: "vfnmadd." ) || // Added in 7.0 |
| 366 | Name.starts_with(Prefix: "vfnmsub." ) || // Added in 7.0 |
| 367 | Name.starts_with(Prefix: "vpdpbusd." ) || // Added in 7.0 |
| 368 | Name.starts_with(Prefix: "vpdpbusds." ) || // Added in 7.0 |
| 369 | Name.starts_with(Prefix: "vpdpwssd." ) || // Added in 7.0 |
| 370 | Name.starts_with(Prefix: "vpdpwssds." ) || // Added in 7.0 |
| 371 | Name.starts_with(Prefix: "vpermi2var." ) || // Added in 7.0 |
| 372 | Name.starts_with(Prefix: "vpermil.p" ) || // Added in 3.9 |
| 373 | Name.starts_with(Prefix: "vpermilvar." ) || // Added in 4.0 |
| 374 | Name.starts_with(Prefix: "vpermt2var." ) || // Added in 7.0 |
| 375 | Name.starts_with(Prefix: "vpmadd52" ) || // Added in 7.0 |
| 376 | Name.starts_with(Prefix: "vpshld." ) || // Added in 7.0 |
| 377 | Name.starts_with(Prefix: "vpshldv." ) || // Added in 8.0 |
| 378 | Name.starts_with(Prefix: "vpshrd." ) || // Added in 7.0 |
| 379 | Name.starts_with(Prefix: "vpshrdv." ) || // Added in 8.0 |
| 380 | Name.starts_with(Prefix: "vpshufbitqmb." ) || // Added in 8.0 |
| 381 | Name.starts_with(Prefix: "xor." )); // Added in 3.9 |
| 382 | |
| 383 | if (Name.consume_front(Prefix: "mask3." )) |
| 384 | // 'avx512.mask3.*' |
| 385 | return (Name.starts_with(Prefix: "vfmadd." ) || // Added in 7.0 |
| 386 | Name.starts_with(Prefix: "vfmaddsub." ) || // Added in 7.0 |
| 387 | Name.starts_with(Prefix: "vfmsub." ) || // Added in 7.0 |
| 388 | Name.starts_with(Prefix: "vfmsubadd." ) || // Added in 7.0 |
| 389 | Name.starts_with(Prefix: "vfnmsub." )); // Added in 7.0 |
| 390 | |
| 391 | if (Name.consume_front(Prefix: "maskz." )) |
| 392 | // 'avx512.maskz.*' |
| 393 | return (Name.starts_with(Prefix: "pternlog." ) || // Added in 7.0 |
| 394 | Name.starts_with(Prefix: "vfmadd." ) || // Added in 7.0 |
| 395 | Name.starts_with(Prefix: "vfmaddsub." ) || // Added in 7.0 |
| 396 | Name.starts_with(Prefix: "vpdpbusd." ) || // Added in 7.0 |
| 397 | Name.starts_with(Prefix: "vpdpbusds." ) || // Added in 7.0 |
| 398 | Name.starts_with(Prefix: "vpdpwssd." ) || // Added in 7.0 |
| 399 | Name.starts_with(Prefix: "vpdpwssds." ) || // Added in 7.0 |
| 400 | Name.starts_with(Prefix: "vpermt2var." ) || // Added in 7.0 |
| 401 | Name.starts_with(Prefix: "vpmadd52" ) || // Added in 7.0 |
| 402 | Name.starts_with(Prefix: "vpshldv." ) || // Added in 8.0 |
| 403 | Name.starts_with(Prefix: "vpshrdv." )); // Added in 8.0 |
| 404 | |
| 405 | // 'avx512.*' |
| 406 | return (Name == "movntdqa" || // Added in 5.0 |
| 407 | Name == "pmul.dq.512" || // Added in 7.0 |
| 408 | Name == "pmulu.dq.512" || // Added in 7.0 |
| 409 | Name.starts_with(Prefix: "broadcastm" ) || // Added in 6.0 |
| 410 | Name.starts_with(Prefix: "cmp.p" ) || // Added in 12.0 |
| 411 | Name.starts_with(Prefix: "cvtb2mask." ) || // Added in 7.0 |
| 412 | Name.starts_with(Prefix: "cvtd2mask." ) || // Added in 7.0 |
| 413 | Name.starts_with(Prefix: "cvtmask2" ) || // Added in 5.0 |
| 414 | Name.starts_with(Prefix: "cvtq2mask." ) || // Added in 7.0 |
| 415 | Name == "cvtusi2sd" || // Added in 7.0 |
| 416 | Name.starts_with(Prefix: "cvtw2mask." ) || // Added in 7.0 |
| 417 | Name == "kand.w" || // Added in 7.0 |
| 418 | Name == "kandn.w" || // Added in 7.0 |
| 419 | Name == "knot.w" || // Added in 7.0 |
| 420 | Name == "kor.w" || // Added in 7.0 |
| 421 | Name == "kortestc.w" || // Added in 7.0 |
| 422 | Name == "kortestz.w" || // Added in 7.0 |
| 423 | Name.starts_with(Prefix: "kunpck" ) || // added in 6.0 |
| 424 | Name == "kxnor.w" || // Added in 7.0 |
| 425 | Name == "kxor.w" || // Added in 7.0 |
| 426 | Name.starts_with(Prefix: "padds." ) || // Added in 8.0 |
| 427 | Name.starts_with(Prefix: "pbroadcast" ) || // Added in 3.9 |
| 428 | Name.starts_with(Prefix: "prol" ) || // Added in 8.0 |
| 429 | Name.starts_with(Prefix: "pror" ) || // Added in 8.0 |
| 430 | Name.starts_with(Prefix: "psll.dq" ) || // Added in 3.9 |
| 431 | Name.starts_with(Prefix: "psrl.dq" ) || // Added in 3.9 |
| 432 | Name.starts_with(Prefix: "psubs." ) || // Added in 8.0 |
| 433 | Name.starts_with(Prefix: "ptestm" ) || // Added in 6.0 |
| 434 | Name.starts_with(Prefix: "ptestnm" ) || // Added in 6.0 |
| 435 | Name.starts_with(Prefix: "storent." ) || // Added in 3.9 |
| 436 | Name.starts_with(Prefix: "vbroadcast.s" ) || // Added in 7.0 |
| 437 | Name.starts_with(Prefix: "vpshld." ) || // Added in 8.0 |
| 438 | Name.starts_with(Prefix: "vpshrd." )); // Added in 8.0 |
| 439 | } |
| 440 | |
| 441 | if (Name.consume_front(Prefix: "fma." )) |
| 442 | return (Name.starts_with(Prefix: "vfmadd." ) || // Added in 7.0 |
| 443 | Name.starts_with(Prefix: "vfmsub." ) || // Added in 7.0 |
| 444 | Name.starts_with(Prefix: "vfmsubadd." ) || // Added in 7.0 |
| 445 | Name.starts_with(Prefix: "vfnmadd." ) || // Added in 7.0 |
| 446 | Name.starts_with(Prefix: "vfnmsub." )); // Added in 7.0 |
| 447 | |
| 448 | if (Name.consume_front(Prefix: "fma4." )) |
| 449 | return Name.starts_with(Prefix: "vfmadd.s" ); // Added in 7.0 |
| 450 | |
| 451 | if (Name.consume_front(Prefix: "sse." )) |
| 452 | return (Name == "add.ss" || // Added in 4.0 |
| 453 | Name == "cvtsi2ss" || // Added in 7.0 |
| 454 | Name == "cvtsi642ss" || // Added in 7.0 |
| 455 | Name == "div.ss" || // Added in 4.0 |
| 456 | Name == "mul.ss" || // Added in 4.0 |
| 457 | Name.starts_with(Prefix: "sqrt.p" ) || // Added in 7.0 |
| 458 | Name == "sqrt.ss" || // Added in 7.0 |
| 459 | Name.starts_with(Prefix: "storeu." ) || // Added in 3.9 |
| 460 | Name == "sub.ss" ); // Added in 4.0 |
| 461 | |
| 462 | if (Name.consume_front(Prefix: "sse2." )) |
| 463 | return (Name == "add.sd" || // Added in 4.0 |
| 464 | Name == "cvtdq2pd" || // Added in 3.9 |
| 465 | Name == "cvtdq2ps" || // Added in 7.0 |
| 466 | Name == "cvtps2pd" || // Added in 3.9 |
| 467 | Name == "cvtsi2sd" || // Added in 7.0 |
| 468 | Name == "cvtsi642sd" || // Added in 7.0 |
| 469 | Name == "cvtss2sd" || // Added in 7.0 |
| 470 | Name == "div.sd" || // Added in 4.0 |
| 471 | Name == "mul.sd" || // Added in 4.0 |
| 472 | Name.starts_with(Prefix: "padds." ) || // Added in 8.0 |
| 473 | Name.starts_with(Prefix: "paddus." ) || // Added in 8.0 |
| 474 | Name.starts_with(Prefix: "pcmpeq." ) || // Added in 3.1 |
| 475 | Name.starts_with(Prefix: "pcmpgt." ) || // Added in 3.1 |
| 476 | Name == "pmaxs.w" || // Added in 3.9 |
| 477 | Name == "pmaxu.b" || // Added in 3.9 |
| 478 | Name == "pmins.w" || // Added in 3.9 |
| 479 | Name == "pminu.b" || // Added in 3.9 |
| 480 | Name == "pmulu.dq" || // Added in 7.0 |
| 481 | Name.starts_with(Prefix: "pshuf" ) || // Added in 3.9 |
| 482 | Name.starts_with(Prefix: "psll.dq" ) || // Added in 3.7 |
| 483 | Name.starts_with(Prefix: "psrl.dq" ) || // Added in 3.7 |
| 484 | Name.starts_with(Prefix: "psubs." ) || // Added in 8.0 |
| 485 | Name.starts_with(Prefix: "psubus." ) || // Added in 8.0 |
| 486 | Name.starts_with(Prefix: "sqrt.p" ) || // Added in 7.0 |
| 487 | Name == "sqrt.sd" || // Added in 7.0 |
| 488 | Name == "storel.dq" || // Added in 3.9 |
| 489 | Name.starts_with(Prefix: "storeu." ) || // Added in 3.9 |
| 490 | Name == "sub.sd" ); // Added in 4.0 |
| 491 | |
| 492 | if (Name.consume_front(Prefix: "sse41." )) |
| 493 | return (Name.starts_with(Prefix: "blendp" ) || // Added in 3.7 |
| 494 | Name == "movntdqa" || // Added in 5.0 |
| 495 | Name == "pblendw" || // Added in 3.7 |
| 496 | Name == "pmaxsb" || // Added in 3.9 |
| 497 | Name == "pmaxsd" || // Added in 3.9 |
| 498 | Name == "pmaxud" || // Added in 3.9 |
| 499 | Name == "pmaxuw" || // Added in 3.9 |
| 500 | Name == "pminsb" || // Added in 3.9 |
| 501 | Name == "pminsd" || // Added in 3.9 |
| 502 | Name == "pminud" || // Added in 3.9 |
| 503 | Name == "pminuw" || // Added in 3.9 |
| 504 | Name.starts_with(Prefix: "pmovsx" ) || // Added in 3.8 |
| 505 | Name.starts_with(Prefix: "pmovzx" ) || // Added in 3.9 |
| 506 | Name == "pmuldq" ); // Added in 7.0 |
| 507 | |
| 508 | if (Name.consume_front(Prefix: "sse42." )) |
| 509 | return Name == "crc32.64.8" ; // Added in 3.4 |
| 510 | |
| 511 | if (Name.consume_front(Prefix: "sse4a." )) |
| 512 | return Name.starts_with(Prefix: "movnt." ); // Added in 3.9 |
| 513 | |
| 514 | if (Name.consume_front(Prefix: "ssse3." )) |
| 515 | return (Name == "pabs.b.128" || // Added in 6.0 |
| 516 | Name == "pabs.d.128" || // Added in 6.0 |
| 517 | Name == "pabs.w.128" ); // Added in 6.0 |
| 518 | |
| 519 | if (Name.consume_front(Prefix: "xop." )) |
| 520 | return (Name == "vpcmov" || // Added in 3.8 |
| 521 | Name == "vpcmov.256" || // Added in 5.0 |
| 522 | Name.starts_with(Prefix: "vpcom" ) || // Added in 3.2, Updated in 9.0 |
| 523 | Name.starts_with(Prefix: "vprot" )); // Added in 8.0 |
| 524 | |
| 525 | return (Name == "addcarry.u32" || // Added in 8.0 |
| 526 | Name == "addcarry.u64" || // Added in 8.0 |
| 527 | Name == "addcarryx.u32" || // Added in 8.0 |
| 528 | Name == "addcarryx.u64" || // Added in 8.0 |
| 529 | Name == "subborrow.u32" || // Added in 8.0 |
| 530 | Name == "subborrow.u64" || // Added in 8.0 |
| 531 | Name.starts_with(Prefix: "vcvtph2ps." )); // Added in 11.0 |
| 532 | } |
| 533 | |
| 534 | static bool upgradeX86IntrinsicFunction(Function *F, StringRef Name, |
| 535 | Function *&NewFn) { |
| 536 | // Only handle intrinsics that start with "x86.". |
| 537 | if (!Name.consume_front(Prefix: "x86." )) |
| 538 | return false; |
| 539 | |
| 540 | if (shouldUpgradeX86Intrinsic(F, Name)) { |
| 541 | NewFn = nullptr; |
| 542 | return true; |
| 543 | } |
| 544 | |
| 545 | if (Name == "rdtscp" ) { // Added in 8.0 |
| 546 | // If this intrinsic has 0 operands, it's the new version. |
| 547 | if (F->getFunctionType()->getNumParams() == 0) |
| 548 | return false; |
| 549 | |
| 550 | rename(GV: F); |
| 551 | NewFn = Intrinsic::getOrInsertDeclaration(M: F->getParent(), |
| 552 | id: Intrinsic::x86_rdtscp); |
| 553 | return true; |
| 554 | } |
| 555 | |
| 556 | Intrinsic::ID ID; |
| 557 | |
| 558 | // SSE4.1 ptest functions may have an old signature. |
| 559 | if (Name.consume_front(Prefix: "sse41.ptest" )) { // Added in 3.2 |
| 560 | ID = StringSwitch<Intrinsic::ID>(Name) |
| 561 | .Case(S: "c" , Value: Intrinsic::x86_sse41_ptestc) |
| 562 | .Case(S: "z" , Value: Intrinsic::x86_sse41_ptestz) |
| 563 | .Case(S: "nzc" , Value: Intrinsic::x86_sse41_ptestnzc) |
| 564 | .Default(Value: Intrinsic::not_intrinsic); |
| 565 | if (ID != Intrinsic::not_intrinsic) |
| 566 | return upgradePTESTIntrinsic(F, IID: ID, NewFn); |
| 567 | |
| 568 | return false; |
| 569 | } |
| 570 | |
| 571 | // Several blend and other instructions with masks used the wrong number of |
| 572 | // bits. |
| 573 | |
| 574 | // Added in 3.6 |
| 575 | ID = StringSwitch<Intrinsic::ID>(Name) |
| 576 | .Case(S: "sse41.insertps" , Value: Intrinsic::x86_sse41_insertps) |
| 577 | .Case(S: "sse41.dppd" , Value: Intrinsic::x86_sse41_dppd) |
| 578 | .Case(S: "sse41.dpps" , Value: Intrinsic::x86_sse41_dpps) |
| 579 | .Case(S: "sse41.mpsadbw" , Value: Intrinsic::x86_sse41_mpsadbw) |
| 580 | .Case(S: "avx.dp.ps.256" , Value: Intrinsic::x86_avx_dp_ps_256) |
| 581 | .Case(S: "avx2.mpsadbw" , Value: Intrinsic::x86_avx2_mpsadbw) |
| 582 | .Default(Value: Intrinsic::not_intrinsic); |
| 583 | if (ID != Intrinsic::not_intrinsic) |
| 584 | return upgradeX86IntrinsicsWith8BitMask(F, IID: ID, NewFn); |
| 585 | |
| 586 | if (Name.consume_front(Prefix: "avx512." )) { |
| 587 | if (Name.consume_front(Prefix: "mask.cmp." )) { |
| 588 | // Added in 7.0 |
| 589 | ID = StringSwitch<Intrinsic::ID>(Name) |
| 590 | .Case(S: "pd.128" , Value: Intrinsic::x86_avx512_mask_cmp_pd_128) |
| 591 | .Case(S: "pd.256" , Value: Intrinsic::x86_avx512_mask_cmp_pd_256) |
| 592 | .Case(S: "pd.512" , Value: Intrinsic::x86_avx512_mask_cmp_pd_512) |
| 593 | .Case(S: "ps.128" , Value: Intrinsic::x86_avx512_mask_cmp_ps_128) |
| 594 | .Case(S: "ps.256" , Value: Intrinsic::x86_avx512_mask_cmp_ps_256) |
| 595 | .Case(S: "ps.512" , Value: Intrinsic::x86_avx512_mask_cmp_ps_512) |
| 596 | .Default(Value: Intrinsic::not_intrinsic); |
| 597 | if (ID != Intrinsic::not_intrinsic) |
| 598 | return upgradeX86MaskedFPCompare(F, IID: ID, NewFn); |
| 599 | } else if (Name.starts_with(Prefix: "vpdpbusd." ) || |
| 600 | Name.starts_with(Prefix: "vpdpbusds." )) { |
| 601 | // Added in 21.1 |
| 602 | ID = StringSwitch<Intrinsic::ID>(Name) |
| 603 | .Case(S: "vpdpbusd.128" , Value: Intrinsic::x86_avx512_vpdpbusd_128) |
| 604 | .Case(S: "vpdpbusd.256" , Value: Intrinsic::x86_avx512_vpdpbusd_256) |
| 605 | .Case(S: "vpdpbusd.512" , Value: Intrinsic::x86_avx512_vpdpbusd_512) |
| 606 | .Case(S: "vpdpbusds.128" , Value: Intrinsic::x86_avx512_vpdpbusds_128) |
| 607 | .Case(S: "vpdpbusds.256" , Value: Intrinsic::x86_avx512_vpdpbusds_256) |
| 608 | .Case(S: "vpdpbusds.512" , Value: Intrinsic::x86_avx512_vpdpbusds_512) |
| 609 | .Default(Value: Intrinsic::not_intrinsic); |
| 610 | if (ID != Intrinsic::not_intrinsic) |
| 611 | return upgradeX86MultiplyAddBytes(F, IID: ID, NewFn); |
| 612 | } else if (Name.starts_with(Prefix: "vpdpwssd." ) || |
| 613 | Name.starts_with(Prefix: "vpdpwssds." )) { |
| 614 | // Added in 21.1 |
| 615 | ID = StringSwitch<Intrinsic::ID>(Name) |
| 616 | .Case(S: "vpdpwssd.128" , Value: Intrinsic::x86_avx512_vpdpwssd_128) |
| 617 | .Case(S: "vpdpwssd.256" , Value: Intrinsic::x86_avx512_vpdpwssd_256) |
| 618 | .Case(S: "vpdpwssd.512" , Value: Intrinsic::x86_avx512_vpdpwssd_512) |
| 619 | .Case(S: "vpdpwssds.128" , Value: Intrinsic::x86_avx512_vpdpwssds_128) |
| 620 | .Case(S: "vpdpwssds.256" , Value: Intrinsic::x86_avx512_vpdpwssds_256) |
| 621 | .Case(S: "vpdpwssds.512" , Value: Intrinsic::x86_avx512_vpdpwssds_512) |
| 622 | .Default(Value: Intrinsic::not_intrinsic); |
| 623 | if (ID != Intrinsic::not_intrinsic) |
| 624 | return upgradeX86MultiplyAddWords(F, IID: ID, NewFn); |
| 625 | } |
| 626 | return false; // No other 'x86.avx512.*'. |
| 627 | } |
| 628 | |
| 629 | if (Name.consume_front(Prefix: "avx2." )) { |
| 630 | if (Name.consume_front(Prefix: "vpdpb" )) { |
| 631 | // Added in 21.1 |
| 632 | ID = StringSwitch<Intrinsic::ID>(Name) |
| 633 | .Case(S: "ssd.128" , Value: Intrinsic::x86_avx2_vpdpbssd_128) |
| 634 | .Case(S: "ssd.256" , Value: Intrinsic::x86_avx2_vpdpbssd_256) |
| 635 | .Case(S: "ssds.128" , Value: Intrinsic::x86_avx2_vpdpbssds_128) |
| 636 | .Case(S: "ssds.256" , Value: Intrinsic::x86_avx2_vpdpbssds_256) |
| 637 | .Case(S: "sud.128" , Value: Intrinsic::x86_avx2_vpdpbsud_128) |
| 638 | .Case(S: "sud.256" , Value: Intrinsic::x86_avx2_vpdpbsud_256) |
| 639 | .Case(S: "suds.128" , Value: Intrinsic::x86_avx2_vpdpbsuds_128) |
| 640 | .Case(S: "suds.256" , Value: Intrinsic::x86_avx2_vpdpbsuds_256) |
| 641 | .Case(S: "uud.128" , Value: Intrinsic::x86_avx2_vpdpbuud_128) |
| 642 | .Case(S: "uud.256" , Value: Intrinsic::x86_avx2_vpdpbuud_256) |
| 643 | .Case(S: "uuds.128" , Value: Intrinsic::x86_avx2_vpdpbuuds_128) |
| 644 | .Case(S: "uuds.256" , Value: Intrinsic::x86_avx2_vpdpbuuds_256) |
| 645 | .Default(Value: Intrinsic::not_intrinsic); |
| 646 | if (ID != Intrinsic::not_intrinsic) |
| 647 | return upgradeX86MultiplyAddBytes(F, IID: ID, NewFn); |
| 648 | } else if (Name.consume_front(Prefix: "vpdpw" )) { |
| 649 | // Added in 21.1 |
| 650 | ID = StringSwitch<Intrinsic::ID>(Name) |
| 651 | .Case(S: "sud.128" , Value: Intrinsic::x86_avx2_vpdpwsud_128) |
| 652 | .Case(S: "sud.256" , Value: Intrinsic::x86_avx2_vpdpwsud_256) |
| 653 | .Case(S: "suds.128" , Value: Intrinsic::x86_avx2_vpdpwsuds_128) |
| 654 | .Case(S: "suds.256" , Value: Intrinsic::x86_avx2_vpdpwsuds_256) |
| 655 | .Case(S: "usd.128" , Value: Intrinsic::x86_avx2_vpdpwusd_128) |
| 656 | .Case(S: "usd.256" , Value: Intrinsic::x86_avx2_vpdpwusd_256) |
| 657 | .Case(S: "usds.128" , Value: Intrinsic::x86_avx2_vpdpwusds_128) |
| 658 | .Case(S: "usds.256" , Value: Intrinsic::x86_avx2_vpdpwusds_256) |
| 659 | .Case(S: "uud.128" , Value: Intrinsic::x86_avx2_vpdpwuud_128) |
| 660 | .Case(S: "uud.256" , Value: Intrinsic::x86_avx2_vpdpwuud_256) |
| 661 | .Case(S: "uuds.128" , Value: Intrinsic::x86_avx2_vpdpwuuds_128) |
| 662 | .Case(S: "uuds.256" , Value: Intrinsic::x86_avx2_vpdpwuuds_256) |
| 663 | .Default(Value: Intrinsic::not_intrinsic); |
| 664 | if (ID != Intrinsic::not_intrinsic) |
| 665 | return upgradeX86MultiplyAddWords(F, IID: ID, NewFn); |
| 666 | } |
| 667 | return false; // No other 'x86.avx2.*' |
| 668 | } |
| 669 | |
| 670 | if (Name.consume_front(Prefix: "avx10." )) { |
| 671 | if (Name.consume_front(Prefix: "vpdpb" )) { |
| 672 | // Added in 21.1 |
| 673 | ID = StringSwitch<Intrinsic::ID>(Name) |
| 674 | .Case(S: "ssd.512" , Value: Intrinsic::x86_avx10_vpdpbssd_512) |
| 675 | .Case(S: "ssds.512" , Value: Intrinsic::x86_avx10_vpdpbssds_512) |
| 676 | .Case(S: "sud.512" , Value: Intrinsic::x86_avx10_vpdpbsud_512) |
| 677 | .Case(S: "suds.512" , Value: Intrinsic::x86_avx10_vpdpbsuds_512) |
| 678 | .Case(S: "uud.512" , Value: Intrinsic::x86_avx10_vpdpbuud_512) |
| 679 | .Case(S: "uuds.512" , Value: Intrinsic::x86_avx10_vpdpbuuds_512) |
| 680 | .Default(Value: Intrinsic::not_intrinsic); |
| 681 | if (ID != Intrinsic::not_intrinsic) |
| 682 | return upgradeX86MultiplyAddBytes(F, IID: ID, NewFn); |
| 683 | } else if (Name.consume_front(Prefix: "vpdpw" )) { |
| 684 | ID = StringSwitch<Intrinsic::ID>(Name) |
| 685 | .Case(S: "sud.512" , Value: Intrinsic::x86_avx10_vpdpwsud_512) |
| 686 | .Case(S: "suds.512" , Value: Intrinsic::x86_avx10_vpdpwsuds_512) |
| 687 | .Case(S: "usd.512" , Value: Intrinsic::x86_avx10_vpdpwusd_512) |
| 688 | .Case(S: "usds.512" , Value: Intrinsic::x86_avx10_vpdpwusds_512) |
| 689 | .Case(S: "uud.512" , Value: Intrinsic::x86_avx10_vpdpwuud_512) |
| 690 | .Case(S: "uuds.512" , Value: Intrinsic::x86_avx10_vpdpwuuds_512) |
| 691 | .Default(Value: Intrinsic::not_intrinsic); |
| 692 | if (ID != Intrinsic::not_intrinsic) |
| 693 | return upgradeX86MultiplyAddWords(F, IID: ID, NewFn); |
| 694 | } |
| 695 | return false; // No other 'x86.avx10.*' |
| 696 | } |
| 697 | |
| 698 | if (Name.consume_front(Prefix: "avx512bf16." )) { |
| 699 | // Added in 9.0 |
| 700 | ID = StringSwitch<Intrinsic::ID>(Name) |
| 701 | .Case(S: "cvtne2ps2bf16.128" , |
| 702 | Value: Intrinsic::x86_avx512bf16_cvtne2ps2bf16_128) |
| 703 | .Case(S: "cvtne2ps2bf16.256" , |
| 704 | Value: Intrinsic::x86_avx512bf16_cvtne2ps2bf16_256) |
| 705 | .Case(S: "cvtne2ps2bf16.512" , |
| 706 | Value: Intrinsic::x86_avx512bf16_cvtne2ps2bf16_512) |
| 707 | .Case(S: "mask.cvtneps2bf16.128" , |
| 708 | Value: Intrinsic::x86_avx512bf16_mask_cvtneps2bf16_128) |
| 709 | .Case(S: "cvtneps2bf16.256" , |
| 710 | Value: Intrinsic::x86_avx512bf16_cvtneps2bf16_256) |
| 711 | .Case(S: "cvtneps2bf16.512" , |
| 712 | Value: Intrinsic::x86_avx512bf16_cvtneps2bf16_512) |
| 713 | .Default(Value: Intrinsic::not_intrinsic); |
| 714 | if (ID != Intrinsic::not_intrinsic) |
| 715 | return upgradeX86BF16Intrinsic(F, IID: ID, NewFn); |
| 716 | |
| 717 | // Added in 9.0 |
| 718 | ID = StringSwitch<Intrinsic::ID>(Name) |
| 719 | .Case(S: "dpbf16ps.128" , Value: Intrinsic::x86_avx512bf16_dpbf16ps_128) |
| 720 | .Case(S: "dpbf16ps.256" , Value: Intrinsic::x86_avx512bf16_dpbf16ps_256) |
| 721 | .Case(S: "dpbf16ps.512" , Value: Intrinsic::x86_avx512bf16_dpbf16ps_512) |
| 722 | .Default(Value: Intrinsic::not_intrinsic); |
| 723 | if (ID != Intrinsic::not_intrinsic) |
| 724 | return upgradeX86BF16DPIntrinsic(F, IID: ID, NewFn); |
| 725 | return false; // No other 'x86.avx512bf16.*'. |
| 726 | } |
| 727 | |
| 728 | if (Name.consume_front(Prefix: "xop." )) { |
| 729 | Intrinsic::ID ID = Intrinsic::not_intrinsic; |
| 730 | if (Name.starts_with(Prefix: "vpermil2" )) { // Added in 3.9 |
| 731 | // Upgrade any XOP PERMIL2 index operand still using a float/double |
| 732 | // vector. |
| 733 | auto Idx = F->getFunctionType()->getParamType(i: 2); |
| 734 | if (Idx->isFPOrFPVectorTy()) { |
| 735 | unsigned IdxSize = Idx->getPrimitiveSizeInBits(); |
| 736 | unsigned EltSize = Idx->getScalarSizeInBits(); |
| 737 | if (EltSize == 64 && IdxSize == 128) |
| 738 | ID = Intrinsic::x86_xop_vpermil2pd; |
| 739 | else if (EltSize == 32 && IdxSize == 128) |
| 740 | ID = Intrinsic::x86_xop_vpermil2ps; |
| 741 | else if (EltSize == 64 && IdxSize == 256) |
| 742 | ID = Intrinsic::x86_xop_vpermil2pd_256; |
| 743 | else |
| 744 | ID = Intrinsic::x86_xop_vpermil2ps_256; |
| 745 | } |
| 746 | } else if (F->arg_size() == 2) |
| 747 | // frcz.ss/sd may need to have an argument dropped. Added in 3.2 |
| 748 | ID = StringSwitch<Intrinsic::ID>(Name) |
| 749 | .Case(S: "vfrcz.ss" , Value: Intrinsic::x86_xop_vfrcz_ss) |
| 750 | .Case(S: "vfrcz.sd" , Value: Intrinsic::x86_xop_vfrcz_sd) |
| 751 | .Default(Value: Intrinsic::not_intrinsic); |
| 752 | |
| 753 | if (ID != Intrinsic::not_intrinsic) { |
| 754 | rename(GV: F); |
| 755 | NewFn = Intrinsic::getOrInsertDeclaration(M: F->getParent(), id: ID); |
| 756 | return true; |
| 757 | } |
| 758 | return false; // No other 'x86.xop.*' |
| 759 | } |
| 760 | |
| 761 | if (Name == "seh.recoverfp" ) { |
| 762 | NewFn = Intrinsic::getOrInsertDeclaration(M: F->getParent(), |
| 763 | id: Intrinsic::eh_recoverfp); |
| 764 | return true; |
| 765 | } |
| 766 | |
| 767 | return false; |
| 768 | } |
| 769 | |
| 770 | // Upgrade ARM (IsArm) or Aarch64 (!IsArm) intrinsic fns. Return true iff so. |
| 771 | // IsArm: 'arm.*', !IsArm: 'aarch64.*'. |
| 772 | static bool upgradeArmOrAarch64IntrinsicFunction(bool IsArm, Function *F, |
| 773 | StringRef Name, |
| 774 | Function *&NewFn) { |
| 775 | if (Name.starts_with(Prefix: "rbit" )) { |
| 776 | // '(arm|aarch64).rbit'. |
| 777 | NewFn = Intrinsic::getOrInsertDeclaration( |
| 778 | M: F->getParent(), id: Intrinsic::bitreverse, Tys: F->arg_begin()->getType()); |
| 779 | return true; |
| 780 | } |
| 781 | |
| 782 | if (Name == "thread.pointer" ) { |
| 783 | // '(arm|aarch64).thread.pointer'. |
| 784 | NewFn = Intrinsic::getOrInsertDeclaration( |
| 785 | M: F->getParent(), id: Intrinsic::thread_pointer, Tys: F->getReturnType()); |
| 786 | return true; |
| 787 | } |
| 788 | |
| 789 | bool Neon = Name.consume_front(Prefix: "neon." ); |
| 790 | if (Neon) { |
| 791 | // '(arm|aarch64).neon.*'. |
| 792 | // Changed in 12.0: bfdot accept v4bf16 and v8bf16 instead of v8i8 and |
| 793 | // v16i8 respectively. |
| 794 | if (Name.consume_front(Prefix: "bfdot." )) { |
| 795 | // (arm|aarch64).neon.bfdot.*'. |
| 796 | Intrinsic::ID ID = |
| 797 | StringSwitch<Intrinsic::ID>(Name) |
| 798 | .Cases(CaseStrings: {"v2f32.v8i8" , "v4f32.v16i8" }, |
| 799 | Value: IsArm ? (Intrinsic::ID)Intrinsic::arm_neon_bfdot |
| 800 | : (Intrinsic::ID)Intrinsic::aarch64_neon_bfdot) |
| 801 | .Default(Value: Intrinsic::not_intrinsic); |
| 802 | if (ID != Intrinsic::not_intrinsic) { |
| 803 | size_t OperandWidth = F->getReturnType()->getPrimitiveSizeInBits(); |
| 804 | assert((OperandWidth == 64 || OperandWidth == 128) && |
| 805 | "Unexpected operand width" ); |
| 806 | LLVMContext &Ctx = F->getParent()->getContext(); |
| 807 | std::array<Type *, 2> Tys{ |
| 808 | ._M_elems: {F->getReturnType(), |
| 809 | FixedVectorType::get(ElementType: Type::getBFloatTy(C&: Ctx), NumElts: OperandWidth / 16)}}; |
| 810 | NewFn = Intrinsic::getOrInsertDeclaration(M: F->getParent(), id: ID, Tys); |
| 811 | return true; |
| 812 | } |
| 813 | return false; // No other '(arm|aarch64).neon.bfdot.*'. |
| 814 | } |
| 815 | |
| 816 | // Changed in 12.0: bfmmla, bfmlalb and bfmlalt are not polymorphic |
| 817 | // anymore and accept v8bf16 instead of v16i8. |
| 818 | if (Name.consume_front(Prefix: "bfm" )) { |
| 819 | // (arm|aarch64).neon.bfm*'. |
| 820 | if (Name.consume_back(Suffix: ".v4f32.v16i8" )) { |
| 821 | // (arm|aarch64).neon.bfm*.v4f32.v16i8'. |
| 822 | Intrinsic::ID ID = |
| 823 | StringSwitch<Intrinsic::ID>(Name) |
| 824 | .Case(S: "mla" , |
| 825 | Value: IsArm ? (Intrinsic::ID)Intrinsic::arm_neon_bfmmla |
| 826 | : (Intrinsic::ID)Intrinsic::aarch64_neon_bfmmla) |
| 827 | .Case(S: "lalb" , |
| 828 | Value: IsArm ? (Intrinsic::ID)Intrinsic::arm_neon_bfmlalb |
| 829 | : (Intrinsic::ID)Intrinsic::aarch64_neon_bfmlalb) |
| 830 | .Case(S: "lalt" , |
| 831 | Value: IsArm ? (Intrinsic::ID)Intrinsic::arm_neon_bfmlalt |
| 832 | : (Intrinsic::ID)Intrinsic::aarch64_neon_bfmlalt) |
| 833 | .Default(Value: Intrinsic::not_intrinsic); |
| 834 | if (ID != Intrinsic::not_intrinsic) { |
| 835 | NewFn = Intrinsic::getOrInsertDeclaration(M: F->getParent(), id: ID); |
| 836 | return true; |
| 837 | } |
| 838 | return false; // No other '(arm|aarch64).neon.bfm*.v16i8'. |
| 839 | } |
| 840 | return false; // No other '(arm|aarch64).neon.bfm*. |
| 841 | } |
| 842 | // Continue on to Aarch64 Neon or Arm Neon. |
| 843 | } |
| 844 | // Continue on to Arm or Aarch64. |
| 845 | |
| 846 | if (IsArm) { |
| 847 | // 'arm.*'. |
| 848 | if (Neon) { |
| 849 | // 'arm.neon.*'. |
| 850 | Intrinsic::ID ID = StringSwitch<Intrinsic::ID>(Name) |
| 851 | .StartsWith(S: "vclz." , Value: Intrinsic::ctlz) |
| 852 | .StartsWith(S: "vcnt." , Value: Intrinsic::ctpop) |
| 853 | .StartsWith(S: "vqadds." , Value: Intrinsic::sadd_sat) |
| 854 | .StartsWith(S: "vqaddu." , Value: Intrinsic::uadd_sat) |
| 855 | .StartsWith(S: "vqsubs." , Value: Intrinsic::ssub_sat) |
| 856 | .StartsWith(S: "vqsubu." , Value: Intrinsic::usub_sat) |
| 857 | .StartsWith(S: "vrinta." , Value: Intrinsic::round) |
| 858 | .StartsWith(S: "vrintn." , Value: Intrinsic::roundeven) |
| 859 | .StartsWith(S: "vrintm." , Value: Intrinsic::floor) |
| 860 | .StartsWith(S: "vrintp." , Value: Intrinsic::ceil) |
| 861 | .StartsWith(S: "vrintx." , Value: Intrinsic::rint) |
| 862 | .StartsWith(S: "vrintz." , Value: Intrinsic::trunc) |
| 863 | .Default(Value: Intrinsic::not_intrinsic); |
| 864 | if (ID != Intrinsic::not_intrinsic) { |
| 865 | NewFn = Intrinsic::getOrInsertDeclaration(M: F->getParent(), id: ID, |
| 866 | Tys: F->arg_begin()->getType()); |
| 867 | return true; |
| 868 | } |
| 869 | |
| 870 | if (Name.consume_front(Prefix: "vst" )) { |
| 871 | // 'arm.neon.vst*'. |
| 872 | static const Regex vstRegex("^([1234]|[234]lane)\\.v[a-z0-9]*$" ); |
| 873 | SmallVector<StringRef, 2> Groups; |
| 874 | if (vstRegex.match(String: Name, Matches: &Groups)) { |
| 875 | static const Intrinsic::ID StoreInts[] = { |
| 876 | Intrinsic::arm_neon_vst1, Intrinsic::arm_neon_vst2, |
| 877 | Intrinsic::arm_neon_vst3, Intrinsic::arm_neon_vst4}; |
| 878 | |
| 879 | static const Intrinsic::ID StoreLaneInts[] = { |
| 880 | Intrinsic::arm_neon_vst2lane, Intrinsic::arm_neon_vst3lane, |
| 881 | Intrinsic::arm_neon_vst4lane}; |
| 882 | |
| 883 | auto fArgs = F->getFunctionType()->params(); |
| 884 | Type *Tys[] = {fArgs[0], fArgs[1]}; |
| 885 | if (Groups[1].size() == 1) |
| 886 | NewFn = Intrinsic::getOrInsertDeclaration( |
| 887 | M: F->getParent(), id: StoreInts[fArgs.size() - 3], Tys); |
| 888 | else |
| 889 | NewFn = Intrinsic::getOrInsertDeclaration( |
| 890 | M: F->getParent(), id: StoreLaneInts[fArgs.size() - 5], Tys); |
| 891 | return true; |
| 892 | } |
| 893 | return false; // No other 'arm.neon.vst*'. |
| 894 | } |
| 895 | |
| 896 | return false; // No other 'arm.neon.*'. |
| 897 | } |
| 898 | |
| 899 | if (Name.consume_front(Prefix: "mve." )) { |
| 900 | // 'arm.mve.*'. |
| 901 | if (Name == "vctp64" ) { |
| 902 | if (cast<FixedVectorType>(Val: F->getReturnType())->getNumElements() == 4) { |
| 903 | // A vctp64 returning a v4i1 is converted to return a v2i1. Rename |
| 904 | // the function and deal with it below in UpgradeIntrinsicCall. |
| 905 | rename(GV: F); |
| 906 | return true; |
| 907 | } |
| 908 | return false; // Not 'arm.mve.vctp64'. |
| 909 | } |
| 910 | |
| 911 | if (Name.starts_with(Prefix: "vrintn.v" )) { |
| 912 | NewFn = Intrinsic::getOrInsertDeclaration( |
| 913 | M: F->getParent(), id: Intrinsic::roundeven, Tys: F->arg_begin()->getType()); |
| 914 | return true; |
| 915 | } |
| 916 | |
| 917 | // These too are changed to accept a v2i1 instead of the old v4i1. |
| 918 | if (Name.consume_back(Suffix: ".v4i1" )) { |
| 919 | // 'arm.mve.*.v4i1'. |
| 920 | if (Name.consume_back(Suffix: ".predicated.v2i64.v4i32" )) |
| 921 | // 'arm.mve.*.predicated.v2i64.v4i32.v4i1' |
| 922 | return Name == "mull.int" || Name == "vqdmull" ; |
| 923 | |
| 924 | if (Name.consume_back(Suffix: ".v2i64" )) { |
| 925 | // 'arm.mve.*.v2i64.v4i1' |
| 926 | bool IsGather = Name.consume_front(Prefix: "vldr.gather." ); |
| 927 | if (IsGather || Name.consume_front(Prefix: "vstr.scatter." )) { |
| 928 | if (Name.consume_front(Prefix: "base." )) { |
| 929 | // Optional 'wb.' prefix. |
| 930 | Name.consume_front(Prefix: "wb." ); |
| 931 | // 'arm.mve.(vldr.gather|vstr.scatter).base.(wb.)? |
| 932 | // predicated.v2i64.v2i64.v4i1'. |
| 933 | return Name == "predicated.v2i64" ; |
| 934 | } |
| 935 | |
| 936 | if (Name.consume_front(Prefix: "offset.predicated." )) |
| 937 | return Name == (IsGather ? "v2i64.p0i64" : "p0i64.v2i64" ) || |
| 938 | Name == (IsGather ? "v2i64.p0" : "p0.v2i64" ); |
| 939 | |
| 940 | // No other 'arm.mve.(vldr.gather|vstr.scatter).*.v2i64.v4i1'. |
| 941 | return false; |
| 942 | } |
| 943 | |
| 944 | return false; // No other 'arm.mve.*.v2i64.v4i1'. |
| 945 | } |
| 946 | return false; // No other 'arm.mve.*.v4i1'. |
| 947 | } |
| 948 | return false; // No other 'arm.mve.*'. |
| 949 | } |
| 950 | |
| 951 | if (Name.consume_front(Prefix: "cde.vcx" )) { |
| 952 | // 'arm.cde.vcx*'. |
| 953 | if (Name.consume_back(Suffix: ".predicated.v2i64.v4i1" )) |
| 954 | // 'arm.cde.vcx*.predicated.v2i64.v4i1'. |
| 955 | return Name == "1q" || Name == "1qa" || Name == "2q" || Name == "2qa" || |
| 956 | Name == "3q" || Name == "3qa" ; |
| 957 | |
| 958 | return false; // No other 'arm.cde.vcx*'. |
| 959 | } |
| 960 | } else { |
| 961 | // 'aarch64.*'. |
| 962 | if (Neon) { |
| 963 | // 'aarch64.neon.*'. |
| 964 | Intrinsic::ID ID = StringSwitch<Intrinsic::ID>(Name) |
| 965 | .StartsWith(S: "frintn" , Value: Intrinsic::roundeven) |
| 966 | .StartsWith(S: "rbit" , Value: Intrinsic::bitreverse) |
| 967 | .Default(Value: Intrinsic::not_intrinsic); |
| 968 | if (ID != Intrinsic::not_intrinsic) { |
| 969 | NewFn = Intrinsic::getOrInsertDeclaration(M: F->getParent(), id: ID, |
| 970 | Tys: F->arg_begin()->getType()); |
| 971 | return true; |
| 972 | } |
| 973 | |
| 974 | if (Name.starts_with(Prefix: "addp" )) { |
| 975 | // 'aarch64.neon.addp*'. |
| 976 | if (F->arg_size() != 2) |
| 977 | return false; // Invalid IR. |
| 978 | VectorType *Ty = dyn_cast<VectorType>(Val: F->getReturnType()); |
| 979 | if (Ty && Ty->getElementType()->isFloatingPointTy()) { |
| 980 | NewFn = Intrinsic::getOrInsertDeclaration( |
| 981 | M: F->getParent(), id: Intrinsic::aarch64_neon_faddp, Tys: Ty); |
| 982 | return true; |
| 983 | } |
| 984 | } |
| 985 | |
| 986 | // Changed in 20.0: bfcvt/bfcvtn/bcvtn2 have been replaced with fptrunc. |
| 987 | if (Name.starts_with(Prefix: "bfcvt" )) { |
| 988 | NewFn = nullptr; |
| 989 | return true; |
| 990 | } |
| 991 | |
| 992 | return false; // No other 'aarch64.neon.*'. |
| 993 | } |
| 994 | if (Name.consume_front(Prefix: "sve." )) { |
| 995 | // 'aarch64.sve.*'. |
| 996 | if (Name.consume_front(Prefix: "bf" )) { |
| 997 | if (Name.consume_back(Suffix: ".lane" )) { |
| 998 | // 'aarch64.sve.bf*.lane'. |
| 999 | Intrinsic::ID ID = |
| 1000 | StringSwitch<Intrinsic::ID>(Name) |
| 1001 | .Case(S: "dot" , Value: Intrinsic::aarch64_sve_bfdot_lane_v2) |
| 1002 | .Case(S: "mlalb" , Value: Intrinsic::aarch64_sve_bfmlalb_lane_v2) |
| 1003 | .Case(S: "mlalt" , Value: Intrinsic::aarch64_sve_bfmlalt_lane_v2) |
| 1004 | .Default(Value: Intrinsic::not_intrinsic); |
| 1005 | if (ID != Intrinsic::not_intrinsic) { |
| 1006 | NewFn = Intrinsic::getOrInsertDeclaration(M: F->getParent(), id: ID); |
| 1007 | return true; |
| 1008 | } |
| 1009 | return false; // No other 'aarch64.sve.bf*.lane'. |
| 1010 | } |
| 1011 | return false; // No other 'aarch64.sve.bf*'. |
| 1012 | } |
| 1013 | |
| 1014 | // 'aarch64.sve.fcvt.bf16f32' || 'aarch64.sve.fcvtnt.bf16f32' |
| 1015 | if (Name == "fcvt.bf16f32" || Name == "fcvtnt.bf16f32" ) { |
| 1016 | NewFn = nullptr; |
| 1017 | return true; |
| 1018 | } |
| 1019 | |
| 1020 | if (Name.consume_front(Prefix: "addqv" )) { |
| 1021 | // 'aarch64.sve.addqv'. |
| 1022 | if (!F->getReturnType()->isFPOrFPVectorTy()) |
| 1023 | return false; |
| 1024 | |
| 1025 | auto Args = F->getFunctionType()->params(); |
| 1026 | Type *Tys[] = {F->getReturnType(), Args[1]}; |
| 1027 | NewFn = Intrinsic::getOrInsertDeclaration( |
| 1028 | M: F->getParent(), id: Intrinsic::aarch64_sve_faddqv, Tys); |
| 1029 | return true; |
| 1030 | } |
| 1031 | |
| 1032 | if (Name.consume_front(Prefix: "ld" )) { |
| 1033 | // 'aarch64.sve.ld*'. |
| 1034 | static const Regex LdRegex("^[234](.nxv[a-z0-9]+|$)" ); |
| 1035 | if (LdRegex.match(String: Name)) { |
| 1036 | Type *ScalarTy = |
| 1037 | cast<VectorType>(Val: F->getReturnType())->getElementType(); |
| 1038 | ElementCount EC = |
| 1039 | cast<VectorType>(Val: F->arg_begin()->getType())->getElementCount(); |
| 1040 | Type *Ty = VectorType::get(ElementType: ScalarTy, EC); |
| 1041 | static const Intrinsic::ID LoadIDs[] = { |
| 1042 | Intrinsic::aarch64_sve_ld2_sret, |
| 1043 | Intrinsic::aarch64_sve_ld3_sret, |
| 1044 | Intrinsic::aarch64_sve_ld4_sret, |
| 1045 | }; |
| 1046 | NewFn = Intrinsic::getOrInsertDeclaration(M: F->getParent(), |
| 1047 | id: LoadIDs[Name[0] - '2'], Tys: Ty); |
| 1048 | return true; |
| 1049 | } |
| 1050 | return false; // No other 'aarch64.sve.ld*'. |
| 1051 | } |
| 1052 | |
| 1053 | if (Name.consume_front(Prefix: "tuple." )) { |
| 1054 | // 'aarch64.sve.tuple.*'. |
| 1055 | if (Name.starts_with(Prefix: "get" )) { |
| 1056 | // 'aarch64.sve.tuple.get*'. |
| 1057 | Type *Tys[] = {F->getReturnType(), F->arg_begin()->getType()}; |
| 1058 | NewFn = Intrinsic::getOrInsertDeclaration( |
| 1059 | M: F->getParent(), id: Intrinsic::vector_extract, Tys); |
| 1060 | return true; |
| 1061 | } |
| 1062 | |
| 1063 | if (Name.starts_with(Prefix: "set" )) { |
| 1064 | // 'aarch64.sve.tuple.set*'. |
| 1065 | auto Args = F->getFunctionType()->params(); |
| 1066 | Type *Tys[] = {Args[0], Args[2], Args[1]}; |
| 1067 | NewFn = Intrinsic::getOrInsertDeclaration( |
| 1068 | M: F->getParent(), id: Intrinsic::vector_insert, Tys); |
| 1069 | return true; |
| 1070 | } |
| 1071 | |
| 1072 | static const Regex CreateTupleRegex("^create[234](.nxv[a-z0-9]+|$)" ); |
| 1073 | if (CreateTupleRegex.match(String: Name)) { |
| 1074 | // 'aarch64.sve.tuple.create*'. |
| 1075 | auto Args = F->getFunctionType()->params(); |
| 1076 | Type *Tys[] = {F->getReturnType(), Args[1]}; |
| 1077 | NewFn = Intrinsic::getOrInsertDeclaration( |
| 1078 | M: F->getParent(), id: Intrinsic::vector_insert, Tys); |
| 1079 | return true; |
| 1080 | } |
| 1081 | return false; // No other 'aarch64.sve.tuple.*'. |
| 1082 | } |
| 1083 | |
| 1084 | if (Name.starts_with(Prefix: "rev.nxv" )) { |
| 1085 | // 'aarch64.sve.rev.<Ty>' |
| 1086 | NewFn = Intrinsic::getOrInsertDeclaration( |
| 1087 | M: F->getParent(), id: Intrinsic::vector_reverse, Tys: F->getReturnType()); |
| 1088 | return true; |
| 1089 | } |
| 1090 | |
| 1091 | return false; // No other 'aarch64.sve.*'. |
| 1092 | } |
| 1093 | } |
| 1094 | return false; // No other 'arm.*', 'aarch64.*'. |
| 1095 | } |
| 1096 | |
| 1097 | static Intrinsic::ID shouldUpgradeNVPTXTMAG2SIntrinsics(Function *F, |
| 1098 | StringRef Name) { |
| 1099 | if (Name.consume_front(Prefix: "cp.async.bulk.tensor.g2s." )) { |
| 1100 | Intrinsic::ID ID = |
| 1101 | StringSwitch<Intrinsic::ID>(Name) |
| 1102 | .Case(S: "im2col.3d" , |
| 1103 | Value: Intrinsic::nvvm_cp_async_bulk_tensor_g2s_im2col_3d) |
| 1104 | .Case(S: "im2col.4d" , |
| 1105 | Value: Intrinsic::nvvm_cp_async_bulk_tensor_g2s_im2col_4d) |
| 1106 | .Case(S: "im2col.5d" , |
| 1107 | Value: Intrinsic::nvvm_cp_async_bulk_tensor_g2s_im2col_5d) |
| 1108 | .Case(S: "tile.1d" , Value: Intrinsic::nvvm_cp_async_bulk_tensor_g2s_tile_1d) |
| 1109 | .Case(S: "tile.2d" , Value: Intrinsic::nvvm_cp_async_bulk_tensor_g2s_tile_2d) |
| 1110 | .Case(S: "tile.3d" , Value: Intrinsic::nvvm_cp_async_bulk_tensor_g2s_tile_3d) |
| 1111 | .Case(S: "tile.4d" , Value: Intrinsic::nvvm_cp_async_bulk_tensor_g2s_tile_4d) |
| 1112 | .Case(S: "tile.5d" , Value: Intrinsic::nvvm_cp_async_bulk_tensor_g2s_tile_5d) |
| 1113 | .Default(Value: Intrinsic::not_intrinsic); |
| 1114 | |
| 1115 | if (ID == Intrinsic::not_intrinsic) |
| 1116 | return ID; |
| 1117 | |
| 1118 | // These intrinsics may need upgrade for two reasons: |
| 1119 | // (1) When the address-space of the first argument is shared[AS=3] |
| 1120 | // (and we upgrade it to use shared_cluster address-space[AS=7]) |
| 1121 | if (F->getArg(i: 0)->getType()->getPointerAddressSpace() == |
| 1122 | NVPTXAS::ADDRESS_SPACE_SHARED) |
| 1123 | return ID; |
| 1124 | |
| 1125 | // (2) When there are only two boolean flag arguments at the end: |
| 1126 | // |
| 1127 | // The last three parameters of the older version of these |
| 1128 | // intrinsics are: arg1, arg2, .. i64 ch, i1 mc_flag, i1 ch_flag |
| 1129 | // |
| 1130 | // The newer version reads as: |
| 1131 | // arg1, arg2, .. i64 ch, i1 mc_flag, i1 ch_flag, i32 cta_group_flag |
| 1132 | // |
| 1133 | // So, when the type of the [N-3]rd argument is "not i1", then |
| 1134 | // it is the older version and we need to upgrade. |
| 1135 | size_t FlagStartIndex = F->getFunctionType()->getNumParams() - 3; |
| 1136 | Type *ArgType = F->getFunctionType()->getParamType(i: FlagStartIndex); |
| 1137 | if (!ArgType->isIntegerTy(Bitwidth: 1)) |
| 1138 | return ID; |
| 1139 | } |
| 1140 | |
| 1141 | return Intrinsic::not_intrinsic; |
| 1142 | } |
| 1143 | |
| 1144 | static Intrinsic::ID shouldUpgradeNVPTXSharedClusterIntrinsic(Function *F, |
| 1145 | StringRef Name) { |
| 1146 | if (Name.consume_front(Prefix: "mapa.shared.cluster" )) |
| 1147 | if (F->getReturnType()->getPointerAddressSpace() == |
| 1148 | NVPTXAS::ADDRESS_SPACE_SHARED) |
| 1149 | return Intrinsic::nvvm_mapa_shared_cluster; |
| 1150 | |
| 1151 | if (Name.consume_front(Prefix: "cp.async.bulk." )) { |
| 1152 | Intrinsic::ID ID = |
| 1153 | StringSwitch<Intrinsic::ID>(Name) |
| 1154 | .Case(S: "global.to.shared.cluster" , |
| 1155 | Value: Intrinsic::nvvm_cp_async_bulk_global_to_shared_cluster) |
| 1156 | .Case(S: "shared.cta.to.cluster" , |
| 1157 | Value: Intrinsic::nvvm_cp_async_bulk_shared_cta_to_cluster) |
| 1158 | .Default(Value: Intrinsic::not_intrinsic); |
| 1159 | |
| 1160 | if (ID != Intrinsic::not_intrinsic) |
| 1161 | if (F->getArg(i: 0)->getType()->getPointerAddressSpace() == |
| 1162 | NVPTXAS::ADDRESS_SPACE_SHARED) |
| 1163 | return ID; |
| 1164 | } |
| 1165 | |
| 1166 | return Intrinsic::not_intrinsic; |
| 1167 | } |
| 1168 | |
| 1169 | static Intrinsic::ID shouldUpgradeNVPTXBF16Intrinsic(StringRef Name) { |
| 1170 | if (Name.consume_front(Prefix: "fma.rn." )) |
| 1171 | return StringSwitch<Intrinsic::ID>(Name) |
| 1172 | .Case(S: "bf16" , Value: Intrinsic::nvvm_fma_rn_bf16) |
| 1173 | .Case(S: "bf16x2" , Value: Intrinsic::nvvm_fma_rn_bf16x2) |
| 1174 | .Case(S: "relu.bf16" , Value: Intrinsic::nvvm_fma_rn_relu_bf16) |
| 1175 | .Case(S: "relu.bf16x2" , Value: Intrinsic::nvvm_fma_rn_relu_bf16x2) |
| 1176 | .Default(Value: Intrinsic::not_intrinsic); |
| 1177 | |
| 1178 | if (Name.consume_front(Prefix: "fmax." )) |
| 1179 | return StringSwitch<Intrinsic::ID>(Name) |
| 1180 | .Case(S: "bf16" , Value: Intrinsic::nvvm_fmax_bf16) |
| 1181 | .Case(S: "bf16x2" , Value: Intrinsic::nvvm_fmax_bf16x2) |
| 1182 | .Case(S: "ftz.bf16" , Value: Intrinsic::nvvm_fmax_ftz_bf16) |
| 1183 | .Case(S: "ftz.bf16x2" , Value: Intrinsic::nvvm_fmax_ftz_bf16x2) |
| 1184 | .Case(S: "ftz.nan.bf16" , Value: Intrinsic::nvvm_fmax_ftz_nan_bf16) |
| 1185 | .Case(S: "ftz.nan.bf16x2" , Value: Intrinsic::nvvm_fmax_ftz_nan_bf16x2) |
| 1186 | .Case(S: "ftz.nan.xorsign.abs.bf16" , |
| 1187 | Value: Intrinsic::nvvm_fmax_ftz_nan_xorsign_abs_bf16) |
| 1188 | .Case(S: "ftz.nan.xorsign.abs.bf16x2" , |
| 1189 | Value: Intrinsic::nvvm_fmax_ftz_nan_xorsign_abs_bf16x2) |
| 1190 | .Case(S: "ftz.xorsign.abs.bf16" , Value: Intrinsic::nvvm_fmax_ftz_xorsign_abs_bf16) |
| 1191 | .Case(S: "ftz.xorsign.abs.bf16x2" , |
| 1192 | Value: Intrinsic::nvvm_fmax_ftz_xorsign_abs_bf16x2) |
| 1193 | .Case(S: "nan.bf16" , Value: Intrinsic::nvvm_fmax_nan_bf16) |
| 1194 | .Case(S: "nan.bf16x2" , Value: Intrinsic::nvvm_fmax_nan_bf16x2) |
| 1195 | .Case(S: "nan.xorsign.abs.bf16" , Value: Intrinsic::nvvm_fmax_nan_xorsign_abs_bf16) |
| 1196 | .Case(S: "nan.xorsign.abs.bf16x2" , |
| 1197 | Value: Intrinsic::nvvm_fmax_nan_xorsign_abs_bf16x2) |
| 1198 | .Case(S: "xorsign.abs.bf16" , Value: Intrinsic::nvvm_fmax_xorsign_abs_bf16) |
| 1199 | .Case(S: "xorsign.abs.bf16x2" , Value: Intrinsic::nvvm_fmax_xorsign_abs_bf16x2) |
| 1200 | .Default(Value: Intrinsic::not_intrinsic); |
| 1201 | |
| 1202 | if (Name.consume_front(Prefix: "fmin." )) |
| 1203 | return StringSwitch<Intrinsic::ID>(Name) |
| 1204 | .Case(S: "bf16" , Value: Intrinsic::nvvm_fmin_bf16) |
| 1205 | .Case(S: "bf16x2" , Value: Intrinsic::nvvm_fmin_bf16x2) |
| 1206 | .Case(S: "ftz.bf16" , Value: Intrinsic::nvvm_fmin_ftz_bf16) |
| 1207 | .Case(S: "ftz.bf16x2" , Value: Intrinsic::nvvm_fmin_ftz_bf16x2) |
| 1208 | .Case(S: "ftz.nan.bf16" , Value: Intrinsic::nvvm_fmin_ftz_nan_bf16) |
| 1209 | .Case(S: "ftz.nan.bf16x2" , Value: Intrinsic::nvvm_fmin_ftz_nan_bf16x2) |
| 1210 | .Case(S: "ftz.nan.xorsign.abs.bf16" , |
| 1211 | Value: Intrinsic::nvvm_fmin_ftz_nan_xorsign_abs_bf16) |
| 1212 | .Case(S: "ftz.nan.xorsign.abs.bf16x2" , |
| 1213 | Value: Intrinsic::nvvm_fmin_ftz_nan_xorsign_abs_bf16x2) |
| 1214 | .Case(S: "ftz.xorsign.abs.bf16" , Value: Intrinsic::nvvm_fmin_ftz_xorsign_abs_bf16) |
| 1215 | .Case(S: "ftz.xorsign.abs.bf16x2" , |
| 1216 | Value: Intrinsic::nvvm_fmin_ftz_xorsign_abs_bf16x2) |
| 1217 | .Case(S: "nan.bf16" , Value: Intrinsic::nvvm_fmin_nan_bf16) |
| 1218 | .Case(S: "nan.bf16x2" , Value: Intrinsic::nvvm_fmin_nan_bf16x2) |
| 1219 | .Case(S: "nan.xorsign.abs.bf16" , Value: Intrinsic::nvvm_fmin_nan_xorsign_abs_bf16) |
| 1220 | .Case(S: "nan.xorsign.abs.bf16x2" , |
| 1221 | Value: Intrinsic::nvvm_fmin_nan_xorsign_abs_bf16x2) |
| 1222 | .Case(S: "xorsign.abs.bf16" , Value: Intrinsic::nvvm_fmin_xorsign_abs_bf16) |
| 1223 | .Case(S: "xorsign.abs.bf16x2" , Value: Intrinsic::nvvm_fmin_xorsign_abs_bf16x2) |
| 1224 | .Default(Value: Intrinsic::not_intrinsic); |
| 1225 | |
| 1226 | if (Name.consume_front(Prefix: "neg." )) |
| 1227 | return StringSwitch<Intrinsic::ID>(Name) |
| 1228 | .Case(S: "bf16" , Value: Intrinsic::nvvm_neg_bf16) |
| 1229 | .Case(S: "bf16x2" , Value: Intrinsic::nvvm_neg_bf16x2) |
| 1230 | .Default(Value: Intrinsic::not_intrinsic); |
| 1231 | |
| 1232 | return Intrinsic::not_intrinsic; |
| 1233 | } |
| 1234 | |
| 1235 | static bool consumeNVVMPtrAddrSpace(StringRef &Name) { |
| 1236 | return Name.consume_front(Prefix: "local" ) || Name.consume_front(Prefix: "shared" ) || |
| 1237 | Name.consume_front(Prefix: "global" ) || Name.consume_front(Prefix: "constant" ) || |
| 1238 | Name.consume_front(Prefix: "param" ); |
| 1239 | } |
| 1240 | |
| 1241 | static bool convertIntrinsicValidType(StringRef Name, |
| 1242 | const FunctionType *FuncTy) { |
| 1243 | Type *HalfTy = Type::getHalfTy(C&: FuncTy->getContext()); |
| 1244 | if (Name.starts_with(Prefix: "to.fp16" )) { |
| 1245 | return CastInst::castIsValid(op: Instruction::FPTrunc, SrcTy: FuncTy->getParamType(i: 0), |
| 1246 | DstTy: HalfTy) && |
| 1247 | CastInst::castIsValid(op: Instruction::BitCast, SrcTy: HalfTy, |
| 1248 | DstTy: FuncTy->getReturnType()); |
| 1249 | } |
| 1250 | |
| 1251 | if (Name.starts_with(Prefix: "from.fp16" )) { |
| 1252 | return CastInst::castIsValid(op: Instruction::BitCast, SrcTy: FuncTy->getParamType(i: 0), |
| 1253 | DstTy: HalfTy) && |
| 1254 | CastInst::castIsValid(op: Instruction::FPExt, SrcTy: HalfTy, |
| 1255 | DstTy: FuncTy->getReturnType()); |
| 1256 | } |
| 1257 | |
| 1258 | return false; |
| 1259 | } |
| 1260 | |
| 1261 | static bool upgradeIntrinsicFunction1(Function *F, Function *&NewFn, |
| 1262 | bool CanUpgradeDebugIntrinsicsToRecords) { |
| 1263 | assert(F && "Illegal to upgrade a non-existent Function." ); |
| 1264 | |
| 1265 | StringRef Name = F->getName(); |
| 1266 | |
| 1267 | // Quickly eliminate it, if it's not a candidate. |
| 1268 | if (!Name.consume_front(Prefix: "llvm." ) || Name.empty()) |
| 1269 | return false; |
| 1270 | |
| 1271 | switch (Name[0]) { |
| 1272 | default: break; |
| 1273 | case 'a': { |
| 1274 | bool IsArm = Name.consume_front(Prefix: "arm." ); |
| 1275 | if (IsArm || Name.consume_front(Prefix: "aarch64." )) { |
| 1276 | if (upgradeArmOrAarch64IntrinsicFunction(IsArm, F, Name, NewFn)) |
| 1277 | return true; |
| 1278 | break; |
| 1279 | } |
| 1280 | |
| 1281 | if (Name.consume_front(Prefix: "amdgcn." )) { |
| 1282 | if (Name == "alignbit" ) { |
| 1283 | // Target specific intrinsic became redundant |
| 1284 | NewFn = Intrinsic::getOrInsertDeclaration( |
| 1285 | M: F->getParent(), id: Intrinsic::fshr, Tys: {F->getReturnType()}); |
| 1286 | return true; |
| 1287 | } |
| 1288 | |
| 1289 | if (Name.consume_front(Prefix: "atomic." )) { |
| 1290 | if (Name.starts_with(Prefix: "inc" ) || Name.starts_with(Prefix: "dec" ) || |
| 1291 | Name.starts_with(Prefix: "cond.sub" ) || Name.starts_with(Prefix: "csub" )) { |
| 1292 | // These were replaced with atomicrmw uinc_wrap, udec_wrap, usub_cond |
| 1293 | // and usub_sat so there's no new declaration. |
| 1294 | NewFn = nullptr; |
| 1295 | return true; |
| 1296 | } |
| 1297 | break; // No other 'amdgcn.atomic.*' |
| 1298 | } |
| 1299 | |
| 1300 | // Legacy wmma iu intrinsics without the optional clamp operand. |
| 1301 | if (F->getIntrinsicID() == Intrinsic::amdgcn_wmma_i32_16x16x64_iu8 && |
| 1302 | F->arg_size() == 7) { |
| 1303 | NewFn = nullptr; |
| 1304 | return true; |
| 1305 | } |
| 1306 | if (F->getIntrinsicID() == Intrinsic::amdgcn_swmmac_i32_16x16x128_iu8 && |
| 1307 | F->arg_size() == 8) { |
| 1308 | NewFn = nullptr; |
| 1309 | return true; |
| 1310 | } |
| 1311 | |
| 1312 | if (Name.consume_front(Prefix: "ds." ) || Name.consume_front(Prefix: "global.atomic." ) || |
| 1313 | Name.consume_front(Prefix: "flat.atomic." )) { |
| 1314 | if (Name.starts_with(Prefix: "fadd" ) || |
| 1315 | // FIXME: We should also remove fmin.num and fmax.num intrinsics. |
| 1316 | (Name.starts_with(Prefix: "fmin" ) && !Name.starts_with(Prefix: "fmin.num" )) || |
| 1317 | (Name.starts_with(Prefix: "fmax" ) && !Name.starts_with(Prefix: "fmax.num" ))) { |
| 1318 | // Replaced with atomicrmw fadd/fmin/fmax, so there's no new |
| 1319 | // declaration. |
| 1320 | NewFn = nullptr; |
| 1321 | return true; |
| 1322 | } |
| 1323 | } |
| 1324 | |
| 1325 | if (Name.starts_with(Prefix: "ldexp." )) { |
| 1326 | // Target specific intrinsic became redundant |
| 1327 | NewFn = Intrinsic::getOrInsertDeclaration( |
| 1328 | M: F->getParent(), id: Intrinsic::ldexp, |
| 1329 | Tys: {F->getReturnType(), F->getArg(i: 1)->getType()}); |
| 1330 | return true; |
| 1331 | } |
| 1332 | break; // No other 'amdgcn.*' |
| 1333 | } |
| 1334 | |
| 1335 | break; |
| 1336 | } |
| 1337 | case 'c': { |
| 1338 | if (F->arg_size() == 1) { |
| 1339 | if (Name.consume_front(Prefix: "convert." )) { |
| 1340 | if (convertIntrinsicValidType(Name, FuncTy: F->getFunctionType())) { |
| 1341 | NewFn = nullptr; |
| 1342 | return true; |
| 1343 | } |
| 1344 | } |
| 1345 | |
| 1346 | Intrinsic::ID ID = StringSwitch<Intrinsic::ID>(Name) |
| 1347 | .StartsWith(S: "ctlz." , Value: Intrinsic::ctlz) |
| 1348 | .StartsWith(S: "cttz." , Value: Intrinsic::cttz) |
| 1349 | .Default(Value: Intrinsic::not_intrinsic); |
| 1350 | if (ID != Intrinsic::not_intrinsic) { |
| 1351 | rename(GV: F); |
| 1352 | NewFn = Intrinsic::getOrInsertDeclaration(M: F->getParent(), id: ID, |
| 1353 | Tys: F->arg_begin()->getType()); |
| 1354 | return true; |
| 1355 | } |
| 1356 | } |
| 1357 | |
| 1358 | if (F->arg_size() == 2 && Name == "coro.end" ) { |
| 1359 | rename(GV: F); |
| 1360 | NewFn = Intrinsic::getOrInsertDeclaration(M: F->getParent(), |
| 1361 | id: Intrinsic::coro_end); |
| 1362 | return true; |
| 1363 | } |
| 1364 | |
| 1365 | break; |
| 1366 | } |
| 1367 | case 'd': |
| 1368 | if (Name.consume_front(Prefix: "dbg." )) { |
| 1369 | // Mark debug intrinsics for upgrade to new debug format. |
| 1370 | if (CanUpgradeDebugIntrinsicsToRecords) { |
| 1371 | if (Name == "addr" || Name == "value" || Name == "assign" || |
| 1372 | Name == "declare" || Name == "label" ) { |
| 1373 | // There's no function to replace these with. |
| 1374 | NewFn = nullptr; |
| 1375 | // But we do want these to get upgraded. |
| 1376 | return true; |
| 1377 | } |
| 1378 | } |
| 1379 | // Update llvm.dbg.addr intrinsics even in "new debug mode"; they'll get |
| 1380 | // converted to DbgVariableRecords later. |
| 1381 | if (Name == "addr" || (Name == "value" && F->arg_size() == 4)) { |
| 1382 | rename(GV: F); |
| 1383 | NewFn = Intrinsic::getOrInsertDeclaration(M: F->getParent(), |
| 1384 | id: Intrinsic::dbg_value); |
| 1385 | return true; |
| 1386 | } |
| 1387 | break; // No other 'dbg.*'. |
| 1388 | } |
| 1389 | break; |
| 1390 | case 'e': |
| 1391 | if (Name.consume_front(Prefix: "experimental.vector." )) { |
| 1392 | Intrinsic::ID ID = |
| 1393 | StringSwitch<Intrinsic::ID>(Name) |
| 1394 | // Skip over extract.last.active, otherwise it will be 'upgraded' |
| 1395 | // to a regular vector extract which is a different operation. |
| 1396 | .StartsWith(S: "extract.last.active." , Value: Intrinsic::not_intrinsic) |
| 1397 | .StartsWith(S: "extract." , Value: Intrinsic::vector_extract) |
| 1398 | .StartsWith(S: "insert." , Value: Intrinsic::vector_insert) |
| 1399 | .StartsWith(S: "reverse." , Value: Intrinsic::vector_reverse) |
| 1400 | .StartsWith(S: "interleave2." , Value: Intrinsic::vector_interleave2) |
| 1401 | .StartsWith(S: "deinterleave2." , Value: Intrinsic::vector_deinterleave2) |
| 1402 | .StartsWith(S: "partial.reduce.add" , |
| 1403 | Value: Intrinsic::vector_partial_reduce_add) |
| 1404 | .Default(Value: Intrinsic::not_intrinsic); |
| 1405 | if (ID != Intrinsic::not_intrinsic) { |
| 1406 | const auto *FT = F->getFunctionType(); |
| 1407 | SmallVector<Type *, 2> Tys; |
| 1408 | if (ID == Intrinsic::vector_extract || |
| 1409 | ID == Intrinsic::vector_interleave2) |
| 1410 | // Extracting overloads the return type. |
| 1411 | Tys.push_back(Elt: FT->getReturnType()); |
| 1412 | if (ID != Intrinsic::vector_interleave2) |
| 1413 | Tys.push_back(Elt: FT->getParamType(i: 0)); |
| 1414 | if (ID == Intrinsic::vector_insert || |
| 1415 | ID == Intrinsic::vector_partial_reduce_add) |
| 1416 | // Inserting overloads the inserted type. |
| 1417 | Tys.push_back(Elt: FT->getParamType(i: 1)); |
| 1418 | rename(GV: F); |
| 1419 | NewFn = Intrinsic::getOrInsertDeclaration(M: F->getParent(), id: ID, Tys); |
| 1420 | return true; |
| 1421 | } |
| 1422 | |
| 1423 | if (Name.consume_front(Prefix: "reduce." )) { |
| 1424 | SmallVector<StringRef, 2> Groups; |
| 1425 | static const Regex R("^([a-z]+)\\.[a-z][0-9]+" ); |
| 1426 | if (R.match(String: Name, Matches: &Groups)) |
| 1427 | ID = StringSwitch<Intrinsic::ID>(Groups[1]) |
| 1428 | .Case(S: "add" , Value: Intrinsic::vector_reduce_add) |
| 1429 | .Case(S: "mul" , Value: Intrinsic::vector_reduce_mul) |
| 1430 | .Case(S: "and" , Value: Intrinsic::vector_reduce_and) |
| 1431 | .Case(S: "or" , Value: Intrinsic::vector_reduce_or) |
| 1432 | .Case(S: "xor" , Value: Intrinsic::vector_reduce_xor) |
| 1433 | .Case(S: "smax" , Value: Intrinsic::vector_reduce_smax) |
| 1434 | .Case(S: "smin" , Value: Intrinsic::vector_reduce_smin) |
| 1435 | .Case(S: "umax" , Value: Intrinsic::vector_reduce_umax) |
| 1436 | .Case(S: "umin" , Value: Intrinsic::vector_reduce_umin) |
| 1437 | .Case(S: "fmax" , Value: Intrinsic::vector_reduce_fmax) |
| 1438 | .Case(S: "fmin" , Value: Intrinsic::vector_reduce_fmin) |
| 1439 | .Default(Value: Intrinsic::not_intrinsic); |
| 1440 | |
| 1441 | bool V2 = false; |
| 1442 | if (ID == Intrinsic::not_intrinsic) { |
| 1443 | static const Regex R2("^v2\\.([a-z]+)\\.[fi][0-9]+" ); |
| 1444 | Groups.clear(); |
| 1445 | V2 = true; |
| 1446 | if (R2.match(String: Name, Matches: &Groups)) |
| 1447 | ID = StringSwitch<Intrinsic::ID>(Groups[1]) |
| 1448 | .Case(S: "fadd" , Value: Intrinsic::vector_reduce_fadd) |
| 1449 | .Case(S: "fmul" , Value: Intrinsic::vector_reduce_fmul) |
| 1450 | .Default(Value: Intrinsic::not_intrinsic); |
| 1451 | } |
| 1452 | if (ID != Intrinsic::not_intrinsic) { |
| 1453 | rename(GV: F); |
| 1454 | auto Args = F->getFunctionType()->params(); |
| 1455 | NewFn = Intrinsic::getOrInsertDeclaration(M: F->getParent(), id: ID, |
| 1456 | Tys: {Args[V2 ? 1 : 0]}); |
| 1457 | return true; |
| 1458 | } |
| 1459 | break; // No other 'expermental.vector.reduce.*'. |
| 1460 | } |
| 1461 | |
| 1462 | if (Name.consume_front(Prefix: "splice" )) |
| 1463 | return true; |
| 1464 | break; // No other 'experimental.vector.*'. |
| 1465 | } |
| 1466 | if (Name.consume_front(Prefix: "experimental.stepvector." )) { |
| 1467 | Intrinsic::ID ID = Intrinsic::stepvector; |
| 1468 | rename(GV: F); |
| 1469 | NewFn = Intrinsic::getOrInsertDeclaration( |
| 1470 | M: F->getParent(), id: ID, Tys: F->getFunctionType()->getReturnType()); |
| 1471 | return true; |
| 1472 | } |
| 1473 | break; // No other 'e*'. |
| 1474 | case 'f': |
| 1475 | if (Name.starts_with(Prefix: "flt.rounds" )) { |
| 1476 | rename(GV: F); |
| 1477 | NewFn = Intrinsic::getOrInsertDeclaration(M: F->getParent(), |
| 1478 | id: Intrinsic::get_rounding); |
| 1479 | return true; |
| 1480 | } |
| 1481 | break; |
| 1482 | case 'i': |
| 1483 | if (Name.starts_with(Prefix: "invariant.group.barrier" )) { |
| 1484 | // Rename invariant.group.barrier to launder.invariant.group |
| 1485 | auto Args = F->getFunctionType()->params(); |
| 1486 | Type* ObjectPtr[1] = {Args[0]}; |
| 1487 | rename(GV: F); |
| 1488 | NewFn = Intrinsic::getOrInsertDeclaration( |
| 1489 | M: F->getParent(), id: Intrinsic::launder_invariant_group, Tys: ObjectPtr); |
| 1490 | return true; |
| 1491 | } |
| 1492 | break; |
| 1493 | case 'l': |
| 1494 | if ((Name.starts_with(Prefix: "lifetime.start" ) || |
| 1495 | Name.starts_with(Prefix: "lifetime.end" )) && |
| 1496 | F->arg_size() == 2) { |
| 1497 | Intrinsic::ID IID = Name.starts_with(Prefix: "lifetime.start" ) |
| 1498 | ? Intrinsic::lifetime_start |
| 1499 | : Intrinsic::lifetime_end; |
| 1500 | rename(GV: F); |
| 1501 | NewFn = Intrinsic::getOrInsertDeclaration(M: F->getParent(), id: IID, |
| 1502 | Tys: F->getArg(i: 0)->getType()); |
| 1503 | return true; |
| 1504 | } |
| 1505 | break; |
| 1506 | case 'm': { |
| 1507 | // Updating the memory intrinsics (memcpy/memmove/memset) that have an |
| 1508 | // alignment parameter to embedding the alignment as an attribute of |
| 1509 | // the pointer args. |
| 1510 | if (unsigned ID = StringSwitch<unsigned>(Name) |
| 1511 | .StartsWith(S: "memcpy." , Value: Intrinsic::memcpy) |
| 1512 | .StartsWith(S: "memmove." , Value: Intrinsic::memmove) |
| 1513 | .Default(Value: 0)) { |
| 1514 | if (F->arg_size() == 5) { |
| 1515 | rename(GV: F); |
| 1516 | // Get the types of dest, src, and len |
| 1517 | ArrayRef<Type *> ParamTypes = |
| 1518 | F->getFunctionType()->params().slice(N: 0, M: 3); |
| 1519 | NewFn = |
| 1520 | Intrinsic::getOrInsertDeclaration(M: F->getParent(), id: ID, Tys: ParamTypes); |
| 1521 | return true; |
| 1522 | } |
| 1523 | } |
| 1524 | if (Name.starts_with(Prefix: "memset." ) && F->arg_size() == 5) { |
| 1525 | rename(GV: F); |
| 1526 | // Get the types of dest, and len |
| 1527 | const auto *FT = F->getFunctionType(); |
| 1528 | Type *ParamTypes[2] = { |
| 1529 | FT->getParamType(i: 0), // Dest |
| 1530 | FT->getParamType(i: 2) // len |
| 1531 | }; |
| 1532 | NewFn = Intrinsic::getOrInsertDeclaration(M: F->getParent(), |
| 1533 | id: Intrinsic::memset, Tys: ParamTypes); |
| 1534 | return true; |
| 1535 | } |
| 1536 | |
| 1537 | unsigned MaskedID = |
| 1538 | StringSwitch<unsigned>(Name) |
| 1539 | .StartsWith(S: "masked.load" , Value: Intrinsic::masked_load) |
| 1540 | .StartsWith(S: "masked.gather" , Value: Intrinsic::masked_gather) |
| 1541 | .StartsWith(S: "masked.store" , Value: Intrinsic::masked_store) |
| 1542 | .StartsWith(S: "masked.scatter" , Value: Intrinsic::masked_scatter) |
| 1543 | .Default(Value: 0); |
| 1544 | if (MaskedID && F->arg_size() == 4) { |
| 1545 | rename(GV: F); |
| 1546 | if (MaskedID == Intrinsic::masked_load || |
| 1547 | MaskedID == Intrinsic::masked_gather) { |
| 1548 | NewFn = Intrinsic::getOrInsertDeclaration( |
| 1549 | M: F->getParent(), id: MaskedID, |
| 1550 | Tys: {F->getReturnType(), F->getArg(i: 0)->getType()}); |
| 1551 | return true; |
| 1552 | } |
| 1553 | NewFn = Intrinsic::getOrInsertDeclaration( |
| 1554 | M: F->getParent(), id: MaskedID, |
| 1555 | Tys: {F->getArg(i: 0)->getType(), F->getArg(i: 1)->getType()}); |
| 1556 | return true; |
| 1557 | } |
| 1558 | break; |
| 1559 | } |
| 1560 | case 'n': { |
| 1561 | if (Name.consume_front(Prefix: "nvvm." )) { |
| 1562 | // Check for nvvm intrinsics corresponding exactly to an LLVM intrinsic. |
| 1563 | if (F->arg_size() == 1) { |
| 1564 | Intrinsic::ID IID = |
| 1565 | StringSwitch<Intrinsic::ID>(Name) |
| 1566 | .Cases(CaseStrings: {"brev32" , "brev64" }, Value: Intrinsic::bitreverse) |
| 1567 | .Case(S: "clz.i" , Value: Intrinsic::ctlz) |
| 1568 | .Case(S: "popc.i" , Value: Intrinsic::ctpop) |
| 1569 | .Default(Value: Intrinsic::not_intrinsic); |
| 1570 | if (IID != Intrinsic::not_intrinsic) { |
| 1571 | NewFn = Intrinsic::getOrInsertDeclaration(M: F->getParent(), id: IID, |
| 1572 | Tys: {F->getReturnType()}); |
| 1573 | return true; |
| 1574 | } |
| 1575 | } else if (F->arg_size() == 2) { |
| 1576 | Intrinsic::ID IID = |
| 1577 | StringSwitch<Intrinsic::ID>(Name) |
| 1578 | .Cases(CaseStrings: {"max.s" , "max.i" , "max.ll" }, Value: Intrinsic::smax) |
| 1579 | .Cases(CaseStrings: {"min.s" , "min.i" , "min.ll" }, Value: Intrinsic::smin) |
| 1580 | .Cases(CaseStrings: {"max.us" , "max.ui" , "max.ull" }, Value: Intrinsic::umax) |
| 1581 | .Cases(CaseStrings: {"min.us" , "min.ui" , "min.ull" }, Value: Intrinsic::umin) |
| 1582 | .Default(Value: Intrinsic::not_intrinsic); |
| 1583 | if (IID != Intrinsic::not_intrinsic) { |
| 1584 | NewFn = Intrinsic::getOrInsertDeclaration(M: F->getParent(), id: IID, |
| 1585 | Tys: {F->getReturnType()}); |
| 1586 | return true; |
| 1587 | } |
| 1588 | } |
| 1589 | |
| 1590 | // Check for nvvm intrinsics that need a return type adjustment. |
| 1591 | if (!F->getReturnType()->getScalarType()->isBFloatTy()) { |
| 1592 | Intrinsic::ID IID = shouldUpgradeNVPTXBF16Intrinsic(Name); |
| 1593 | if (IID != Intrinsic::not_intrinsic) { |
| 1594 | NewFn = nullptr; |
| 1595 | return true; |
| 1596 | } |
| 1597 | } |
| 1598 | |
| 1599 | // Upgrade Distributed Shared Memory Intrinsics |
| 1600 | Intrinsic::ID IID = shouldUpgradeNVPTXSharedClusterIntrinsic(F, Name); |
| 1601 | if (IID != Intrinsic::not_intrinsic) { |
| 1602 | rename(GV: F); |
| 1603 | NewFn = Intrinsic::getOrInsertDeclaration(M: F->getParent(), id: IID); |
| 1604 | return true; |
| 1605 | } |
| 1606 | |
| 1607 | // Upgrade TMA copy G2S Intrinsics |
| 1608 | IID = shouldUpgradeNVPTXTMAG2SIntrinsics(F, Name); |
| 1609 | if (IID != Intrinsic::not_intrinsic) { |
| 1610 | rename(GV: F); |
| 1611 | NewFn = Intrinsic::getOrInsertDeclaration(M: F->getParent(), id: IID); |
| 1612 | return true; |
| 1613 | } |
| 1614 | |
| 1615 | // The following nvvm intrinsics correspond exactly to an LLVM idiom, but |
| 1616 | // not to an intrinsic alone. We expand them in UpgradeIntrinsicCall. |
| 1617 | // |
| 1618 | // TODO: We could add lohi.i2d. |
| 1619 | bool Expand = false; |
| 1620 | if (Name.consume_front(Prefix: "abs." )) |
| 1621 | // nvvm.abs.{i,ii} |
| 1622 | Expand = |
| 1623 | Name == "i" || Name == "ll" || Name == "bf16" || Name == "bf16x2" ; |
| 1624 | else if (Name.consume_front(Prefix: "fabs." )) |
| 1625 | // nvvm.fabs.{f,ftz.f,d} |
| 1626 | Expand = Name == "f" || Name == "ftz.f" || Name == "d" ; |
| 1627 | else if (Name.consume_front(Prefix: "ex2.approx." )) |
| 1628 | // nvvm.ex2.approx.{f,ftz.f,d,f16x2} |
| 1629 | Expand = |
| 1630 | Name == "f" || Name == "ftz.f" || Name == "d" || Name == "f16x2" ; |
| 1631 | else if (Name.consume_front(Prefix: "atomic.load." )) |
| 1632 | // nvvm.atomic.load.add.{f32,f64}.p |
| 1633 | // nvvm.atomic.load.{inc,dec}.32.p |
| 1634 | Expand = StringSwitch<bool>(Name) |
| 1635 | .StartsWith(S: "add.f32.p" , Value: true) |
| 1636 | .StartsWith(S: "add.f64.p" , Value: true) |
| 1637 | .StartsWith(S: "inc.32.p" , Value: true) |
| 1638 | .StartsWith(S: "dec.32.p" , Value: true) |
| 1639 | .Default(Value: false); |
| 1640 | else if (Name.consume_front(Prefix: "bitcast." )) |
| 1641 | // nvvm.bitcast.{f2i,i2f,ll2d,d2ll} |
| 1642 | Expand = |
| 1643 | Name == "f2i" || Name == "i2f" || Name == "ll2d" || Name == "d2ll" ; |
| 1644 | else if (Name.consume_front(Prefix: "rotate." )) |
| 1645 | // nvvm.rotate.{b32,b64,right.b64} |
| 1646 | Expand = Name == "b32" || Name == "b64" || Name == "right.b64" ; |
| 1647 | else if (Name.consume_front(Prefix: "ptr.gen.to." )) |
| 1648 | // nvvm.ptr.gen.to.{local,shared,global,constant,param} |
| 1649 | Expand = consumeNVVMPtrAddrSpace(Name); |
| 1650 | else if (Name.consume_front(Prefix: "ptr." )) |
| 1651 | // nvvm.ptr.{local,shared,global,constant,param}.to.gen |
| 1652 | Expand = consumeNVVMPtrAddrSpace(Name) && Name.starts_with(Prefix: ".to.gen" ); |
| 1653 | else if (Name.consume_front(Prefix: "ldg.global." )) |
| 1654 | // nvvm.ldg.global.{i,p,f} |
| 1655 | Expand = (Name.starts_with(Prefix: "i." ) || Name.starts_with(Prefix: "f." ) || |
| 1656 | Name.starts_with(Prefix: "p." )); |
| 1657 | else |
| 1658 | Expand = StringSwitch<bool>(Name) |
| 1659 | .Case(S: "barrier0" , Value: true) |
| 1660 | .Case(S: "barrier.n" , Value: true) |
| 1661 | .Case(S: "barrier.sync.cnt" , Value: true) |
| 1662 | .Case(S: "barrier.sync" , Value: true) |
| 1663 | .Case(S: "barrier" , Value: true) |
| 1664 | .Case(S: "bar.sync" , Value: true) |
| 1665 | .Case(S: "barrier0.popc" , Value: true) |
| 1666 | .Case(S: "barrier0.and" , Value: true) |
| 1667 | .Case(S: "barrier0.or" , Value: true) |
| 1668 | .Case(S: "clz.ll" , Value: true) |
| 1669 | .Case(S: "popc.ll" , Value: true) |
| 1670 | .Case(S: "h2f" , Value: true) |
| 1671 | .Case(S: "swap.lo.hi.b64" , Value: true) |
| 1672 | .Case(S: "tanh.approx.f32" , Value: true) |
| 1673 | .Default(Value: false); |
| 1674 | |
| 1675 | if (Expand) { |
| 1676 | NewFn = nullptr; |
| 1677 | return true; |
| 1678 | } |
| 1679 | break; // No other 'nvvm.*'. |
| 1680 | } |
| 1681 | break; |
| 1682 | } |
| 1683 | case 'o': |
| 1684 | if (Name.starts_with(Prefix: "objectsize." )) { |
| 1685 | Type *Tys[2] = { F->getReturnType(), F->arg_begin()->getType() }; |
| 1686 | if (F->arg_size() == 2 || F->arg_size() == 3) { |
| 1687 | rename(GV: F); |
| 1688 | NewFn = Intrinsic::getOrInsertDeclaration(M: F->getParent(), |
| 1689 | id: Intrinsic::objectsize, Tys); |
| 1690 | return true; |
| 1691 | } |
| 1692 | } |
| 1693 | break; |
| 1694 | |
| 1695 | case 'p': |
| 1696 | if (Name.starts_with(Prefix: "ptr.annotation." ) && F->arg_size() == 4) { |
| 1697 | rename(GV: F); |
| 1698 | NewFn = Intrinsic::getOrInsertDeclaration( |
| 1699 | M: F->getParent(), id: Intrinsic::ptr_annotation, |
| 1700 | Tys: {F->arg_begin()->getType(), F->getArg(i: 1)->getType()}); |
| 1701 | return true; |
| 1702 | } |
| 1703 | break; |
| 1704 | |
| 1705 | case 'r': { |
| 1706 | if (Name.consume_front(Prefix: "riscv." )) { |
| 1707 | Intrinsic::ID ID; |
| 1708 | ID = StringSwitch<Intrinsic::ID>(Name) |
| 1709 | .Case(S: "aes32dsi" , Value: Intrinsic::riscv_aes32dsi) |
| 1710 | .Case(S: "aes32dsmi" , Value: Intrinsic::riscv_aes32dsmi) |
| 1711 | .Case(S: "aes32esi" , Value: Intrinsic::riscv_aes32esi) |
| 1712 | .Case(S: "aes32esmi" , Value: Intrinsic::riscv_aes32esmi) |
| 1713 | .Default(Value: Intrinsic::not_intrinsic); |
| 1714 | if (ID != Intrinsic::not_intrinsic) { |
| 1715 | if (!F->getFunctionType()->getParamType(i: 2)->isIntegerTy(Bitwidth: 32)) { |
| 1716 | rename(GV: F); |
| 1717 | NewFn = Intrinsic::getOrInsertDeclaration(M: F->getParent(), id: ID); |
| 1718 | return true; |
| 1719 | } |
| 1720 | break; // No other applicable upgrades. |
| 1721 | } |
| 1722 | |
| 1723 | ID = StringSwitch<Intrinsic::ID>(Name) |
| 1724 | .StartsWith(S: "sm4ks" , Value: Intrinsic::riscv_sm4ks) |
| 1725 | .StartsWith(S: "sm4ed" , Value: Intrinsic::riscv_sm4ed) |
| 1726 | .Default(Value: Intrinsic::not_intrinsic); |
| 1727 | if (ID != Intrinsic::not_intrinsic) { |
| 1728 | if (!F->getFunctionType()->getParamType(i: 2)->isIntegerTy(Bitwidth: 32) || |
| 1729 | F->getFunctionType()->getReturnType()->isIntegerTy(Bitwidth: 64)) { |
| 1730 | rename(GV: F); |
| 1731 | NewFn = Intrinsic::getOrInsertDeclaration(M: F->getParent(), id: ID); |
| 1732 | return true; |
| 1733 | } |
| 1734 | break; // No other applicable upgrades. |
| 1735 | } |
| 1736 | |
| 1737 | ID = StringSwitch<Intrinsic::ID>(Name) |
| 1738 | .StartsWith(S: "sha256sig0" , Value: Intrinsic::riscv_sha256sig0) |
| 1739 | .StartsWith(S: "sha256sig1" , Value: Intrinsic::riscv_sha256sig1) |
| 1740 | .StartsWith(S: "sha256sum0" , Value: Intrinsic::riscv_sha256sum0) |
| 1741 | .StartsWith(S: "sha256sum1" , Value: Intrinsic::riscv_sha256sum1) |
| 1742 | .StartsWith(S: "sm3p0" , Value: Intrinsic::riscv_sm3p0) |
| 1743 | .StartsWith(S: "sm3p1" , Value: Intrinsic::riscv_sm3p1) |
| 1744 | .Default(Value: Intrinsic::not_intrinsic); |
| 1745 | if (ID != Intrinsic::not_intrinsic) { |
| 1746 | if (F->getFunctionType()->getReturnType()->isIntegerTy(Bitwidth: 64)) { |
| 1747 | rename(GV: F); |
| 1748 | NewFn = Intrinsic::getOrInsertDeclaration(M: F->getParent(), id: ID); |
| 1749 | return true; |
| 1750 | } |
| 1751 | break; // No other applicable upgrades. |
| 1752 | } |
| 1753 | |
| 1754 | // Replace llvm.riscv.clmul with llvm.clmul. |
| 1755 | if (Name == "clmul.i32" || Name == "clmul.i64" ) { |
| 1756 | NewFn = Intrinsic::getOrInsertDeclaration( |
| 1757 | M: F->getParent(), id: Intrinsic::clmul, Tys: {F->getReturnType()}); |
| 1758 | return true; |
| 1759 | } |
| 1760 | |
| 1761 | break; // No other 'riscv.*' intrinsics |
| 1762 | } |
| 1763 | } break; |
| 1764 | |
| 1765 | case 's': |
| 1766 | if (Name == "stackprotectorcheck" ) { |
| 1767 | NewFn = nullptr; |
| 1768 | return true; |
| 1769 | } |
| 1770 | break; |
| 1771 | |
| 1772 | case 't': |
| 1773 | if (Name == "thread.pointer" ) { |
| 1774 | NewFn = Intrinsic::getOrInsertDeclaration( |
| 1775 | M: F->getParent(), id: Intrinsic::thread_pointer, Tys: F->getReturnType()); |
| 1776 | return true; |
| 1777 | } |
| 1778 | break; |
| 1779 | |
| 1780 | case 'v': { |
| 1781 | if (Name == "var.annotation" && F->arg_size() == 4) { |
| 1782 | rename(GV: F); |
| 1783 | NewFn = Intrinsic::getOrInsertDeclaration( |
| 1784 | M: F->getParent(), id: Intrinsic::var_annotation, |
| 1785 | Tys: {{F->arg_begin()->getType(), F->getArg(i: 1)->getType()}}); |
| 1786 | return true; |
| 1787 | } |
| 1788 | if (Name.consume_front(Prefix: "vector.splice" )) { |
| 1789 | if (Name.starts_with(Prefix: ".left" ) || Name.starts_with(Prefix: ".right" )) |
| 1790 | break; |
| 1791 | return true; |
| 1792 | } |
| 1793 | break; |
| 1794 | } |
| 1795 | |
| 1796 | case 'w': |
| 1797 | if (Name.consume_front(Prefix: "wasm." )) { |
| 1798 | Intrinsic::ID ID = |
| 1799 | StringSwitch<Intrinsic::ID>(Name) |
| 1800 | .StartsWith(S: "fma." , Value: Intrinsic::wasm_relaxed_madd) |
| 1801 | .StartsWith(S: "fms." , Value: Intrinsic::wasm_relaxed_nmadd) |
| 1802 | .StartsWith(S: "laneselect." , Value: Intrinsic::wasm_relaxed_laneselect) |
| 1803 | .Default(Value: Intrinsic::not_intrinsic); |
| 1804 | if (ID != Intrinsic::not_intrinsic) { |
| 1805 | rename(GV: F); |
| 1806 | NewFn = Intrinsic::getOrInsertDeclaration(M: F->getParent(), id: ID, |
| 1807 | Tys: F->getReturnType()); |
| 1808 | return true; |
| 1809 | } |
| 1810 | |
| 1811 | if (Name.consume_front(Prefix: "dot.i8x16.i7x16." )) { |
| 1812 | ID = StringSwitch<Intrinsic::ID>(Name) |
| 1813 | .Case(S: "signed" , Value: Intrinsic::wasm_relaxed_dot_i8x16_i7x16_signed) |
| 1814 | .Case(S: "add.signed" , |
| 1815 | Value: Intrinsic::wasm_relaxed_dot_i8x16_i7x16_add_signed) |
| 1816 | .Default(Value: Intrinsic::not_intrinsic); |
| 1817 | if (ID != Intrinsic::not_intrinsic) { |
| 1818 | rename(GV: F); |
| 1819 | NewFn = Intrinsic::getOrInsertDeclaration(M: F->getParent(), id: ID); |
| 1820 | return true; |
| 1821 | } |
| 1822 | break; // No other 'wasm.dot.i8x16.i7x16.*'. |
| 1823 | } |
| 1824 | break; // No other 'wasm.*'. |
| 1825 | } |
| 1826 | break; |
| 1827 | |
| 1828 | case 'x': |
| 1829 | if (upgradeX86IntrinsicFunction(F, Name, NewFn)) |
| 1830 | return true; |
| 1831 | } |
| 1832 | |
| 1833 | auto *ST = dyn_cast<StructType>(Val: F->getReturnType()); |
| 1834 | if (ST && (!ST->isLiteral() || ST->isPacked()) && |
| 1835 | F->getIntrinsicID() != Intrinsic::not_intrinsic) { |
| 1836 | // Replace return type with literal non-packed struct. Only do this for |
| 1837 | // intrinsics declared to return a struct, not for intrinsics with |
| 1838 | // overloaded return type, in which case the exact struct type will be |
| 1839 | // mangled into the name. |
| 1840 | SmallVector<Intrinsic::IITDescriptor> Desc; |
| 1841 | Intrinsic::getIntrinsicInfoTableEntries(id: F->getIntrinsicID(), T&: Desc); |
| 1842 | if (Desc.front().Kind == Intrinsic::IITDescriptor::Struct) { |
| 1843 | auto *FT = F->getFunctionType(); |
| 1844 | auto *NewST = StructType::get(Context&: ST->getContext(), Elements: ST->elements()); |
| 1845 | auto *NewFT = FunctionType::get(Result: NewST, Params: FT->params(), isVarArg: FT->isVarArg()); |
| 1846 | std::string Name = F->getName().str(); |
| 1847 | rename(GV: F); |
| 1848 | NewFn = Function::Create(Ty: NewFT, Linkage: F->getLinkage(), AddrSpace: F->getAddressSpace(), |
| 1849 | N: Name, M: F->getParent()); |
| 1850 | |
| 1851 | // The new function may also need remangling. |
| 1852 | if (auto Result = llvm::Intrinsic::remangleIntrinsicFunction(F: NewFn)) |
| 1853 | NewFn = *Result; |
| 1854 | return true; |
| 1855 | } |
| 1856 | } |
| 1857 | |
| 1858 | // Remangle our intrinsic since we upgrade the mangling |
| 1859 | auto Result = llvm::Intrinsic::remangleIntrinsicFunction(F); |
| 1860 | if (Result != std::nullopt) { |
| 1861 | NewFn = *Result; |
| 1862 | return true; |
| 1863 | } |
| 1864 | |
| 1865 | // This may not belong here. This function is effectively being overloaded |
| 1866 | // to both detect an intrinsic which needs upgrading, and to provide the |
| 1867 | // upgraded form of the intrinsic. We should perhaps have two separate |
| 1868 | // functions for this. |
| 1869 | return false; |
| 1870 | } |
| 1871 | |
| 1872 | bool llvm::UpgradeIntrinsicFunction(Function *F, Function *&NewFn, |
| 1873 | bool CanUpgradeDebugIntrinsicsToRecords) { |
| 1874 | NewFn = nullptr; |
| 1875 | bool Upgraded = |
| 1876 | upgradeIntrinsicFunction1(F, NewFn, CanUpgradeDebugIntrinsicsToRecords); |
| 1877 | |
| 1878 | // Upgrade intrinsic attributes. This does not change the function. |
| 1879 | if (NewFn) |
| 1880 | F = NewFn; |
| 1881 | if (Intrinsic::ID id = F->getIntrinsicID()) { |
| 1882 | // Only do this if the intrinsic signature is valid. |
| 1883 | SmallVector<Type *> OverloadTys; |
| 1884 | if (Intrinsic::getIntrinsicSignature(id, FT: F->getFunctionType(), ArgTys&: OverloadTys)) |
| 1885 | F->setAttributes( |
| 1886 | Intrinsic::getAttributes(C&: F->getContext(), id, FT: F->getFunctionType())); |
| 1887 | } |
| 1888 | return Upgraded; |
| 1889 | } |
| 1890 | |
| 1891 | GlobalVariable *llvm::UpgradeGlobalVariable(GlobalVariable *GV) { |
| 1892 | if (!(GV->hasName() && (GV->getName() == "llvm.global_ctors" || |
| 1893 | GV->getName() == "llvm.global_dtors" )) || |
| 1894 | !GV->hasInitializer()) |
| 1895 | return nullptr; |
| 1896 | ArrayType *ATy = dyn_cast<ArrayType>(Val: GV->getValueType()); |
| 1897 | if (!ATy) |
| 1898 | return nullptr; |
| 1899 | StructType *STy = dyn_cast<StructType>(Val: ATy->getElementType()); |
| 1900 | if (!STy || STy->getNumElements() != 2) |
| 1901 | return nullptr; |
| 1902 | |
| 1903 | LLVMContext &C = GV->getContext(); |
| 1904 | IRBuilder<> IRB(C); |
| 1905 | auto EltTy = StructType::get(elt1: STy->getElementType(N: 0), elts: STy->getElementType(N: 1), |
| 1906 | elts: IRB.getPtrTy()); |
| 1907 | Constant *Init = GV->getInitializer(); |
| 1908 | unsigned N = Init->getNumOperands(); |
| 1909 | std::vector<Constant *> NewCtors(N); |
| 1910 | for (unsigned i = 0; i != N; ++i) { |
| 1911 | auto Ctor = cast<Constant>(Val: Init->getOperand(i)); |
| 1912 | NewCtors[i] = ConstantStruct::get(T: EltTy, Vs: Ctor->getAggregateElement(Elt: 0u), |
| 1913 | Vs: Ctor->getAggregateElement(Elt: 1), |
| 1914 | Vs: ConstantPointerNull::get(T: IRB.getPtrTy())); |
| 1915 | } |
| 1916 | Constant *NewInit = ConstantArray::get(T: ArrayType::get(ElementType: EltTy, NumElements: N), V: NewCtors); |
| 1917 | |
| 1918 | return new GlobalVariable(NewInit->getType(), false, GV->getLinkage(), |
| 1919 | NewInit, GV->getName()); |
| 1920 | } |
| 1921 | |
| 1922 | // Handles upgrading SSE2/AVX2/AVX512BW PSLLDQ intrinsics by converting them |
| 1923 | // to byte shuffles. |
| 1924 | static Value *upgradeX86PSLLDQIntrinsics(IRBuilder<> &Builder, Value *Op, |
| 1925 | unsigned Shift) { |
| 1926 | auto *ResultTy = cast<FixedVectorType>(Val: Op->getType()); |
| 1927 | unsigned NumElts = ResultTy->getNumElements() * 8; |
| 1928 | |
| 1929 | // Bitcast from a 64-bit element type to a byte element type. |
| 1930 | Type *VecTy = FixedVectorType::get(ElementType: Builder.getInt8Ty(), NumElts); |
| 1931 | Op = Builder.CreateBitCast(V: Op, DestTy: VecTy, Name: "cast" ); |
| 1932 | |
| 1933 | // We'll be shuffling in zeroes. |
| 1934 | Value *Res = Constant::getNullValue(Ty: VecTy); |
| 1935 | |
| 1936 | // If shift is less than 16, emit a shuffle to move the bytes. Otherwise, |
| 1937 | // we'll just return the zero vector. |
| 1938 | if (Shift < 16) { |
| 1939 | int Idxs[64]; |
| 1940 | // 256/512-bit version is split into 2/4 16-byte lanes. |
| 1941 | for (unsigned l = 0; l != NumElts; l += 16) |
| 1942 | for (unsigned i = 0; i != 16; ++i) { |
| 1943 | unsigned Idx = NumElts + i - Shift; |
| 1944 | if (Idx < NumElts) |
| 1945 | Idx -= NumElts - 16; // end of lane, switch operand. |
| 1946 | Idxs[l + i] = Idx + l; |
| 1947 | } |
| 1948 | |
| 1949 | Res = Builder.CreateShuffleVector(V1: Res, V2: Op, Mask: ArrayRef(Idxs, NumElts)); |
| 1950 | } |
| 1951 | |
| 1952 | // Bitcast back to a 64-bit element type. |
| 1953 | return Builder.CreateBitCast(V: Res, DestTy: ResultTy, Name: "cast" ); |
| 1954 | } |
| 1955 | |
| 1956 | // Handles upgrading SSE2/AVX2/AVX512BW PSRLDQ intrinsics by converting them |
| 1957 | // to byte shuffles. |
| 1958 | static Value *upgradeX86PSRLDQIntrinsics(IRBuilder<> &Builder, Value *Op, |
| 1959 | unsigned Shift) { |
| 1960 | auto *ResultTy = cast<FixedVectorType>(Val: Op->getType()); |
| 1961 | unsigned NumElts = ResultTy->getNumElements() * 8; |
| 1962 | |
| 1963 | // Bitcast from a 64-bit element type to a byte element type. |
| 1964 | Type *VecTy = FixedVectorType::get(ElementType: Builder.getInt8Ty(), NumElts); |
| 1965 | Op = Builder.CreateBitCast(V: Op, DestTy: VecTy, Name: "cast" ); |
| 1966 | |
| 1967 | // We'll be shuffling in zeroes. |
| 1968 | Value *Res = Constant::getNullValue(Ty: VecTy); |
| 1969 | |
| 1970 | // If shift is less than 16, emit a shuffle to move the bytes. Otherwise, |
| 1971 | // we'll just return the zero vector. |
| 1972 | if (Shift < 16) { |
| 1973 | int Idxs[64]; |
| 1974 | // 256/512-bit version is split into 2/4 16-byte lanes. |
| 1975 | for (unsigned l = 0; l != NumElts; l += 16) |
| 1976 | for (unsigned i = 0; i != 16; ++i) { |
| 1977 | unsigned Idx = i + Shift; |
| 1978 | if (Idx >= 16) |
| 1979 | Idx += NumElts - 16; // end of lane, switch operand. |
| 1980 | Idxs[l + i] = Idx + l; |
| 1981 | } |
| 1982 | |
| 1983 | Res = Builder.CreateShuffleVector(V1: Op, V2: Res, Mask: ArrayRef(Idxs, NumElts)); |
| 1984 | } |
| 1985 | |
| 1986 | // Bitcast back to a 64-bit element type. |
| 1987 | return Builder.CreateBitCast(V: Res, DestTy: ResultTy, Name: "cast" ); |
| 1988 | } |
| 1989 | |
| 1990 | static Value *getX86MaskVec(IRBuilder<> &Builder, Value *Mask, |
| 1991 | unsigned NumElts) { |
| 1992 | assert(isPowerOf2_32(NumElts) && "Expected power-of-2 mask elements" ); |
| 1993 | llvm::VectorType *MaskTy = FixedVectorType::get( |
| 1994 | ElementType: Builder.getInt1Ty(), NumElts: cast<IntegerType>(Val: Mask->getType())->getBitWidth()); |
| 1995 | Mask = Builder.CreateBitCast(V: Mask, DestTy: MaskTy); |
| 1996 | |
| 1997 | // If we have less than 8 elements (1, 2 or 4), then the starting mask was an |
| 1998 | // i8 and we need to extract down to the right number of elements. |
| 1999 | if (NumElts <= 4) { |
| 2000 | int Indices[4]; |
| 2001 | for (unsigned i = 0; i != NumElts; ++i) |
| 2002 | Indices[i] = i; |
| 2003 | Mask = Builder.CreateShuffleVector(V1: Mask, V2: Mask, Mask: ArrayRef(Indices, NumElts), |
| 2004 | Name: "extract" ); |
| 2005 | } |
| 2006 | |
| 2007 | return Mask; |
| 2008 | } |
| 2009 | |
| 2010 | static Value *emitX86Select(IRBuilder<> &Builder, Value *Mask, Value *Op0, |
| 2011 | Value *Op1) { |
| 2012 | // If the mask is all ones just emit the first operation. |
| 2013 | if (const auto *C = dyn_cast<Constant>(Val: Mask)) |
| 2014 | if (C->isAllOnesValue()) |
| 2015 | return Op0; |
| 2016 | |
| 2017 | Mask = getX86MaskVec(Builder, Mask, |
| 2018 | NumElts: cast<FixedVectorType>(Val: Op0->getType())->getNumElements()); |
| 2019 | return Builder.CreateSelect(C: Mask, True: Op0, False: Op1); |
| 2020 | } |
| 2021 | |
| 2022 | static Value *emitX86ScalarSelect(IRBuilder<> &Builder, Value *Mask, Value *Op0, |
| 2023 | Value *Op1) { |
| 2024 | // If the mask is all ones just emit the first operation. |
| 2025 | if (const auto *C = dyn_cast<Constant>(Val: Mask)) |
| 2026 | if (C->isAllOnesValue()) |
| 2027 | return Op0; |
| 2028 | |
| 2029 | auto *MaskTy = FixedVectorType::get(ElementType: Builder.getInt1Ty(), |
| 2030 | NumElts: Mask->getType()->getIntegerBitWidth()); |
| 2031 | Mask = Builder.CreateBitCast(V: Mask, DestTy: MaskTy); |
| 2032 | Mask = Builder.CreateExtractElement(Vec: Mask, Idx: (uint64_t)0); |
| 2033 | return Builder.CreateSelect(C: Mask, True: Op0, False: Op1); |
| 2034 | } |
| 2035 | |
| 2036 | // Handle autoupgrade for masked PALIGNR and VALIGND/Q intrinsics. |
| 2037 | // PALIGNR handles large immediates by shifting while VALIGN masks the immediate |
| 2038 | // so we need to handle both cases. VALIGN also doesn't have 128-bit lanes. |
| 2039 | static Value *upgradeX86ALIGNIntrinsics(IRBuilder<> &Builder, Value *Op0, |
| 2040 | Value *Op1, Value *Shift, |
| 2041 | Value *Passthru, Value *Mask, |
| 2042 | bool IsVALIGN) { |
| 2043 | unsigned ShiftVal = cast<llvm::ConstantInt>(Val: Shift)->getZExtValue(); |
| 2044 | |
| 2045 | unsigned NumElts = cast<FixedVectorType>(Val: Op0->getType())->getNumElements(); |
| 2046 | assert((IsVALIGN || NumElts % 16 == 0) && "Illegal NumElts for PALIGNR!" ); |
| 2047 | assert((!IsVALIGN || NumElts <= 16) && "NumElts too large for VALIGN!" ); |
| 2048 | assert(isPowerOf2_32(NumElts) && "NumElts not a power of 2!" ); |
| 2049 | |
| 2050 | // Mask the immediate for VALIGN. |
| 2051 | if (IsVALIGN) |
| 2052 | ShiftVal &= (NumElts - 1); |
| 2053 | |
| 2054 | // If palignr is shifting the pair of vectors more than the size of two |
| 2055 | // lanes, emit zero. |
| 2056 | if (ShiftVal >= 32) |
| 2057 | return llvm::Constant::getNullValue(Ty: Op0->getType()); |
| 2058 | |
| 2059 | // If palignr is shifting the pair of input vectors more than one lane, |
| 2060 | // but less than two lanes, convert to shifting in zeroes. |
| 2061 | if (ShiftVal > 16) { |
| 2062 | ShiftVal -= 16; |
| 2063 | Op1 = Op0; |
| 2064 | Op0 = llvm::Constant::getNullValue(Ty: Op0->getType()); |
| 2065 | } |
| 2066 | |
| 2067 | int Indices[64]; |
| 2068 | // 256-bit palignr operates on 128-bit lanes so we need to handle that |
| 2069 | for (unsigned l = 0; l < NumElts; l += 16) { |
| 2070 | for (unsigned i = 0; i != 16; ++i) { |
| 2071 | unsigned Idx = ShiftVal + i; |
| 2072 | if (!IsVALIGN && Idx >= 16) // Disable wrap for VALIGN. |
| 2073 | Idx += NumElts - 16; // End of lane, switch operand. |
| 2074 | Indices[l + i] = Idx + l; |
| 2075 | } |
| 2076 | } |
| 2077 | |
| 2078 | Value *Align = Builder.CreateShuffleVector( |
| 2079 | V1: Op1, V2: Op0, Mask: ArrayRef(Indices, NumElts), Name: "palignr" ); |
| 2080 | |
| 2081 | return emitX86Select(Builder, Mask, Op0: Align, Op1: Passthru); |
| 2082 | } |
| 2083 | |
| 2084 | static Value *upgradeX86VPERMT2Intrinsics(IRBuilder<> &Builder, CallBase &CI, |
| 2085 | bool ZeroMask, bool IndexForm) { |
| 2086 | Type *Ty = CI.getType(); |
| 2087 | unsigned VecWidth = Ty->getPrimitiveSizeInBits(); |
| 2088 | unsigned EltWidth = Ty->getScalarSizeInBits(); |
| 2089 | bool IsFloat = Ty->isFPOrFPVectorTy(); |
| 2090 | Intrinsic::ID IID; |
| 2091 | if (VecWidth == 128 && EltWidth == 32 && IsFloat) |
| 2092 | IID = Intrinsic::x86_avx512_vpermi2var_ps_128; |
| 2093 | else if (VecWidth == 128 && EltWidth == 32 && !IsFloat) |
| 2094 | IID = Intrinsic::x86_avx512_vpermi2var_d_128; |
| 2095 | else if (VecWidth == 128 && EltWidth == 64 && IsFloat) |
| 2096 | IID = Intrinsic::x86_avx512_vpermi2var_pd_128; |
| 2097 | else if (VecWidth == 128 && EltWidth == 64 && !IsFloat) |
| 2098 | IID = Intrinsic::x86_avx512_vpermi2var_q_128; |
| 2099 | else if (VecWidth == 256 && EltWidth == 32 && IsFloat) |
| 2100 | IID = Intrinsic::x86_avx512_vpermi2var_ps_256; |
| 2101 | else if (VecWidth == 256 && EltWidth == 32 && !IsFloat) |
| 2102 | IID = Intrinsic::x86_avx512_vpermi2var_d_256; |
| 2103 | else if (VecWidth == 256 && EltWidth == 64 && IsFloat) |
| 2104 | IID = Intrinsic::x86_avx512_vpermi2var_pd_256; |
| 2105 | else if (VecWidth == 256 && EltWidth == 64 && !IsFloat) |
| 2106 | IID = Intrinsic::x86_avx512_vpermi2var_q_256; |
| 2107 | else if (VecWidth == 512 && EltWidth == 32 && IsFloat) |
| 2108 | IID = Intrinsic::x86_avx512_vpermi2var_ps_512; |
| 2109 | else if (VecWidth == 512 && EltWidth == 32 && !IsFloat) |
| 2110 | IID = Intrinsic::x86_avx512_vpermi2var_d_512; |
| 2111 | else if (VecWidth == 512 && EltWidth == 64 && IsFloat) |
| 2112 | IID = Intrinsic::x86_avx512_vpermi2var_pd_512; |
| 2113 | else if (VecWidth == 512 && EltWidth == 64 && !IsFloat) |
| 2114 | IID = Intrinsic::x86_avx512_vpermi2var_q_512; |
| 2115 | else if (VecWidth == 128 && EltWidth == 16) |
| 2116 | IID = Intrinsic::x86_avx512_vpermi2var_hi_128; |
| 2117 | else if (VecWidth == 256 && EltWidth == 16) |
| 2118 | IID = Intrinsic::x86_avx512_vpermi2var_hi_256; |
| 2119 | else if (VecWidth == 512 && EltWidth == 16) |
| 2120 | IID = Intrinsic::x86_avx512_vpermi2var_hi_512; |
| 2121 | else if (VecWidth == 128 && EltWidth == 8) |
| 2122 | IID = Intrinsic::x86_avx512_vpermi2var_qi_128; |
| 2123 | else if (VecWidth == 256 && EltWidth == 8) |
| 2124 | IID = Intrinsic::x86_avx512_vpermi2var_qi_256; |
| 2125 | else if (VecWidth == 512 && EltWidth == 8) |
| 2126 | IID = Intrinsic::x86_avx512_vpermi2var_qi_512; |
| 2127 | else |
| 2128 | llvm_unreachable("Unexpected intrinsic" ); |
| 2129 | |
| 2130 | Value *Args[] = { CI.getArgOperand(i: 0) , CI.getArgOperand(i: 1), |
| 2131 | CI.getArgOperand(i: 2) }; |
| 2132 | |
| 2133 | // If this isn't index form we need to swap operand 0 and 1. |
| 2134 | if (!IndexForm) |
| 2135 | std::swap(a&: Args[0], b&: Args[1]); |
| 2136 | |
| 2137 | Value *V = Builder.CreateIntrinsic(ID: IID, Args); |
| 2138 | Value *PassThru = ZeroMask ? ConstantAggregateZero::get(Ty) |
| 2139 | : Builder.CreateBitCast(V: CI.getArgOperand(i: 1), |
| 2140 | DestTy: Ty); |
| 2141 | return emitX86Select(Builder, Mask: CI.getArgOperand(i: 3), Op0: V, Op1: PassThru); |
| 2142 | } |
| 2143 | |
| 2144 | static Value *upgradeX86BinaryIntrinsics(IRBuilder<> &Builder, CallBase &CI, |
| 2145 | Intrinsic::ID IID) { |
| 2146 | Type *Ty = CI.getType(); |
| 2147 | Value *Op0 = CI.getOperand(i_nocapture: 0); |
| 2148 | Value *Op1 = CI.getOperand(i_nocapture: 1); |
| 2149 | Value *Res = Builder.CreateIntrinsic(ID: IID, Types: Ty, Args: {Op0, Op1}); |
| 2150 | |
| 2151 | if (CI.arg_size() == 4) { // For masked intrinsics. |
| 2152 | Value *VecSrc = CI.getOperand(i_nocapture: 2); |
| 2153 | Value *Mask = CI.getOperand(i_nocapture: 3); |
| 2154 | Res = emitX86Select(Builder, Mask, Op0: Res, Op1: VecSrc); |
| 2155 | } |
| 2156 | return Res; |
| 2157 | } |
| 2158 | |
| 2159 | static Value *upgradeX86Rotate(IRBuilder<> &Builder, CallBase &CI, |
| 2160 | bool IsRotateRight) { |
| 2161 | Type *Ty = CI.getType(); |
| 2162 | Value *Src = CI.getArgOperand(i: 0); |
| 2163 | Value *Amt = CI.getArgOperand(i: 1); |
| 2164 | |
| 2165 | // Amount may be scalar immediate, in which case create a splat vector. |
| 2166 | // Funnel shifts amounts are treated as modulo and types are all power-of-2 so |
| 2167 | // we only care about the lowest log2 bits anyway. |
| 2168 | if (Amt->getType() != Ty) { |
| 2169 | unsigned NumElts = cast<FixedVectorType>(Val: Ty)->getNumElements(); |
| 2170 | Amt = Builder.CreateIntCast(V: Amt, DestTy: Ty->getScalarType(), isSigned: false); |
| 2171 | Amt = Builder.CreateVectorSplat(NumElts, V: Amt); |
| 2172 | } |
| 2173 | |
| 2174 | Intrinsic::ID IID = IsRotateRight ? Intrinsic::fshr : Intrinsic::fshl; |
| 2175 | Value *Res = Builder.CreateIntrinsic(ID: IID, Types: Ty, Args: {Src, Src, Amt}); |
| 2176 | |
| 2177 | if (CI.arg_size() == 4) { // For masked intrinsics. |
| 2178 | Value *VecSrc = CI.getOperand(i_nocapture: 2); |
| 2179 | Value *Mask = CI.getOperand(i_nocapture: 3); |
| 2180 | Res = emitX86Select(Builder, Mask, Op0: Res, Op1: VecSrc); |
| 2181 | } |
| 2182 | return Res; |
| 2183 | } |
| 2184 | |
| 2185 | static Value *upgradeX86vpcom(IRBuilder<> &Builder, CallBase &CI, unsigned Imm, |
| 2186 | bool IsSigned) { |
| 2187 | Type *Ty = CI.getType(); |
| 2188 | Value *LHS = CI.getArgOperand(i: 0); |
| 2189 | Value *RHS = CI.getArgOperand(i: 1); |
| 2190 | |
| 2191 | CmpInst::Predicate Pred; |
| 2192 | switch (Imm) { |
| 2193 | case 0x0: |
| 2194 | Pred = IsSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; |
| 2195 | break; |
| 2196 | case 0x1: |
| 2197 | Pred = IsSigned ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE; |
| 2198 | break; |
| 2199 | case 0x2: |
| 2200 | Pred = IsSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; |
| 2201 | break; |
| 2202 | case 0x3: |
| 2203 | Pred = IsSigned ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE; |
| 2204 | break; |
| 2205 | case 0x4: |
| 2206 | Pred = ICmpInst::ICMP_EQ; |
| 2207 | break; |
| 2208 | case 0x5: |
| 2209 | Pred = ICmpInst::ICMP_NE; |
| 2210 | break; |
| 2211 | case 0x6: |
| 2212 | return Constant::getNullValue(Ty); // FALSE |
| 2213 | case 0x7: |
| 2214 | return Constant::getAllOnesValue(Ty); // TRUE |
| 2215 | default: |
| 2216 | llvm_unreachable("Unknown XOP vpcom/vpcomu predicate" ); |
| 2217 | } |
| 2218 | |
| 2219 | Value *Cmp = Builder.CreateICmp(P: Pred, LHS, RHS); |
| 2220 | Value *Ext = Builder.CreateSExt(V: Cmp, DestTy: Ty); |
| 2221 | return Ext; |
| 2222 | } |
| 2223 | |
| 2224 | static Value *upgradeX86ConcatShift(IRBuilder<> &Builder, CallBase &CI, |
| 2225 | bool IsShiftRight, bool ZeroMask) { |
| 2226 | Type *Ty = CI.getType(); |
| 2227 | Value *Op0 = CI.getArgOperand(i: 0); |
| 2228 | Value *Op1 = CI.getArgOperand(i: 1); |
| 2229 | Value *Amt = CI.getArgOperand(i: 2); |
| 2230 | |
| 2231 | if (IsShiftRight) |
| 2232 | std::swap(a&: Op0, b&: Op1); |
| 2233 | |
| 2234 | // Amount may be scalar immediate, in which case create a splat vector. |
| 2235 | // Funnel shifts amounts are treated as modulo and types are all power-of-2 so |
| 2236 | // we only care about the lowest log2 bits anyway. |
| 2237 | if (Amt->getType() != Ty) { |
| 2238 | unsigned NumElts = cast<FixedVectorType>(Val: Ty)->getNumElements(); |
| 2239 | Amt = Builder.CreateIntCast(V: Amt, DestTy: Ty->getScalarType(), isSigned: false); |
| 2240 | Amt = Builder.CreateVectorSplat(NumElts, V: Amt); |
| 2241 | } |
| 2242 | |
| 2243 | Intrinsic::ID IID = IsShiftRight ? Intrinsic::fshr : Intrinsic::fshl; |
| 2244 | Value *Res = Builder.CreateIntrinsic(ID: IID, Types: Ty, Args: {Op0, Op1, Amt}); |
| 2245 | |
| 2246 | unsigned NumArgs = CI.arg_size(); |
| 2247 | if (NumArgs >= 4) { // For masked intrinsics. |
| 2248 | Value *VecSrc = NumArgs == 5 ? CI.getArgOperand(i: 3) : |
| 2249 | ZeroMask ? ConstantAggregateZero::get(Ty: CI.getType()) : |
| 2250 | CI.getArgOperand(i: 0); |
| 2251 | Value *Mask = CI.getOperand(i_nocapture: NumArgs - 1); |
| 2252 | Res = emitX86Select(Builder, Mask, Op0: Res, Op1: VecSrc); |
| 2253 | } |
| 2254 | return Res; |
| 2255 | } |
| 2256 | |
| 2257 | static Value *upgradeMaskedStore(IRBuilder<> &Builder, Value *Ptr, Value *Data, |
| 2258 | Value *Mask, bool Aligned) { |
| 2259 | const Align Alignment = |
| 2260 | Aligned |
| 2261 | ? Align(Data->getType()->getPrimitiveSizeInBits().getFixedValue() / 8) |
| 2262 | : Align(1); |
| 2263 | |
| 2264 | // If the mask is all ones just emit a regular store. |
| 2265 | if (const auto *C = dyn_cast<Constant>(Val: Mask)) |
| 2266 | if (C->isAllOnesValue()) |
| 2267 | return Builder.CreateAlignedStore(Val: Data, Ptr, Align: Alignment); |
| 2268 | |
| 2269 | // Convert the mask from an integer type to a vector of i1. |
| 2270 | unsigned NumElts = cast<FixedVectorType>(Val: Data->getType())->getNumElements(); |
| 2271 | Mask = getX86MaskVec(Builder, Mask, NumElts); |
| 2272 | return Builder.CreateMaskedStore(Val: Data, Ptr, Alignment, Mask); |
| 2273 | } |
| 2274 | |
| 2275 | static Value *upgradeMaskedLoad(IRBuilder<> &Builder, Value *Ptr, |
| 2276 | Value *Passthru, Value *Mask, bool Aligned) { |
| 2277 | Type *ValTy = Passthru->getType(); |
| 2278 | const Align Alignment = |
| 2279 | Aligned |
| 2280 | ? Align( |
| 2281 | Passthru->getType()->getPrimitiveSizeInBits().getFixedValue() / |
| 2282 | 8) |
| 2283 | : Align(1); |
| 2284 | |
| 2285 | // If the mask is all ones just emit a regular store. |
| 2286 | if (const auto *C = dyn_cast<Constant>(Val: Mask)) |
| 2287 | if (C->isAllOnesValue()) |
| 2288 | return Builder.CreateAlignedLoad(Ty: ValTy, Ptr, Align: Alignment); |
| 2289 | |
| 2290 | // Convert the mask from an integer type to a vector of i1. |
| 2291 | unsigned NumElts = cast<FixedVectorType>(Val: ValTy)->getNumElements(); |
| 2292 | Mask = getX86MaskVec(Builder, Mask, NumElts); |
| 2293 | return Builder.CreateMaskedLoad(Ty: ValTy, Ptr, Alignment, Mask, PassThru: Passthru); |
| 2294 | } |
| 2295 | |
| 2296 | static Value *upgradeAbs(IRBuilder<> &Builder, CallBase &CI) { |
| 2297 | Type *Ty = CI.getType(); |
| 2298 | Value *Op0 = CI.getArgOperand(i: 0); |
| 2299 | Value *Res = Builder.CreateIntrinsic(ID: Intrinsic::abs, Types: Ty, |
| 2300 | Args: {Op0, Builder.getInt1(V: false)}); |
| 2301 | if (CI.arg_size() == 3) |
| 2302 | Res = emitX86Select(Builder, Mask: CI.getArgOperand(i: 2), Op0: Res, Op1: CI.getArgOperand(i: 1)); |
| 2303 | return Res; |
| 2304 | } |
| 2305 | |
| 2306 | static Value *upgradePMULDQ(IRBuilder<> &Builder, CallBase &CI, bool IsSigned) { |
| 2307 | Type *Ty = CI.getType(); |
| 2308 | |
| 2309 | // Arguments have a vXi32 type so cast to vXi64. |
| 2310 | Value *LHS = Builder.CreateBitCast(V: CI.getArgOperand(i: 0), DestTy: Ty); |
| 2311 | Value *RHS = Builder.CreateBitCast(V: CI.getArgOperand(i: 1), DestTy: Ty); |
| 2312 | |
| 2313 | if (IsSigned) { |
| 2314 | // Shift left then arithmetic shift right. |
| 2315 | Constant *ShiftAmt = ConstantInt::get(Ty, V: 32); |
| 2316 | LHS = Builder.CreateShl(LHS, RHS: ShiftAmt); |
| 2317 | LHS = Builder.CreateAShr(LHS, RHS: ShiftAmt); |
| 2318 | RHS = Builder.CreateShl(LHS: RHS, RHS: ShiftAmt); |
| 2319 | RHS = Builder.CreateAShr(LHS: RHS, RHS: ShiftAmt); |
| 2320 | } else { |
| 2321 | // Clear the upper bits. |
| 2322 | Constant *Mask = ConstantInt::get(Ty, V: 0xffffffff); |
| 2323 | LHS = Builder.CreateAnd(LHS, RHS: Mask); |
| 2324 | RHS = Builder.CreateAnd(LHS: RHS, RHS: Mask); |
| 2325 | } |
| 2326 | |
| 2327 | Value *Res = Builder.CreateMul(LHS, RHS); |
| 2328 | |
| 2329 | if (CI.arg_size() == 4) |
| 2330 | Res = emitX86Select(Builder, Mask: CI.getArgOperand(i: 3), Op0: Res, Op1: CI.getArgOperand(i: 2)); |
| 2331 | |
| 2332 | return Res; |
| 2333 | } |
| 2334 | |
| 2335 | // Applying mask on vector of i1's and make sure result is at least 8 bits wide. |
| 2336 | static Value *applyX86MaskOn1BitsVec(IRBuilder<> &Builder, Value *Vec, |
| 2337 | Value *Mask) { |
| 2338 | unsigned NumElts = cast<FixedVectorType>(Val: Vec->getType())->getNumElements(); |
| 2339 | if (Mask) { |
| 2340 | const auto *C = dyn_cast<Constant>(Val: Mask); |
| 2341 | if (!C || !C->isAllOnesValue()) |
| 2342 | Vec = Builder.CreateAnd(LHS: Vec, RHS: getX86MaskVec(Builder, Mask, NumElts)); |
| 2343 | } |
| 2344 | |
| 2345 | if (NumElts < 8) { |
| 2346 | int Indices[8]; |
| 2347 | for (unsigned i = 0; i != NumElts; ++i) |
| 2348 | Indices[i] = i; |
| 2349 | for (unsigned i = NumElts; i != 8; ++i) |
| 2350 | Indices[i] = NumElts + i % NumElts; |
| 2351 | Vec = Builder.CreateShuffleVector(V1: Vec, |
| 2352 | V2: Constant::getNullValue(Ty: Vec->getType()), |
| 2353 | Mask: Indices); |
| 2354 | } |
| 2355 | return Builder.CreateBitCast(V: Vec, DestTy: Builder.getIntNTy(N: std::max(a: NumElts, b: 8U))); |
| 2356 | } |
| 2357 | |
| 2358 | static Value *upgradeMaskedCompare(IRBuilder<> &Builder, CallBase &CI, |
| 2359 | unsigned CC, bool Signed) { |
| 2360 | Value *Op0 = CI.getArgOperand(i: 0); |
| 2361 | unsigned NumElts = cast<FixedVectorType>(Val: Op0->getType())->getNumElements(); |
| 2362 | |
| 2363 | Value *Cmp; |
| 2364 | if (CC == 3) { |
| 2365 | Cmp = Constant::getNullValue( |
| 2366 | Ty: FixedVectorType::get(ElementType: Builder.getInt1Ty(), NumElts)); |
| 2367 | } else if (CC == 7) { |
| 2368 | Cmp = Constant::getAllOnesValue( |
| 2369 | Ty: FixedVectorType::get(ElementType: Builder.getInt1Ty(), NumElts)); |
| 2370 | } else { |
| 2371 | ICmpInst::Predicate Pred; |
| 2372 | switch (CC) { |
| 2373 | default: llvm_unreachable("Unknown condition code" ); |
| 2374 | case 0: Pred = ICmpInst::ICMP_EQ; break; |
| 2375 | case 1: Pred = Signed ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; break; |
| 2376 | case 2: Pred = Signed ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE; break; |
| 2377 | case 4: Pred = ICmpInst::ICMP_NE; break; |
| 2378 | case 5: Pred = Signed ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE; break; |
| 2379 | case 6: Pred = Signed ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; break; |
| 2380 | } |
| 2381 | Cmp = Builder.CreateICmp(P: Pred, LHS: Op0, RHS: CI.getArgOperand(i: 1)); |
| 2382 | } |
| 2383 | |
| 2384 | Value *Mask = CI.getArgOperand(i: CI.arg_size() - 1); |
| 2385 | |
| 2386 | return applyX86MaskOn1BitsVec(Builder, Vec: Cmp, Mask); |
| 2387 | } |
| 2388 | |
| 2389 | // Replace a masked intrinsic with an older unmasked intrinsic. |
| 2390 | static Value *upgradeX86MaskedShift(IRBuilder<> &Builder, CallBase &CI, |
| 2391 | Intrinsic::ID IID) { |
| 2392 | Value *Rep = |
| 2393 | Builder.CreateIntrinsic(ID: IID, Args: {CI.getArgOperand(i: 0), CI.getArgOperand(i: 1)}); |
| 2394 | return emitX86Select(Builder, Mask: CI.getArgOperand(i: 3), Op0: Rep, Op1: CI.getArgOperand(i: 2)); |
| 2395 | } |
| 2396 | |
| 2397 | static Value *upgradeMaskedMove(IRBuilder<> &Builder, CallBase &CI) { |
| 2398 | Value* A = CI.getArgOperand(i: 0); |
| 2399 | Value* B = CI.getArgOperand(i: 1); |
| 2400 | Value* Src = CI.getArgOperand(i: 2); |
| 2401 | Value* Mask = CI.getArgOperand(i: 3); |
| 2402 | |
| 2403 | Value* AndNode = Builder.CreateAnd(LHS: Mask, RHS: APInt(8, 1)); |
| 2404 | Value* Cmp = Builder.CreateIsNotNull(Arg: AndNode); |
| 2405 | Value* = Builder.CreateExtractElement(Vec: B, Idx: (uint64_t)0); |
| 2406 | Value* = Builder.CreateExtractElement(Vec: Src, Idx: (uint64_t)0); |
| 2407 | Value* Select = Builder.CreateSelect(C: Cmp, True: Extract1, False: Extract2); |
| 2408 | return Builder.CreateInsertElement(Vec: A, NewElt: Select, Idx: (uint64_t)0); |
| 2409 | } |
| 2410 | |
| 2411 | static Value *upgradeMaskToInt(IRBuilder<> &Builder, CallBase &CI) { |
| 2412 | Value* Op = CI.getArgOperand(i: 0); |
| 2413 | Type* ReturnOp = CI.getType(); |
| 2414 | unsigned NumElts = cast<FixedVectorType>(Val: CI.getType())->getNumElements(); |
| 2415 | Value *Mask = getX86MaskVec(Builder, Mask: Op, NumElts); |
| 2416 | return Builder.CreateSExt(V: Mask, DestTy: ReturnOp, Name: "vpmovm2" ); |
| 2417 | } |
| 2418 | |
| 2419 | // Replace intrinsic with unmasked version and a select. |
| 2420 | static bool upgradeAVX512MaskToSelect(StringRef Name, IRBuilder<> &Builder, |
| 2421 | CallBase &CI, Value *&Rep) { |
| 2422 | Name = Name.substr(Start: 12); // Remove avx512.mask. |
| 2423 | |
| 2424 | unsigned VecWidth = CI.getType()->getPrimitiveSizeInBits(); |
| 2425 | unsigned EltWidth = CI.getType()->getScalarSizeInBits(); |
| 2426 | Intrinsic::ID IID; |
| 2427 | if (Name.starts_with(Prefix: "max.p" )) { |
| 2428 | if (VecWidth == 128 && EltWidth == 32) |
| 2429 | IID = Intrinsic::x86_sse_max_ps; |
| 2430 | else if (VecWidth == 128 && EltWidth == 64) |
| 2431 | IID = Intrinsic::x86_sse2_max_pd; |
| 2432 | else if (VecWidth == 256 && EltWidth == 32) |
| 2433 | IID = Intrinsic::x86_avx_max_ps_256; |
| 2434 | else if (VecWidth == 256 && EltWidth == 64) |
| 2435 | IID = Intrinsic::x86_avx_max_pd_256; |
| 2436 | else |
| 2437 | llvm_unreachable("Unexpected intrinsic" ); |
| 2438 | } else if (Name.starts_with(Prefix: "min.p" )) { |
| 2439 | if (VecWidth == 128 && EltWidth == 32) |
| 2440 | IID = Intrinsic::x86_sse_min_ps; |
| 2441 | else if (VecWidth == 128 && EltWidth == 64) |
| 2442 | IID = Intrinsic::x86_sse2_min_pd; |
| 2443 | else if (VecWidth == 256 && EltWidth == 32) |
| 2444 | IID = Intrinsic::x86_avx_min_ps_256; |
| 2445 | else if (VecWidth == 256 && EltWidth == 64) |
| 2446 | IID = Intrinsic::x86_avx_min_pd_256; |
| 2447 | else |
| 2448 | llvm_unreachable("Unexpected intrinsic" ); |
| 2449 | } else if (Name.starts_with(Prefix: "pshuf.b." )) { |
| 2450 | if (VecWidth == 128) |
| 2451 | IID = Intrinsic::x86_ssse3_pshuf_b_128; |
| 2452 | else if (VecWidth == 256) |
| 2453 | IID = Intrinsic::x86_avx2_pshuf_b; |
| 2454 | else if (VecWidth == 512) |
| 2455 | IID = Intrinsic::x86_avx512_pshuf_b_512; |
| 2456 | else |
| 2457 | llvm_unreachable("Unexpected intrinsic" ); |
| 2458 | } else if (Name.starts_with(Prefix: "pmul.hr.sw." )) { |
| 2459 | if (VecWidth == 128) |
| 2460 | IID = Intrinsic::x86_ssse3_pmul_hr_sw_128; |
| 2461 | else if (VecWidth == 256) |
| 2462 | IID = Intrinsic::x86_avx2_pmul_hr_sw; |
| 2463 | else if (VecWidth == 512) |
| 2464 | IID = Intrinsic::x86_avx512_pmul_hr_sw_512; |
| 2465 | else |
| 2466 | llvm_unreachable("Unexpected intrinsic" ); |
| 2467 | } else if (Name.starts_with(Prefix: "pmulh.w." )) { |
| 2468 | if (VecWidth == 128) |
| 2469 | IID = Intrinsic::x86_sse2_pmulh_w; |
| 2470 | else if (VecWidth == 256) |
| 2471 | IID = Intrinsic::x86_avx2_pmulh_w; |
| 2472 | else if (VecWidth == 512) |
| 2473 | IID = Intrinsic::x86_avx512_pmulh_w_512; |
| 2474 | else |
| 2475 | llvm_unreachable("Unexpected intrinsic" ); |
| 2476 | } else if (Name.starts_with(Prefix: "pmulhu.w." )) { |
| 2477 | if (VecWidth == 128) |
| 2478 | IID = Intrinsic::x86_sse2_pmulhu_w; |
| 2479 | else if (VecWidth == 256) |
| 2480 | IID = Intrinsic::x86_avx2_pmulhu_w; |
| 2481 | else if (VecWidth == 512) |
| 2482 | IID = Intrinsic::x86_avx512_pmulhu_w_512; |
| 2483 | else |
| 2484 | llvm_unreachable("Unexpected intrinsic" ); |
| 2485 | } else if (Name.starts_with(Prefix: "pmaddw.d." )) { |
| 2486 | if (VecWidth == 128) |
| 2487 | IID = Intrinsic::x86_sse2_pmadd_wd; |
| 2488 | else if (VecWidth == 256) |
| 2489 | IID = Intrinsic::x86_avx2_pmadd_wd; |
| 2490 | else if (VecWidth == 512) |
| 2491 | IID = Intrinsic::x86_avx512_pmaddw_d_512; |
| 2492 | else |
| 2493 | llvm_unreachable("Unexpected intrinsic" ); |
| 2494 | } else if (Name.starts_with(Prefix: "pmaddubs.w." )) { |
| 2495 | if (VecWidth == 128) |
| 2496 | IID = Intrinsic::x86_ssse3_pmadd_ub_sw_128; |
| 2497 | else if (VecWidth == 256) |
| 2498 | IID = Intrinsic::x86_avx2_pmadd_ub_sw; |
| 2499 | else if (VecWidth == 512) |
| 2500 | IID = Intrinsic::x86_avx512_pmaddubs_w_512; |
| 2501 | else |
| 2502 | llvm_unreachable("Unexpected intrinsic" ); |
| 2503 | } else if (Name.starts_with(Prefix: "packsswb." )) { |
| 2504 | if (VecWidth == 128) |
| 2505 | IID = Intrinsic::x86_sse2_packsswb_128; |
| 2506 | else if (VecWidth == 256) |
| 2507 | IID = Intrinsic::x86_avx2_packsswb; |
| 2508 | else if (VecWidth == 512) |
| 2509 | IID = Intrinsic::x86_avx512_packsswb_512; |
| 2510 | else |
| 2511 | llvm_unreachable("Unexpected intrinsic" ); |
| 2512 | } else if (Name.starts_with(Prefix: "packssdw." )) { |
| 2513 | if (VecWidth == 128) |
| 2514 | IID = Intrinsic::x86_sse2_packssdw_128; |
| 2515 | else if (VecWidth == 256) |
| 2516 | IID = Intrinsic::x86_avx2_packssdw; |
| 2517 | else if (VecWidth == 512) |
| 2518 | IID = Intrinsic::x86_avx512_packssdw_512; |
| 2519 | else |
| 2520 | llvm_unreachable("Unexpected intrinsic" ); |
| 2521 | } else if (Name.starts_with(Prefix: "packuswb." )) { |
| 2522 | if (VecWidth == 128) |
| 2523 | IID = Intrinsic::x86_sse2_packuswb_128; |
| 2524 | else if (VecWidth == 256) |
| 2525 | IID = Intrinsic::x86_avx2_packuswb; |
| 2526 | else if (VecWidth == 512) |
| 2527 | IID = Intrinsic::x86_avx512_packuswb_512; |
| 2528 | else |
| 2529 | llvm_unreachable("Unexpected intrinsic" ); |
| 2530 | } else if (Name.starts_with(Prefix: "packusdw." )) { |
| 2531 | if (VecWidth == 128) |
| 2532 | IID = Intrinsic::x86_sse41_packusdw; |
| 2533 | else if (VecWidth == 256) |
| 2534 | IID = Intrinsic::x86_avx2_packusdw; |
| 2535 | else if (VecWidth == 512) |
| 2536 | IID = Intrinsic::x86_avx512_packusdw_512; |
| 2537 | else |
| 2538 | llvm_unreachable("Unexpected intrinsic" ); |
| 2539 | } else if (Name.starts_with(Prefix: "vpermilvar." )) { |
| 2540 | if (VecWidth == 128 && EltWidth == 32) |
| 2541 | IID = Intrinsic::x86_avx_vpermilvar_ps; |
| 2542 | else if (VecWidth == 128 && EltWidth == 64) |
| 2543 | IID = Intrinsic::x86_avx_vpermilvar_pd; |
| 2544 | else if (VecWidth == 256 && EltWidth == 32) |
| 2545 | IID = Intrinsic::x86_avx_vpermilvar_ps_256; |
| 2546 | else if (VecWidth == 256 && EltWidth == 64) |
| 2547 | IID = Intrinsic::x86_avx_vpermilvar_pd_256; |
| 2548 | else if (VecWidth == 512 && EltWidth == 32) |
| 2549 | IID = Intrinsic::x86_avx512_vpermilvar_ps_512; |
| 2550 | else if (VecWidth == 512 && EltWidth == 64) |
| 2551 | IID = Intrinsic::x86_avx512_vpermilvar_pd_512; |
| 2552 | else |
| 2553 | llvm_unreachable("Unexpected intrinsic" ); |
| 2554 | } else if (Name == "cvtpd2dq.256" ) { |
| 2555 | IID = Intrinsic::x86_avx_cvt_pd2dq_256; |
| 2556 | } else if (Name == "cvtpd2ps.256" ) { |
| 2557 | IID = Intrinsic::x86_avx_cvt_pd2_ps_256; |
| 2558 | } else if (Name == "cvttpd2dq.256" ) { |
| 2559 | IID = Intrinsic::x86_avx_cvtt_pd2dq_256; |
| 2560 | } else if (Name == "cvttps2dq.128" ) { |
| 2561 | IID = Intrinsic::x86_sse2_cvttps2dq; |
| 2562 | } else if (Name == "cvttps2dq.256" ) { |
| 2563 | IID = Intrinsic::x86_avx_cvtt_ps2dq_256; |
| 2564 | } else if (Name.starts_with(Prefix: "permvar." )) { |
| 2565 | bool IsFloat = CI.getType()->isFPOrFPVectorTy(); |
| 2566 | if (VecWidth == 256 && EltWidth == 32 && IsFloat) |
| 2567 | IID = Intrinsic::x86_avx2_permps; |
| 2568 | else if (VecWidth == 256 && EltWidth == 32 && !IsFloat) |
| 2569 | IID = Intrinsic::x86_avx2_permd; |
| 2570 | else if (VecWidth == 256 && EltWidth == 64 && IsFloat) |
| 2571 | IID = Intrinsic::x86_avx512_permvar_df_256; |
| 2572 | else if (VecWidth == 256 && EltWidth == 64 && !IsFloat) |
| 2573 | IID = Intrinsic::x86_avx512_permvar_di_256; |
| 2574 | else if (VecWidth == 512 && EltWidth == 32 && IsFloat) |
| 2575 | IID = Intrinsic::x86_avx512_permvar_sf_512; |
| 2576 | else if (VecWidth == 512 && EltWidth == 32 && !IsFloat) |
| 2577 | IID = Intrinsic::x86_avx512_permvar_si_512; |
| 2578 | else if (VecWidth == 512 && EltWidth == 64 && IsFloat) |
| 2579 | IID = Intrinsic::x86_avx512_permvar_df_512; |
| 2580 | else if (VecWidth == 512 && EltWidth == 64 && !IsFloat) |
| 2581 | IID = Intrinsic::x86_avx512_permvar_di_512; |
| 2582 | else if (VecWidth == 128 && EltWidth == 16) |
| 2583 | IID = Intrinsic::x86_avx512_permvar_hi_128; |
| 2584 | else if (VecWidth == 256 && EltWidth == 16) |
| 2585 | IID = Intrinsic::x86_avx512_permvar_hi_256; |
| 2586 | else if (VecWidth == 512 && EltWidth == 16) |
| 2587 | IID = Intrinsic::x86_avx512_permvar_hi_512; |
| 2588 | else if (VecWidth == 128 && EltWidth == 8) |
| 2589 | IID = Intrinsic::x86_avx512_permvar_qi_128; |
| 2590 | else if (VecWidth == 256 && EltWidth == 8) |
| 2591 | IID = Intrinsic::x86_avx512_permvar_qi_256; |
| 2592 | else if (VecWidth == 512 && EltWidth == 8) |
| 2593 | IID = Intrinsic::x86_avx512_permvar_qi_512; |
| 2594 | else |
| 2595 | llvm_unreachable("Unexpected intrinsic" ); |
| 2596 | } else if (Name.starts_with(Prefix: "dbpsadbw." )) { |
| 2597 | if (VecWidth == 128) |
| 2598 | IID = Intrinsic::x86_avx512_dbpsadbw_128; |
| 2599 | else if (VecWidth == 256) |
| 2600 | IID = Intrinsic::x86_avx512_dbpsadbw_256; |
| 2601 | else if (VecWidth == 512) |
| 2602 | IID = Intrinsic::x86_avx512_dbpsadbw_512; |
| 2603 | else |
| 2604 | llvm_unreachable("Unexpected intrinsic" ); |
| 2605 | } else if (Name.starts_with(Prefix: "pmultishift.qb." )) { |
| 2606 | if (VecWidth == 128) |
| 2607 | IID = Intrinsic::x86_avx512_pmultishift_qb_128; |
| 2608 | else if (VecWidth == 256) |
| 2609 | IID = Intrinsic::x86_avx512_pmultishift_qb_256; |
| 2610 | else if (VecWidth == 512) |
| 2611 | IID = Intrinsic::x86_avx512_pmultishift_qb_512; |
| 2612 | else |
| 2613 | llvm_unreachable("Unexpected intrinsic" ); |
| 2614 | } else if (Name.starts_with(Prefix: "conflict." )) { |
| 2615 | if (Name[9] == 'd' && VecWidth == 128) |
| 2616 | IID = Intrinsic::x86_avx512_conflict_d_128; |
| 2617 | else if (Name[9] == 'd' && VecWidth == 256) |
| 2618 | IID = Intrinsic::x86_avx512_conflict_d_256; |
| 2619 | else if (Name[9] == 'd' && VecWidth == 512) |
| 2620 | IID = Intrinsic::x86_avx512_conflict_d_512; |
| 2621 | else if (Name[9] == 'q' && VecWidth == 128) |
| 2622 | IID = Intrinsic::x86_avx512_conflict_q_128; |
| 2623 | else if (Name[9] == 'q' && VecWidth == 256) |
| 2624 | IID = Intrinsic::x86_avx512_conflict_q_256; |
| 2625 | else if (Name[9] == 'q' && VecWidth == 512) |
| 2626 | IID = Intrinsic::x86_avx512_conflict_q_512; |
| 2627 | else |
| 2628 | llvm_unreachable("Unexpected intrinsic" ); |
| 2629 | } else if (Name.starts_with(Prefix: "pavg." )) { |
| 2630 | if (Name[5] == 'b' && VecWidth == 128) |
| 2631 | IID = Intrinsic::x86_sse2_pavg_b; |
| 2632 | else if (Name[5] == 'b' && VecWidth == 256) |
| 2633 | IID = Intrinsic::x86_avx2_pavg_b; |
| 2634 | else if (Name[5] == 'b' && VecWidth == 512) |
| 2635 | IID = Intrinsic::x86_avx512_pavg_b_512; |
| 2636 | else if (Name[5] == 'w' && VecWidth == 128) |
| 2637 | IID = Intrinsic::x86_sse2_pavg_w; |
| 2638 | else if (Name[5] == 'w' && VecWidth == 256) |
| 2639 | IID = Intrinsic::x86_avx2_pavg_w; |
| 2640 | else if (Name[5] == 'w' && VecWidth == 512) |
| 2641 | IID = Intrinsic::x86_avx512_pavg_w_512; |
| 2642 | else |
| 2643 | llvm_unreachable("Unexpected intrinsic" ); |
| 2644 | } else |
| 2645 | return false; |
| 2646 | |
| 2647 | SmallVector<Value *, 4> Args(CI.args()); |
| 2648 | Args.pop_back(); |
| 2649 | Args.pop_back(); |
| 2650 | Rep = Builder.CreateIntrinsic(ID: IID, Args); |
| 2651 | unsigned NumArgs = CI.arg_size(); |
| 2652 | Rep = emitX86Select(Builder, Mask: CI.getArgOperand(i: NumArgs - 1), Op0: Rep, |
| 2653 | Op1: CI.getArgOperand(i: NumArgs - 2)); |
| 2654 | return true; |
| 2655 | } |
| 2656 | |
| 2657 | /// Upgrade comment in call to inline asm that represents an objc retain release |
| 2658 | /// marker. |
| 2659 | void llvm::UpgradeInlineAsmString(std::string *AsmStr) { |
| 2660 | size_t Pos; |
| 2661 | if (AsmStr->find(s: "mov\tfp" ) == 0 && |
| 2662 | AsmStr->find(s: "objc_retainAutoreleaseReturnValue" ) != std::string::npos && |
| 2663 | (Pos = AsmStr->find(s: "# marker" )) != std::string::npos) { |
| 2664 | AsmStr->replace(pos: Pos, n1: 1, s: ";" ); |
| 2665 | } |
| 2666 | } |
| 2667 | |
| 2668 | static Value *upgradeNVVMIntrinsicCall(StringRef Name, CallBase *CI, |
| 2669 | Function *F, IRBuilder<> &Builder) { |
| 2670 | Value *Rep = nullptr; |
| 2671 | |
| 2672 | if (Name == "abs.i" || Name == "abs.ll" ) { |
| 2673 | Value *Arg = CI->getArgOperand(i: 0); |
| 2674 | Value *Neg = Builder.CreateNeg(V: Arg, Name: "neg" ); |
| 2675 | Value *Cmp = Builder.CreateICmpSGE( |
| 2676 | LHS: Arg, RHS: llvm::Constant::getNullValue(Ty: Arg->getType()), Name: "abs.cond" ); |
| 2677 | Rep = Builder.CreateSelect(C: Cmp, True: Arg, False: Neg, Name: "abs" ); |
| 2678 | } else if (Name == "abs.bf16" || Name == "abs.bf16x2" ) { |
| 2679 | Type *Ty = (Name == "abs.bf16" ) |
| 2680 | ? Builder.getBFloatTy() |
| 2681 | : FixedVectorType::get(ElementType: Builder.getBFloatTy(), NumElts: 2); |
| 2682 | Value *Arg = Builder.CreateBitCast(V: CI->getArgOperand(i: 0), DestTy: Ty); |
| 2683 | Value *Abs = Builder.CreateUnaryIntrinsic(ID: Intrinsic::nvvm_fabs, V: Arg); |
| 2684 | Rep = Builder.CreateBitCast(V: Abs, DestTy: CI->getType()); |
| 2685 | } else if (Name == "fabs.f" || Name == "fabs.ftz.f" || Name == "fabs.d" ) { |
| 2686 | Intrinsic::ID IID = (Name == "fabs.ftz.f" ) ? Intrinsic::nvvm_fabs_ftz |
| 2687 | : Intrinsic::nvvm_fabs; |
| 2688 | Rep = Builder.CreateUnaryIntrinsic(ID: IID, V: CI->getArgOperand(i: 0)); |
| 2689 | } else if (Name.consume_front(Prefix: "ex2.approx." )) { |
| 2690 | // nvvm.ex2.approx.{f,ftz.f,d,f16x2} |
| 2691 | Intrinsic::ID IID = Name.starts_with(Prefix: "ftz" ) ? Intrinsic::nvvm_ex2_approx_ftz |
| 2692 | : Intrinsic::nvvm_ex2_approx; |
| 2693 | Rep = Builder.CreateUnaryIntrinsic(ID: IID, V: CI->getArgOperand(i: 0)); |
| 2694 | } else if (Name.starts_with(Prefix: "atomic.load.add.f32.p" ) || |
| 2695 | Name.starts_with(Prefix: "atomic.load.add.f64.p" )) { |
| 2696 | Value *Ptr = CI->getArgOperand(i: 0); |
| 2697 | Value *Val = CI->getArgOperand(i: 1); |
| 2698 | Rep = Builder.CreateAtomicRMW(Op: AtomicRMWInst::FAdd, Ptr, Val, Align: MaybeAlign(), |
| 2699 | Ordering: AtomicOrdering::SequentiallyConsistent); |
| 2700 | } else if (Name.starts_with(Prefix: "atomic.load.inc.32.p" ) || |
| 2701 | Name.starts_with(Prefix: "atomic.load.dec.32.p" )) { |
| 2702 | Value *Ptr = CI->getArgOperand(i: 0); |
| 2703 | Value *Val = CI->getArgOperand(i: 1); |
| 2704 | auto Op = Name.starts_with(Prefix: "atomic.load.inc" ) ? AtomicRMWInst::UIncWrap |
| 2705 | : AtomicRMWInst::UDecWrap; |
| 2706 | Rep = Builder.CreateAtomicRMW(Op, Ptr, Val, Align: MaybeAlign(), |
| 2707 | Ordering: AtomicOrdering::SequentiallyConsistent); |
| 2708 | } else if (Name == "clz.ll" ) { |
| 2709 | // llvm.nvvm.clz.ll returns an i32, but llvm.ctlz.i64 returns an i64. |
| 2710 | Value *Arg = CI->getArgOperand(i: 0); |
| 2711 | Value *Ctlz = Builder.CreateIntrinsic(ID: Intrinsic::ctlz, Types: {Arg->getType()}, |
| 2712 | Args: {Arg, Builder.getFalse()}, |
| 2713 | /*FMFSource=*/nullptr, Name: "ctlz" ); |
| 2714 | Rep = Builder.CreateTrunc(V: Ctlz, DestTy: Builder.getInt32Ty(), Name: "ctlz.trunc" ); |
| 2715 | } else if (Name == "popc.ll" ) { |
| 2716 | // llvm.nvvm.popc.ll returns an i32, but llvm.ctpop.i64 returns an |
| 2717 | // i64. |
| 2718 | Value *Arg = CI->getArgOperand(i: 0); |
| 2719 | Value *Popc = Builder.CreateIntrinsic(ID: Intrinsic::ctpop, Types: {Arg->getType()}, |
| 2720 | Args: Arg, /*FMFSource=*/nullptr, Name: "ctpop" ); |
| 2721 | Rep = Builder.CreateTrunc(V: Popc, DestTy: Builder.getInt32Ty(), Name: "ctpop.trunc" ); |
| 2722 | } else if (Name == "h2f" ) { |
| 2723 | Value *Cast = |
| 2724 | Builder.CreateBitCast(V: CI->getArgOperand(i: 0), DestTy: Builder.getHalfTy()); |
| 2725 | Rep = Builder.CreateFPExt(V: Cast, DestTy: Builder.getFloatTy()); |
| 2726 | } else if (Name.consume_front(Prefix: "bitcast." ) && |
| 2727 | (Name == "f2i" || Name == "i2f" || Name == "ll2d" || |
| 2728 | Name == "d2ll" )) { |
| 2729 | Rep = Builder.CreateBitCast(V: CI->getArgOperand(i: 0), DestTy: CI->getType()); |
| 2730 | } else if (Name == "rotate.b32" ) { |
| 2731 | Value *Arg = CI->getOperand(i_nocapture: 0); |
| 2732 | Value *ShiftAmt = CI->getOperand(i_nocapture: 1); |
| 2733 | Rep = Builder.CreateIntrinsic(RetTy: Builder.getInt32Ty(), ID: Intrinsic::fshl, |
| 2734 | Args: {Arg, Arg, ShiftAmt}); |
| 2735 | } else if (Name == "rotate.b64" ) { |
| 2736 | Type *Int64Ty = Builder.getInt64Ty(); |
| 2737 | Value *Arg = CI->getOperand(i_nocapture: 0); |
| 2738 | Value *ZExtShiftAmt = Builder.CreateZExt(V: CI->getOperand(i_nocapture: 1), DestTy: Int64Ty); |
| 2739 | Rep = Builder.CreateIntrinsic(RetTy: Int64Ty, ID: Intrinsic::fshl, |
| 2740 | Args: {Arg, Arg, ZExtShiftAmt}); |
| 2741 | } else if (Name == "rotate.right.b64" ) { |
| 2742 | Type *Int64Ty = Builder.getInt64Ty(); |
| 2743 | Value *Arg = CI->getOperand(i_nocapture: 0); |
| 2744 | Value *ZExtShiftAmt = Builder.CreateZExt(V: CI->getOperand(i_nocapture: 1), DestTy: Int64Ty); |
| 2745 | Rep = Builder.CreateIntrinsic(RetTy: Int64Ty, ID: Intrinsic::fshr, |
| 2746 | Args: {Arg, Arg, ZExtShiftAmt}); |
| 2747 | } else if (Name == "swap.lo.hi.b64" ) { |
| 2748 | Type *Int64Ty = Builder.getInt64Ty(); |
| 2749 | Value *Arg = CI->getOperand(i_nocapture: 0); |
| 2750 | Rep = Builder.CreateIntrinsic(RetTy: Int64Ty, ID: Intrinsic::fshl, |
| 2751 | Args: {Arg, Arg, Builder.getInt64(C: 32)}); |
| 2752 | } else if ((Name.consume_front(Prefix: "ptr.gen.to." ) && |
| 2753 | consumeNVVMPtrAddrSpace(Name)) || |
| 2754 | (Name.consume_front(Prefix: "ptr." ) && consumeNVVMPtrAddrSpace(Name) && |
| 2755 | Name.starts_with(Prefix: ".to.gen" ))) { |
| 2756 | Rep = Builder.CreateAddrSpaceCast(V: CI->getArgOperand(i: 0), DestTy: CI->getType()); |
| 2757 | } else if (Name.consume_front(Prefix: "ldg.global" )) { |
| 2758 | Value *Ptr = CI->getArgOperand(i: 0); |
| 2759 | Align PtrAlign = cast<ConstantInt>(Val: CI->getArgOperand(i: 1))->getAlignValue(); |
| 2760 | // Use addrspace(1) for NVPTX ADDRESS_SPACE_GLOBAL |
| 2761 | Value *ASC = Builder.CreateAddrSpaceCast(V: Ptr, DestTy: Builder.getPtrTy(AddrSpace: 1)); |
| 2762 | Instruction *LD = Builder.CreateAlignedLoad(Ty: CI->getType(), Ptr: ASC, Align: PtrAlign); |
| 2763 | MDNode *MD = MDNode::get(Context&: Builder.getContext(), MDs: {}); |
| 2764 | LD->setMetadata(KindID: LLVMContext::MD_invariant_load, Node: MD); |
| 2765 | return LD; |
| 2766 | } else if (Name == "tanh.approx.f32" ) { |
| 2767 | // nvvm.tanh.approx.f32 -> afn llvm.tanh.f32 |
| 2768 | FastMathFlags FMF; |
| 2769 | FMF.setApproxFunc(); |
| 2770 | Rep = Builder.CreateUnaryIntrinsic(ID: Intrinsic::tanh, V: CI->getArgOperand(i: 0), |
| 2771 | FMFSource: FMF); |
| 2772 | } else if (Name == "barrier0" || Name == "barrier.n" || Name == "bar.sync" ) { |
| 2773 | Value *Arg = |
| 2774 | Name.ends_with(Suffix: '0') ? Builder.getInt32(C: 0) : CI->getArgOperand(i: 0); |
| 2775 | Rep = Builder.CreateIntrinsic(ID: Intrinsic::nvvm_barrier_cta_sync_aligned_all, |
| 2776 | Types: {}, Args: {Arg}); |
| 2777 | } else if (Name == "barrier" ) { |
| 2778 | Rep = Builder.CreateIntrinsic( |
| 2779 | ID: Intrinsic::nvvm_barrier_cta_sync_aligned_count, Types: {}, |
| 2780 | Args: {CI->getArgOperand(i: 0), CI->getArgOperand(i: 1)}); |
| 2781 | } else if (Name == "barrier.sync" ) { |
| 2782 | Rep = Builder.CreateIntrinsic(ID: Intrinsic::nvvm_barrier_cta_sync_all, Types: {}, |
| 2783 | Args: {CI->getArgOperand(i: 0)}); |
| 2784 | } else if (Name == "barrier.sync.cnt" ) { |
| 2785 | Rep = Builder.CreateIntrinsic(ID: Intrinsic::nvvm_barrier_cta_sync_count, Types: {}, |
| 2786 | Args: {CI->getArgOperand(i: 0), CI->getArgOperand(i: 1)}); |
| 2787 | } else if (Name == "barrier0.popc" || Name == "barrier0.and" || |
| 2788 | Name == "barrier0.or" ) { |
| 2789 | Value *C = CI->getArgOperand(i: 0); |
| 2790 | C = Builder.CreateICmpNE(LHS: C, RHS: Builder.getInt32(C: 0)); |
| 2791 | |
| 2792 | Intrinsic::ID IID = |
| 2793 | StringSwitch<Intrinsic::ID>(Name) |
| 2794 | .Case(S: "barrier0.popc" , |
| 2795 | Value: Intrinsic::nvvm_barrier_cta_red_popc_aligned_all) |
| 2796 | .Case(S: "barrier0.and" , |
| 2797 | Value: Intrinsic::nvvm_barrier_cta_red_and_aligned_all) |
| 2798 | .Case(S: "barrier0.or" , |
| 2799 | Value: Intrinsic::nvvm_barrier_cta_red_or_aligned_all); |
| 2800 | Value *Bar = Builder.CreateIntrinsic(ID: IID, Types: {}, Args: {Builder.getInt32(C: 0), C}); |
| 2801 | Rep = Builder.CreateZExt(V: Bar, DestTy: CI->getType()); |
| 2802 | } else { |
| 2803 | Intrinsic::ID IID = shouldUpgradeNVPTXBF16Intrinsic(Name); |
| 2804 | if (IID != Intrinsic::not_intrinsic && |
| 2805 | !F->getReturnType()->getScalarType()->isBFloatTy()) { |
| 2806 | rename(GV: F); |
| 2807 | Function *NewFn = Intrinsic::getOrInsertDeclaration(M: F->getParent(), id: IID); |
| 2808 | SmallVector<Value *, 2> Args; |
| 2809 | for (size_t I = 0; I < NewFn->arg_size(); ++I) { |
| 2810 | Value *Arg = CI->getArgOperand(i: I); |
| 2811 | Type *OldType = Arg->getType(); |
| 2812 | Type *NewType = NewFn->getArg(i: I)->getType(); |
| 2813 | Args.push_back( |
| 2814 | Elt: (OldType->isIntegerTy() && NewType->getScalarType()->isBFloatTy()) |
| 2815 | ? Builder.CreateBitCast(V: Arg, DestTy: NewType) |
| 2816 | : Arg); |
| 2817 | } |
| 2818 | Rep = Builder.CreateCall(Callee: NewFn, Args); |
| 2819 | if (F->getReturnType()->isIntegerTy()) |
| 2820 | Rep = Builder.CreateBitCast(V: Rep, DestTy: F->getReturnType()); |
| 2821 | } |
| 2822 | } |
| 2823 | |
| 2824 | return Rep; |
| 2825 | } |
| 2826 | |
| 2827 | static Value *upgradeX86IntrinsicCall(StringRef Name, CallBase *CI, Function *F, |
| 2828 | IRBuilder<> &Builder) { |
| 2829 | LLVMContext &C = F->getContext(); |
| 2830 | Value *Rep = nullptr; |
| 2831 | |
| 2832 | if (Name.starts_with(Prefix: "sse4a.movnt." )) { |
| 2833 | SmallVector<Metadata *, 1> Elts; |
| 2834 | Elts.push_back( |
| 2835 | Elt: ConstantAsMetadata::get(C: ConstantInt::get(Ty: Type::getInt32Ty(C), V: 1))); |
| 2836 | MDNode *Node = MDNode::get(Context&: C, MDs: Elts); |
| 2837 | |
| 2838 | Value *Arg0 = CI->getArgOperand(i: 0); |
| 2839 | Value *Arg1 = CI->getArgOperand(i: 1); |
| 2840 | |
| 2841 | // Nontemporal (unaligned) store of the 0'th element of the float/double |
| 2842 | // vector. |
| 2843 | Value * = |
| 2844 | Builder.CreateExtractElement(Vec: Arg1, Idx: (uint64_t)0, Name: "extractelement" ); |
| 2845 | |
| 2846 | StoreInst *SI = Builder.CreateAlignedStore(Val: Extract, Ptr: Arg0, Align: Align(1)); |
| 2847 | SI->setMetadata(KindID: LLVMContext::MD_nontemporal, Node); |
| 2848 | } else if (Name.starts_with(Prefix: "avx.movnt." ) || |
| 2849 | Name.starts_with(Prefix: "avx512.storent." )) { |
| 2850 | SmallVector<Metadata *, 1> Elts; |
| 2851 | Elts.push_back( |
| 2852 | Elt: ConstantAsMetadata::get(C: ConstantInt::get(Ty: Type::getInt32Ty(C), V: 1))); |
| 2853 | MDNode *Node = MDNode::get(Context&: C, MDs: Elts); |
| 2854 | |
| 2855 | Value *Arg0 = CI->getArgOperand(i: 0); |
| 2856 | Value *Arg1 = CI->getArgOperand(i: 1); |
| 2857 | |
| 2858 | StoreInst *SI = Builder.CreateAlignedStore( |
| 2859 | Val: Arg1, Ptr: Arg0, |
| 2860 | Align: Align(Arg1->getType()->getPrimitiveSizeInBits().getFixedValue() / 8)); |
| 2861 | SI->setMetadata(KindID: LLVMContext::MD_nontemporal, Node); |
| 2862 | } else if (Name == "sse2.storel.dq" ) { |
| 2863 | Value *Arg0 = CI->getArgOperand(i: 0); |
| 2864 | Value *Arg1 = CI->getArgOperand(i: 1); |
| 2865 | |
| 2866 | auto *NewVecTy = FixedVectorType::get(ElementType: Type::getInt64Ty(C), NumElts: 2); |
| 2867 | Value *BC0 = Builder.CreateBitCast(V: Arg1, DestTy: NewVecTy, Name: "cast" ); |
| 2868 | Value *Elt = Builder.CreateExtractElement(Vec: BC0, Idx: (uint64_t)0); |
| 2869 | Builder.CreateAlignedStore(Val: Elt, Ptr: Arg0, Align: Align(1)); |
| 2870 | } else if (Name.starts_with(Prefix: "sse.storeu." ) || |
| 2871 | Name.starts_with(Prefix: "sse2.storeu." ) || |
| 2872 | Name.starts_with(Prefix: "avx.storeu." )) { |
| 2873 | Value *Arg0 = CI->getArgOperand(i: 0); |
| 2874 | Value *Arg1 = CI->getArgOperand(i: 1); |
| 2875 | Builder.CreateAlignedStore(Val: Arg1, Ptr: Arg0, Align: Align(1)); |
| 2876 | } else if (Name == "avx512.mask.store.ss" ) { |
| 2877 | Value *Mask = Builder.CreateAnd(LHS: CI->getArgOperand(i: 2), RHS: Builder.getInt8(C: 1)); |
| 2878 | upgradeMaskedStore(Builder, Ptr: CI->getArgOperand(i: 0), Data: CI->getArgOperand(i: 1), |
| 2879 | Mask, Aligned: false); |
| 2880 | } else if (Name.starts_with(Prefix: "avx512.mask.store" )) { |
| 2881 | // "avx512.mask.storeu." or "avx512.mask.store." |
| 2882 | bool Aligned = Name[17] != 'u'; // "avx512.mask.storeu". |
| 2883 | upgradeMaskedStore(Builder, Ptr: CI->getArgOperand(i: 0), Data: CI->getArgOperand(i: 1), |
| 2884 | Mask: CI->getArgOperand(i: 2), Aligned); |
| 2885 | } else if (Name.starts_with(Prefix: "sse2.pcmp" ) || Name.starts_with(Prefix: "avx2.pcmp" )) { |
| 2886 | // Upgrade packed integer vector compare intrinsics to compare instructions. |
| 2887 | // "sse2.pcpmpeq." "sse2.pcmpgt." "avx2.pcmpeq." or "avx2.pcmpgt." |
| 2888 | bool CmpEq = Name[9] == 'e'; |
| 2889 | Rep = Builder.CreateICmp(P: CmpEq ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_SGT, |
| 2890 | LHS: CI->getArgOperand(i: 0), RHS: CI->getArgOperand(i: 1)); |
| 2891 | Rep = Builder.CreateSExt(V: Rep, DestTy: CI->getType(), Name: "" ); |
| 2892 | } else if (Name.starts_with(Prefix: "avx512.broadcastm" )) { |
| 2893 | Type *ExtTy = Type::getInt32Ty(C); |
| 2894 | if (CI->getOperand(i_nocapture: 0)->getType()->isIntegerTy(Bitwidth: 8)) |
| 2895 | ExtTy = Type::getInt64Ty(C); |
| 2896 | unsigned NumElts = CI->getType()->getPrimitiveSizeInBits() / |
| 2897 | ExtTy->getPrimitiveSizeInBits(); |
| 2898 | Rep = Builder.CreateZExt(V: CI->getArgOperand(i: 0), DestTy: ExtTy); |
| 2899 | Rep = Builder.CreateVectorSplat(NumElts, V: Rep); |
| 2900 | } else if (Name == "sse.sqrt.ss" || Name == "sse2.sqrt.sd" ) { |
| 2901 | Value *Vec = CI->getArgOperand(i: 0); |
| 2902 | Value *Elt0 = Builder.CreateExtractElement(Vec, Idx: (uint64_t)0); |
| 2903 | Elt0 = Builder.CreateIntrinsic(ID: Intrinsic::sqrt, Types: Elt0->getType(), Args: Elt0); |
| 2904 | Rep = Builder.CreateInsertElement(Vec, NewElt: Elt0, Idx: (uint64_t)0); |
| 2905 | } else if (Name.starts_with(Prefix: "avx.sqrt.p" ) || |
| 2906 | Name.starts_with(Prefix: "sse2.sqrt.p" ) || |
| 2907 | Name.starts_with(Prefix: "sse.sqrt.p" )) { |
| 2908 | Rep = Builder.CreateIntrinsic(ID: Intrinsic::sqrt, Types: CI->getType(), |
| 2909 | Args: {CI->getArgOperand(i: 0)}); |
| 2910 | } else if (Name.starts_with(Prefix: "avx512.mask.sqrt.p" )) { |
| 2911 | if (CI->arg_size() == 4 && |
| 2912 | (!isa<ConstantInt>(Val: CI->getArgOperand(i: 3)) || |
| 2913 | cast<ConstantInt>(Val: CI->getArgOperand(i: 3))->getZExtValue() != 4)) { |
| 2914 | Intrinsic::ID IID = Name[18] == 's' ? Intrinsic::x86_avx512_sqrt_ps_512 |
| 2915 | : Intrinsic::x86_avx512_sqrt_pd_512; |
| 2916 | |
| 2917 | Value *Args[] = {CI->getArgOperand(i: 0), CI->getArgOperand(i: 3)}; |
| 2918 | Rep = Builder.CreateIntrinsic(ID: IID, Args); |
| 2919 | } else { |
| 2920 | Rep = Builder.CreateIntrinsic(ID: Intrinsic::sqrt, Types: CI->getType(), |
| 2921 | Args: {CI->getArgOperand(i: 0)}); |
| 2922 | } |
| 2923 | Rep = |
| 2924 | emitX86Select(Builder, Mask: CI->getArgOperand(i: 2), Op0: Rep, Op1: CI->getArgOperand(i: 1)); |
| 2925 | } else if (Name.starts_with(Prefix: "avx512.ptestm" ) || |
| 2926 | Name.starts_with(Prefix: "avx512.ptestnm" )) { |
| 2927 | Value *Op0 = CI->getArgOperand(i: 0); |
| 2928 | Value *Op1 = CI->getArgOperand(i: 1); |
| 2929 | Value *Mask = CI->getArgOperand(i: 2); |
| 2930 | Rep = Builder.CreateAnd(LHS: Op0, RHS: Op1); |
| 2931 | llvm::Type *Ty = Op0->getType(); |
| 2932 | Value *Zero = llvm::Constant::getNullValue(Ty); |
| 2933 | ICmpInst::Predicate Pred = Name.starts_with(Prefix: "avx512.ptestm" ) |
| 2934 | ? ICmpInst::ICMP_NE |
| 2935 | : ICmpInst::ICMP_EQ; |
| 2936 | Rep = Builder.CreateICmp(P: Pred, LHS: Rep, RHS: Zero); |
| 2937 | Rep = applyX86MaskOn1BitsVec(Builder, Vec: Rep, Mask); |
| 2938 | } else if (Name.starts_with(Prefix: "avx512.mask.pbroadcast" )) { |
| 2939 | unsigned NumElts = cast<FixedVectorType>(Val: CI->getArgOperand(i: 1)->getType()) |
| 2940 | ->getNumElements(); |
| 2941 | Rep = Builder.CreateVectorSplat(NumElts, V: CI->getArgOperand(i: 0)); |
| 2942 | Rep = |
| 2943 | emitX86Select(Builder, Mask: CI->getArgOperand(i: 2), Op0: Rep, Op1: CI->getArgOperand(i: 1)); |
| 2944 | } else if (Name.starts_with(Prefix: "avx512.kunpck" )) { |
| 2945 | unsigned NumElts = CI->getType()->getScalarSizeInBits(); |
| 2946 | Value *LHS = getX86MaskVec(Builder, Mask: CI->getArgOperand(i: 0), NumElts); |
| 2947 | Value *RHS = getX86MaskVec(Builder, Mask: CI->getArgOperand(i: 1), NumElts); |
| 2948 | int Indices[64]; |
| 2949 | for (unsigned i = 0; i != NumElts; ++i) |
| 2950 | Indices[i] = i; |
| 2951 | |
| 2952 | // First extract half of each vector. This gives better codegen than |
| 2953 | // doing it in a single shuffle. |
| 2954 | LHS = Builder.CreateShuffleVector(V1: LHS, V2: LHS, Mask: ArrayRef(Indices, NumElts / 2)); |
| 2955 | RHS = Builder.CreateShuffleVector(V1: RHS, V2: RHS, Mask: ArrayRef(Indices, NumElts / 2)); |
| 2956 | // Concat the vectors. |
| 2957 | // NOTE: Operands have to be swapped to match intrinsic definition. |
| 2958 | Rep = Builder.CreateShuffleVector(V1: RHS, V2: LHS, Mask: ArrayRef(Indices, NumElts)); |
| 2959 | Rep = Builder.CreateBitCast(V: Rep, DestTy: CI->getType()); |
| 2960 | } else if (Name == "avx512.kand.w" ) { |
| 2961 | Value *LHS = getX86MaskVec(Builder, Mask: CI->getArgOperand(i: 0), NumElts: 16); |
| 2962 | Value *RHS = getX86MaskVec(Builder, Mask: CI->getArgOperand(i: 1), NumElts: 16); |
| 2963 | Rep = Builder.CreateAnd(LHS, RHS); |
| 2964 | Rep = Builder.CreateBitCast(V: Rep, DestTy: CI->getType()); |
| 2965 | } else if (Name == "avx512.kandn.w" ) { |
| 2966 | Value *LHS = getX86MaskVec(Builder, Mask: CI->getArgOperand(i: 0), NumElts: 16); |
| 2967 | Value *RHS = getX86MaskVec(Builder, Mask: CI->getArgOperand(i: 1), NumElts: 16); |
| 2968 | LHS = Builder.CreateNot(V: LHS); |
| 2969 | Rep = Builder.CreateAnd(LHS, RHS); |
| 2970 | Rep = Builder.CreateBitCast(V: Rep, DestTy: CI->getType()); |
| 2971 | } else if (Name == "avx512.kor.w" ) { |
| 2972 | Value *LHS = getX86MaskVec(Builder, Mask: CI->getArgOperand(i: 0), NumElts: 16); |
| 2973 | Value *RHS = getX86MaskVec(Builder, Mask: CI->getArgOperand(i: 1), NumElts: 16); |
| 2974 | Rep = Builder.CreateOr(LHS, RHS); |
| 2975 | Rep = Builder.CreateBitCast(V: Rep, DestTy: CI->getType()); |
| 2976 | } else if (Name == "avx512.kxor.w" ) { |
| 2977 | Value *LHS = getX86MaskVec(Builder, Mask: CI->getArgOperand(i: 0), NumElts: 16); |
| 2978 | Value *RHS = getX86MaskVec(Builder, Mask: CI->getArgOperand(i: 1), NumElts: 16); |
| 2979 | Rep = Builder.CreateXor(LHS, RHS); |
| 2980 | Rep = Builder.CreateBitCast(V: Rep, DestTy: CI->getType()); |
| 2981 | } else if (Name == "avx512.kxnor.w" ) { |
| 2982 | Value *LHS = getX86MaskVec(Builder, Mask: CI->getArgOperand(i: 0), NumElts: 16); |
| 2983 | Value *RHS = getX86MaskVec(Builder, Mask: CI->getArgOperand(i: 1), NumElts: 16); |
| 2984 | LHS = Builder.CreateNot(V: LHS); |
| 2985 | Rep = Builder.CreateXor(LHS, RHS); |
| 2986 | Rep = Builder.CreateBitCast(V: Rep, DestTy: CI->getType()); |
| 2987 | } else if (Name == "avx512.knot.w" ) { |
| 2988 | Rep = getX86MaskVec(Builder, Mask: CI->getArgOperand(i: 0), NumElts: 16); |
| 2989 | Rep = Builder.CreateNot(V: Rep); |
| 2990 | Rep = Builder.CreateBitCast(V: Rep, DestTy: CI->getType()); |
| 2991 | } else if (Name == "avx512.kortestz.w" || Name == "avx512.kortestc.w" ) { |
| 2992 | Value *LHS = getX86MaskVec(Builder, Mask: CI->getArgOperand(i: 0), NumElts: 16); |
| 2993 | Value *RHS = getX86MaskVec(Builder, Mask: CI->getArgOperand(i: 1), NumElts: 16); |
| 2994 | Rep = Builder.CreateOr(LHS, RHS); |
| 2995 | Rep = Builder.CreateBitCast(V: Rep, DestTy: Builder.getInt16Ty()); |
| 2996 | Value *C; |
| 2997 | if (Name[14] == 'c') |
| 2998 | C = ConstantInt::getAllOnesValue(Ty: Builder.getInt16Ty()); |
| 2999 | else |
| 3000 | C = ConstantInt::getNullValue(Ty: Builder.getInt16Ty()); |
| 3001 | Rep = Builder.CreateICmpEQ(LHS: Rep, RHS: C); |
| 3002 | Rep = Builder.CreateZExt(V: Rep, DestTy: Builder.getInt32Ty()); |
| 3003 | } else if (Name == "sse.add.ss" || Name == "sse2.add.sd" || |
| 3004 | Name == "sse.sub.ss" || Name == "sse2.sub.sd" || |
| 3005 | Name == "sse.mul.ss" || Name == "sse2.mul.sd" || |
| 3006 | Name == "sse.div.ss" || Name == "sse2.div.sd" ) { |
| 3007 | Type *I32Ty = Type::getInt32Ty(C); |
| 3008 | Value *Elt0 = Builder.CreateExtractElement(Vec: CI->getArgOperand(i: 0), |
| 3009 | Idx: ConstantInt::get(Ty: I32Ty, V: 0)); |
| 3010 | Value *Elt1 = Builder.CreateExtractElement(Vec: CI->getArgOperand(i: 1), |
| 3011 | Idx: ConstantInt::get(Ty: I32Ty, V: 0)); |
| 3012 | Value *EltOp; |
| 3013 | if (Name.contains(Other: ".add." )) |
| 3014 | EltOp = Builder.CreateFAdd(L: Elt0, R: Elt1); |
| 3015 | else if (Name.contains(Other: ".sub." )) |
| 3016 | EltOp = Builder.CreateFSub(L: Elt0, R: Elt1); |
| 3017 | else if (Name.contains(Other: ".mul." )) |
| 3018 | EltOp = Builder.CreateFMul(L: Elt0, R: Elt1); |
| 3019 | else |
| 3020 | EltOp = Builder.CreateFDiv(L: Elt0, R: Elt1); |
| 3021 | Rep = Builder.CreateInsertElement(Vec: CI->getArgOperand(i: 0), NewElt: EltOp, |
| 3022 | Idx: ConstantInt::get(Ty: I32Ty, V: 0)); |
| 3023 | } else if (Name.starts_with(Prefix: "avx512.mask.pcmp" )) { |
| 3024 | // "avx512.mask.pcmpeq." or "avx512.mask.pcmpgt." |
| 3025 | bool CmpEq = Name[16] == 'e'; |
| 3026 | Rep = upgradeMaskedCompare(Builder, CI&: *CI, CC: CmpEq ? 0 : 6, Signed: true); |
| 3027 | } else if (Name.starts_with(Prefix: "avx512.mask.vpshufbitqmb." )) { |
| 3028 | Type *OpTy = CI->getArgOperand(i: 0)->getType(); |
| 3029 | unsigned VecWidth = OpTy->getPrimitiveSizeInBits(); |
| 3030 | Intrinsic::ID IID; |
| 3031 | switch (VecWidth) { |
| 3032 | default: |
| 3033 | llvm_unreachable("Unexpected intrinsic" ); |
| 3034 | case 128: |
| 3035 | IID = Intrinsic::x86_avx512_vpshufbitqmb_128; |
| 3036 | break; |
| 3037 | case 256: |
| 3038 | IID = Intrinsic::x86_avx512_vpshufbitqmb_256; |
| 3039 | break; |
| 3040 | case 512: |
| 3041 | IID = Intrinsic::x86_avx512_vpshufbitqmb_512; |
| 3042 | break; |
| 3043 | } |
| 3044 | |
| 3045 | Rep = |
| 3046 | Builder.CreateIntrinsic(ID: IID, Args: {CI->getOperand(i_nocapture: 0), CI->getArgOperand(i: 1)}); |
| 3047 | Rep = applyX86MaskOn1BitsVec(Builder, Vec: Rep, Mask: CI->getArgOperand(i: 2)); |
| 3048 | } else if (Name.starts_with(Prefix: "avx512.mask.fpclass.p" )) { |
| 3049 | Type *OpTy = CI->getArgOperand(i: 0)->getType(); |
| 3050 | unsigned VecWidth = OpTy->getPrimitiveSizeInBits(); |
| 3051 | unsigned EltWidth = OpTy->getScalarSizeInBits(); |
| 3052 | Intrinsic::ID IID; |
| 3053 | if (VecWidth == 128 && EltWidth == 32) |
| 3054 | IID = Intrinsic::x86_avx512_fpclass_ps_128; |
| 3055 | else if (VecWidth == 256 && EltWidth == 32) |
| 3056 | IID = Intrinsic::x86_avx512_fpclass_ps_256; |
| 3057 | else if (VecWidth == 512 && EltWidth == 32) |
| 3058 | IID = Intrinsic::x86_avx512_fpclass_ps_512; |
| 3059 | else if (VecWidth == 128 && EltWidth == 64) |
| 3060 | IID = Intrinsic::x86_avx512_fpclass_pd_128; |
| 3061 | else if (VecWidth == 256 && EltWidth == 64) |
| 3062 | IID = Intrinsic::x86_avx512_fpclass_pd_256; |
| 3063 | else if (VecWidth == 512 && EltWidth == 64) |
| 3064 | IID = Intrinsic::x86_avx512_fpclass_pd_512; |
| 3065 | else |
| 3066 | llvm_unreachable("Unexpected intrinsic" ); |
| 3067 | |
| 3068 | Rep = |
| 3069 | Builder.CreateIntrinsic(ID: IID, Args: {CI->getOperand(i_nocapture: 0), CI->getArgOperand(i: 1)}); |
| 3070 | Rep = applyX86MaskOn1BitsVec(Builder, Vec: Rep, Mask: CI->getArgOperand(i: 2)); |
| 3071 | } else if (Name.starts_with(Prefix: "avx512.cmp.p" )) { |
| 3072 | SmallVector<Value *, 4> Args(CI->args()); |
| 3073 | Type *OpTy = Args[0]->getType(); |
| 3074 | unsigned VecWidth = OpTy->getPrimitiveSizeInBits(); |
| 3075 | unsigned EltWidth = OpTy->getScalarSizeInBits(); |
| 3076 | Intrinsic::ID IID; |
| 3077 | if (VecWidth == 128 && EltWidth == 32) |
| 3078 | IID = Intrinsic::x86_avx512_mask_cmp_ps_128; |
| 3079 | else if (VecWidth == 256 && EltWidth == 32) |
| 3080 | IID = Intrinsic::x86_avx512_mask_cmp_ps_256; |
| 3081 | else if (VecWidth == 512 && EltWidth == 32) |
| 3082 | IID = Intrinsic::x86_avx512_mask_cmp_ps_512; |
| 3083 | else if (VecWidth == 128 && EltWidth == 64) |
| 3084 | IID = Intrinsic::x86_avx512_mask_cmp_pd_128; |
| 3085 | else if (VecWidth == 256 && EltWidth == 64) |
| 3086 | IID = Intrinsic::x86_avx512_mask_cmp_pd_256; |
| 3087 | else if (VecWidth == 512 && EltWidth == 64) |
| 3088 | IID = Intrinsic::x86_avx512_mask_cmp_pd_512; |
| 3089 | else |
| 3090 | llvm_unreachable("Unexpected intrinsic" ); |
| 3091 | |
| 3092 | Value *Mask = Constant::getAllOnesValue(Ty: CI->getType()); |
| 3093 | if (VecWidth == 512) |
| 3094 | std::swap(a&: Mask, b&: Args.back()); |
| 3095 | Args.push_back(Elt: Mask); |
| 3096 | |
| 3097 | Rep = Builder.CreateIntrinsic(ID: IID, Args); |
| 3098 | } else if (Name.starts_with(Prefix: "avx512.mask.cmp." )) { |
| 3099 | // Integer compare intrinsics. |
| 3100 | unsigned Imm = cast<ConstantInt>(Val: CI->getArgOperand(i: 2))->getZExtValue(); |
| 3101 | Rep = upgradeMaskedCompare(Builder, CI&: *CI, CC: Imm, Signed: true); |
| 3102 | } else if (Name.starts_with(Prefix: "avx512.mask.ucmp." )) { |
| 3103 | unsigned Imm = cast<ConstantInt>(Val: CI->getArgOperand(i: 2))->getZExtValue(); |
| 3104 | Rep = upgradeMaskedCompare(Builder, CI&: *CI, CC: Imm, Signed: false); |
| 3105 | } else if (Name.starts_with(Prefix: "avx512.cvtb2mask." ) || |
| 3106 | Name.starts_with(Prefix: "avx512.cvtw2mask." ) || |
| 3107 | Name.starts_with(Prefix: "avx512.cvtd2mask." ) || |
| 3108 | Name.starts_with(Prefix: "avx512.cvtq2mask." )) { |
| 3109 | Value *Op = CI->getArgOperand(i: 0); |
| 3110 | Value *Zero = llvm::Constant::getNullValue(Ty: Op->getType()); |
| 3111 | Rep = Builder.CreateICmp(P: ICmpInst::ICMP_SLT, LHS: Op, RHS: Zero); |
| 3112 | Rep = applyX86MaskOn1BitsVec(Builder, Vec: Rep, Mask: nullptr); |
| 3113 | } else if (Name == "ssse3.pabs.b.128" || Name == "ssse3.pabs.w.128" || |
| 3114 | Name == "ssse3.pabs.d.128" || Name.starts_with(Prefix: "avx2.pabs" ) || |
| 3115 | Name.starts_with(Prefix: "avx512.mask.pabs" )) { |
| 3116 | Rep = upgradeAbs(Builder, CI&: *CI); |
| 3117 | } else if (Name == "sse41.pmaxsb" || Name == "sse2.pmaxs.w" || |
| 3118 | Name == "sse41.pmaxsd" || Name.starts_with(Prefix: "avx2.pmaxs" ) || |
| 3119 | Name.starts_with(Prefix: "avx512.mask.pmaxs" )) { |
| 3120 | Rep = upgradeX86BinaryIntrinsics(Builder, CI&: *CI, IID: Intrinsic::smax); |
| 3121 | } else if (Name == "sse2.pmaxu.b" || Name == "sse41.pmaxuw" || |
| 3122 | Name == "sse41.pmaxud" || Name.starts_with(Prefix: "avx2.pmaxu" ) || |
| 3123 | Name.starts_with(Prefix: "avx512.mask.pmaxu" )) { |
| 3124 | Rep = upgradeX86BinaryIntrinsics(Builder, CI&: *CI, IID: Intrinsic::umax); |
| 3125 | } else if (Name == "sse41.pminsb" || Name == "sse2.pmins.w" || |
| 3126 | Name == "sse41.pminsd" || Name.starts_with(Prefix: "avx2.pmins" ) || |
| 3127 | Name.starts_with(Prefix: "avx512.mask.pmins" )) { |
| 3128 | Rep = upgradeX86BinaryIntrinsics(Builder, CI&: *CI, IID: Intrinsic::smin); |
| 3129 | } else if (Name == "sse2.pminu.b" || Name == "sse41.pminuw" || |
| 3130 | Name == "sse41.pminud" || Name.starts_with(Prefix: "avx2.pminu" ) || |
| 3131 | Name.starts_with(Prefix: "avx512.mask.pminu" )) { |
| 3132 | Rep = upgradeX86BinaryIntrinsics(Builder, CI&: *CI, IID: Intrinsic::umin); |
| 3133 | } else if (Name == "sse2.pmulu.dq" || Name == "avx2.pmulu.dq" || |
| 3134 | Name == "avx512.pmulu.dq.512" || |
| 3135 | Name.starts_with(Prefix: "avx512.mask.pmulu.dq." )) { |
| 3136 | Rep = upgradePMULDQ(Builder, CI&: *CI, /*Signed*/ IsSigned: false); |
| 3137 | } else if (Name == "sse41.pmuldq" || Name == "avx2.pmul.dq" || |
| 3138 | Name == "avx512.pmul.dq.512" || |
| 3139 | Name.starts_with(Prefix: "avx512.mask.pmul.dq." )) { |
| 3140 | Rep = upgradePMULDQ(Builder, CI&: *CI, /*Signed*/ IsSigned: true); |
| 3141 | } else if (Name == "sse.cvtsi2ss" || Name == "sse2.cvtsi2sd" || |
| 3142 | Name == "sse.cvtsi642ss" || Name == "sse2.cvtsi642sd" ) { |
| 3143 | Rep = |
| 3144 | Builder.CreateSIToFP(V: CI->getArgOperand(i: 1), |
| 3145 | DestTy: cast<VectorType>(Val: CI->getType())->getElementType()); |
| 3146 | Rep = Builder.CreateInsertElement(Vec: CI->getArgOperand(i: 0), NewElt: Rep, Idx: (uint64_t)0); |
| 3147 | } else if (Name == "avx512.cvtusi2sd" ) { |
| 3148 | Rep = |
| 3149 | Builder.CreateUIToFP(V: CI->getArgOperand(i: 1), |
| 3150 | DestTy: cast<VectorType>(Val: CI->getType())->getElementType()); |
| 3151 | Rep = Builder.CreateInsertElement(Vec: CI->getArgOperand(i: 0), NewElt: Rep, Idx: (uint64_t)0); |
| 3152 | } else if (Name == "sse2.cvtss2sd" ) { |
| 3153 | Rep = Builder.CreateExtractElement(Vec: CI->getArgOperand(i: 1), Idx: (uint64_t)0); |
| 3154 | Rep = Builder.CreateFPExt( |
| 3155 | V: Rep, DestTy: cast<VectorType>(Val: CI->getType())->getElementType()); |
| 3156 | Rep = Builder.CreateInsertElement(Vec: CI->getArgOperand(i: 0), NewElt: Rep, Idx: (uint64_t)0); |
| 3157 | } else if (Name == "sse2.cvtdq2pd" || Name == "sse2.cvtdq2ps" || |
| 3158 | Name == "avx.cvtdq2.pd.256" || Name == "avx.cvtdq2.ps.256" || |
| 3159 | Name.starts_with(Prefix: "avx512.mask.cvtdq2pd." ) || |
| 3160 | Name.starts_with(Prefix: "avx512.mask.cvtudq2pd." ) || |
| 3161 | Name.starts_with(Prefix: "avx512.mask.cvtdq2ps." ) || |
| 3162 | Name.starts_with(Prefix: "avx512.mask.cvtudq2ps." ) || |
| 3163 | Name.starts_with(Prefix: "avx512.mask.cvtqq2pd." ) || |
| 3164 | Name.starts_with(Prefix: "avx512.mask.cvtuqq2pd." ) || |
| 3165 | Name == "avx512.mask.cvtqq2ps.256" || |
| 3166 | Name == "avx512.mask.cvtqq2ps.512" || |
| 3167 | Name == "avx512.mask.cvtuqq2ps.256" || |
| 3168 | Name == "avx512.mask.cvtuqq2ps.512" || Name == "sse2.cvtps2pd" || |
| 3169 | Name == "avx.cvt.ps2.pd.256" || |
| 3170 | Name == "avx512.mask.cvtps2pd.128" || |
| 3171 | Name == "avx512.mask.cvtps2pd.256" ) { |
| 3172 | auto *DstTy = cast<FixedVectorType>(Val: CI->getType()); |
| 3173 | Rep = CI->getArgOperand(i: 0); |
| 3174 | auto *SrcTy = cast<FixedVectorType>(Val: Rep->getType()); |
| 3175 | |
| 3176 | unsigned NumDstElts = DstTy->getNumElements(); |
| 3177 | if (NumDstElts < SrcTy->getNumElements()) { |
| 3178 | assert(NumDstElts == 2 && "Unexpected vector size" ); |
| 3179 | Rep = Builder.CreateShuffleVector(V1: Rep, V2: Rep, Mask: ArrayRef<int>{0, 1}); |
| 3180 | } |
| 3181 | |
| 3182 | bool IsPS2PD = SrcTy->getElementType()->isFloatTy(); |
| 3183 | bool IsUnsigned = Name.contains(Other: "cvtu" ); |
| 3184 | if (IsPS2PD) |
| 3185 | Rep = Builder.CreateFPExt(V: Rep, DestTy: DstTy, Name: "cvtps2pd" ); |
| 3186 | else if (CI->arg_size() == 4 && |
| 3187 | (!isa<ConstantInt>(Val: CI->getArgOperand(i: 3)) || |
| 3188 | cast<ConstantInt>(Val: CI->getArgOperand(i: 3))->getZExtValue() != 4)) { |
| 3189 | Intrinsic::ID IID = IsUnsigned ? Intrinsic::x86_avx512_uitofp_round |
| 3190 | : Intrinsic::x86_avx512_sitofp_round; |
| 3191 | Rep = Builder.CreateIntrinsic(ID: IID, Types: {DstTy, SrcTy}, |
| 3192 | Args: {Rep, CI->getArgOperand(i: 3)}); |
| 3193 | } else { |
| 3194 | Rep = IsUnsigned ? Builder.CreateUIToFP(V: Rep, DestTy: DstTy, Name: "cvt" ) |
| 3195 | : Builder.CreateSIToFP(V: Rep, DestTy: DstTy, Name: "cvt" ); |
| 3196 | } |
| 3197 | |
| 3198 | if (CI->arg_size() >= 3) |
| 3199 | Rep = emitX86Select(Builder, Mask: CI->getArgOperand(i: 2), Op0: Rep, |
| 3200 | Op1: CI->getArgOperand(i: 1)); |
| 3201 | } else if (Name.starts_with(Prefix: "avx512.mask.vcvtph2ps." ) || |
| 3202 | Name.starts_with(Prefix: "vcvtph2ps." )) { |
| 3203 | auto *DstTy = cast<FixedVectorType>(Val: CI->getType()); |
| 3204 | Rep = CI->getArgOperand(i: 0); |
| 3205 | auto *SrcTy = cast<FixedVectorType>(Val: Rep->getType()); |
| 3206 | unsigned NumDstElts = DstTy->getNumElements(); |
| 3207 | if (NumDstElts != SrcTy->getNumElements()) { |
| 3208 | assert(NumDstElts == 4 && "Unexpected vector size" ); |
| 3209 | Rep = Builder.CreateShuffleVector(V1: Rep, V2: Rep, Mask: ArrayRef<int>{0, 1, 2, 3}); |
| 3210 | } |
| 3211 | Rep = Builder.CreateBitCast( |
| 3212 | V: Rep, DestTy: FixedVectorType::get(ElementType: Type::getHalfTy(C), NumElts: NumDstElts)); |
| 3213 | Rep = Builder.CreateFPExt(V: Rep, DestTy: DstTy, Name: "cvtph2ps" ); |
| 3214 | if (CI->arg_size() >= 3) |
| 3215 | Rep = emitX86Select(Builder, Mask: CI->getArgOperand(i: 2), Op0: Rep, |
| 3216 | Op1: CI->getArgOperand(i: 1)); |
| 3217 | } else if (Name.starts_with(Prefix: "avx512.mask.load" )) { |
| 3218 | // "avx512.mask.loadu." or "avx512.mask.load." |
| 3219 | bool Aligned = Name[16] != 'u'; // "avx512.mask.loadu". |
| 3220 | Rep = upgradeMaskedLoad(Builder, Ptr: CI->getArgOperand(i: 0), Passthru: CI->getArgOperand(i: 1), |
| 3221 | Mask: CI->getArgOperand(i: 2), Aligned); |
| 3222 | } else if (Name.starts_with(Prefix: "avx512.mask.expand.load." )) { |
| 3223 | auto *ResultTy = cast<FixedVectorType>(Val: CI->getType()); |
| 3224 | Value *MaskVec = getX86MaskVec(Builder, Mask: CI->getArgOperand(i: 2), |
| 3225 | NumElts: ResultTy->getNumElements()); |
| 3226 | |
| 3227 | Rep = Builder.CreateIntrinsic( |
| 3228 | ID: Intrinsic::masked_expandload, Types: ResultTy, |
| 3229 | Args: {CI->getOperand(i_nocapture: 0), MaskVec, CI->getOperand(i_nocapture: 1)}); |
| 3230 | } else if (Name.starts_with(Prefix: "avx512.mask.compress.store." )) { |
| 3231 | auto *ResultTy = cast<VectorType>(Val: CI->getArgOperand(i: 1)->getType()); |
| 3232 | Value *MaskVec = |
| 3233 | getX86MaskVec(Builder, Mask: CI->getArgOperand(i: 2), |
| 3234 | NumElts: cast<FixedVectorType>(Val: ResultTy)->getNumElements()); |
| 3235 | |
| 3236 | Rep = Builder.CreateIntrinsic( |
| 3237 | ID: Intrinsic::masked_compressstore, Types: ResultTy, |
| 3238 | Args: {CI->getArgOperand(i: 1), CI->getArgOperand(i: 0), MaskVec}); |
| 3239 | } else if (Name.starts_with(Prefix: "avx512.mask.compress." ) || |
| 3240 | Name.starts_with(Prefix: "avx512.mask.expand." )) { |
| 3241 | auto *ResultTy = cast<FixedVectorType>(Val: CI->getType()); |
| 3242 | |
| 3243 | Value *MaskVec = getX86MaskVec(Builder, Mask: CI->getArgOperand(i: 2), |
| 3244 | NumElts: ResultTy->getNumElements()); |
| 3245 | |
| 3246 | bool IsCompress = Name[12] == 'c'; |
| 3247 | Intrinsic::ID IID = IsCompress ? Intrinsic::x86_avx512_mask_compress |
| 3248 | : Intrinsic::x86_avx512_mask_expand; |
| 3249 | Rep = Builder.CreateIntrinsic( |
| 3250 | ID: IID, Types: ResultTy, Args: {CI->getOperand(i_nocapture: 0), CI->getOperand(i_nocapture: 1), MaskVec}); |
| 3251 | } else if (Name.starts_with(Prefix: "xop.vpcom" )) { |
| 3252 | bool IsSigned; |
| 3253 | if (Name.ends_with(Suffix: "ub" ) || Name.ends_with(Suffix: "uw" ) || Name.ends_with(Suffix: "ud" ) || |
| 3254 | Name.ends_with(Suffix: "uq" )) |
| 3255 | IsSigned = false; |
| 3256 | else if (Name.ends_with(Suffix: "b" ) || Name.ends_with(Suffix: "w" ) || |
| 3257 | Name.ends_with(Suffix: "d" ) || Name.ends_with(Suffix: "q" )) |
| 3258 | IsSigned = true; |
| 3259 | else |
| 3260 | llvm_unreachable("Unknown suffix" ); |
| 3261 | |
| 3262 | unsigned Imm; |
| 3263 | if (CI->arg_size() == 3) { |
| 3264 | Imm = cast<ConstantInt>(Val: CI->getArgOperand(i: 2))->getZExtValue(); |
| 3265 | } else { |
| 3266 | Name = Name.substr(Start: 9); // strip off "xop.vpcom" |
| 3267 | if (Name.starts_with(Prefix: "lt" )) |
| 3268 | Imm = 0; |
| 3269 | else if (Name.starts_with(Prefix: "le" )) |
| 3270 | Imm = 1; |
| 3271 | else if (Name.starts_with(Prefix: "gt" )) |
| 3272 | Imm = 2; |
| 3273 | else if (Name.starts_with(Prefix: "ge" )) |
| 3274 | Imm = 3; |
| 3275 | else if (Name.starts_with(Prefix: "eq" )) |
| 3276 | Imm = 4; |
| 3277 | else if (Name.starts_with(Prefix: "ne" )) |
| 3278 | Imm = 5; |
| 3279 | else if (Name.starts_with(Prefix: "false" )) |
| 3280 | Imm = 6; |
| 3281 | else if (Name.starts_with(Prefix: "true" )) |
| 3282 | Imm = 7; |
| 3283 | else |
| 3284 | llvm_unreachable("Unknown condition" ); |
| 3285 | } |
| 3286 | |
| 3287 | Rep = upgradeX86vpcom(Builder, CI&: *CI, Imm, IsSigned); |
| 3288 | } else if (Name.starts_with(Prefix: "xop.vpcmov" )) { |
| 3289 | Value *Sel = CI->getArgOperand(i: 2); |
| 3290 | Value *NotSel = Builder.CreateNot(V: Sel); |
| 3291 | Value *Sel0 = Builder.CreateAnd(LHS: CI->getArgOperand(i: 0), RHS: Sel); |
| 3292 | Value *Sel1 = Builder.CreateAnd(LHS: CI->getArgOperand(i: 1), RHS: NotSel); |
| 3293 | Rep = Builder.CreateOr(LHS: Sel0, RHS: Sel1); |
| 3294 | } else if (Name.starts_with(Prefix: "xop.vprot" ) || Name.starts_with(Prefix: "avx512.prol" ) || |
| 3295 | Name.starts_with(Prefix: "avx512.mask.prol" )) { |
| 3296 | Rep = upgradeX86Rotate(Builder, CI&: *CI, IsRotateRight: false); |
| 3297 | } else if (Name.starts_with(Prefix: "avx512.pror" ) || |
| 3298 | Name.starts_with(Prefix: "avx512.mask.pror" )) { |
| 3299 | Rep = upgradeX86Rotate(Builder, CI&: *CI, IsRotateRight: true); |
| 3300 | } else if (Name.starts_with(Prefix: "avx512.vpshld." ) || |
| 3301 | Name.starts_with(Prefix: "avx512.mask.vpshld" ) || |
| 3302 | Name.starts_with(Prefix: "avx512.maskz.vpshld" )) { |
| 3303 | bool ZeroMask = Name[11] == 'z'; |
| 3304 | Rep = upgradeX86ConcatShift(Builder, CI&: *CI, IsShiftRight: false, ZeroMask); |
| 3305 | } else if (Name.starts_with(Prefix: "avx512.vpshrd." ) || |
| 3306 | Name.starts_with(Prefix: "avx512.mask.vpshrd" ) || |
| 3307 | Name.starts_with(Prefix: "avx512.maskz.vpshrd" )) { |
| 3308 | bool ZeroMask = Name[11] == 'z'; |
| 3309 | Rep = upgradeX86ConcatShift(Builder, CI&: *CI, IsShiftRight: true, ZeroMask); |
| 3310 | } else if (Name == "sse42.crc32.64.8" ) { |
| 3311 | Value *Trunc0 = |
| 3312 | Builder.CreateTrunc(V: CI->getArgOperand(i: 0), DestTy: Type::getInt32Ty(C)); |
| 3313 | Rep = Builder.CreateIntrinsic(ID: Intrinsic::x86_sse42_crc32_32_8, |
| 3314 | Args: {Trunc0, CI->getArgOperand(i: 1)}); |
| 3315 | Rep = Builder.CreateZExt(V: Rep, DestTy: CI->getType(), Name: "" ); |
| 3316 | } else if (Name.starts_with(Prefix: "avx.vbroadcast.s" ) || |
| 3317 | Name.starts_with(Prefix: "avx512.vbroadcast.s" )) { |
| 3318 | // Replace broadcasts with a series of insertelements. |
| 3319 | auto *VecTy = cast<FixedVectorType>(Val: CI->getType()); |
| 3320 | Type *EltTy = VecTy->getElementType(); |
| 3321 | unsigned EltNum = VecTy->getNumElements(); |
| 3322 | Value *Load = Builder.CreateLoad(Ty: EltTy, Ptr: CI->getArgOperand(i: 0)); |
| 3323 | Type *I32Ty = Type::getInt32Ty(C); |
| 3324 | Rep = PoisonValue::get(T: VecTy); |
| 3325 | for (unsigned I = 0; I < EltNum; ++I) |
| 3326 | Rep = Builder.CreateInsertElement(Vec: Rep, NewElt: Load, Idx: ConstantInt::get(Ty: I32Ty, V: I)); |
| 3327 | } else if (Name.starts_with(Prefix: "sse41.pmovsx" ) || |
| 3328 | Name.starts_with(Prefix: "sse41.pmovzx" ) || |
| 3329 | Name.starts_with(Prefix: "avx2.pmovsx" ) || |
| 3330 | Name.starts_with(Prefix: "avx2.pmovzx" ) || |
| 3331 | Name.starts_with(Prefix: "avx512.mask.pmovsx" ) || |
| 3332 | Name.starts_with(Prefix: "avx512.mask.pmovzx" )) { |
| 3333 | auto *DstTy = cast<FixedVectorType>(Val: CI->getType()); |
| 3334 | unsigned NumDstElts = DstTy->getNumElements(); |
| 3335 | |
| 3336 | // Extract a subvector of the first NumDstElts lanes and sign/zero extend. |
| 3337 | SmallVector<int, 8> ShuffleMask(NumDstElts); |
| 3338 | for (unsigned i = 0; i != NumDstElts; ++i) |
| 3339 | ShuffleMask[i] = i; |
| 3340 | |
| 3341 | Value *SV = Builder.CreateShuffleVector(V: CI->getArgOperand(i: 0), Mask: ShuffleMask); |
| 3342 | |
| 3343 | bool DoSext = Name.contains(Other: "pmovsx" ); |
| 3344 | Rep = |
| 3345 | DoSext ? Builder.CreateSExt(V: SV, DestTy: DstTy) : Builder.CreateZExt(V: SV, DestTy: DstTy); |
| 3346 | // If there are 3 arguments, it's a masked intrinsic so we need a select. |
| 3347 | if (CI->arg_size() == 3) |
| 3348 | Rep = emitX86Select(Builder, Mask: CI->getArgOperand(i: 2), Op0: Rep, |
| 3349 | Op1: CI->getArgOperand(i: 1)); |
| 3350 | } else if (Name == "avx512.mask.pmov.qd.256" || |
| 3351 | Name == "avx512.mask.pmov.qd.512" || |
| 3352 | Name == "avx512.mask.pmov.wb.256" || |
| 3353 | Name == "avx512.mask.pmov.wb.512" ) { |
| 3354 | Type *Ty = CI->getArgOperand(i: 1)->getType(); |
| 3355 | Rep = Builder.CreateTrunc(V: CI->getArgOperand(i: 0), DestTy: Ty); |
| 3356 | Rep = |
| 3357 | emitX86Select(Builder, Mask: CI->getArgOperand(i: 2), Op0: Rep, Op1: CI->getArgOperand(i: 1)); |
| 3358 | } else if (Name.starts_with(Prefix: "avx.vbroadcastf128" ) || |
| 3359 | Name == "avx2.vbroadcasti128" ) { |
| 3360 | // Replace vbroadcastf128/vbroadcasti128 with a vector load+shuffle. |
| 3361 | Type *EltTy = cast<VectorType>(Val: CI->getType())->getElementType(); |
| 3362 | unsigned NumSrcElts = 128 / EltTy->getPrimitiveSizeInBits(); |
| 3363 | auto *VT = FixedVectorType::get(ElementType: EltTy, NumElts: NumSrcElts); |
| 3364 | Value *Load = Builder.CreateAlignedLoad(Ty: VT, Ptr: CI->getArgOperand(i: 0), Align: Align(1)); |
| 3365 | if (NumSrcElts == 2) |
| 3366 | Rep = Builder.CreateShuffleVector(V: Load, Mask: ArrayRef<int>{0, 1, 0, 1}); |
| 3367 | else |
| 3368 | Rep = Builder.CreateShuffleVector(V: Load, |
| 3369 | Mask: ArrayRef<int>{0, 1, 2, 3, 0, 1, 2, 3}); |
| 3370 | } else if (Name.starts_with(Prefix: "avx512.mask.shuf.i" ) || |
| 3371 | Name.starts_with(Prefix: "avx512.mask.shuf.f" )) { |
| 3372 | unsigned Imm = cast<ConstantInt>(Val: CI->getArgOperand(i: 2))->getZExtValue(); |
| 3373 | Type *VT = CI->getType(); |
| 3374 | unsigned NumLanes = VT->getPrimitiveSizeInBits() / 128; |
| 3375 | unsigned NumElementsInLane = 128 / VT->getScalarSizeInBits(); |
| 3376 | unsigned ControlBitsMask = NumLanes - 1; |
| 3377 | unsigned NumControlBits = NumLanes / 2; |
| 3378 | SmallVector<int, 8> ShuffleMask(0); |
| 3379 | |
| 3380 | for (unsigned l = 0; l != NumLanes; ++l) { |
| 3381 | unsigned LaneMask = (Imm >> (l * NumControlBits)) & ControlBitsMask; |
| 3382 | // We actually need the other source. |
| 3383 | if (l >= NumLanes / 2) |
| 3384 | LaneMask += NumLanes; |
| 3385 | for (unsigned i = 0; i != NumElementsInLane; ++i) |
| 3386 | ShuffleMask.push_back(Elt: LaneMask * NumElementsInLane + i); |
| 3387 | } |
| 3388 | Rep = Builder.CreateShuffleVector(V1: CI->getArgOperand(i: 0), |
| 3389 | V2: CI->getArgOperand(i: 1), Mask: ShuffleMask); |
| 3390 | Rep = |
| 3391 | emitX86Select(Builder, Mask: CI->getArgOperand(i: 4), Op0: Rep, Op1: CI->getArgOperand(i: 3)); |
| 3392 | } else if (Name.starts_with(Prefix: "avx512.mask.broadcastf" ) || |
| 3393 | Name.starts_with(Prefix: "avx512.mask.broadcasti" )) { |
| 3394 | unsigned NumSrcElts = cast<FixedVectorType>(Val: CI->getArgOperand(i: 0)->getType()) |
| 3395 | ->getNumElements(); |
| 3396 | unsigned NumDstElts = |
| 3397 | cast<FixedVectorType>(Val: CI->getType())->getNumElements(); |
| 3398 | |
| 3399 | SmallVector<int, 8> ShuffleMask(NumDstElts); |
| 3400 | for (unsigned i = 0; i != NumDstElts; ++i) |
| 3401 | ShuffleMask[i] = i % NumSrcElts; |
| 3402 | |
| 3403 | Rep = Builder.CreateShuffleVector(V1: CI->getArgOperand(i: 0), |
| 3404 | V2: CI->getArgOperand(i: 0), Mask: ShuffleMask); |
| 3405 | Rep = |
| 3406 | emitX86Select(Builder, Mask: CI->getArgOperand(i: 2), Op0: Rep, Op1: CI->getArgOperand(i: 1)); |
| 3407 | } else if (Name.starts_with(Prefix: "avx2.pbroadcast" ) || |
| 3408 | Name.starts_with(Prefix: "avx2.vbroadcast" ) || |
| 3409 | Name.starts_with(Prefix: "avx512.pbroadcast" ) || |
| 3410 | Name.starts_with(Prefix: "avx512.mask.broadcast.s" )) { |
| 3411 | // Replace vp?broadcasts with a vector shuffle. |
| 3412 | Value *Op = CI->getArgOperand(i: 0); |
| 3413 | ElementCount EC = cast<VectorType>(Val: CI->getType())->getElementCount(); |
| 3414 | Type *MaskTy = VectorType::get(ElementType: Type::getInt32Ty(C), EC); |
| 3415 | SmallVector<int, 8> M; |
| 3416 | ShuffleVectorInst::getShuffleMask(Mask: Constant::getNullValue(Ty: MaskTy), Result&: M); |
| 3417 | Rep = Builder.CreateShuffleVector(V: Op, Mask: M); |
| 3418 | |
| 3419 | if (CI->arg_size() == 3) |
| 3420 | Rep = emitX86Select(Builder, Mask: CI->getArgOperand(i: 2), Op0: Rep, |
| 3421 | Op1: CI->getArgOperand(i: 1)); |
| 3422 | } else if (Name.starts_with(Prefix: "sse2.padds." ) || |
| 3423 | Name.starts_with(Prefix: "avx2.padds." ) || |
| 3424 | Name.starts_with(Prefix: "avx512.padds." ) || |
| 3425 | Name.starts_with(Prefix: "avx512.mask.padds." )) { |
| 3426 | Rep = upgradeX86BinaryIntrinsics(Builder, CI&: *CI, IID: Intrinsic::sadd_sat); |
| 3427 | } else if (Name.starts_with(Prefix: "sse2.psubs." ) || |
| 3428 | Name.starts_with(Prefix: "avx2.psubs." ) || |
| 3429 | Name.starts_with(Prefix: "avx512.psubs." ) || |
| 3430 | Name.starts_with(Prefix: "avx512.mask.psubs." )) { |
| 3431 | Rep = upgradeX86BinaryIntrinsics(Builder, CI&: *CI, IID: Intrinsic::ssub_sat); |
| 3432 | } else if (Name.starts_with(Prefix: "sse2.paddus." ) || |
| 3433 | Name.starts_with(Prefix: "avx2.paddus." ) || |
| 3434 | Name.starts_with(Prefix: "avx512.mask.paddus." )) { |
| 3435 | Rep = upgradeX86BinaryIntrinsics(Builder, CI&: *CI, IID: Intrinsic::uadd_sat); |
| 3436 | } else if (Name.starts_with(Prefix: "sse2.psubus." ) || |
| 3437 | Name.starts_with(Prefix: "avx2.psubus." ) || |
| 3438 | Name.starts_with(Prefix: "avx512.mask.psubus." )) { |
| 3439 | Rep = upgradeX86BinaryIntrinsics(Builder, CI&: *CI, IID: Intrinsic::usub_sat); |
| 3440 | } else if (Name.starts_with(Prefix: "avx512.mask.palignr." )) { |
| 3441 | Rep = upgradeX86ALIGNIntrinsics(Builder, Op0: CI->getArgOperand(i: 0), |
| 3442 | Op1: CI->getArgOperand(i: 1), Shift: CI->getArgOperand(i: 2), |
| 3443 | Passthru: CI->getArgOperand(i: 3), Mask: CI->getArgOperand(i: 4), |
| 3444 | IsVALIGN: false); |
| 3445 | } else if (Name.starts_with(Prefix: "avx512.mask.valign." )) { |
| 3446 | Rep = upgradeX86ALIGNIntrinsics( |
| 3447 | Builder, Op0: CI->getArgOperand(i: 0), Op1: CI->getArgOperand(i: 1), |
| 3448 | Shift: CI->getArgOperand(i: 2), Passthru: CI->getArgOperand(i: 3), Mask: CI->getArgOperand(i: 4), IsVALIGN: true); |
| 3449 | } else if (Name == "sse2.psll.dq" || Name == "avx2.psll.dq" ) { |
| 3450 | // 128/256-bit shift left specified in bits. |
| 3451 | unsigned Shift = cast<ConstantInt>(Val: CI->getArgOperand(i: 1))->getZExtValue(); |
| 3452 | Rep = upgradeX86PSLLDQIntrinsics(Builder, Op: CI->getArgOperand(i: 0), |
| 3453 | Shift: Shift / 8); // Shift is in bits. |
| 3454 | } else if (Name == "sse2.psrl.dq" || Name == "avx2.psrl.dq" ) { |
| 3455 | // 128/256-bit shift right specified in bits. |
| 3456 | unsigned Shift = cast<ConstantInt>(Val: CI->getArgOperand(i: 1))->getZExtValue(); |
| 3457 | Rep = upgradeX86PSRLDQIntrinsics(Builder, Op: CI->getArgOperand(i: 0), |
| 3458 | Shift: Shift / 8); // Shift is in bits. |
| 3459 | } else if (Name == "sse2.psll.dq.bs" || Name == "avx2.psll.dq.bs" || |
| 3460 | Name == "avx512.psll.dq.512" ) { |
| 3461 | // 128/256/512-bit shift left specified in bytes. |
| 3462 | unsigned Shift = cast<ConstantInt>(Val: CI->getArgOperand(i: 1))->getZExtValue(); |
| 3463 | Rep = upgradeX86PSLLDQIntrinsics(Builder, Op: CI->getArgOperand(i: 0), Shift); |
| 3464 | } else if (Name == "sse2.psrl.dq.bs" || Name == "avx2.psrl.dq.bs" || |
| 3465 | Name == "avx512.psrl.dq.512" ) { |
| 3466 | // 128/256/512-bit shift right specified in bytes. |
| 3467 | unsigned Shift = cast<ConstantInt>(Val: CI->getArgOperand(i: 1))->getZExtValue(); |
| 3468 | Rep = upgradeX86PSRLDQIntrinsics(Builder, Op: CI->getArgOperand(i: 0), Shift); |
| 3469 | } else if (Name == "sse41.pblendw" || Name.starts_with(Prefix: "sse41.blendp" ) || |
| 3470 | Name.starts_with(Prefix: "avx.blend.p" ) || Name == "avx2.pblendw" || |
| 3471 | Name.starts_with(Prefix: "avx2.pblendd." )) { |
| 3472 | Value *Op0 = CI->getArgOperand(i: 0); |
| 3473 | Value *Op1 = CI->getArgOperand(i: 1); |
| 3474 | unsigned Imm = cast<ConstantInt>(Val: CI->getArgOperand(i: 2))->getZExtValue(); |
| 3475 | auto *VecTy = cast<FixedVectorType>(Val: CI->getType()); |
| 3476 | unsigned NumElts = VecTy->getNumElements(); |
| 3477 | |
| 3478 | SmallVector<int, 16> Idxs(NumElts); |
| 3479 | for (unsigned i = 0; i != NumElts; ++i) |
| 3480 | Idxs[i] = ((Imm >> (i % 8)) & 1) ? i + NumElts : i; |
| 3481 | |
| 3482 | Rep = Builder.CreateShuffleVector(V1: Op0, V2: Op1, Mask: Idxs); |
| 3483 | } else if (Name.starts_with(Prefix: "avx.vinsertf128." ) || |
| 3484 | Name == "avx2.vinserti128" || |
| 3485 | Name.starts_with(Prefix: "avx512.mask.insert" )) { |
| 3486 | Value *Op0 = CI->getArgOperand(i: 0); |
| 3487 | Value *Op1 = CI->getArgOperand(i: 1); |
| 3488 | unsigned Imm = cast<ConstantInt>(Val: CI->getArgOperand(i: 2))->getZExtValue(); |
| 3489 | unsigned DstNumElts = |
| 3490 | cast<FixedVectorType>(Val: CI->getType())->getNumElements(); |
| 3491 | unsigned SrcNumElts = |
| 3492 | cast<FixedVectorType>(Val: Op1->getType())->getNumElements(); |
| 3493 | unsigned Scale = DstNumElts / SrcNumElts; |
| 3494 | |
| 3495 | // Mask off the high bits of the immediate value; hardware ignores those. |
| 3496 | Imm = Imm % Scale; |
| 3497 | |
| 3498 | // Extend the second operand into a vector the size of the destination. |
| 3499 | SmallVector<int, 8> Idxs(DstNumElts); |
| 3500 | for (unsigned i = 0; i != SrcNumElts; ++i) |
| 3501 | Idxs[i] = i; |
| 3502 | for (unsigned i = SrcNumElts; i != DstNumElts; ++i) |
| 3503 | Idxs[i] = SrcNumElts; |
| 3504 | Rep = Builder.CreateShuffleVector(V: Op1, Mask: Idxs); |
| 3505 | |
| 3506 | // Insert the second operand into the first operand. |
| 3507 | |
| 3508 | // Note that there is no guarantee that instruction lowering will actually |
| 3509 | // produce a vinsertf128 instruction for the created shuffles. In |
| 3510 | // particular, the 0 immediate case involves no lane changes, so it can |
| 3511 | // be handled as a blend. |
| 3512 | |
| 3513 | // Example of shuffle mask for 32-bit elements: |
| 3514 | // Imm = 1 <i32 0, i32 1, i32 2, i32 3, i32 8, i32 9, i32 10, i32 11> |
| 3515 | // Imm = 0 <i32 8, i32 9, i32 10, i32 11, i32 4, i32 5, i32 6, i32 7 > |
| 3516 | |
| 3517 | // First fill with identify mask. |
| 3518 | for (unsigned i = 0; i != DstNumElts; ++i) |
| 3519 | Idxs[i] = i; |
| 3520 | // Then replace the elements where we need to insert. |
| 3521 | for (unsigned i = 0; i != SrcNumElts; ++i) |
| 3522 | Idxs[i + Imm * SrcNumElts] = i + DstNumElts; |
| 3523 | Rep = Builder.CreateShuffleVector(V1: Op0, V2: Rep, Mask: Idxs); |
| 3524 | |
| 3525 | // If the intrinsic has a mask operand, handle that. |
| 3526 | if (CI->arg_size() == 5) |
| 3527 | Rep = emitX86Select(Builder, Mask: CI->getArgOperand(i: 4), Op0: Rep, |
| 3528 | Op1: CI->getArgOperand(i: 3)); |
| 3529 | } else if (Name.starts_with(Prefix: "avx.vextractf128." ) || |
| 3530 | Name == "avx2.vextracti128" || |
| 3531 | Name.starts_with(Prefix: "avx512.mask.vextract" )) { |
| 3532 | Value *Op0 = CI->getArgOperand(i: 0); |
| 3533 | unsigned Imm = cast<ConstantInt>(Val: CI->getArgOperand(i: 1))->getZExtValue(); |
| 3534 | unsigned DstNumElts = |
| 3535 | cast<FixedVectorType>(Val: CI->getType())->getNumElements(); |
| 3536 | unsigned SrcNumElts = |
| 3537 | cast<FixedVectorType>(Val: Op0->getType())->getNumElements(); |
| 3538 | unsigned Scale = SrcNumElts / DstNumElts; |
| 3539 | |
| 3540 | // Mask off the high bits of the immediate value; hardware ignores those. |
| 3541 | Imm = Imm % Scale; |
| 3542 | |
| 3543 | // Get indexes for the subvector of the input vector. |
| 3544 | SmallVector<int, 8> Idxs(DstNumElts); |
| 3545 | for (unsigned i = 0; i != DstNumElts; ++i) { |
| 3546 | Idxs[i] = i + (Imm * DstNumElts); |
| 3547 | } |
| 3548 | Rep = Builder.CreateShuffleVector(V1: Op0, V2: Op0, Mask: Idxs); |
| 3549 | |
| 3550 | // If the intrinsic has a mask operand, handle that. |
| 3551 | if (CI->arg_size() == 4) |
| 3552 | Rep = emitX86Select(Builder, Mask: CI->getArgOperand(i: 3), Op0: Rep, |
| 3553 | Op1: CI->getArgOperand(i: 2)); |
| 3554 | } else if (Name.starts_with(Prefix: "avx512.mask.perm.df." ) || |
| 3555 | Name.starts_with(Prefix: "avx512.mask.perm.di." )) { |
| 3556 | Value *Op0 = CI->getArgOperand(i: 0); |
| 3557 | unsigned Imm = cast<ConstantInt>(Val: CI->getArgOperand(i: 1))->getZExtValue(); |
| 3558 | auto *VecTy = cast<FixedVectorType>(Val: CI->getType()); |
| 3559 | unsigned NumElts = VecTy->getNumElements(); |
| 3560 | |
| 3561 | SmallVector<int, 8> Idxs(NumElts); |
| 3562 | for (unsigned i = 0; i != NumElts; ++i) |
| 3563 | Idxs[i] = (i & ~0x3) + ((Imm >> (2 * (i & 0x3))) & 3); |
| 3564 | |
| 3565 | Rep = Builder.CreateShuffleVector(V1: Op0, V2: Op0, Mask: Idxs); |
| 3566 | |
| 3567 | if (CI->arg_size() == 4) |
| 3568 | Rep = emitX86Select(Builder, Mask: CI->getArgOperand(i: 3), Op0: Rep, |
| 3569 | Op1: CI->getArgOperand(i: 2)); |
| 3570 | } else if (Name.starts_with(Prefix: "avx.vperm2f128." ) || Name == "avx2.vperm2i128" ) { |
| 3571 | // The immediate permute control byte looks like this: |
| 3572 | // [1:0] - select 128 bits from sources for low half of destination |
| 3573 | // [2] - ignore |
| 3574 | // [3] - zero low half of destination |
| 3575 | // [5:4] - select 128 bits from sources for high half of destination |
| 3576 | // [6] - ignore |
| 3577 | // [7] - zero high half of destination |
| 3578 | |
| 3579 | uint8_t Imm = cast<ConstantInt>(Val: CI->getArgOperand(i: 2))->getZExtValue(); |
| 3580 | |
| 3581 | unsigned NumElts = cast<FixedVectorType>(Val: CI->getType())->getNumElements(); |
| 3582 | unsigned HalfSize = NumElts / 2; |
| 3583 | SmallVector<int, 8> ShuffleMask(NumElts); |
| 3584 | |
| 3585 | // Determine which operand(s) are actually in use for this instruction. |
| 3586 | Value *V0 = (Imm & 0x02) ? CI->getArgOperand(i: 1) : CI->getArgOperand(i: 0); |
| 3587 | Value *V1 = (Imm & 0x20) ? CI->getArgOperand(i: 1) : CI->getArgOperand(i: 0); |
| 3588 | |
| 3589 | // If needed, replace operands based on zero mask. |
| 3590 | V0 = (Imm & 0x08) ? ConstantAggregateZero::get(Ty: CI->getType()) : V0; |
| 3591 | V1 = (Imm & 0x80) ? ConstantAggregateZero::get(Ty: CI->getType()) : V1; |
| 3592 | |
| 3593 | // Permute low half of result. |
| 3594 | unsigned StartIndex = (Imm & 0x01) ? HalfSize : 0; |
| 3595 | for (unsigned i = 0; i < HalfSize; ++i) |
| 3596 | ShuffleMask[i] = StartIndex + i; |
| 3597 | |
| 3598 | // Permute high half of result. |
| 3599 | StartIndex = (Imm & 0x10) ? HalfSize : 0; |
| 3600 | for (unsigned i = 0; i < HalfSize; ++i) |
| 3601 | ShuffleMask[i + HalfSize] = NumElts + StartIndex + i; |
| 3602 | |
| 3603 | Rep = Builder.CreateShuffleVector(V1: V0, V2: V1, Mask: ShuffleMask); |
| 3604 | |
| 3605 | } else if (Name.starts_with(Prefix: "avx.vpermil." ) || Name == "sse2.pshuf.d" || |
| 3606 | Name.starts_with(Prefix: "avx512.mask.vpermil.p" ) || |
| 3607 | Name.starts_with(Prefix: "avx512.mask.pshuf.d." )) { |
| 3608 | Value *Op0 = CI->getArgOperand(i: 0); |
| 3609 | unsigned Imm = cast<ConstantInt>(Val: CI->getArgOperand(i: 1))->getZExtValue(); |
| 3610 | auto *VecTy = cast<FixedVectorType>(Val: CI->getType()); |
| 3611 | unsigned NumElts = VecTy->getNumElements(); |
| 3612 | // Calculate the size of each index in the immediate. |
| 3613 | unsigned IdxSize = 64 / VecTy->getScalarSizeInBits(); |
| 3614 | unsigned IdxMask = ((1 << IdxSize) - 1); |
| 3615 | |
| 3616 | SmallVector<int, 8> Idxs(NumElts); |
| 3617 | // Lookup the bits for this element, wrapping around the immediate every |
| 3618 | // 8-bits. Elements are grouped into sets of 2 or 4 elements so we need |
| 3619 | // to offset by the first index of each group. |
| 3620 | for (unsigned i = 0; i != NumElts; ++i) |
| 3621 | Idxs[i] = ((Imm >> ((i * IdxSize) % 8)) & IdxMask) | (i & ~IdxMask); |
| 3622 | |
| 3623 | Rep = Builder.CreateShuffleVector(V1: Op0, V2: Op0, Mask: Idxs); |
| 3624 | |
| 3625 | if (CI->arg_size() == 4) |
| 3626 | Rep = emitX86Select(Builder, Mask: CI->getArgOperand(i: 3), Op0: Rep, |
| 3627 | Op1: CI->getArgOperand(i: 2)); |
| 3628 | } else if (Name == "sse2.pshufl.w" || |
| 3629 | Name.starts_with(Prefix: "avx512.mask.pshufl.w." )) { |
| 3630 | Value *Op0 = CI->getArgOperand(i: 0); |
| 3631 | unsigned Imm = cast<ConstantInt>(Val: CI->getArgOperand(i: 1))->getZExtValue(); |
| 3632 | unsigned NumElts = cast<FixedVectorType>(Val: CI->getType())->getNumElements(); |
| 3633 | |
| 3634 | SmallVector<int, 16> Idxs(NumElts); |
| 3635 | for (unsigned l = 0; l != NumElts; l += 8) { |
| 3636 | for (unsigned i = 0; i != 4; ++i) |
| 3637 | Idxs[i + l] = ((Imm >> (2 * i)) & 0x3) + l; |
| 3638 | for (unsigned i = 4; i != 8; ++i) |
| 3639 | Idxs[i + l] = i + l; |
| 3640 | } |
| 3641 | |
| 3642 | Rep = Builder.CreateShuffleVector(V1: Op0, V2: Op0, Mask: Idxs); |
| 3643 | |
| 3644 | if (CI->arg_size() == 4) |
| 3645 | Rep = emitX86Select(Builder, Mask: CI->getArgOperand(i: 3), Op0: Rep, |
| 3646 | Op1: CI->getArgOperand(i: 2)); |
| 3647 | } else if (Name == "sse2.pshufh.w" || |
| 3648 | Name.starts_with(Prefix: "avx512.mask.pshufh.w." )) { |
| 3649 | Value *Op0 = CI->getArgOperand(i: 0); |
| 3650 | unsigned Imm = cast<ConstantInt>(Val: CI->getArgOperand(i: 1))->getZExtValue(); |
| 3651 | unsigned NumElts = cast<FixedVectorType>(Val: CI->getType())->getNumElements(); |
| 3652 | |
| 3653 | SmallVector<int, 16> Idxs(NumElts); |
| 3654 | for (unsigned l = 0; l != NumElts; l += 8) { |
| 3655 | for (unsigned i = 0; i != 4; ++i) |
| 3656 | Idxs[i + l] = i + l; |
| 3657 | for (unsigned i = 0; i != 4; ++i) |
| 3658 | Idxs[i + l + 4] = ((Imm >> (2 * i)) & 0x3) + 4 + l; |
| 3659 | } |
| 3660 | |
| 3661 | Rep = Builder.CreateShuffleVector(V1: Op0, V2: Op0, Mask: Idxs); |
| 3662 | |
| 3663 | if (CI->arg_size() == 4) |
| 3664 | Rep = emitX86Select(Builder, Mask: CI->getArgOperand(i: 3), Op0: Rep, |
| 3665 | Op1: CI->getArgOperand(i: 2)); |
| 3666 | } else if (Name.starts_with(Prefix: "avx512.mask.shuf.p" )) { |
| 3667 | Value *Op0 = CI->getArgOperand(i: 0); |
| 3668 | Value *Op1 = CI->getArgOperand(i: 1); |
| 3669 | unsigned Imm = cast<ConstantInt>(Val: CI->getArgOperand(i: 2))->getZExtValue(); |
| 3670 | unsigned NumElts = cast<FixedVectorType>(Val: CI->getType())->getNumElements(); |
| 3671 | |
| 3672 | unsigned NumLaneElts = 128 / CI->getType()->getScalarSizeInBits(); |
| 3673 | unsigned HalfLaneElts = NumLaneElts / 2; |
| 3674 | |
| 3675 | SmallVector<int, 16> Idxs(NumElts); |
| 3676 | for (unsigned i = 0; i != NumElts; ++i) { |
| 3677 | // Base index is the starting element of the lane. |
| 3678 | Idxs[i] = i - (i % NumLaneElts); |
| 3679 | // If we are half way through the lane switch to the other source. |
| 3680 | if ((i % NumLaneElts) >= HalfLaneElts) |
| 3681 | Idxs[i] += NumElts; |
| 3682 | // Now select the specific element. By adding HalfLaneElts bits from |
| 3683 | // the immediate. Wrapping around the immediate every 8-bits. |
| 3684 | Idxs[i] += (Imm >> ((i * HalfLaneElts) % 8)) & ((1 << HalfLaneElts) - 1); |
| 3685 | } |
| 3686 | |
| 3687 | Rep = Builder.CreateShuffleVector(V1: Op0, V2: Op1, Mask: Idxs); |
| 3688 | |
| 3689 | Rep = |
| 3690 | emitX86Select(Builder, Mask: CI->getArgOperand(i: 4), Op0: Rep, Op1: CI->getArgOperand(i: 3)); |
| 3691 | } else if (Name.starts_with(Prefix: "avx512.mask.movddup" ) || |
| 3692 | Name.starts_with(Prefix: "avx512.mask.movshdup" ) || |
| 3693 | Name.starts_with(Prefix: "avx512.mask.movsldup" )) { |
| 3694 | Value *Op0 = CI->getArgOperand(i: 0); |
| 3695 | unsigned NumElts = cast<FixedVectorType>(Val: CI->getType())->getNumElements(); |
| 3696 | unsigned NumLaneElts = 128 / CI->getType()->getScalarSizeInBits(); |
| 3697 | |
| 3698 | unsigned Offset = 0; |
| 3699 | if (Name.starts_with(Prefix: "avx512.mask.movshdup." )) |
| 3700 | Offset = 1; |
| 3701 | |
| 3702 | SmallVector<int, 16> Idxs(NumElts); |
| 3703 | for (unsigned l = 0; l != NumElts; l += NumLaneElts) |
| 3704 | for (unsigned i = 0; i != NumLaneElts; i += 2) { |
| 3705 | Idxs[i + l + 0] = i + l + Offset; |
| 3706 | Idxs[i + l + 1] = i + l + Offset; |
| 3707 | } |
| 3708 | |
| 3709 | Rep = Builder.CreateShuffleVector(V1: Op0, V2: Op0, Mask: Idxs); |
| 3710 | |
| 3711 | Rep = |
| 3712 | emitX86Select(Builder, Mask: CI->getArgOperand(i: 2), Op0: Rep, Op1: CI->getArgOperand(i: 1)); |
| 3713 | } else if (Name.starts_with(Prefix: "avx512.mask.punpckl" ) || |
| 3714 | Name.starts_with(Prefix: "avx512.mask.unpckl." )) { |
| 3715 | Value *Op0 = CI->getArgOperand(i: 0); |
| 3716 | Value *Op1 = CI->getArgOperand(i: 1); |
| 3717 | int NumElts = cast<FixedVectorType>(Val: CI->getType())->getNumElements(); |
| 3718 | int NumLaneElts = 128 / CI->getType()->getScalarSizeInBits(); |
| 3719 | |
| 3720 | SmallVector<int, 64> Idxs(NumElts); |
| 3721 | for (int l = 0; l != NumElts; l += NumLaneElts) |
| 3722 | for (int i = 0; i != NumLaneElts; ++i) |
| 3723 | Idxs[i + l] = l + (i / 2) + NumElts * (i % 2); |
| 3724 | |
| 3725 | Rep = Builder.CreateShuffleVector(V1: Op0, V2: Op1, Mask: Idxs); |
| 3726 | |
| 3727 | Rep = |
| 3728 | emitX86Select(Builder, Mask: CI->getArgOperand(i: 3), Op0: Rep, Op1: CI->getArgOperand(i: 2)); |
| 3729 | } else if (Name.starts_with(Prefix: "avx512.mask.punpckh" ) || |
| 3730 | Name.starts_with(Prefix: "avx512.mask.unpckh." )) { |
| 3731 | Value *Op0 = CI->getArgOperand(i: 0); |
| 3732 | Value *Op1 = CI->getArgOperand(i: 1); |
| 3733 | int NumElts = cast<FixedVectorType>(Val: CI->getType())->getNumElements(); |
| 3734 | int NumLaneElts = 128 / CI->getType()->getScalarSizeInBits(); |
| 3735 | |
| 3736 | SmallVector<int, 64> Idxs(NumElts); |
| 3737 | for (int l = 0; l != NumElts; l += NumLaneElts) |
| 3738 | for (int i = 0; i != NumLaneElts; ++i) |
| 3739 | Idxs[i + l] = (NumLaneElts / 2) + l + (i / 2) + NumElts * (i % 2); |
| 3740 | |
| 3741 | Rep = Builder.CreateShuffleVector(V1: Op0, V2: Op1, Mask: Idxs); |
| 3742 | |
| 3743 | Rep = |
| 3744 | emitX86Select(Builder, Mask: CI->getArgOperand(i: 3), Op0: Rep, Op1: CI->getArgOperand(i: 2)); |
| 3745 | } else if (Name.starts_with(Prefix: "avx512.mask.and." ) || |
| 3746 | Name.starts_with(Prefix: "avx512.mask.pand." )) { |
| 3747 | VectorType *FTy = cast<VectorType>(Val: CI->getType()); |
| 3748 | VectorType *ITy = VectorType::getInteger(VTy: FTy); |
| 3749 | Rep = Builder.CreateAnd(LHS: Builder.CreateBitCast(V: CI->getArgOperand(i: 0), DestTy: ITy), |
| 3750 | RHS: Builder.CreateBitCast(V: CI->getArgOperand(i: 1), DestTy: ITy)); |
| 3751 | Rep = Builder.CreateBitCast(V: Rep, DestTy: FTy); |
| 3752 | Rep = |
| 3753 | emitX86Select(Builder, Mask: CI->getArgOperand(i: 3), Op0: Rep, Op1: CI->getArgOperand(i: 2)); |
| 3754 | } else if (Name.starts_with(Prefix: "avx512.mask.andn." ) || |
| 3755 | Name.starts_with(Prefix: "avx512.mask.pandn." )) { |
| 3756 | VectorType *FTy = cast<VectorType>(Val: CI->getType()); |
| 3757 | VectorType *ITy = VectorType::getInteger(VTy: FTy); |
| 3758 | Rep = Builder.CreateNot(V: Builder.CreateBitCast(V: CI->getArgOperand(i: 0), DestTy: ITy)); |
| 3759 | Rep = Builder.CreateAnd(LHS: Rep, |
| 3760 | RHS: Builder.CreateBitCast(V: CI->getArgOperand(i: 1), DestTy: ITy)); |
| 3761 | Rep = Builder.CreateBitCast(V: Rep, DestTy: FTy); |
| 3762 | Rep = |
| 3763 | emitX86Select(Builder, Mask: CI->getArgOperand(i: 3), Op0: Rep, Op1: CI->getArgOperand(i: 2)); |
| 3764 | } else if (Name.starts_with(Prefix: "avx512.mask.or." ) || |
| 3765 | Name.starts_with(Prefix: "avx512.mask.por." )) { |
| 3766 | VectorType *FTy = cast<VectorType>(Val: CI->getType()); |
| 3767 | VectorType *ITy = VectorType::getInteger(VTy: FTy); |
| 3768 | Rep = Builder.CreateOr(LHS: Builder.CreateBitCast(V: CI->getArgOperand(i: 0), DestTy: ITy), |
| 3769 | RHS: Builder.CreateBitCast(V: CI->getArgOperand(i: 1), DestTy: ITy)); |
| 3770 | Rep = Builder.CreateBitCast(V: Rep, DestTy: FTy); |
| 3771 | Rep = |
| 3772 | emitX86Select(Builder, Mask: CI->getArgOperand(i: 3), Op0: Rep, Op1: CI->getArgOperand(i: 2)); |
| 3773 | } else if (Name.starts_with(Prefix: "avx512.mask.xor." ) || |
| 3774 | Name.starts_with(Prefix: "avx512.mask.pxor." )) { |
| 3775 | VectorType *FTy = cast<VectorType>(Val: CI->getType()); |
| 3776 | VectorType *ITy = VectorType::getInteger(VTy: FTy); |
| 3777 | Rep = Builder.CreateXor(LHS: Builder.CreateBitCast(V: CI->getArgOperand(i: 0), DestTy: ITy), |
| 3778 | RHS: Builder.CreateBitCast(V: CI->getArgOperand(i: 1), DestTy: ITy)); |
| 3779 | Rep = Builder.CreateBitCast(V: Rep, DestTy: FTy); |
| 3780 | Rep = |
| 3781 | emitX86Select(Builder, Mask: CI->getArgOperand(i: 3), Op0: Rep, Op1: CI->getArgOperand(i: 2)); |
| 3782 | } else if (Name.starts_with(Prefix: "avx512.mask.padd." )) { |
| 3783 | Rep = Builder.CreateAdd(LHS: CI->getArgOperand(i: 0), RHS: CI->getArgOperand(i: 1)); |
| 3784 | Rep = |
| 3785 | emitX86Select(Builder, Mask: CI->getArgOperand(i: 3), Op0: Rep, Op1: CI->getArgOperand(i: 2)); |
| 3786 | } else if (Name.starts_with(Prefix: "avx512.mask.psub." )) { |
| 3787 | Rep = Builder.CreateSub(LHS: CI->getArgOperand(i: 0), RHS: CI->getArgOperand(i: 1)); |
| 3788 | Rep = |
| 3789 | emitX86Select(Builder, Mask: CI->getArgOperand(i: 3), Op0: Rep, Op1: CI->getArgOperand(i: 2)); |
| 3790 | } else if (Name.starts_with(Prefix: "avx512.mask.pmull." )) { |
| 3791 | Rep = Builder.CreateMul(LHS: CI->getArgOperand(i: 0), RHS: CI->getArgOperand(i: 1)); |
| 3792 | Rep = |
| 3793 | emitX86Select(Builder, Mask: CI->getArgOperand(i: 3), Op0: Rep, Op1: CI->getArgOperand(i: 2)); |
| 3794 | } else if (Name.starts_with(Prefix: "avx512.mask.add.p" )) { |
| 3795 | if (Name.ends_with(Suffix: ".512" )) { |
| 3796 | Intrinsic::ID IID; |
| 3797 | if (Name[17] == 's') |
| 3798 | IID = Intrinsic::x86_avx512_add_ps_512; |
| 3799 | else |
| 3800 | IID = Intrinsic::x86_avx512_add_pd_512; |
| 3801 | |
| 3802 | Rep = Builder.CreateIntrinsic( |
| 3803 | ID: IID, |
| 3804 | Args: {CI->getArgOperand(i: 0), CI->getArgOperand(i: 1), CI->getArgOperand(i: 4)}); |
| 3805 | } else { |
| 3806 | Rep = Builder.CreateFAdd(L: CI->getArgOperand(i: 0), R: CI->getArgOperand(i: 1)); |
| 3807 | } |
| 3808 | Rep = |
| 3809 | emitX86Select(Builder, Mask: CI->getArgOperand(i: 3), Op0: Rep, Op1: CI->getArgOperand(i: 2)); |
| 3810 | } else if (Name.starts_with(Prefix: "avx512.mask.div.p" )) { |
| 3811 | if (Name.ends_with(Suffix: ".512" )) { |
| 3812 | Intrinsic::ID IID; |
| 3813 | if (Name[17] == 's') |
| 3814 | IID = Intrinsic::x86_avx512_div_ps_512; |
| 3815 | else |
| 3816 | IID = Intrinsic::x86_avx512_div_pd_512; |
| 3817 | |
| 3818 | Rep = Builder.CreateIntrinsic( |
| 3819 | ID: IID, |
| 3820 | Args: {CI->getArgOperand(i: 0), CI->getArgOperand(i: 1), CI->getArgOperand(i: 4)}); |
| 3821 | } else { |
| 3822 | Rep = Builder.CreateFDiv(L: CI->getArgOperand(i: 0), R: CI->getArgOperand(i: 1)); |
| 3823 | } |
| 3824 | Rep = |
| 3825 | emitX86Select(Builder, Mask: CI->getArgOperand(i: 3), Op0: Rep, Op1: CI->getArgOperand(i: 2)); |
| 3826 | } else if (Name.starts_with(Prefix: "avx512.mask.mul.p" )) { |
| 3827 | if (Name.ends_with(Suffix: ".512" )) { |
| 3828 | Intrinsic::ID IID; |
| 3829 | if (Name[17] == 's') |
| 3830 | IID = Intrinsic::x86_avx512_mul_ps_512; |
| 3831 | else |
| 3832 | IID = Intrinsic::x86_avx512_mul_pd_512; |
| 3833 | |
| 3834 | Rep = Builder.CreateIntrinsic( |
| 3835 | ID: IID, |
| 3836 | Args: {CI->getArgOperand(i: 0), CI->getArgOperand(i: 1), CI->getArgOperand(i: 4)}); |
| 3837 | } else { |
| 3838 | Rep = Builder.CreateFMul(L: CI->getArgOperand(i: 0), R: CI->getArgOperand(i: 1)); |
| 3839 | } |
| 3840 | Rep = |
| 3841 | emitX86Select(Builder, Mask: CI->getArgOperand(i: 3), Op0: Rep, Op1: CI->getArgOperand(i: 2)); |
| 3842 | } else if (Name.starts_with(Prefix: "avx512.mask.sub.p" )) { |
| 3843 | if (Name.ends_with(Suffix: ".512" )) { |
| 3844 | Intrinsic::ID IID; |
| 3845 | if (Name[17] == 's') |
| 3846 | IID = Intrinsic::x86_avx512_sub_ps_512; |
| 3847 | else |
| 3848 | IID = Intrinsic::x86_avx512_sub_pd_512; |
| 3849 | |
| 3850 | Rep = Builder.CreateIntrinsic( |
| 3851 | ID: IID, |
| 3852 | Args: {CI->getArgOperand(i: 0), CI->getArgOperand(i: 1), CI->getArgOperand(i: 4)}); |
| 3853 | } else { |
| 3854 | Rep = Builder.CreateFSub(L: CI->getArgOperand(i: 0), R: CI->getArgOperand(i: 1)); |
| 3855 | } |
| 3856 | Rep = |
| 3857 | emitX86Select(Builder, Mask: CI->getArgOperand(i: 3), Op0: Rep, Op1: CI->getArgOperand(i: 2)); |
| 3858 | } else if ((Name.starts_with(Prefix: "avx512.mask.max.p" ) || |
| 3859 | Name.starts_with(Prefix: "avx512.mask.min.p" )) && |
| 3860 | Name.drop_front(N: 18) == ".512" ) { |
| 3861 | bool IsDouble = Name[17] == 'd'; |
| 3862 | bool IsMin = Name[13] == 'i'; |
| 3863 | static const Intrinsic::ID MinMaxTbl[2][2] = { |
| 3864 | {Intrinsic::x86_avx512_max_ps_512, Intrinsic::x86_avx512_max_pd_512}, |
| 3865 | {Intrinsic::x86_avx512_min_ps_512, Intrinsic::x86_avx512_min_pd_512}}; |
| 3866 | Intrinsic::ID IID = MinMaxTbl[IsMin][IsDouble]; |
| 3867 | |
| 3868 | Rep = Builder.CreateIntrinsic( |
| 3869 | ID: IID, |
| 3870 | Args: {CI->getArgOperand(i: 0), CI->getArgOperand(i: 1), CI->getArgOperand(i: 4)}); |
| 3871 | Rep = |
| 3872 | emitX86Select(Builder, Mask: CI->getArgOperand(i: 3), Op0: Rep, Op1: CI->getArgOperand(i: 2)); |
| 3873 | } else if (Name.starts_with(Prefix: "avx512.mask.lzcnt." )) { |
| 3874 | Rep = |
| 3875 | Builder.CreateIntrinsic(ID: Intrinsic::ctlz, Types: CI->getType(), |
| 3876 | Args: {CI->getArgOperand(i: 0), Builder.getInt1(V: false)}); |
| 3877 | Rep = |
| 3878 | emitX86Select(Builder, Mask: CI->getArgOperand(i: 2), Op0: Rep, Op1: CI->getArgOperand(i: 1)); |
| 3879 | } else if (Name.starts_with(Prefix: "avx512.mask.psll" )) { |
| 3880 | bool IsImmediate = Name[16] == 'i' || (Name.size() > 18 && Name[18] == 'i'); |
| 3881 | bool IsVariable = Name[16] == 'v'; |
| 3882 | char Size = Name[16] == '.' ? Name[17] |
| 3883 | : Name[17] == '.' ? Name[18] |
| 3884 | : Name[18] == '.' ? Name[19] |
| 3885 | : Name[20]; |
| 3886 | |
| 3887 | Intrinsic::ID IID; |
| 3888 | if (IsVariable && Name[17] != '.') { |
| 3889 | if (Size == 'd' && Name[17] == '2') // avx512.mask.psllv2.di |
| 3890 | IID = Intrinsic::x86_avx2_psllv_q; |
| 3891 | else if (Size == 'd' && Name[17] == '4') // avx512.mask.psllv4.di |
| 3892 | IID = Intrinsic::x86_avx2_psllv_q_256; |
| 3893 | else if (Size == 's' && Name[17] == '4') // avx512.mask.psllv4.si |
| 3894 | IID = Intrinsic::x86_avx2_psllv_d; |
| 3895 | else if (Size == 's' && Name[17] == '8') // avx512.mask.psllv8.si |
| 3896 | IID = Intrinsic::x86_avx2_psllv_d_256; |
| 3897 | else if (Size == 'h' && Name[17] == '8') // avx512.mask.psllv8.hi |
| 3898 | IID = Intrinsic::x86_avx512_psllv_w_128; |
| 3899 | else if (Size == 'h' && Name[17] == '1') // avx512.mask.psllv16.hi |
| 3900 | IID = Intrinsic::x86_avx512_psllv_w_256; |
| 3901 | else if (Name[17] == '3' && Name[18] == '2') // avx512.mask.psllv32hi |
| 3902 | IID = Intrinsic::x86_avx512_psllv_w_512; |
| 3903 | else |
| 3904 | llvm_unreachable("Unexpected size" ); |
| 3905 | } else if (Name.ends_with(Suffix: ".128" )) { |
| 3906 | if (Size == 'd') // avx512.mask.psll.d.128, avx512.mask.psll.di.128 |
| 3907 | IID = IsImmediate ? Intrinsic::x86_sse2_pslli_d |
| 3908 | : Intrinsic::x86_sse2_psll_d; |
| 3909 | else if (Size == 'q') // avx512.mask.psll.q.128, avx512.mask.psll.qi.128 |
| 3910 | IID = IsImmediate ? Intrinsic::x86_sse2_pslli_q |
| 3911 | : Intrinsic::x86_sse2_psll_q; |
| 3912 | else if (Size == 'w') // avx512.mask.psll.w.128, avx512.mask.psll.wi.128 |
| 3913 | IID = IsImmediate ? Intrinsic::x86_sse2_pslli_w |
| 3914 | : Intrinsic::x86_sse2_psll_w; |
| 3915 | else |
| 3916 | llvm_unreachable("Unexpected size" ); |
| 3917 | } else if (Name.ends_with(Suffix: ".256" )) { |
| 3918 | if (Size == 'd') // avx512.mask.psll.d.256, avx512.mask.psll.di.256 |
| 3919 | IID = IsImmediate ? Intrinsic::x86_avx2_pslli_d |
| 3920 | : Intrinsic::x86_avx2_psll_d; |
| 3921 | else if (Size == 'q') // avx512.mask.psll.q.256, avx512.mask.psll.qi.256 |
| 3922 | IID = IsImmediate ? Intrinsic::x86_avx2_pslli_q |
| 3923 | : Intrinsic::x86_avx2_psll_q; |
| 3924 | else if (Size == 'w') // avx512.mask.psll.w.256, avx512.mask.psll.wi.256 |
| 3925 | IID = IsImmediate ? Intrinsic::x86_avx2_pslli_w |
| 3926 | : Intrinsic::x86_avx2_psll_w; |
| 3927 | else |
| 3928 | llvm_unreachable("Unexpected size" ); |
| 3929 | } else { |
| 3930 | if (Size == 'd') // psll.di.512, pslli.d, psll.d, psllv.d.512 |
| 3931 | IID = IsImmediate ? Intrinsic::x86_avx512_pslli_d_512 |
| 3932 | : IsVariable ? Intrinsic::x86_avx512_psllv_d_512 |
| 3933 | : Intrinsic::x86_avx512_psll_d_512; |
| 3934 | else if (Size == 'q') // psll.qi.512, pslli.q, psll.q, psllv.q.512 |
| 3935 | IID = IsImmediate ? Intrinsic::x86_avx512_pslli_q_512 |
| 3936 | : IsVariable ? Intrinsic::x86_avx512_psllv_q_512 |
| 3937 | : Intrinsic::x86_avx512_psll_q_512; |
| 3938 | else if (Size == 'w') // psll.wi.512, pslli.w, psll.w |
| 3939 | IID = IsImmediate ? Intrinsic::x86_avx512_pslli_w_512 |
| 3940 | : Intrinsic::x86_avx512_psll_w_512; |
| 3941 | else |
| 3942 | llvm_unreachable("Unexpected size" ); |
| 3943 | } |
| 3944 | |
| 3945 | Rep = upgradeX86MaskedShift(Builder, CI&: *CI, IID); |
| 3946 | } else if (Name.starts_with(Prefix: "avx512.mask.psrl" )) { |
| 3947 | bool IsImmediate = Name[16] == 'i' || (Name.size() > 18 && Name[18] == 'i'); |
| 3948 | bool IsVariable = Name[16] == 'v'; |
| 3949 | char Size = Name[16] == '.' ? Name[17] |
| 3950 | : Name[17] == '.' ? Name[18] |
| 3951 | : Name[18] == '.' ? Name[19] |
| 3952 | : Name[20]; |
| 3953 | |
| 3954 | Intrinsic::ID IID; |
| 3955 | if (IsVariable && Name[17] != '.') { |
| 3956 | if (Size == 'd' && Name[17] == '2') // avx512.mask.psrlv2.di |
| 3957 | IID = Intrinsic::x86_avx2_psrlv_q; |
| 3958 | else if (Size == 'd' && Name[17] == '4') // avx512.mask.psrlv4.di |
| 3959 | IID = Intrinsic::x86_avx2_psrlv_q_256; |
| 3960 | else if (Size == 's' && Name[17] == '4') // avx512.mask.psrlv4.si |
| 3961 | IID = Intrinsic::x86_avx2_psrlv_d; |
| 3962 | else if (Size == 's' && Name[17] == '8') // avx512.mask.psrlv8.si |
| 3963 | IID = Intrinsic::x86_avx2_psrlv_d_256; |
| 3964 | else if (Size == 'h' && Name[17] == '8') // avx512.mask.psrlv8.hi |
| 3965 | IID = Intrinsic::x86_avx512_psrlv_w_128; |
| 3966 | else if (Size == 'h' && Name[17] == '1') // avx512.mask.psrlv16.hi |
| 3967 | IID = Intrinsic::x86_avx512_psrlv_w_256; |
| 3968 | else if (Name[17] == '3' && Name[18] == '2') // avx512.mask.psrlv32hi |
| 3969 | IID = Intrinsic::x86_avx512_psrlv_w_512; |
| 3970 | else |
| 3971 | llvm_unreachable("Unexpected size" ); |
| 3972 | } else if (Name.ends_with(Suffix: ".128" )) { |
| 3973 | if (Size == 'd') // avx512.mask.psrl.d.128, avx512.mask.psrl.di.128 |
| 3974 | IID = IsImmediate ? Intrinsic::x86_sse2_psrli_d |
| 3975 | : Intrinsic::x86_sse2_psrl_d; |
| 3976 | else if (Size == 'q') // avx512.mask.psrl.q.128, avx512.mask.psrl.qi.128 |
| 3977 | IID = IsImmediate ? Intrinsic::x86_sse2_psrli_q |
| 3978 | : Intrinsic::x86_sse2_psrl_q; |
| 3979 | else if (Size == 'w') // avx512.mask.psrl.w.128, avx512.mask.psrl.wi.128 |
| 3980 | IID = IsImmediate ? Intrinsic::x86_sse2_psrli_w |
| 3981 | : Intrinsic::x86_sse2_psrl_w; |
| 3982 | else |
| 3983 | llvm_unreachable("Unexpected size" ); |
| 3984 | } else if (Name.ends_with(Suffix: ".256" )) { |
| 3985 | if (Size == 'd') // avx512.mask.psrl.d.256, avx512.mask.psrl.di.256 |
| 3986 | IID = IsImmediate ? Intrinsic::x86_avx2_psrli_d |
| 3987 | : Intrinsic::x86_avx2_psrl_d; |
| 3988 | else if (Size == 'q') // avx512.mask.psrl.q.256, avx512.mask.psrl.qi.256 |
| 3989 | IID = IsImmediate ? Intrinsic::x86_avx2_psrli_q |
| 3990 | : Intrinsic::x86_avx2_psrl_q; |
| 3991 | else if (Size == 'w') // avx512.mask.psrl.w.256, avx512.mask.psrl.wi.256 |
| 3992 | IID = IsImmediate ? Intrinsic::x86_avx2_psrli_w |
| 3993 | : Intrinsic::x86_avx2_psrl_w; |
| 3994 | else |
| 3995 | llvm_unreachable("Unexpected size" ); |
| 3996 | } else { |
| 3997 | if (Size == 'd') // psrl.di.512, psrli.d, psrl.d, psrl.d.512 |
| 3998 | IID = IsImmediate ? Intrinsic::x86_avx512_psrli_d_512 |
| 3999 | : IsVariable ? Intrinsic::x86_avx512_psrlv_d_512 |
| 4000 | : Intrinsic::x86_avx512_psrl_d_512; |
| 4001 | else if (Size == 'q') // psrl.qi.512, psrli.q, psrl.q, psrl.q.512 |
| 4002 | IID = IsImmediate ? Intrinsic::x86_avx512_psrli_q_512 |
| 4003 | : IsVariable ? Intrinsic::x86_avx512_psrlv_q_512 |
| 4004 | : Intrinsic::x86_avx512_psrl_q_512; |
| 4005 | else if (Size == 'w') // psrl.wi.512, psrli.w, psrl.w) |
| 4006 | IID = IsImmediate ? Intrinsic::x86_avx512_psrli_w_512 |
| 4007 | : Intrinsic::x86_avx512_psrl_w_512; |
| 4008 | else |
| 4009 | llvm_unreachable("Unexpected size" ); |
| 4010 | } |
| 4011 | |
| 4012 | Rep = upgradeX86MaskedShift(Builder, CI&: *CI, IID); |
| 4013 | } else if (Name.starts_with(Prefix: "avx512.mask.psra" )) { |
| 4014 | bool IsImmediate = Name[16] == 'i' || (Name.size() > 18 && Name[18] == 'i'); |
| 4015 | bool IsVariable = Name[16] == 'v'; |
| 4016 | char Size = Name[16] == '.' ? Name[17] |
| 4017 | : Name[17] == '.' ? Name[18] |
| 4018 | : Name[18] == '.' ? Name[19] |
| 4019 | : Name[20]; |
| 4020 | |
| 4021 | Intrinsic::ID IID; |
| 4022 | if (IsVariable && Name[17] != '.') { |
| 4023 | if (Size == 's' && Name[17] == '4') // avx512.mask.psrav4.si |
| 4024 | IID = Intrinsic::x86_avx2_psrav_d; |
| 4025 | else if (Size == 's' && Name[17] == '8') // avx512.mask.psrav8.si |
| 4026 | IID = Intrinsic::x86_avx2_psrav_d_256; |
| 4027 | else if (Size == 'h' && Name[17] == '8') // avx512.mask.psrav8.hi |
| 4028 | IID = Intrinsic::x86_avx512_psrav_w_128; |
| 4029 | else if (Size == 'h' && Name[17] == '1') // avx512.mask.psrav16.hi |
| 4030 | IID = Intrinsic::x86_avx512_psrav_w_256; |
| 4031 | else if (Name[17] == '3' && Name[18] == '2') // avx512.mask.psrav32hi |
| 4032 | IID = Intrinsic::x86_avx512_psrav_w_512; |
| 4033 | else |
| 4034 | llvm_unreachable("Unexpected size" ); |
| 4035 | } else if (Name.ends_with(Suffix: ".128" )) { |
| 4036 | if (Size == 'd') // avx512.mask.psra.d.128, avx512.mask.psra.di.128 |
| 4037 | IID = IsImmediate ? Intrinsic::x86_sse2_psrai_d |
| 4038 | : Intrinsic::x86_sse2_psra_d; |
| 4039 | else if (Size == 'q') // avx512.mask.psra.q.128, avx512.mask.psra.qi.128 |
| 4040 | IID = IsImmediate ? Intrinsic::x86_avx512_psrai_q_128 |
| 4041 | : IsVariable ? Intrinsic::x86_avx512_psrav_q_128 |
| 4042 | : Intrinsic::x86_avx512_psra_q_128; |
| 4043 | else if (Size == 'w') // avx512.mask.psra.w.128, avx512.mask.psra.wi.128 |
| 4044 | IID = IsImmediate ? Intrinsic::x86_sse2_psrai_w |
| 4045 | : Intrinsic::x86_sse2_psra_w; |
| 4046 | else |
| 4047 | llvm_unreachable("Unexpected size" ); |
| 4048 | } else if (Name.ends_with(Suffix: ".256" )) { |
| 4049 | if (Size == 'd') // avx512.mask.psra.d.256, avx512.mask.psra.di.256 |
| 4050 | IID = IsImmediate ? Intrinsic::x86_avx2_psrai_d |
| 4051 | : Intrinsic::x86_avx2_psra_d; |
| 4052 | else if (Size == 'q') // avx512.mask.psra.q.256, avx512.mask.psra.qi.256 |
| 4053 | IID = IsImmediate ? Intrinsic::x86_avx512_psrai_q_256 |
| 4054 | : IsVariable ? Intrinsic::x86_avx512_psrav_q_256 |
| 4055 | : Intrinsic::x86_avx512_psra_q_256; |
| 4056 | else if (Size == 'w') // avx512.mask.psra.w.256, avx512.mask.psra.wi.256 |
| 4057 | IID = IsImmediate ? Intrinsic::x86_avx2_psrai_w |
| 4058 | : Intrinsic::x86_avx2_psra_w; |
| 4059 | else |
| 4060 | llvm_unreachable("Unexpected size" ); |
| 4061 | } else { |
| 4062 | if (Size == 'd') // psra.di.512, psrai.d, psra.d, psrav.d.512 |
| 4063 | IID = IsImmediate ? Intrinsic::x86_avx512_psrai_d_512 |
| 4064 | : IsVariable ? Intrinsic::x86_avx512_psrav_d_512 |
| 4065 | : Intrinsic::x86_avx512_psra_d_512; |
| 4066 | else if (Size == 'q') // psra.qi.512, psrai.q, psra.q |
| 4067 | IID = IsImmediate ? Intrinsic::x86_avx512_psrai_q_512 |
| 4068 | : IsVariable ? Intrinsic::x86_avx512_psrav_q_512 |
| 4069 | : Intrinsic::x86_avx512_psra_q_512; |
| 4070 | else if (Size == 'w') // psra.wi.512, psrai.w, psra.w |
| 4071 | IID = IsImmediate ? Intrinsic::x86_avx512_psrai_w_512 |
| 4072 | : Intrinsic::x86_avx512_psra_w_512; |
| 4073 | else |
| 4074 | llvm_unreachable("Unexpected size" ); |
| 4075 | } |
| 4076 | |
| 4077 | Rep = upgradeX86MaskedShift(Builder, CI&: *CI, IID); |
| 4078 | } else if (Name.starts_with(Prefix: "avx512.mask.move.s" )) { |
| 4079 | Rep = upgradeMaskedMove(Builder, CI&: *CI); |
| 4080 | } else if (Name.starts_with(Prefix: "avx512.cvtmask2" )) { |
| 4081 | Rep = upgradeMaskToInt(Builder, CI&: *CI); |
| 4082 | } else if (Name.ends_with(Suffix: ".movntdqa" )) { |
| 4083 | MDNode *Node = MDNode::get( |
| 4084 | Context&: C, MDs: ConstantAsMetadata::get(C: ConstantInt::get(Ty: Type::getInt32Ty(C), V: 1))); |
| 4085 | |
| 4086 | LoadInst *LI = Builder.CreateAlignedLoad( |
| 4087 | Ty: CI->getType(), Ptr: CI->getArgOperand(i: 0), |
| 4088 | Align: Align(CI->getType()->getPrimitiveSizeInBits().getFixedValue() / 8)); |
| 4089 | LI->setMetadata(KindID: LLVMContext::MD_nontemporal, Node); |
| 4090 | Rep = LI; |
| 4091 | } else if (Name.starts_with(Prefix: "fma.vfmadd." ) || |
| 4092 | Name.starts_with(Prefix: "fma.vfmsub." ) || |
| 4093 | Name.starts_with(Prefix: "fma.vfnmadd." ) || |
| 4094 | Name.starts_with(Prefix: "fma.vfnmsub." )) { |
| 4095 | bool NegMul = Name[6] == 'n'; |
| 4096 | bool NegAcc = NegMul ? Name[8] == 's' : Name[7] == 's'; |
| 4097 | bool IsScalar = NegMul ? Name[12] == 's' : Name[11] == 's'; |
| 4098 | |
| 4099 | Value *Ops[] = {CI->getArgOperand(i: 0), CI->getArgOperand(i: 1), |
| 4100 | CI->getArgOperand(i: 2)}; |
| 4101 | |
| 4102 | if (IsScalar) { |
| 4103 | Ops[0] = Builder.CreateExtractElement(Vec: Ops[0], Idx: (uint64_t)0); |
| 4104 | Ops[1] = Builder.CreateExtractElement(Vec: Ops[1], Idx: (uint64_t)0); |
| 4105 | Ops[2] = Builder.CreateExtractElement(Vec: Ops[2], Idx: (uint64_t)0); |
| 4106 | } |
| 4107 | |
| 4108 | if (NegMul && !IsScalar) |
| 4109 | Ops[0] = Builder.CreateFNeg(V: Ops[0]); |
| 4110 | if (NegMul && IsScalar) |
| 4111 | Ops[1] = Builder.CreateFNeg(V: Ops[1]); |
| 4112 | if (NegAcc) |
| 4113 | Ops[2] = Builder.CreateFNeg(V: Ops[2]); |
| 4114 | |
| 4115 | Rep = Builder.CreateIntrinsic(ID: Intrinsic::fma, Types: Ops[0]->getType(), Args: Ops); |
| 4116 | |
| 4117 | if (IsScalar) |
| 4118 | Rep = Builder.CreateInsertElement(Vec: CI->getArgOperand(i: 0), NewElt: Rep, Idx: (uint64_t)0); |
| 4119 | } else if (Name.starts_with(Prefix: "fma4.vfmadd.s" )) { |
| 4120 | Value *Ops[] = {CI->getArgOperand(i: 0), CI->getArgOperand(i: 1), |
| 4121 | CI->getArgOperand(i: 2)}; |
| 4122 | |
| 4123 | Ops[0] = Builder.CreateExtractElement(Vec: Ops[0], Idx: (uint64_t)0); |
| 4124 | Ops[1] = Builder.CreateExtractElement(Vec: Ops[1], Idx: (uint64_t)0); |
| 4125 | Ops[2] = Builder.CreateExtractElement(Vec: Ops[2], Idx: (uint64_t)0); |
| 4126 | |
| 4127 | Rep = Builder.CreateIntrinsic(ID: Intrinsic::fma, Types: Ops[0]->getType(), Args: Ops); |
| 4128 | |
| 4129 | Rep = Builder.CreateInsertElement(Vec: Constant::getNullValue(Ty: CI->getType()), |
| 4130 | NewElt: Rep, Idx: (uint64_t)0); |
| 4131 | } else if (Name.starts_with(Prefix: "avx512.mask.vfmadd.s" ) || |
| 4132 | Name.starts_with(Prefix: "avx512.maskz.vfmadd.s" ) || |
| 4133 | Name.starts_with(Prefix: "avx512.mask3.vfmadd.s" ) || |
| 4134 | Name.starts_with(Prefix: "avx512.mask3.vfmsub.s" ) || |
| 4135 | Name.starts_with(Prefix: "avx512.mask3.vfnmsub.s" )) { |
| 4136 | bool IsMask3 = Name[11] == '3'; |
| 4137 | bool IsMaskZ = Name[11] == 'z'; |
| 4138 | // Drop the "avx512.mask." to make it easier. |
| 4139 | Name = Name.drop_front(N: IsMask3 || IsMaskZ ? 13 : 12); |
| 4140 | bool NegMul = Name[2] == 'n'; |
| 4141 | bool NegAcc = NegMul ? Name[4] == 's' : Name[3] == 's'; |
| 4142 | |
| 4143 | Value *A = CI->getArgOperand(i: 0); |
| 4144 | Value *B = CI->getArgOperand(i: 1); |
| 4145 | Value *C = CI->getArgOperand(i: 2); |
| 4146 | |
| 4147 | if (NegMul && (IsMask3 || IsMaskZ)) |
| 4148 | A = Builder.CreateFNeg(V: A); |
| 4149 | if (NegMul && !(IsMask3 || IsMaskZ)) |
| 4150 | B = Builder.CreateFNeg(V: B); |
| 4151 | if (NegAcc) |
| 4152 | C = Builder.CreateFNeg(V: C); |
| 4153 | |
| 4154 | A = Builder.CreateExtractElement(Vec: A, Idx: (uint64_t)0); |
| 4155 | B = Builder.CreateExtractElement(Vec: B, Idx: (uint64_t)0); |
| 4156 | C = Builder.CreateExtractElement(Vec: C, Idx: (uint64_t)0); |
| 4157 | |
| 4158 | if (!isa<ConstantInt>(Val: CI->getArgOperand(i: 4)) || |
| 4159 | cast<ConstantInt>(Val: CI->getArgOperand(i: 4))->getZExtValue() != 4) { |
| 4160 | Value *Ops[] = {A, B, C, CI->getArgOperand(i: 4)}; |
| 4161 | |
| 4162 | Intrinsic::ID IID; |
| 4163 | if (Name.back() == 'd') |
| 4164 | IID = Intrinsic::x86_avx512_vfmadd_f64; |
| 4165 | else |
| 4166 | IID = Intrinsic::x86_avx512_vfmadd_f32; |
| 4167 | Rep = Builder.CreateIntrinsic(ID: IID, Args: Ops); |
| 4168 | } else { |
| 4169 | Rep = Builder.CreateFMA(Factor1: A, Factor2: B, Summand: C); |
| 4170 | } |
| 4171 | |
| 4172 | Value *PassThru = IsMaskZ ? Constant::getNullValue(Ty: Rep->getType()) |
| 4173 | : IsMask3 ? C |
| 4174 | : A; |
| 4175 | |
| 4176 | // For Mask3 with NegAcc, we need to create a new extractelement that |
| 4177 | // avoids the negation above. |
| 4178 | if (NegAcc && IsMask3) |
| 4179 | PassThru = |
| 4180 | Builder.CreateExtractElement(Vec: CI->getArgOperand(i: 2), Idx: (uint64_t)0); |
| 4181 | |
| 4182 | Rep = emitX86ScalarSelect(Builder, Mask: CI->getArgOperand(i: 3), Op0: Rep, Op1: PassThru); |
| 4183 | Rep = Builder.CreateInsertElement(Vec: CI->getArgOperand(i: IsMask3 ? 2 : 0), NewElt: Rep, |
| 4184 | Idx: (uint64_t)0); |
| 4185 | } else if (Name.starts_with(Prefix: "avx512.mask.vfmadd.p" ) || |
| 4186 | Name.starts_with(Prefix: "avx512.mask.vfnmadd.p" ) || |
| 4187 | Name.starts_with(Prefix: "avx512.mask.vfnmsub.p" ) || |
| 4188 | Name.starts_with(Prefix: "avx512.mask3.vfmadd.p" ) || |
| 4189 | Name.starts_with(Prefix: "avx512.mask3.vfmsub.p" ) || |
| 4190 | Name.starts_with(Prefix: "avx512.mask3.vfnmsub.p" ) || |
| 4191 | Name.starts_with(Prefix: "avx512.maskz.vfmadd.p" )) { |
| 4192 | bool IsMask3 = Name[11] == '3'; |
| 4193 | bool IsMaskZ = Name[11] == 'z'; |
| 4194 | // Drop the "avx512.mask." to make it easier. |
| 4195 | Name = Name.drop_front(N: IsMask3 || IsMaskZ ? 13 : 12); |
| 4196 | bool NegMul = Name[2] == 'n'; |
| 4197 | bool NegAcc = NegMul ? Name[4] == 's' : Name[3] == 's'; |
| 4198 | |
| 4199 | Value *A = CI->getArgOperand(i: 0); |
| 4200 | Value *B = CI->getArgOperand(i: 1); |
| 4201 | Value *C = CI->getArgOperand(i: 2); |
| 4202 | |
| 4203 | if (NegMul && (IsMask3 || IsMaskZ)) |
| 4204 | A = Builder.CreateFNeg(V: A); |
| 4205 | if (NegMul && !(IsMask3 || IsMaskZ)) |
| 4206 | B = Builder.CreateFNeg(V: B); |
| 4207 | if (NegAcc) |
| 4208 | C = Builder.CreateFNeg(V: C); |
| 4209 | |
| 4210 | if (CI->arg_size() == 5 && |
| 4211 | (!isa<ConstantInt>(Val: CI->getArgOperand(i: 4)) || |
| 4212 | cast<ConstantInt>(Val: CI->getArgOperand(i: 4))->getZExtValue() != 4)) { |
| 4213 | Intrinsic::ID IID; |
| 4214 | // Check the character before ".512" in string. |
| 4215 | if (Name[Name.size() - 5] == 's') |
| 4216 | IID = Intrinsic::x86_avx512_vfmadd_ps_512; |
| 4217 | else |
| 4218 | IID = Intrinsic::x86_avx512_vfmadd_pd_512; |
| 4219 | |
| 4220 | Rep = Builder.CreateIntrinsic(ID: IID, Args: {A, B, C, CI->getArgOperand(i: 4)}); |
| 4221 | } else { |
| 4222 | Rep = Builder.CreateFMA(Factor1: A, Factor2: B, Summand: C); |
| 4223 | } |
| 4224 | |
| 4225 | Value *PassThru = IsMaskZ ? llvm::Constant::getNullValue(Ty: CI->getType()) |
| 4226 | : IsMask3 ? CI->getArgOperand(i: 2) |
| 4227 | : CI->getArgOperand(i: 0); |
| 4228 | |
| 4229 | Rep = emitX86Select(Builder, Mask: CI->getArgOperand(i: 3), Op0: Rep, Op1: PassThru); |
| 4230 | } else if (Name.starts_with(Prefix: "fma.vfmsubadd.p" )) { |
| 4231 | unsigned VecWidth = CI->getType()->getPrimitiveSizeInBits(); |
| 4232 | unsigned EltWidth = CI->getType()->getScalarSizeInBits(); |
| 4233 | Intrinsic::ID IID; |
| 4234 | if (VecWidth == 128 && EltWidth == 32) |
| 4235 | IID = Intrinsic::x86_fma_vfmaddsub_ps; |
| 4236 | else if (VecWidth == 256 && EltWidth == 32) |
| 4237 | IID = Intrinsic::x86_fma_vfmaddsub_ps_256; |
| 4238 | else if (VecWidth == 128 && EltWidth == 64) |
| 4239 | IID = Intrinsic::x86_fma_vfmaddsub_pd; |
| 4240 | else if (VecWidth == 256 && EltWidth == 64) |
| 4241 | IID = Intrinsic::x86_fma_vfmaddsub_pd_256; |
| 4242 | else |
| 4243 | llvm_unreachable("Unexpected intrinsic" ); |
| 4244 | |
| 4245 | Value *Ops[] = {CI->getArgOperand(i: 0), CI->getArgOperand(i: 1), |
| 4246 | CI->getArgOperand(i: 2)}; |
| 4247 | Ops[2] = Builder.CreateFNeg(V: Ops[2]); |
| 4248 | Rep = Builder.CreateIntrinsic(ID: IID, Args: Ops); |
| 4249 | } else if (Name.starts_with(Prefix: "avx512.mask.vfmaddsub.p" ) || |
| 4250 | Name.starts_with(Prefix: "avx512.mask3.vfmaddsub.p" ) || |
| 4251 | Name.starts_with(Prefix: "avx512.maskz.vfmaddsub.p" ) || |
| 4252 | Name.starts_with(Prefix: "avx512.mask3.vfmsubadd.p" )) { |
| 4253 | bool IsMask3 = Name[11] == '3'; |
| 4254 | bool IsMaskZ = Name[11] == 'z'; |
| 4255 | // Drop the "avx512.mask." to make it easier. |
| 4256 | Name = Name.drop_front(N: IsMask3 || IsMaskZ ? 13 : 12); |
| 4257 | bool IsSubAdd = Name[3] == 's'; |
| 4258 | if (CI->arg_size() == 5) { |
| 4259 | Intrinsic::ID IID; |
| 4260 | // Check the character before ".512" in string. |
| 4261 | if (Name[Name.size() - 5] == 's') |
| 4262 | IID = Intrinsic::x86_avx512_vfmaddsub_ps_512; |
| 4263 | else |
| 4264 | IID = Intrinsic::x86_avx512_vfmaddsub_pd_512; |
| 4265 | |
| 4266 | Value *Ops[] = {CI->getArgOperand(i: 0), CI->getArgOperand(i: 1), |
| 4267 | CI->getArgOperand(i: 2), CI->getArgOperand(i: 4)}; |
| 4268 | if (IsSubAdd) |
| 4269 | Ops[2] = Builder.CreateFNeg(V: Ops[2]); |
| 4270 | |
| 4271 | Rep = Builder.CreateIntrinsic(ID: IID, Args: Ops); |
| 4272 | } else { |
| 4273 | int NumElts = cast<FixedVectorType>(Val: CI->getType())->getNumElements(); |
| 4274 | |
| 4275 | Value *Ops[] = {CI->getArgOperand(i: 0), CI->getArgOperand(i: 1), |
| 4276 | CI->getArgOperand(i: 2)}; |
| 4277 | |
| 4278 | Function *FMA = Intrinsic::getOrInsertDeclaration( |
| 4279 | M: CI->getModule(), id: Intrinsic::fma, Tys: Ops[0]->getType()); |
| 4280 | Value *Odd = Builder.CreateCall(Callee: FMA, Args: Ops); |
| 4281 | Ops[2] = Builder.CreateFNeg(V: Ops[2]); |
| 4282 | Value *Even = Builder.CreateCall(Callee: FMA, Args: Ops); |
| 4283 | |
| 4284 | if (IsSubAdd) |
| 4285 | std::swap(a&: Even, b&: Odd); |
| 4286 | |
| 4287 | SmallVector<int, 32> Idxs(NumElts); |
| 4288 | for (int i = 0; i != NumElts; ++i) |
| 4289 | Idxs[i] = i + (i % 2) * NumElts; |
| 4290 | |
| 4291 | Rep = Builder.CreateShuffleVector(V1: Even, V2: Odd, Mask: Idxs); |
| 4292 | } |
| 4293 | |
| 4294 | Value *PassThru = IsMaskZ ? llvm::Constant::getNullValue(Ty: CI->getType()) |
| 4295 | : IsMask3 ? CI->getArgOperand(i: 2) |
| 4296 | : CI->getArgOperand(i: 0); |
| 4297 | |
| 4298 | Rep = emitX86Select(Builder, Mask: CI->getArgOperand(i: 3), Op0: Rep, Op1: PassThru); |
| 4299 | } else if (Name.starts_with(Prefix: "avx512.mask.pternlog." ) || |
| 4300 | Name.starts_with(Prefix: "avx512.maskz.pternlog." )) { |
| 4301 | bool ZeroMask = Name[11] == 'z'; |
| 4302 | unsigned VecWidth = CI->getType()->getPrimitiveSizeInBits(); |
| 4303 | unsigned EltWidth = CI->getType()->getScalarSizeInBits(); |
| 4304 | Intrinsic::ID IID; |
| 4305 | if (VecWidth == 128 && EltWidth == 32) |
| 4306 | IID = Intrinsic::x86_avx512_pternlog_d_128; |
| 4307 | else if (VecWidth == 256 && EltWidth == 32) |
| 4308 | IID = Intrinsic::x86_avx512_pternlog_d_256; |
| 4309 | else if (VecWidth == 512 && EltWidth == 32) |
| 4310 | IID = Intrinsic::x86_avx512_pternlog_d_512; |
| 4311 | else if (VecWidth == 128 && EltWidth == 64) |
| 4312 | IID = Intrinsic::x86_avx512_pternlog_q_128; |
| 4313 | else if (VecWidth == 256 && EltWidth == 64) |
| 4314 | IID = Intrinsic::x86_avx512_pternlog_q_256; |
| 4315 | else if (VecWidth == 512 && EltWidth == 64) |
| 4316 | IID = Intrinsic::x86_avx512_pternlog_q_512; |
| 4317 | else |
| 4318 | llvm_unreachable("Unexpected intrinsic" ); |
| 4319 | |
| 4320 | Value *Args[] = {CI->getArgOperand(i: 0), CI->getArgOperand(i: 1), |
| 4321 | CI->getArgOperand(i: 2), CI->getArgOperand(i: 3)}; |
| 4322 | Rep = Builder.CreateIntrinsic(ID: IID, Args); |
| 4323 | Value *PassThru = ZeroMask ? ConstantAggregateZero::get(Ty: CI->getType()) |
| 4324 | : CI->getArgOperand(i: 0); |
| 4325 | Rep = emitX86Select(Builder, Mask: CI->getArgOperand(i: 4), Op0: Rep, Op1: PassThru); |
| 4326 | } else if (Name.starts_with(Prefix: "avx512.mask.vpmadd52" ) || |
| 4327 | Name.starts_with(Prefix: "avx512.maskz.vpmadd52" )) { |
| 4328 | bool ZeroMask = Name[11] == 'z'; |
| 4329 | bool High = Name[20] == 'h' || Name[21] == 'h'; |
| 4330 | unsigned VecWidth = CI->getType()->getPrimitiveSizeInBits(); |
| 4331 | Intrinsic::ID IID; |
| 4332 | if (VecWidth == 128 && !High) |
| 4333 | IID = Intrinsic::x86_avx512_vpmadd52l_uq_128; |
| 4334 | else if (VecWidth == 256 && !High) |
| 4335 | IID = Intrinsic::x86_avx512_vpmadd52l_uq_256; |
| 4336 | else if (VecWidth == 512 && !High) |
| 4337 | IID = Intrinsic::x86_avx512_vpmadd52l_uq_512; |
| 4338 | else if (VecWidth == 128 && High) |
| 4339 | IID = Intrinsic::x86_avx512_vpmadd52h_uq_128; |
| 4340 | else if (VecWidth == 256 && High) |
| 4341 | IID = Intrinsic::x86_avx512_vpmadd52h_uq_256; |
| 4342 | else if (VecWidth == 512 && High) |
| 4343 | IID = Intrinsic::x86_avx512_vpmadd52h_uq_512; |
| 4344 | else |
| 4345 | llvm_unreachable("Unexpected intrinsic" ); |
| 4346 | |
| 4347 | Value *Args[] = {CI->getArgOperand(i: 0), CI->getArgOperand(i: 1), |
| 4348 | CI->getArgOperand(i: 2)}; |
| 4349 | Rep = Builder.CreateIntrinsic(ID: IID, Args); |
| 4350 | Value *PassThru = ZeroMask ? ConstantAggregateZero::get(Ty: CI->getType()) |
| 4351 | : CI->getArgOperand(i: 0); |
| 4352 | Rep = emitX86Select(Builder, Mask: CI->getArgOperand(i: 3), Op0: Rep, Op1: PassThru); |
| 4353 | } else if (Name.starts_with(Prefix: "avx512.mask.vpermi2var." ) || |
| 4354 | Name.starts_with(Prefix: "avx512.mask.vpermt2var." ) || |
| 4355 | Name.starts_with(Prefix: "avx512.maskz.vpermt2var." )) { |
| 4356 | bool ZeroMask = Name[11] == 'z'; |
| 4357 | bool IndexForm = Name[17] == 'i'; |
| 4358 | Rep = upgradeX86VPERMT2Intrinsics(Builder, CI&: *CI, ZeroMask, IndexForm); |
| 4359 | } else if (Name.starts_with(Prefix: "avx512.mask.vpdpbusd." ) || |
| 4360 | Name.starts_with(Prefix: "avx512.maskz.vpdpbusd." ) || |
| 4361 | Name.starts_with(Prefix: "avx512.mask.vpdpbusds." ) || |
| 4362 | Name.starts_with(Prefix: "avx512.maskz.vpdpbusds." )) { |
| 4363 | bool ZeroMask = Name[11] == 'z'; |
| 4364 | bool IsSaturating = Name[ZeroMask ? 21 : 20] == 's'; |
| 4365 | unsigned VecWidth = CI->getType()->getPrimitiveSizeInBits(); |
| 4366 | Intrinsic::ID IID; |
| 4367 | if (VecWidth == 128 && !IsSaturating) |
| 4368 | IID = Intrinsic::x86_avx512_vpdpbusd_128; |
| 4369 | else if (VecWidth == 256 && !IsSaturating) |
| 4370 | IID = Intrinsic::x86_avx512_vpdpbusd_256; |
| 4371 | else if (VecWidth == 512 && !IsSaturating) |
| 4372 | IID = Intrinsic::x86_avx512_vpdpbusd_512; |
| 4373 | else if (VecWidth == 128 && IsSaturating) |
| 4374 | IID = Intrinsic::x86_avx512_vpdpbusds_128; |
| 4375 | else if (VecWidth == 256 && IsSaturating) |
| 4376 | IID = Intrinsic::x86_avx512_vpdpbusds_256; |
| 4377 | else if (VecWidth == 512 && IsSaturating) |
| 4378 | IID = Intrinsic::x86_avx512_vpdpbusds_512; |
| 4379 | else |
| 4380 | llvm_unreachable("Unexpected intrinsic" ); |
| 4381 | |
| 4382 | Value *Args[] = {CI->getArgOperand(i: 0), CI->getArgOperand(i: 1), |
| 4383 | CI->getArgOperand(i: 2)}; |
| 4384 | |
| 4385 | // Input arguments types were incorrectly set to vectors of i32 before but |
| 4386 | // they should be vectors of i8. Insert bit cast when encountering the old |
| 4387 | // types |
| 4388 | if (Args[1]->getType()->isVectorTy() && |
| 4389 | cast<VectorType>(Val: Args[1]->getType()) |
| 4390 | ->getElementType() |
| 4391 | ->isIntegerTy(Bitwidth: 32) && |
| 4392 | Args[2]->getType()->isVectorTy() && |
| 4393 | cast<VectorType>(Val: Args[2]->getType()) |
| 4394 | ->getElementType() |
| 4395 | ->isIntegerTy(Bitwidth: 32)) { |
| 4396 | Type *NewArgType = nullptr; |
| 4397 | if (VecWidth == 128) |
| 4398 | NewArgType = VectorType::get(ElementType: Builder.getInt8Ty(), NumElements: 16, Scalable: false); |
| 4399 | else if (VecWidth == 256) |
| 4400 | NewArgType = VectorType::get(ElementType: Builder.getInt8Ty(), NumElements: 32, Scalable: false); |
| 4401 | else if (VecWidth == 512) |
| 4402 | NewArgType = VectorType::get(ElementType: Builder.getInt8Ty(), NumElements: 64, Scalable: false); |
| 4403 | else |
| 4404 | llvm_unreachable("Unexpected vector bit width" ); |
| 4405 | |
| 4406 | Args[1] = Builder.CreateBitCast(V: Args[1], DestTy: NewArgType); |
| 4407 | Args[2] = Builder.CreateBitCast(V: Args[2], DestTy: NewArgType); |
| 4408 | } |
| 4409 | |
| 4410 | Rep = Builder.CreateIntrinsic(ID: IID, Args); |
| 4411 | Value *PassThru = ZeroMask ? ConstantAggregateZero::get(Ty: CI->getType()) |
| 4412 | : CI->getArgOperand(i: 0); |
| 4413 | Rep = emitX86Select(Builder, Mask: CI->getArgOperand(i: 3), Op0: Rep, Op1: PassThru); |
| 4414 | } else if (Name.starts_with(Prefix: "avx512.mask.vpdpwssd." ) || |
| 4415 | Name.starts_with(Prefix: "avx512.maskz.vpdpwssd." ) || |
| 4416 | Name.starts_with(Prefix: "avx512.mask.vpdpwssds." ) || |
| 4417 | Name.starts_with(Prefix: "avx512.maskz.vpdpwssds." )) { |
| 4418 | bool ZeroMask = Name[11] == 'z'; |
| 4419 | bool IsSaturating = Name[ZeroMask ? 21 : 20] == 's'; |
| 4420 | unsigned VecWidth = CI->getType()->getPrimitiveSizeInBits(); |
| 4421 | Intrinsic::ID IID; |
| 4422 | if (VecWidth == 128 && !IsSaturating) |
| 4423 | IID = Intrinsic::x86_avx512_vpdpwssd_128; |
| 4424 | else if (VecWidth == 256 && !IsSaturating) |
| 4425 | IID = Intrinsic::x86_avx512_vpdpwssd_256; |
| 4426 | else if (VecWidth == 512 && !IsSaturating) |
| 4427 | IID = Intrinsic::x86_avx512_vpdpwssd_512; |
| 4428 | else if (VecWidth == 128 && IsSaturating) |
| 4429 | IID = Intrinsic::x86_avx512_vpdpwssds_128; |
| 4430 | else if (VecWidth == 256 && IsSaturating) |
| 4431 | IID = Intrinsic::x86_avx512_vpdpwssds_256; |
| 4432 | else if (VecWidth == 512 && IsSaturating) |
| 4433 | IID = Intrinsic::x86_avx512_vpdpwssds_512; |
| 4434 | else |
| 4435 | llvm_unreachable("Unexpected intrinsic" ); |
| 4436 | |
| 4437 | Value *Args[] = {CI->getArgOperand(i: 0), CI->getArgOperand(i: 1), |
| 4438 | CI->getArgOperand(i: 2)}; |
| 4439 | |
| 4440 | // Input arguments types were incorrectly set to vectors of i32 before but |
| 4441 | // they should be vectors of i16. Insert bit cast when encountering the old |
| 4442 | // types |
| 4443 | if (Args[1]->getType()->isVectorTy() && |
| 4444 | cast<VectorType>(Val: Args[1]->getType()) |
| 4445 | ->getElementType() |
| 4446 | ->isIntegerTy(Bitwidth: 32) && |
| 4447 | Args[2]->getType()->isVectorTy() && |
| 4448 | cast<VectorType>(Val: Args[2]->getType()) |
| 4449 | ->getElementType() |
| 4450 | ->isIntegerTy(Bitwidth: 32)) { |
| 4451 | Type *NewArgType = nullptr; |
| 4452 | if (VecWidth == 128) |
| 4453 | NewArgType = VectorType::get(ElementType: Builder.getInt16Ty(), NumElements: 8, Scalable: false); |
| 4454 | else if (VecWidth == 256) |
| 4455 | NewArgType = VectorType::get(ElementType: Builder.getInt16Ty(), NumElements: 16, Scalable: false); |
| 4456 | else if (VecWidth == 512) |
| 4457 | NewArgType = VectorType::get(ElementType: Builder.getInt16Ty(), NumElements: 32, Scalable: false); |
| 4458 | else |
| 4459 | llvm_unreachable("Unexpected vector bit width" ); |
| 4460 | |
| 4461 | Args[1] = Builder.CreateBitCast(V: Args[1], DestTy: NewArgType); |
| 4462 | Args[2] = Builder.CreateBitCast(V: Args[2], DestTy: NewArgType); |
| 4463 | } |
| 4464 | |
| 4465 | Rep = Builder.CreateIntrinsic(ID: IID, Args); |
| 4466 | Value *PassThru = ZeroMask ? ConstantAggregateZero::get(Ty: CI->getType()) |
| 4467 | : CI->getArgOperand(i: 0); |
| 4468 | Rep = emitX86Select(Builder, Mask: CI->getArgOperand(i: 3), Op0: Rep, Op1: PassThru); |
| 4469 | } else if (Name == "addcarryx.u32" || Name == "addcarryx.u64" || |
| 4470 | Name == "addcarry.u32" || Name == "addcarry.u64" || |
| 4471 | Name == "subborrow.u32" || Name == "subborrow.u64" ) { |
| 4472 | Intrinsic::ID IID; |
| 4473 | if (Name[0] == 'a' && Name.back() == '2') |
| 4474 | IID = Intrinsic::x86_addcarry_32; |
| 4475 | else if (Name[0] == 'a' && Name.back() == '4') |
| 4476 | IID = Intrinsic::x86_addcarry_64; |
| 4477 | else if (Name[0] == 's' && Name.back() == '2') |
| 4478 | IID = Intrinsic::x86_subborrow_32; |
| 4479 | else if (Name[0] == 's' && Name.back() == '4') |
| 4480 | IID = Intrinsic::x86_subborrow_64; |
| 4481 | else |
| 4482 | llvm_unreachable("Unexpected intrinsic" ); |
| 4483 | |
| 4484 | // Make a call with 3 operands. |
| 4485 | Value *Args[] = {CI->getArgOperand(i: 0), CI->getArgOperand(i: 1), |
| 4486 | CI->getArgOperand(i: 2)}; |
| 4487 | Value *NewCall = Builder.CreateIntrinsic(ID: IID, Args); |
| 4488 | |
| 4489 | // Extract the second result and store it. |
| 4490 | Value *Data = Builder.CreateExtractValue(Agg: NewCall, Idxs: 1); |
| 4491 | Builder.CreateAlignedStore(Val: Data, Ptr: CI->getArgOperand(i: 3), Align: Align(1)); |
| 4492 | // Replace the original call result with the first result of the new call. |
| 4493 | Value *CF = Builder.CreateExtractValue(Agg: NewCall, Idxs: 0); |
| 4494 | |
| 4495 | CI->replaceAllUsesWith(V: CF); |
| 4496 | Rep = nullptr; |
| 4497 | } else if (Name.starts_with(Prefix: "avx512.mask." ) && |
| 4498 | upgradeAVX512MaskToSelect(Name, Builder, CI&: *CI, Rep)) { |
| 4499 | // Rep will be updated by the call in the condition. |
| 4500 | } |
| 4501 | |
| 4502 | return Rep; |
| 4503 | } |
| 4504 | |
| 4505 | static Value *upgradeAArch64IntrinsicCall(StringRef Name, CallBase *CI, |
| 4506 | Function *F, IRBuilder<> &Builder) { |
| 4507 | if (Name.starts_with(Prefix: "neon.bfcvt" )) { |
| 4508 | if (Name.starts_with(Prefix: "neon.bfcvtn2" )) { |
| 4509 | SmallVector<int, 32> LoMask(4); |
| 4510 | std::iota(first: LoMask.begin(), last: LoMask.end(), value: 0); |
| 4511 | SmallVector<int, 32> ConcatMask(8); |
| 4512 | std::iota(first: ConcatMask.begin(), last: ConcatMask.end(), value: 0); |
| 4513 | Value *Inactive = Builder.CreateShuffleVector(V: CI->getOperand(i_nocapture: 0), Mask: LoMask); |
| 4514 | Value *Trunc = |
| 4515 | Builder.CreateFPTrunc(V: CI->getOperand(i_nocapture: 1), DestTy: Inactive->getType()); |
| 4516 | return Builder.CreateShuffleVector(V1: Inactive, V2: Trunc, Mask: ConcatMask); |
| 4517 | } else if (Name.starts_with(Prefix: "neon.bfcvtn" )) { |
| 4518 | SmallVector<int, 32> ConcatMask(8); |
| 4519 | std::iota(first: ConcatMask.begin(), last: ConcatMask.end(), value: 0); |
| 4520 | Type *V4BF16 = |
| 4521 | FixedVectorType::get(ElementType: Type::getBFloatTy(C&: F->getContext()), NumElts: 4); |
| 4522 | Value *Trunc = Builder.CreateFPTrunc(V: CI->getOperand(i_nocapture: 0), DestTy: V4BF16); |
| 4523 | dbgs() << "Trunc: " << *Trunc << "\n" ; |
| 4524 | return Builder.CreateShuffleVector( |
| 4525 | V1: Trunc, V2: ConstantAggregateZero::get(Ty: V4BF16), Mask: ConcatMask); |
| 4526 | } else { |
| 4527 | return Builder.CreateFPTrunc(V: CI->getOperand(i_nocapture: 0), |
| 4528 | DestTy: Type::getBFloatTy(C&: F->getContext())); |
| 4529 | } |
| 4530 | } else if (Name.starts_with(Prefix: "sve.fcvt" )) { |
| 4531 | Intrinsic::ID NewID = |
| 4532 | StringSwitch<Intrinsic::ID>(Name) |
| 4533 | .Case(S: "sve.fcvt.bf16f32" , Value: Intrinsic::aarch64_sve_fcvt_bf16f32_v2) |
| 4534 | .Case(S: "sve.fcvtnt.bf16f32" , |
| 4535 | Value: Intrinsic::aarch64_sve_fcvtnt_bf16f32_v2) |
| 4536 | .Default(Value: Intrinsic::not_intrinsic); |
| 4537 | if (NewID == Intrinsic::not_intrinsic) |
| 4538 | llvm_unreachable("Unhandled Intrinsic!" ); |
| 4539 | |
| 4540 | SmallVector<Value *, 3> Args(CI->args()); |
| 4541 | |
| 4542 | // The original intrinsics incorrectly used a predicate based on the |
| 4543 | // smallest element type rather than the largest. |
| 4544 | Type *BadPredTy = ScalableVectorType::get(ElementType: Builder.getInt1Ty(), MinNumElts: 8); |
| 4545 | Type *GoodPredTy = ScalableVectorType::get(ElementType: Builder.getInt1Ty(), MinNumElts: 4); |
| 4546 | |
| 4547 | if (Args[1]->getType() != BadPredTy) |
| 4548 | llvm_unreachable("Unexpected predicate type!" ); |
| 4549 | |
| 4550 | Args[1] = Builder.CreateIntrinsic(ID: Intrinsic::aarch64_sve_convert_to_svbool, |
| 4551 | Types: BadPredTy, Args: Args[1]); |
| 4552 | Args[1] = Builder.CreateIntrinsic( |
| 4553 | ID: Intrinsic::aarch64_sve_convert_from_svbool, Types: GoodPredTy, Args: Args[1]); |
| 4554 | |
| 4555 | return Builder.CreateIntrinsic(ID: NewID, Args, /*FMFSource=*/nullptr, |
| 4556 | Name: CI->getName()); |
| 4557 | } |
| 4558 | |
| 4559 | llvm_unreachable("Unhandled Intrinsic!" ); |
| 4560 | } |
| 4561 | |
| 4562 | static Value *upgradeARMIntrinsicCall(StringRef Name, CallBase *CI, Function *F, |
| 4563 | IRBuilder<> &Builder) { |
| 4564 | if (Name == "mve.vctp64.old" ) { |
| 4565 | // Replace the old v4i1 vctp64 with a v2i1 vctp and predicate-casts to the |
| 4566 | // correct type. |
| 4567 | Value *VCTP = Builder.CreateIntrinsic(ID: Intrinsic::arm_mve_vctp64, Types: {}, |
| 4568 | Args: CI->getArgOperand(i: 0), |
| 4569 | /*FMFSource=*/nullptr, Name: CI->getName()); |
| 4570 | Value *C1 = Builder.CreateIntrinsic( |
| 4571 | ID: Intrinsic::arm_mve_pred_v2i, |
| 4572 | Types: {VectorType::get(ElementType: Builder.getInt1Ty(), NumElements: 2, Scalable: false)}, Args: VCTP); |
| 4573 | return Builder.CreateIntrinsic( |
| 4574 | ID: Intrinsic::arm_mve_pred_i2v, |
| 4575 | Types: {VectorType::get(ElementType: Builder.getInt1Ty(), NumElements: 4, Scalable: false)}, Args: C1); |
| 4576 | } else if (Name == "mve.mull.int.predicated.v2i64.v4i32.v4i1" || |
| 4577 | Name == "mve.vqdmull.predicated.v2i64.v4i32.v4i1" || |
| 4578 | Name == "mve.vldr.gather.base.predicated.v2i64.v2i64.v4i1" || |
| 4579 | Name == "mve.vldr.gather.base.wb.predicated.v2i64.v2i64.v4i1" || |
| 4580 | Name == |
| 4581 | "mve.vldr.gather.offset.predicated.v2i64.p0i64.v2i64.v4i1" || |
| 4582 | Name == "mve.vldr.gather.offset.predicated.v2i64.p0.v2i64.v4i1" || |
| 4583 | Name == "mve.vstr.scatter.base.predicated.v2i64.v2i64.v4i1" || |
| 4584 | Name == "mve.vstr.scatter.base.wb.predicated.v2i64.v2i64.v4i1" || |
| 4585 | Name == |
| 4586 | "mve.vstr.scatter.offset.predicated.p0i64.v2i64.v2i64.v4i1" || |
| 4587 | Name == "mve.vstr.scatter.offset.predicated.p0.v2i64.v2i64.v4i1" || |
| 4588 | Name == "cde.vcx1q.predicated.v2i64.v4i1" || |
| 4589 | Name == "cde.vcx1qa.predicated.v2i64.v4i1" || |
| 4590 | Name == "cde.vcx2q.predicated.v2i64.v4i1" || |
| 4591 | Name == "cde.vcx2qa.predicated.v2i64.v4i1" || |
| 4592 | Name == "cde.vcx3q.predicated.v2i64.v4i1" || |
| 4593 | Name == "cde.vcx3qa.predicated.v2i64.v4i1" ) { |
| 4594 | std::vector<Type *> Tys; |
| 4595 | unsigned ID = CI->getIntrinsicID(); |
| 4596 | Type *V2I1Ty = FixedVectorType::get(ElementType: Builder.getInt1Ty(), NumElts: 2); |
| 4597 | switch (ID) { |
| 4598 | case Intrinsic::arm_mve_mull_int_predicated: |
| 4599 | case Intrinsic::arm_mve_vqdmull_predicated: |
| 4600 | case Intrinsic::arm_mve_vldr_gather_base_predicated: |
| 4601 | Tys = {CI->getType(), CI->getOperand(i_nocapture: 0)->getType(), V2I1Ty}; |
| 4602 | break; |
| 4603 | case Intrinsic::arm_mve_vldr_gather_base_wb_predicated: |
| 4604 | case Intrinsic::arm_mve_vstr_scatter_base_predicated: |
| 4605 | case Intrinsic::arm_mve_vstr_scatter_base_wb_predicated: |
| 4606 | Tys = {CI->getOperand(i_nocapture: 0)->getType(), CI->getOperand(i_nocapture: 0)->getType(), |
| 4607 | V2I1Ty}; |
| 4608 | break; |
| 4609 | case Intrinsic::arm_mve_vldr_gather_offset_predicated: |
| 4610 | Tys = {CI->getType(), CI->getOperand(i_nocapture: 0)->getType(), |
| 4611 | CI->getOperand(i_nocapture: 1)->getType(), V2I1Ty}; |
| 4612 | break; |
| 4613 | case Intrinsic::arm_mve_vstr_scatter_offset_predicated: |
| 4614 | Tys = {CI->getOperand(i_nocapture: 0)->getType(), CI->getOperand(i_nocapture: 1)->getType(), |
| 4615 | CI->getOperand(i_nocapture: 2)->getType(), V2I1Ty}; |
| 4616 | break; |
| 4617 | case Intrinsic::arm_cde_vcx1q_predicated: |
| 4618 | case Intrinsic::arm_cde_vcx1qa_predicated: |
| 4619 | case Intrinsic::arm_cde_vcx2q_predicated: |
| 4620 | case Intrinsic::arm_cde_vcx2qa_predicated: |
| 4621 | case Intrinsic::arm_cde_vcx3q_predicated: |
| 4622 | case Intrinsic::arm_cde_vcx3qa_predicated: |
| 4623 | Tys = {CI->getOperand(i_nocapture: 1)->getType(), V2I1Ty}; |
| 4624 | break; |
| 4625 | default: |
| 4626 | llvm_unreachable("Unhandled Intrinsic!" ); |
| 4627 | } |
| 4628 | |
| 4629 | std::vector<Value *> Ops; |
| 4630 | for (Value *Op : CI->args()) { |
| 4631 | Type *Ty = Op->getType(); |
| 4632 | if (Ty->getScalarSizeInBits() == 1) { |
| 4633 | Value *C1 = Builder.CreateIntrinsic( |
| 4634 | ID: Intrinsic::arm_mve_pred_v2i, |
| 4635 | Types: {VectorType::get(ElementType: Builder.getInt1Ty(), NumElements: 4, Scalable: false)}, Args: Op); |
| 4636 | Op = Builder.CreateIntrinsic(ID: Intrinsic::arm_mve_pred_i2v, Types: {V2I1Ty}, Args: C1); |
| 4637 | } |
| 4638 | Ops.push_back(x: Op); |
| 4639 | } |
| 4640 | |
| 4641 | return Builder.CreateIntrinsic(ID, Types: Tys, Args: Ops, /*FMFSource=*/nullptr, |
| 4642 | Name: CI->getName()); |
| 4643 | } |
| 4644 | llvm_unreachable("Unknown function for ARM CallBase upgrade." ); |
| 4645 | } |
| 4646 | |
| 4647 | // These are expected to have the arguments: |
| 4648 | // atomic.intrin (ptr, rmw_value, ordering, scope, isVolatile) |
| 4649 | // |
| 4650 | // Except for int_amdgcn_ds_fadd_v2bf16 which only has (ptr, rmw_value). |
| 4651 | // |
| 4652 | static Value *upgradeAMDGCNIntrinsicCall(StringRef Name, CallBase *CI, |
| 4653 | Function *F, IRBuilder<> &Builder) { |
| 4654 | // Legacy WMMA iu intrinsics missed the optional clamp operand. Append clamp=0 |
| 4655 | // for compatibility. |
| 4656 | auto UpgradeLegacyWMMAIUIntrinsicCall = |
| 4657 | [](Function *F, CallBase *CI, IRBuilder<> &Builder, |
| 4658 | ArrayRef<Type *> OverloadTys) -> Value * { |
| 4659 | // Prepare arguments, append clamp=0 for compatibility |
| 4660 | SmallVector<Value *, 10> Args(CI->args().begin(), CI->args().end()); |
| 4661 | Args.push_back(Elt: Builder.getFalse()); |
| 4662 | |
| 4663 | // Insert the declaration for the right overload types |
| 4664 | Function *NewDecl = Intrinsic::getOrInsertDeclaration( |
| 4665 | M: F->getParent(), id: F->getIntrinsicID(), Tys: OverloadTys); |
| 4666 | |
| 4667 | // Copy operand bundles if any |
| 4668 | SmallVector<OperandBundleDef, 1> Bundles; |
| 4669 | CI->getOperandBundlesAsDefs(Defs&: Bundles); |
| 4670 | |
| 4671 | // Create the new call and copy calling properties |
| 4672 | auto *NewCall = cast<CallInst>(Val: Builder.CreateCall(Callee: NewDecl, Args, OpBundles: Bundles)); |
| 4673 | NewCall->setTailCallKind(cast<CallInst>(Val: CI)->getTailCallKind()); |
| 4674 | NewCall->setCallingConv(CI->getCallingConv()); |
| 4675 | NewCall->setAttributes(CI->getAttributes()); |
| 4676 | NewCall->setDebugLoc(CI->getDebugLoc()); |
| 4677 | NewCall->copyMetadata(SrcInst: *CI); |
| 4678 | return NewCall; |
| 4679 | }; |
| 4680 | |
| 4681 | if (F->getIntrinsicID() == Intrinsic::amdgcn_wmma_i32_16x16x64_iu8) { |
| 4682 | assert(CI->arg_size() == 7 && "Legacy int_amdgcn_wmma_i32_16x16x64_iu8 " |
| 4683 | "intrinsic should have 7 arguments" ); |
| 4684 | Type *T1 = CI->getArgOperand(i: 4)->getType(); |
| 4685 | Type *T2 = CI->getArgOperand(i: 1)->getType(); |
| 4686 | return UpgradeLegacyWMMAIUIntrinsicCall(F, CI, Builder, {T1, T2}); |
| 4687 | } |
| 4688 | if (F->getIntrinsicID() == Intrinsic::amdgcn_swmmac_i32_16x16x128_iu8) { |
| 4689 | assert(CI->arg_size() == 8 && "Legacy int_amdgcn_swmmac_i32_16x16x128_iu8 " |
| 4690 | "intrinsic should have 8 arguments" ); |
| 4691 | Type *T1 = CI->getArgOperand(i: 4)->getType(); |
| 4692 | Type *T2 = CI->getArgOperand(i: 1)->getType(); |
| 4693 | Type *T3 = CI->getArgOperand(i: 3)->getType(); |
| 4694 | Type *T4 = CI->getArgOperand(i: 5)->getType(); |
| 4695 | return UpgradeLegacyWMMAIUIntrinsicCall(F, CI, Builder, {T1, T2, T3, T4}); |
| 4696 | } |
| 4697 | |
| 4698 | AtomicRMWInst::BinOp RMWOp = |
| 4699 | StringSwitch<AtomicRMWInst::BinOp>(Name) |
| 4700 | .StartsWith(S: "ds.fadd" , Value: AtomicRMWInst::FAdd) |
| 4701 | .StartsWith(S: "ds.fmin" , Value: AtomicRMWInst::FMin) |
| 4702 | .StartsWith(S: "ds.fmax" , Value: AtomicRMWInst::FMax) |
| 4703 | .StartsWith(S: "atomic.inc." , Value: AtomicRMWInst::UIncWrap) |
| 4704 | .StartsWith(S: "atomic.dec." , Value: AtomicRMWInst::UDecWrap) |
| 4705 | .StartsWith(S: "global.atomic.fadd" , Value: AtomicRMWInst::FAdd) |
| 4706 | .StartsWith(S: "flat.atomic.fadd" , Value: AtomicRMWInst::FAdd) |
| 4707 | .StartsWith(S: "global.atomic.fmin" , Value: AtomicRMWInst::FMin) |
| 4708 | .StartsWith(S: "flat.atomic.fmin" , Value: AtomicRMWInst::FMin) |
| 4709 | .StartsWith(S: "global.atomic.fmax" , Value: AtomicRMWInst::FMax) |
| 4710 | .StartsWith(S: "flat.atomic.fmax" , Value: AtomicRMWInst::FMax) |
| 4711 | .StartsWith(S: "atomic.cond.sub" , Value: AtomicRMWInst::USubCond) |
| 4712 | .StartsWith(S: "atomic.csub" , Value: AtomicRMWInst::USubSat); |
| 4713 | |
| 4714 | unsigned NumOperands = CI->getNumOperands(); |
| 4715 | if (NumOperands < 3) // Malformed bitcode. |
| 4716 | return nullptr; |
| 4717 | |
| 4718 | Value *Ptr = CI->getArgOperand(i: 0); |
| 4719 | PointerType *PtrTy = dyn_cast<PointerType>(Val: Ptr->getType()); |
| 4720 | if (!PtrTy) // Malformed. |
| 4721 | return nullptr; |
| 4722 | |
| 4723 | Value *Val = CI->getArgOperand(i: 1); |
| 4724 | if (Val->getType() != CI->getType()) // Malformed. |
| 4725 | return nullptr; |
| 4726 | |
| 4727 | ConstantInt *OrderArg = nullptr; |
| 4728 | bool IsVolatile = false; |
| 4729 | |
| 4730 | // These should have 5 arguments (plus the callee). A separate version of the |
| 4731 | // ds_fadd intrinsic was defined for bf16 which was missing arguments. |
| 4732 | if (NumOperands > 3) |
| 4733 | OrderArg = dyn_cast<ConstantInt>(Val: CI->getArgOperand(i: 2)); |
| 4734 | |
| 4735 | // Ignore scope argument at 3 |
| 4736 | |
| 4737 | if (NumOperands > 5) { |
| 4738 | ConstantInt *VolatileArg = dyn_cast<ConstantInt>(Val: CI->getArgOperand(i: 4)); |
| 4739 | IsVolatile = !VolatileArg || !VolatileArg->isZero(); |
| 4740 | } |
| 4741 | |
| 4742 | AtomicOrdering Order = AtomicOrdering::SequentiallyConsistent; |
| 4743 | if (OrderArg && isValidAtomicOrdering(I: OrderArg->getZExtValue())) |
| 4744 | Order = static_cast<AtomicOrdering>(OrderArg->getZExtValue()); |
| 4745 | if (Order == AtomicOrdering::NotAtomic || Order == AtomicOrdering::Unordered) |
| 4746 | Order = AtomicOrdering::SequentiallyConsistent; |
| 4747 | |
| 4748 | LLVMContext &Ctx = F->getContext(); |
| 4749 | |
| 4750 | // Handle the v2bf16 intrinsic which used <2 x i16> instead of <2 x bfloat> |
| 4751 | Type *RetTy = CI->getType(); |
| 4752 | if (VectorType *VT = dyn_cast<VectorType>(Val: RetTy)) { |
| 4753 | if (VT->getElementType()->isIntegerTy(Bitwidth: 16)) { |
| 4754 | VectorType *AsBF16 = |
| 4755 | VectorType::get(ElementType: Type::getBFloatTy(C&: Ctx), EC: VT->getElementCount()); |
| 4756 | Val = Builder.CreateBitCast(V: Val, DestTy: AsBF16); |
| 4757 | } |
| 4758 | } |
| 4759 | |
| 4760 | // The scope argument never really worked correctly. Use agent as the most |
| 4761 | // conservative option which should still always produce the instruction. |
| 4762 | SyncScope::ID SSID = Ctx.getOrInsertSyncScopeID(SSN: "agent" ); |
| 4763 | AtomicRMWInst *RMW = |
| 4764 | Builder.CreateAtomicRMW(Op: RMWOp, Ptr, Val, Align: std::nullopt, Ordering: Order, SSID); |
| 4765 | |
| 4766 | unsigned AddrSpace = PtrTy->getAddressSpace(); |
| 4767 | if (AddrSpace != AMDGPUAS::LOCAL_ADDRESS) { |
| 4768 | MDNode *EmptyMD = MDNode::get(Context&: F->getContext(), MDs: {}); |
| 4769 | RMW->setMetadata(Kind: "amdgpu.no.fine.grained.memory" , Node: EmptyMD); |
| 4770 | if (RMWOp == AtomicRMWInst::FAdd && RetTy->isFloatTy()) |
| 4771 | RMW->setMetadata(Kind: "amdgpu.ignore.denormal.mode" , Node: EmptyMD); |
| 4772 | } |
| 4773 | |
| 4774 | if (AddrSpace == AMDGPUAS::FLAT_ADDRESS) { |
| 4775 | MDBuilder MDB(F->getContext()); |
| 4776 | MDNode *RangeNotPrivate = |
| 4777 | MDB.createRange(Lo: APInt(32, AMDGPUAS::PRIVATE_ADDRESS), |
| 4778 | Hi: APInt(32, AMDGPUAS::PRIVATE_ADDRESS + 1)); |
| 4779 | RMW->setMetadata(KindID: LLVMContext::MD_noalias_addrspace, Node: RangeNotPrivate); |
| 4780 | } |
| 4781 | |
| 4782 | if (IsVolatile) |
| 4783 | RMW->setVolatile(true); |
| 4784 | |
| 4785 | return Builder.CreateBitCast(V: RMW, DestTy: RetTy); |
| 4786 | } |
| 4787 | |
| 4788 | /// Helper to unwrap intrinsic call MetadataAsValue operands. Return as a |
| 4789 | /// plain MDNode, as it's the verifier's job to check these are the correct |
| 4790 | /// types later. |
| 4791 | static MDNode *unwrapMAVOp(CallBase *CI, unsigned Op) { |
| 4792 | if (Op < CI->arg_size()) { |
| 4793 | if (MetadataAsValue *MAV = |
| 4794 | dyn_cast<MetadataAsValue>(Val: CI->getArgOperand(i: Op))) { |
| 4795 | Metadata *MD = MAV->getMetadata(); |
| 4796 | return dyn_cast_if_present<MDNode>(Val: MD); |
| 4797 | } |
| 4798 | } |
| 4799 | return nullptr; |
| 4800 | } |
| 4801 | |
| 4802 | /// Helper to unwrap Metadata MetadataAsValue operands, such as the Value field. |
| 4803 | static Metadata *unwrapMAVMetadataOp(CallBase *CI, unsigned Op) { |
| 4804 | if (Op < CI->arg_size()) |
| 4805 | if (MetadataAsValue *MAV = dyn_cast<MetadataAsValue>(Val: CI->getArgOperand(i: Op))) |
| 4806 | return MAV->getMetadata(); |
| 4807 | return nullptr; |
| 4808 | } |
| 4809 | |
| 4810 | static MDNode *getDebugLocSafe(const Instruction *I) { |
| 4811 | // The MDNode attached to this instruction might not be the correct type, |
| 4812 | // as the verifier has not yet be run. Fetch it as a bare MDNode. |
| 4813 | return I->getDebugLoc().getAsMDNode(); |
| 4814 | } |
| 4815 | |
| 4816 | /// Convert debug intrinsic calls to non-instruction debug records. |
| 4817 | /// \p Name - Final part of the intrinsic name, e.g. 'value' in llvm.dbg.value. |
| 4818 | /// \p CI - The debug intrinsic call. |
| 4819 | static void upgradeDbgIntrinsicToDbgRecord(StringRef Name, CallBase *CI) { |
| 4820 | DbgRecord *DR = nullptr; |
| 4821 | if (Name == "label" ) { |
| 4822 | DR = DbgLabelRecord::createUnresolvedDbgLabelRecord(Label: unwrapMAVOp(CI, Op: 0), |
| 4823 | DL: CI->getDebugLoc()); |
| 4824 | } else if (Name == "assign" ) { |
| 4825 | DR = DbgVariableRecord::createUnresolvedDbgVariableRecord( |
| 4826 | Type: DbgVariableRecord::LocationType::Assign, Val: unwrapMAVMetadataOp(CI, Op: 0), |
| 4827 | Variable: unwrapMAVOp(CI, Op: 1), Expression: unwrapMAVOp(CI, Op: 2), AssignID: unwrapMAVOp(CI, Op: 3), |
| 4828 | Address: unwrapMAVMetadataOp(CI, Op: 4), |
| 4829 | /*The address is a Value ref, it will be stored as a Metadata */ |
| 4830 | AddressExpression: unwrapMAVOp(CI, Op: 5), DI: getDebugLocSafe(I: CI)); |
| 4831 | } else if (Name == "declare" ) { |
| 4832 | DR = DbgVariableRecord::createUnresolvedDbgVariableRecord( |
| 4833 | Type: DbgVariableRecord::LocationType::Declare, Val: unwrapMAVMetadataOp(CI, Op: 0), |
| 4834 | Variable: unwrapMAVOp(CI, Op: 1), Expression: unwrapMAVOp(CI, Op: 2), AssignID: nullptr, Address: nullptr, AddressExpression: nullptr, |
| 4835 | DI: getDebugLocSafe(I: CI)); |
| 4836 | } else if (Name == "addr" ) { |
| 4837 | // Upgrade dbg.addr to dbg.value with DW_OP_deref. |
| 4838 | MDNode *ExprNode = unwrapMAVOp(CI, Op: 2); |
| 4839 | // Don't try to add something to the expression if it's not an expression. |
| 4840 | // Instead, allow the verifier to fail later. |
| 4841 | if (DIExpression *Expr = dyn_cast<DIExpression>(Val: ExprNode)) { |
| 4842 | ExprNode = DIExpression::append(Expr, Ops: dwarf::DW_OP_deref); |
| 4843 | } |
| 4844 | DR = DbgVariableRecord::createUnresolvedDbgVariableRecord( |
| 4845 | Type: DbgVariableRecord::LocationType::Value, Val: unwrapMAVMetadataOp(CI, Op: 0), |
| 4846 | Variable: unwrapMAVOp(CI, Op: 1), Expression: ExprNode, AssignID: nullptr, Address: nullptr, AddressExpression: nullptr, |
| 4847 | DI: getDebugLocSafe(I: CI)); |
| 4848 | } else if (Name == "value" ) { |
| 4849 | // An old version of dbg.value had an extra offset argument. |
| 4850 | unsigned VarOp = 1; |
| 4851 | unsigned ExprOp = 2; |
| 4852 | if (CI->arg_size() == 4) { |
| 4853 | auto *Offset = dyn_cast_or_null<Constant>(Val: CI->getArgOperand(i: 1)); |
| 4854 | // Nonzero offset dbg.values get dropped without a replacement. |
| 4855 | if (!Offset || !Offset->isZeroValue()) |
| 4856 | return; |
| 4857 | VarOp = 2; |
| 4858 | ExprOp = 3; |
| 4859 | } |
| 4860 | DR = DbgVariableRecord::createUnresolvedDbgVariableRecord( |
| 4861 | Type: DbgVariableRecord::LocationType::Value, Val: unwrapMAVMetadataOp(CI, Op: 0), |
| 4862 | Variable: unwrapMAVOp(CI, Op: VarOp), Expression: unwrapMAVOp(CI, Op: ExprOp), AssignID: nullptr, Address: nullptr, |
| 4863 | AddressExpression: nullptr, DI: getDebugLocSafe(I: CI)); |
| 4864 | } |
| 4865 | assert(DR && "Unhandled intrinsic kind in upgrade to DbgRecord" ); |
| 4866 | CI->getParent()->insertDbgRecordBefore(DR, Here: CI->getIterator()); |
| 4867 | } |
| 4868 | |
| 4869 | static Value *upgradeVectorSplice(CallBase *CI, IRBuilder<> &Builder) { |
| 4870 | auto *Offset = dyn_cast<ConstantInt>(Val: CI->getArgOperand(i: 2)); |
| 4871 | if (!Offset) |
| 4872 | reportFatalUsageError(reason: "Invalid llvm.vector.splice offset argument" ); |
| 4873 | int64_t OffsetVal = Offset->getSExtValue(); |
| 4874 | return Builder.CreateIntrinsic(ID: OffsetVal >= 0 |
| 4875 | ? Intrinsic::vector_splice_left |
| 4876 | : Intrinsic::vector_splice_right, |
| 4877 | Types: CI->getType(), |
| 4878 | Args: {CI->getArgOperand(i: 0), CI->getArgOperand(i: 1), |
| 4879 | Builder.getInt32(C: std::abs(i: OffsetVal))}); |
| 4880 | } |
| 4881 | |
| 4882 | static Value *upgradeConvertIntrinsicCall(StringRef Name, CallBase *CI, |
| 4883 | Function *F, IRBuilder<> &Builder) { |
| 4884 | if (Name.starts_with(Prefix: "to.fp16" )) { |
| 4885 | Value *Cast = |
| 4886 | Builder.CreateFPTrunc(V: CI->getArgOperand(i: 0), DestTy: Builder.getHalfTy()); |
| 4887 | return Builder.CreateBitCast(V: Cast, DestTy: CI->getType()); |
| 4888 | } |
| 4889 | |
| 4890 | if (Name.starts_with(Prefix: "from.fp16" )) { |
| 4891 | Value *Cast = |
| 4892 | Builder.CreateBitCast(V: CI->getArgOperand(i: 0), DestTy: Builder.getHalfTy()); |
| 4893 | return Builder.CreateFPExt(V: Cast, DestTy: CI->getType()); |
| 4894 | } |
| 4895 | |
| 4896 | return nullptr; |
| 4897 | } |
| 4898 | |
| 4899 | /// Upgrade a call to an old intrinsic. All argument and return casting must be |
| 4900 | /// provided to seamlessly integrate with existing context. |
| 4901 | void llvm::UpgradeIntrinsicCall(CallBase *CI, Function *NewFn) { |
| 4902 | // Note dyn_cast to Function is not quite the same as getCalledFunction, which |
| 4903 | // checks the callee's function type matches. It's likely we need to handle |
| 4904 | // type changes here. |
| 4905 | Function *F = dyn_cast<Function>(Val: CI->getCalledOperand()); |
| 4906 | if (!F) |
| 4907 | return; |
| 4908 | |
| 4909 | LLVMContext &C = CI->getContext(); |
| 4910 | IRBuilder<> Builder(C); |
| 4911 | if (isa<FPMathOperator>(Val: CI)) |
| 4912 | Builder.setFastMathFlags(CI->getFastMathFlags()); |
| 4913 | Builder.SetInsertPoint(TheBB: CI->getParent(), IP: CI->getIterator()); |
| 4914 | |
| 4915 | if (!NewFn) { |
| 4916 | // Get the Function's name. |
| 4917 | StringRef Name = F->getName(); |
| 4918 | if (!Name.consume_front(Prefix: "llvm." )) |
| 4919 | llvm_unreachable("intrinsic doesn't start with 'llvm.'" ); |
| 4920 | |
| 4921 | bool IsX86 = Name.consume_front(Prefix: "x86." ); |
| 4922 | bool IsNVVM = Name.consume_front(Prefix: "nvvm." ); |
| 4923 | bool IsAArch64 = Name.consume_front(Prefix: "aarch64." ); |
| 4924 | bool IsARM = Name.consume_front(Prefix: "arm." ); |
| 4925 | bool IsAMDGCN = Name.consume_front(Prefix: "amdgcn." ); |
| 4926 | bool IsDbg = Name.consume_front(Prefix: "dbg." ); |
| 4927 | bool IsOldSplice = |
| 4928 | (Name.consume_front(Prefix: "experimental.vector.splice" ) || |
| 4929 | Name.consume_front(Prefix: "vector.splice" )) && |
| 4930 | !(Name.starts_with(Prefix: ".left" ) || Name.starts_with(Prefix: ".right" )); |
| 4931 | Value *Rep = nullptr; |
| 4932 | |
| 4933 | if (!IsX86 && Name == "stackprotectorcheck" ) { |
| 4934 | Rep = nullptr; |
| 4935 | } else if (IsNVVM) { |
| 4936 | Rep = upgradeNVVMIntrinsicCall(Name, CI, F, Builder); |
| 4937 | } else if (IsX86) { |
| 4938 | Rep = upgradeX86IntrinsicCall(Name, CI, F, Builder); |
| 4939 | } else if (IsAArch64) { |
| 4940 | Rep = upgradeAArch64IntrinsicCall(Name, CI, F, Builder); |
| 4941 | } else if (IsARM) { |
| 4942 | Rep = upgradeARMIntrinsicCall(Name, CI, F, Builder); |
| 4943 | } else if (IsAMDGCN) { |
| 4944 | Rep = upgradeAMDGCNIntrinsicCall(Name, CI, F, Builder); |
| 4945 | } else if (IsDbg) { |
| 4946 | upgradeDbgIntrinsicToDbgRecord(Name, CI); |
| 4947 | } else if (IsOldSplice) { |
| 4948 | Rep = upgradeVectorSplice(CI, Builder); |
| 4949 | } else if (Name.consume_front(Prefix: "convert." )) { |
| 4950 | Rep = upgradeConvertIntrinsicCall(Name, CI, F, Builder); |
| 4951 | } else { |
| 4952 | llvm_unreachable("Unknown function for CallBase upgrade." ); |
| 4953 | } |
| 4954 | |
| 4955 | if (Rep) |
| 4956 | CI->replaceAllUsesWith(V: Rep); |
| 4957 | CI->eraseFromParent(); |
| 4958 | return; |
| 4959 | } |
| 4960 | |
| 4961 | const auto &DefaultCase = [&]() -> void { |
| 4962 | if (F == NewFn) |
| 4963 | return; |
| 4964 | |
| 4965 | if (CI->getFunctionType() == NewFn->getFunctionType()) { |
| 4966 | // Handle generic mangling change. |
| 4967 | assert( |
| 4968 | (CI->getCalledFunction()->getName() != NewFn->getName()) && |
| 4969 | "Unknown function for CallBase upgrade and isn't just a name change" ); |
| 4970 | CI->setCalledFunction(NewFn); |
| 4971 | return; |
| 4972 | } |
| 4973 | |
| 4974 | // This must be an upgrade from a named to a literal struct. |
| 4975 | if (auto *OldST = dyn_cast<StructType>(Val: CI->getType())) { |
| 4976 | assert(OldST != NewFn->getReturnType() && |
| 4977 | "Return type must have changed" ); |
| 4978 | assert(OldST->getNumElements() == |
| 4979 | cast<StructType>(NewFn->getReturnType())->getNumElements() && |
| 4980 | "Must have same number of elements" ); |
| 4981 | |
| 4982 | SmallVector<Value *> Args(CI->args()); |
| 4983 | CallInst *NewCI = Builder.CreateCall(Callee: NewFn, Args); |
| 4984 | NewCI->setAttributes(CI->getAttributes()); |
| 4985 | Value *Res = PoisonValue::get(T: OldST); |
| 4986 | for (unsigned Idx = 0; Idx < OldST->getNumElements(); ++Idx) { |
| 4987 | Value *Elem = Builder.CreateExtractValue(Agg: NewCI, Idxs: Idx); |
| 4988 | Res = Builder.CreateInsertValue(Agg: Res, Val: Elem, Idxs: Idx); |
| 4989 | } |
| 4990 | CI->replaceAllUsesWith(V: Res); |
| 4991 | CI->eraseFromParent(); |
| 4992 | return; |
| 4993 | } |
| 4994 | |
| 4995 | // We're probably about to produce something invalid. Let the verifier catch |
| 4996 | // it instead of dying here. |
| 4997 | CI->setCalledOperand( |
| 4998 | ConstantExpr::getPointerCast(C: NewFn, Ty: CI->getCalledOperand()->getType())); |
| 4999 | return; |
| 5000 | }; |
| 5001 | CallInst *NewCall = nullptr; |
| 5002 | switch (NewFn->getIntrinsicID()) { |
| 5003 | default: { |
| 5004 | DefaultCase(); |
| 5005 | return; |
| 5006 | } |
| 5007 | case Intrinsic::arm_neon_vst1: |
| 5008 | case Intrinsic::arm_neon_vst2: |
| 5009 | case Intrinsic::arm_neon_vst3: |
| 5010 | case Intrinsic::arm_neon_vst4: |
| 5011 | case Intrinsic::arm_neon_vst2lane: |
| 5012 | case Intrinsic::arm_neon_vst3lane: |
| 5013 | case Intrinsic::arm_neon_vst4lane: { |
| 5014 | SmallVector<Value *, 4> Args(CI->args()); |
| 5015 | NewCall = Builder.CreateCall(Callee: NewFn, Args); |
| 5016 | break; |
| 5017 | } |
| 5018 | case Intrinsic::aarch64_sve_bfmlalb_lane_v2: |
| 5019 | case Intrinsic::aarch64_sve_bfmlalt_lane_v2: |
| 5020 | case Intrinsic::aarch64_sve_bfdot_lane_v2: { |
| 5021 | LLVMContext &Ctx = F->getParent()->getContext(); |
| 5022 | SmallVector<Value *, 4> Args(CI->args()); |
| 5023 | Args[3] = ConstantInt::get(Ty: Type::getInt32Ty(C&: Ctx), |
| 5024 | V: cast<ConstantInt>(Val: Args[3])->getZExtValue()); |
| 5025 | NewCall = Builder.CreateCall(Callee: NewFn, Args); |
| 5026 | break; |
| 5027 | } |
| 5028 | case Intrinsic::aarch64_sve_ld3_sret: |
| 5029 | case Intrinsic::aarch64_sve_ld4_sret: |
| 5030 | case Intrinsic::aarch64_sve_ld2_sret: { |
| 5031 | StringRef Name = F->getName(); |
| 5032 | Name = Name.substr(Start: 5); |
| 5033 | unsigned N = StringSwitch<unsigned>(Name) |
| 5034 | .StartsWith(S: "aarch64.sve.ld2" , Value: 2) |
| 5035 | .StartsWith(S: "aarch64.sve.ld3" , Value: 3) |
| 5036 | .StartsWith(S: "aarch64.sve.ld4" , Value: 4) |
| 5037 | .Default(Value: 0); |
| 5038 | auto *RetTy = cast<ScalableVectorType>(Val: F->getReturnType()); |
| 5039 | unsigned MinElts = RetTy->getMinNumElements() / N; |
| 5040 | SmallVector<Value *, 2> Args(CI->args()); |
| 5041 | Value *NewLdCall = Builder.CreateCall(Callee: NewFn, Args); |
| 5042 | Value *Ret = llvm::PoisonValue::get(T: RetTy); |
| 5043 | for (unsigned I = 0; I < N; I++) { |
| 5044 | Value *SRet = Builder.CreateExtractValue(Agg: NewLdCall, Idxs: I); |
| 5045 | Ret = Builder.CreateInsertVector(DstType: RetTy, SrcVec: Ret, SubVec: SRet, Idx: I * MinElts); |
| 5046 | } |
| 5047 | NewCall = dyn_cast<CallInst>(Val: Ret); |
| 5048 | break; |
| 5049 | } |
| 5050 | |
| 5051 | case Intrinsic::coro_end: { |
| 5052 | SmallVector<Value *, 3> Args(CI->args()); |
| 5053 | Args.push_back(Elt: ConstantTokenNone::get(Context&: CI->getContext())); |
| 5054 | NewCall = Builder.CreateCall(Callee: NewFn, Args); |
| 5055 | break; |
| 5056 | } |
| 5057 | |
| 5058 | case Intrinsic::vector_extract: { |
| 5059 | StringRef Name = F->getName(); |
| 5060 | Name = Name.substr(Start: 5); // Strip llvm |
| 5061 | if (!Name.starts_with(Prefix: "aarch64.sve.tuple.get" )) { |
| 5062 | DefaultCase(); |
| 5063 | return; |
| 5064 | } |
| 5065 | auto *RetTy = cast<ScalableVectorType>(Val: F->getReturnType()); |
| 5066 | unsigned MinElts = RetTy->getMinNumElements(); |
| 5067 | unsigned I = cast<ConstantInt>(Val: CI->getArgOperand(i: 1))->getZExtValue(); |
| 5068 | Value *NewIdx = ConstantInt::get(Ty: Type::getInt64Ty(C), V: I * MinElts); |
| 5069 | NewCall = Builder.CreateCall(Callee: NewFn, Args: {CI->getArgOperand(i: 0), NewIdx}); |
| 5070 | break; |
| 5071 | } |
| 5072 | |
| 5073 | case Intrinsic::vector_insert: { |
| 5074 | StringRef Name = F->getName(); |
| 5075 | Name = Name.substr(Start: 5); |
| 5076 | if (!Name.starts_with(Prefix: "aarch64.sve.tuple" )) { |
| 5077 | DefaultCase(); |
| 5078 | return; |
| 5079 | } |
| 5080 | if (Name.starts_with(Prefix: "aarch64.sve.tuple.set" )) { |
| 5081 | unsigned I = cast<ConstantInt>(Val: CI->getArgOperand(i: 1))->getZExtValue(); |
| 5082 | auto *Ty = cast<ScalableVectorType>(Val: CI->getArgOperand(i: 2)->getType()); |
| 5083 | Value *NewIdx = |
| 5084 | ConstantInt::get(Ty: Type::getInt64Ty(C), V: I * Ty->getMinNumElements()); |
| 5085 | NewCall = Builder.CreateCall( |
| 5086 | Callee: NewFn, Args: {CI->getArgOperand(i: 0), CI->getArgOperand(i: 2), NewIdx}); |
| 5087 | break; |
| 5088 | } |
| 5089 | if (Name.starts_with(Prefix: "aarch64.sve.tuple.create" )) { |
| 5090 | unsigned N = StringSwitch<unsigned>(Name) |
| 5091 | .StartsWith(S: "aarch64.sve.tuple.create2" , Value: 2) |
| 5092 | .StartsWith(S: "aarch64.sve.tuple.create3" , Value: 3) |
| 5093 | .StartsWith(S: "aarch64.sve.tuple.create4" , Value: 4) |
| 5094 | .Default(Value: 0); |
| 5095 | assert(N > 1 && "Create is expected to be between 2-4" ); |
| 5096 | auto *RetTy = cast<ScalableVectorType>(Val: F->getReturnType()); |
| 5097 | Value *Ret = llvm::PoisonValue::get(T: RetTy); |
| 5098 | unsigned MinElts = RetTy->getMinNumElements() / N; |
| 5099 | for (unsigned I = 0; I < N; I++) { |
| 5100 | Value *V = CI->getArgOperand(i: I); |
| 5101 | Ret = Builder.CreateInsertVector(DstType: RetTy, SrcVec: Ret, SubVec: V, Idx: I * MinElts); |
| 5102 | } |
| 5103 | NewCall = dyn_cast<CallInst>(Val: Ret); |
| 5104 | } |
| 5105 | break; |
| 5106 | } |
| 5107 | |
| 5108 | case Intrinsic::arm_neon_bfdot: |
| 5109 | case Intrinsic::arm_neon_bfmmla: |
| 5110 | case Intrinsic::arm_neon_bfmlalb: |
| 5111 | case Intrinsic::arm_neon_bfmlalt: |
| 5112 | case Intrinsic::aarch64_neon_bfdot: |
| 5113 | case Intrinsic::aarch64_neon_bfmmla: |
| 5114 | case Intrinsic::aarch64_neon_bfmlalb: |
| 5115 | case Intrinsic::aarch64_neon_bfmlalt: { |
| 5116 | SmallVector<Value *, 3> Args; |
| 5117 | assert(CI->arg_size() == 3 && |
| 5118 | "Mismatch between function args and call args" ); |
| 5119 | size_t OperandWidth = |
| 5120 | CI->getArgOperand(i: 1)->getType()->getPrimitiveSizeInBits(); |
| 5121 | assert((OperandWidth == 64 || OperandWidth == 128) && |
| 5122 | "Unexpected operand width" ); |
| 5123 | Type *NewTy = FixedVectorType::get(ElementType: Type::getBFloatTy(C), NumElts: OperandWidth / 16); |
| 5124 | auto Iter = CI->args().begin(); |
| 5125 | Args.push_back(Elt: *Iter++); |
| 5126 | Args.push_back(Elt: Builder.CreateBitCast(V: *Iter++, DestTy: NewTy)); |
| 5127 | Args.push_back(Elt: Builder.CreateBitCast(V: *Iter++, DestTy: NewTy)); |
| 5128 | NewCall = Builder.CreateCall(Callee: NewFn, Args); |
| 5129 | break; |
| 5130 | } |
| 5131 | |
| 5132 | case Intrinsic::bitreverse: |
| 5133 | NewCall = Builder.CreateCall(Callee: NewFn, Args: {CI->getArgOperand(i: 0)}); |
| 5134 | break; |
| 5135 | |
| 5136 | case Intrinsic::ctlz: |
| 5137 | case Intrinsic::cttz: { |
| 5138 | if (CI->arg_size() != 1) { |
| 5139 | DefaultCase(); |
| 5140 | return; |
| 5141 | } |
| 5142 | |
| 5143 | NewCall = |
| 5144 | Builder.CreateCall(Callee: NewFn, Args: {CI->getArgOperand(i: 0), Builder.getFalse()}); |
| 5145 | break; |
| 5146 | } |
| 5147 | |
| 5148 | case Intrinsic::objectsize: { |
| 5149 | Value *NullIsUnknownSize = |
| 5150 | CI->arg_size() == 2 ? Builder.getFalse() : CI->getArgOperand(i: 2); |
| 5151 | Value *Dynamic = |
| 5152 | CI->arg_size() < 4 ? Builder.getFalse() : CI->getArgOperand(i: 3); |
| 5153 | NewCall = Builder.CreateCall( |
| 5154 | Callee: NewFn, Args: {CI->getArgOperand(i: 0), CI->getArgOperand(i: 1), NullIsUnknownSize, Dynamic}); |
| 5155 | break; |
| 5156 | } |
| 5157 | |
| 5158 | case Intrinsic::ctpop: |
| 5159 | NewCall = Builder.CreateCall(Callee: NewFn, Args: {CI->getArgOperand(i: 0)}); |
| 5160 | break; |
| 5161 | case Intrinsic::dbg_value: { |
| 5162 | StringRef Name = F->getName(); |
| 5163 | Name = Name.substr(Start: 5); // Strip llvm. |
| 5164 | // Upgrade `dbg.addr` to `dbg.value` with `DW_OP_deref`. |
| 5165 | if (Name.starts_with(Prefix: "dbg.addr" )) { |
| 5166 | DIExpression *Expr = cast<DIExpression>( |
| 5167 | Val: cast<MetadataAsValue>(Val: CI->getArgOperand(i: 2))->getMetadata()); |
| 5168 | Expr = DIExpression::append(Expr, Ops: dwarf::DW_OP_deref); |
| 5169 | NewCall = |
| 5170 | Builder.CreateCall(Callee: NewFn, Args: {CI->getArgOperand(i: 0), CI->getArgOperand(i: 1), |
| 5171 | MetadataAsValue::get(Context&: C, MD: Expr)}); |
| 5172 | break; |
| 5173 | } |
| 5174 | |
| 5175 | // Upgrade from the old version that had an extra offset argument. |
| 5176 | assert(CI->arg_size() == 4); |
| 5177 | // Drop nonzero offsets instead of attempting to upgrade them. |
| 5178 | if (auto *Offset = dyn_cast_or_null<Constant>(Val: CI->getArgOperand(i: 1))) |
| 5179 | if (Offset->isZeroValue()) { |
| 5180 | NewCall = Builder.CreateCall( |
| 5181 | Callee: NewFn, |
| 5182 | Args: {CI->getArgOperand(i: 0), CI->getArgOperand(i: 2), CI->getArgOperand(i: 3)}); |
| 5183 | break; |
| 5184 | } |
| 5185 | CI->eraseFromParent(); |
| 5186 | return; |
| 5187 | } |
| 5188 | |
| 5189 | case Intrinsic::ptr_annotation: |
| 5190 | // Upgrade from versions that lacked the annotation attribute argument. |
| 5191 | if (CI->arg_size() != 4) { |
| 5192 | DefaultCase(); |
| 5193 | return; |
| 5194 | } |
| 5195 | |
| 5196 | // Create a new call with an added null annotation attribute argument. |
| 5197 | NewCall = Builder.CreateCall( |
| 5198 | Callee: NewFn, |
| 5199 | Args: {CI->getArgOperand(i: 0), CI->getArgOperand(i: 1), CI->getArgOperand(i: 2), |
| 5200 | CI->getArgOperand(i: 3), ConstantPointerNull::get(T: Builder.getPtrTy())}); |
| 5201 | NewCall->takeName(V: CI); |
| 5202 | CI->replaceAllUsesWith(V: NewCall); |
| 5203 | CI->eraseFromParent(); |
| 5204 | return; |
| 5205 | |
| 5206 | case Intrinsic::var_annotation: |
| 5207 | // Upgrade from versions that lacked the annotation attribute argument. |
| 5208 | if (CI->arg_size() != 4) { |
| 5209 | DefaultCase(); |
| 5210 | return; |
| 5211 | } |
| 5212 | // Create a new call with an added null annotation attribute argument. |
| 5213 | NewCall = Builder.CreateCall( |
| 5214 | Callee: NewFn, |
| 5215 | Args: {CI->getArgOperand(i: 0), CI->getArgOperand(i: 1), CI->getArgOperand(i: 2), |
| 5216 | CI->getArgOperand(i: 3), ConstantPointerNull::get(T: Builder.getPtrTy())}); |
| 5217 | NewCall->takeName(V: CI); |
| 5218 | CI->replaceAllUsesWith(V: NewCall); |
| 5219 | CI->eraseFromParent(); |
| 5220 | return; |
| 5221 | |
| 5222 | case Intrinsic::riscv_aes32dsi: |
| 5223 | case Intrinsic::riscv_aes32dsmi: |
| 5224 | case Intrinsic::riscv_aes32esi: |
| 5225 | case Intrinsic::riscv_aes32esmi: |
| 5226 | case Intrinsic::riscv_sm4ks: |
| 5227 | case Intrinsic::riscv_sm4ed: { |
| 5228 | // The last argument to these intrinsics used to be i8 and changed to i32. |
| 5229 | // The type overload for sm4ks and sm4ed was removed. |
| 5230 | Value *Arg2 = CI->getArgOperand(i: 2); |
| 5231 | if (Arg2->getType()->isIntegerTy(Bitwidth: 32) && !CI->getType()->isIntegerTy(Bitwidth: 64)) |
| 5232 | return; |
| 5233 | |
| 5234 | Value *Arg0 = CI->getArgOperand(i: 0); |
| 5235 | Value *Arg1 = CI->getArgOperand(i: 1); |
| 5236 | if (CI->getType()->isIntegerTy(Bitwidth: 64)) { |
| 5237 | Arg0 = Builder.CreateTrunc(V: Arg0, DestTy: Builder.getInt32Ty()); |
| 5238 | Arg1 = Builder.CreateTrunc(V: Arg1, DestTy: Builder.getInt32Ty()); |
| 5239 | } |
| 5240 | |
| 5241 | Arg2 = ConstantInt::get(Ty: Type::getInt32Ty(C), |
| 5242 | V: cast<ConstantInt>(Val: Arg2)->getZExtValue()); |
| 5243 | |
| 5244 | NewCall = Builder.CreateCall(Callee: NewFn, Args: {Arg0, Arg1, Arg2}); |
| 5245 | Value *Res = NewCall; |
| 5246 | if (Res->getType() != CI->getType()) |
| 5247 | Res = Builder.CreateIntCast(V: NewCall, DestTy: CI->getType(), /*isSigned*/ true); |
| 5248 | NewCall->takeName(V: CI); |
| 5249 | CI->replaceAllUsesWith(V: Res); |
| 5250 | CI->eraseFromParent(); |
| 5251 | return; |
| 5252 | } |
| 5253 | case Intrinsic::nvvm_mapa_shared_cluster: { |
| 5254 | // Create a new call with the correct address space. |
| 5255 | NewCall = |
| 5256 | Builder.CreateCall(Callee: NewFn, Args: {CI->getArgOperand(i: 0), CI->getArgOperand(i: 1)}); |
| 5257 | Value *Res = NewCall; |
| 5258 | Res = Builder.CreateAddrSpaceCast( |
| 5259 | V: Res, DestTy: Builder.getPtrTy(AddrSpace: NVPTXAS::ADDRESS_SPACE_SHARED)); |
| 5260 | NewCall->takeName(V: CI); |
| 5261 | CI->replaceAllUsesWith(V: Res); |
| 5262 | CI->eraseFromParent(); |
| 5263 | return; |
| 5264 | } |
| 5265 | case Intrinsic::nvvm_cp_async_bulk_global_to_shared_cluster: |
| 5266 | case Intrinsic::nvvm_cp_async_bulk_shared_cta_to_cluster: { |
| 5267 | // Create a new call with the correct address space. |
| 5268 | SmallVector<Value *, 4> Args(CI->args()); |
| 5269 | Args[0] = Builder.CreateAddrSpaceCast( |
| 5270 | V: Args[0], DestTy: Builder.getPtrTy(AddrSpace: NVPTXAS::ADDRESS_SPACE_SHARED_CLUSTER)); |
| 5271 | |
| 5272 | NewCall = Builder.CreateCall(Callee: NewFn, Args); |
| 5273 | NewCall->takeName(V: CI); |
| 5274 | CI->replaceAllUsesWith(V: NewCall); |
| 5275 | CI->eraseFromParent(); |
| 5276 | return; |
| 5277 | } |
| 5278 | case Intrinsic::nvvm_cp_async_bulk_tensor_g2s_im2col_3d: |
| 5279 | case Intrinsic::nvvm_cp_async_bulk_tensor_g2s_im2col_4d: |
| 5280 | case Intrinsic::nvvm_cp_async_bulk_tensor_g2s_im2col_5d: |
| 5281 | case Intrinsic::nvvm_cp_async_bulk_tensor_g2s_tile_1d: |
| 5282 | case Intrinsic::nvvm_cp_async_bulk_tensor_g2s_tile_2d: |
| 5283 | case Intrinsic::nvvm_cp_async_bulk_tensor_g2s_tile_3d: |
| 5284 | case Intrinsic::nvvm_cp_async_bulk_tensor_g2s_tile_4d: |
| 5285 | case Intrinsic::nvvm_cp_async_bulk_tensor_g2s_tile_5d: { |
| 5286 | SmallVector<Value *, 16> Args(CI->args()); |
| 5287 | |
| 5288 | // Create AddrSpaceCast to shared_cluster if needed. |
| 5289 | // This handles case (1) in shouldUpgradeNVPTXTMAG2SIntrinsics(). |
| 5290 | unsigned AS = CI->getArgOperand(i: 0)->getType()->getPointerAddressSpace(); |
| 5291 | if (AS == NVPTXAS::ADDRESS_SPACE_SHARED) |
| 5292 | Args[0] = Builder.CreateAddrSpaceCast( |
| 5293 | V: Args[0], DestTy: Builder.getPtrTy(AddrSpace: NVPTXAS::ADDRESS_SPACE_SHARED_CLUSTER)); |
| 5294 | |
| 5295 | // Attach the flag argument for cta_group, with a |
| 5296 | // default value of 0. This handles case (2) in |
| 5297 | // shouldUpgradeNVPTXTMAG2SIntrinsics(). |
| 5298 | size_t NumArgs = CI->arg_size(); |
| 5299 | Value *FlagArg = CI->getArgOperand(i: NumArgs - 3); |
| 5300 | if (!FlagArg->getType()->isIntegerTy(Bitwidth: 1)) |
| 5301 | Args.push_back(Elt: ConstantInt::get(Ty: Builder.getInt32Ty(), V: 0)); |
| 5302 | |
| 5303 | NewCall = Builder.CreateCall(Callee: NewFn, Args); |
| 5304 | NewCall->takeName(V: CI); |
| 5305 | CI->replaceAllUsesWith(V: NewCall); |
| 5306 | CI->eraseFromParent(); |
| 5307 | return; |
| 5308 | } |
| 5309 | case Intrinsic::riscv_sha256sig0: |
| 5310 | case Intrinsic::riscv_sha256sig1: |
| 5311 | case Intrinsic::riscv_sha256sum0: |
| 5312 | case Intrinsic::riscv_sha256sum1: |
| 5313 | case Intrinsic::riscv_sm3p0: |
| 5314 | case Intrinsic::riscv_sm3p1: { |
| 5315 | // The last argument to these intrinsics used to be i8 and changed to i32. |
| 5316 | // The type overload for sm4ks and sm4ed was removed. |
| 5317 | if (!CI->getType()->isIntegerTy(Bitwidth: 64)) |
| 5318 | return; |
| 5319 | |
| 5320 | Value *Arg = |
| 5321 | Builder.CreateTrunc(V: CI->getArgOperand(i: 0), DestTy: Builder.getInt32Ty()); |
| 5322 | |
| 5323 | NewCall = Builder.CreateCall(Callee: NewFn, Args: Arg); |
| 5324 | Value *Res = |
| 5325 | Builder.CreateIntCast(V: NewCall, DestTy: CI->getType(), /*isSigned*/ true); |
| 5326 | NewCall->takeName(V: CI); |
| 5327 | CI->replaceAllUsesWith(V: Res); |
| 5328 | CI->eraseFromParent(); |
| 5329 | return; |
| 5330 | } |
| 5331 | |
| 5332 | case Intrinsic::x86_xop_vfrcz_ss: |
| 5333 | case Intrinsic::x86_xop_vfrcz_sd: |
| 5334 | NewCall = Builder.CreateCall(Callee: NewFn, Args: {CI->getArgOperand(i: 1)}); |
| 5335 | break; |
| 5336 | |
| 5337 | case Intrinsic::x86_xop_vpermil2pd: |
| 5338 | case Intrinsic::x86_xop_vpermil2ps: |
| 5339 | case Intrinsic::x86_xop_vpermil2pd_256: |
| 5340 | case Intrinsic::x86_xop_vpermil2ps_256: { |
| 5341 | SmallVector<Value *, 4> Args(CI->args()); |
| 5342 | VectorType *FltIdxTy = cast<VectorType>(Val: Args[2]->getType()); |
| 5343 | VectorType *IntIdxTy = VectorType::getInteger(VTy: FltIdxTy); |
| 5344 | Args[2] = Builder.CreateBitCast(V: Args[2], DestTy: IntIdxTy); |
| 5345 | NewCall = Builder.CreateCall(Callee: NewFn, Args); |
| 5346 | break; |
| 5347 | } |
| 5348 | |
| 5349 | case Intrinsic::x86_sse41_ptestc: |
| 5350 | case Intrinsic::x86_sse41_ptestz: |
| 5351 | case Intrinsic::x86_sse41_ptestnzc: { |
| 5352 | // The arguments for these intrinsics used to be v4f32, and changed |
| 5353 | // to v2i64. This is purely a nop, since those are bitwise intrinsics. |
| 5354 | // So, the only thing required is a bitcast for both arguments. |
| 5355 | // First, check the arguments have the old type. |
| 5356 | Value *Arg0 = CI->getArgOperand(i: 0); |
| 5357 | if (Arg0->getType() != FixedVectorType::get(ElementType: Type::getFloatTy(C), NumElts: 4)) |
| 5358 | return; |
| 5359 | |
| 5360 | // Old intrinsic, add bitcasts |
| 5361 | Value *Arg1 = CI->getArgOperand(i: 1); |
| 5362 | |
| 5363 | auto *NewVecTy = FixedVectorType::get(ElementType: Type::getInt64Ty(C), NumElts: 2); |
| 5364 | |
| 5365 | Value *BC0 = Builder.CreateBitCast(V: Arg0, DestTy: NewVecTy, Name: "cast" ); |
| 5366 | Value *BC1 = Builder.CreateBitCast(V: Arg1, DestTy: NewVecTy, Name: "cast" ); |
| 5367 | |
| 5368 | NewCall = Builder.CreateCall(Callee: NewFn, Args: {BC0, BC1}); |
| 5369 | break; |
| 5370 | } |
| 5371 | |
| 5372 | case Intrinsic::x86_rdtscp: { |
| 5373 | // This used to take 1 arguments. If we have no arguments, it is already |
| 5374 | // upgraded. |
| 5375 | if (CI->getNumOperands() == 0) |
| 5376 | return; |
| 5377 | |
| 5378 | NewCall = Builder.CreateCall(Callee: NewFn); |
| 5379 | // Extract the second result and store it. |
| 5380 | Value *Data = Builder.CreateExtractValue(Agg: NewCall, Idxs: 1); |
| 5381 | Builder.CreateAlignedStore(Val: Data, Ptr: CI->getArgOperand(i: 0), Align: Align(1)); |
| 5382 | // Replace the original call result with the first result of the new call. |
| 5383 | Value *TSC = Builder.CreateExtractValue(Agg: NewCall, Idxs: 0); |
| 5384 | |
| 5385 | NewCall->takeName(V: CI); |
| 5386 | CI->replaceAllUsesWith(V: TSC); |
| 5387 | CI->eraseFromParent(); |
| 5388 | return; |
| 5389 | } |
| 5390 | |
| 5391 | case Intrinsic::x86_sse41_insertps: |
| 5392 | case Intrinsic::x86_sse41_dppd: |
| 5393 | case Intrinsic::x86_sse41_dpps: |
| 5394 | case Intrinsic::x86_sse41_mpsadbw: |
| 5395 | case Intrinsic::x86_avx_dp_ps_256: |
| 5396 | case Intrinsic::x86_avx2_mpsadbw: { |
| 5397 | // Need to truncate the last argument from i32 to i8 -- this argument models |
| 5398 | // an inherently 8-bit immediate operand to these x86 instructions. |
| 5399 | SmallVector<Value *, 4> Args(CI->args()); |
| 5400 | |
| 5401 | // Replace the last argument with a trunc. |
| 5402 | Args.back() = Builder.CreateTrunc(V: Args.back(), DestTy: Type::getInt8Ty(C), Name: "trunc" ); |
| 5403 | NewCall = Builder.CreateCall(Callee: NewFn, Args); |
| 5404 | break; |
| 5405 | } |
| 5406 | |
| 5407 | case Intrinsic::x86_avx512_mask_cmp_pd_128: |
| 5408 | case Intrinsic::x86_avx512_mask_cmp_pd_256: |
| 5409 | case Intrinsic::x86_avx512_mask_cmp_pd_512: |
| 5410 | case Intrinsic::x86_avx512_mask_cmp_ps_128: |
| 5411 | case Intrinsic::x86_avx512_mask_cmp_ps_256: |
| 5412 | case Intrinsic::x86_avx512_mask_cmp_ps_512: { |
| 5413 | SmallVector<Value *, 4> Args(CI->args()); |
| 5414 | unsigned NumElts = |
| 5415 | cast<FixedVectorType>(Val: Args[0]->getType())->getNumElements(); |
| 5416 | Args[3] = getX86MaskVec(Builder, Mask: Args[3], NumElts); |
| 5417 | |
| 5418 | NewCall = Builder.CreateCall(Callee: NewFn, Args); |
| 5419 | Value *Res = applyX86MaskOn1BitsVec(Builder, Vec: NewCall, Mask: nullptr); |
| 5420 | |
| 5421 | NewCall->takeName(V: CI); |
| 5422 | CI->replaceAllUsesWith(V: Res); |
| 5423 | CI->eraseFromParent(); |
| 5424 | return; |
| 5425 | } |
| 5426 | |
| 5427 | case Intrinsic::x86_avx512bf16_cvtne2ps2bf16_128: |
| 5428 | case Intrinsic::x86_avx512bf16_cvtne2ps2bf16_256: |
| 5429 | case Intrinsic::x86_avx512bf16_cvtne2ps2bf16_512: |
| 5430 | case Intrinsic::x86_avx512bf16_mask_cvtneps2bf16_128: |
| 5431 | case Intrinsic::x86_avx512bf16_cvtneps2bf16_256: |
| 5432 | case Intrinsic::x86_avx512bf16_cvtneps2bf16_512: { |
| 5433 | SmallVector<Value *, 4> Args(CI->args()); |
| 5434 | unsigned NumElts = cast<FixedVectorType>(Val: CI->getType())->getNumElements(); |
| 5435 | if (NewFn->getIntrinsicID() == |
| 5436 | Intrinsic::x86_avx512bf16_mask_cvtneps2bf16_128) |
| 5437 | Args[1] = Builder.CreateBitCast( |
| 5438 | V: Args[1], DestTy: FixedVectorType::get(ElementType: Builder.getBFloatTy(), NumElts)); |
| 5439 | |
| 5440 | NewCall = Builder.CreateCall(Callee: NewFn, Args); |
| 5441 | Value *Res = Builder.CreateBitCast( |
| 5442 | V: NewCall, DestTy: FixedVectorType::get(ElementType: Builder.getInt16Ty(), NumElts)); |
| 5443 | |
| 5444 | NewCall->takeName(V: CI); |
| 5445 | CI->replaceAllUsesWith(V: Res); |
| 5446 | CI->eraseFromParent(); |
| 5447 | return; |
| 5448 | } |
| 5449 | case Intrinsic::x86_avx512bf16_dpbf16ps_128: |
| 5450 | case Intrinsic::x86_avx512bf16_dpbf16ps_256: |
| 5451 | case Intrinsic::x86_avx512bf16_dpbf16ps_512:{ |
| 5452 | SmallVector<Value *, 4> Args(CI->args()); |
| 5453 | unsigned NumElts = |
| 5454 | cast<FixedVectorType>(Val: CI->getType())->getNumElements() * 2; |
| 5455 | Args[1] = Builder.CreateBitCast( |
| 5456 | V: Args[1], DestTy: FixedVectorType::get(ElementType: Builder.getBFloatTy(), NumElts)); |
| 5457 | Args[2] = Builder.CreateBitCast( |
| 5458 | V: Args[2], DestTy: FixedVectorType::get(ElementType: Builder.getBFloatTy(), NumElts)); |
| 5459 | |
| 5460 | NewCall = Builder.CreateCall(Callee: NewFn, Args); |
| 5461 | break; |
| 5462 | } |
| 5463 | |
| 5464 | case Intrinsic::thread_pointer: { |
| 5465 | NewCall = Builder.CreateCall(Callee: NewFn, Args: {}); |
| 5466 | break; |
| 5467 | } |
| 5468 | |
| 5469 | case Intrinsic::memcpy: |
| 5470 | case Intrinsic::memmove: |
| 5471 | case Intrinsic::memset: { |
| 5472 | // We have to make sure that the call signature is what we're expecting. |
| 5473 | // We only want to change the old signatures by removing the alignment arg: |
| 5474 | // @llvm.mem[cpy|move]...(i8*, i8*, i[32|i64], i32, i1) |
| 5475 | // -> @llvm.mem[cpy|move]...(i8*, i8*, i[32|i64], i1) |
| 5476 | // @llvm.memset...(i8*, i8, i[32|64], i32, i1) |
| 5477 | // -> @llvm.memset...(i8*, i8, i[32|64], i1) |
| 5478 | // Note: i8*'s in the above can be any pointer type |
| 5479 | if (CI->arg_size() != 5) { |
| 5480 | DefaultCase(); |
| 5481 | return; |
| 5482 | } |
| 5483 | // Remove alignment argument (3), and add alignment attributes to the |
| 5484 | // dest/src pointers. |
| 5485 | Value *Args[4] = {CI->getArgOperand(i: 0), CI->getArgOperand(i: 1), |
| 5486 | CI->getArgOperand(i: 2), CI->getArgOperand(i: 4)}; |
| 5487 | NewCall = Builder.CreateCall(Callee: NewFn, Args); |
| 5488 | AttributeList OldAttrs = CI->getAttributes(); |
| 5489 | AttributeList NewAttrs = AttributeList::get( |
| 5490 | C, FnAttrs: OldAttrs.getFnAttrs(), RetAttrs: OldAttrs.getRetAttrs(), |
| 5491 | ArgAttrs: {OldAttrs.getParamAttrs(ArgNo: 0), OldAttrs.getParamAttrs(ArgNo: 1), |
| 5492 | OldAttrs.getParamAttrs(ArgNo: 2), OldAttrs.getParamAttrs(ArgNo: 4)}); |
| 5493 | NewCall->setAttributes(NewAttrs); |
| 5494 | auto *MemCI = cast<MemIntrinsic>(Val: NewCall); |
| 5495 | // All mem intrinsics support dest alignment. |
| 5496 | const ConstantInt *Align = cast<ConstantInt>(Val: CI->getArgOperand(i: 3)); |
| 5497 | MemCI->setDestAlignment(Align->getMaybeAlignValue()); |
| 5498 | // Memcpy/Memmove also support source alignment. |
| 5499 | if (auto *MTI = dyn_cast<MemTransferInst>(Val: MemCI)) |
| 5500 | MTI->setSourceAlignment(Align->getMaybeAlignValue()); |
| 5501 | break; |
| 5502 | } |
| 5503 | |
| 5504 | case Intrinsic::masked_load: |
| 5505 | case Intrinsic::masked_gather: |
| 5506 | case Intrinsic::masked_store: |
| 5507 | case Intrinsic::masked_scatter: { |
| 5508 | if (CI->arg_size() != 4) { |
| 5509 | DefaultCase(); |
| 5510 | return; |
| 5511 | } |
| 5512 | |
| 5513 | auto GetMaybeAlign = [](Value *Op) { |
| 5514 | if (auto *CI = dyn_cast<ConstantInt>(Val: Op)) { |
| 5515 | uint64_t Val = CI->getZExtValue(); |
| 5516 | if (Val == 0) |
| 5517 | return MaybeAlign(); |
| 5518 | if (isPowerOf2_64(Value: Val)) |
| 5519 | return MaybeAlign(Val); |
| 5520 | } |
| 5521 | reportFatalUsageError(reason: "Invalid alignment argument" ); |
| 5522 | }; |
| 5523 | auto GetAlign = [&](Value *Op) { |
| 5524 | MaybeAlign Align = GetMaybeAlign(Op); |
| 5525 | if (Align) |
| 5526 | return *Align; |
| 5527 | reportFatalUsageError(reason: "Invalid zero alignment argument" ); |
| 5528 | }; |
| 5529 | |
| 5530 | const DataLayout &DL = CI->getDataLayout(); |
| 5531 | switch (NewFn->getIntrinsicID()) { |
| 5532 | case Intrinsic::masked_load: |
| 5533 | NewCall = Builder.CreateMaskedLoad( |
| 5534 | Ty: CI->getType(), Ptr: CI->getArgOperand(i: 0), Alignment: GetAlign(CI->getArgOperand(i: 1)), |
| 5535 | Mask: CI->getArgOperand(i: 2), PassThru: CI->getArgOperand(i: 3)); |
| 5536 | break; |
| 5537 | case Intrinsic::masked_gather: |
| 5538 | NewCall = Builder.CreateMaskedGather( |
| 5539 | Ty: CI->getType(), Ptrs: CI->getArgOperand(i: 0), |
| 5540 | Alignment: DL.getValueOrABITypeAlignment(Alignment: GetMaybeAlign(CI->getArgOperand(i: 1)), |
| 5541 | Ty: CI->getType()->getScalarType()), |
| 5542 | Mask: CI->getArgOperand(i: 2), PassThru: CI->getArgOperand(i: 3)); |
| 5543 | break; |
| 5544 | case Intrinsic::masked_store: |
| 5545 | NewCall = Builder.CreateMaskedStore( |
| 5546 | Val: CI->getArgOperand(i: 0), Ptr: CI->getArgOperand(i: 1), |
| 5547 | Alignment: GetAlign(CI->getArgOperand(i: 2)), Mask: CI->getArgOperand(i: 3)); |
| 5548 | break; |
| 5549 | case Intrinsic::masked_scatter: |
| 5550 | NewCall = Builder.CreateMaskedScatter( |
| 5551 | Val: CI->getArgOperand(i: 0), Ptrs: CI->getArgOperand(i: 1), |
| 5552 | Alignment: DL.getValueOrABITypeAlignment( |
| 5553 | Alignment: GetMaybeAlign(CI->getArgOperand(i: 2)), |
| 5554 | Ty: CI->getArgOperand(i: 0)->getType()->getScalarType()), |
| 5555 | Mask: CI->getArgOperand(i: 3)); |
| 5556 | break; |
| 5557 | default: |
| 5558 | llvm_unreachable("Unexpected intrinsic ID" ); |
| 5559 | } |
| 5560 | // Previous metadata is still valid. |
| 5561 | NewCall->copyMetadata(SrcInst: *CI); |
| 5562 | NewCall->setTailCallKind(cast<CallInst>(Val: CI)->getTailCallKind()); |
| 5563 | break; |
| 5564 | } |
| 5565 | |
| 5566 | case Intrinsic::lifetime_start: |
| 5567 | case Intrinsic::lifetime_end: { |
| 5568 | if (CI->arg_size() != 2) { |
| 5569 | DefaultCase(); |
| 5570 | return; |
| 5571 | } |
| 5572 | |
| 5573 | Value *Ptr = CI->getArgOperand(i: 1); |
| 5574 | // Try to strip pointer casts, such that the lifetime works on an alloca. |
| 5575 | Ptr = Ptr->stripPointerCasts(); |
| 5576 | if (isa<AllocaInst>(Val: Ptr)) { |
| 5577 | // Don't use NewFn, as we might have looked through an addrspacecast. |
| 5578 | if (NewFn->getIntrinsicID() == Intrinsic::lifetime_start) |
| 5579 | NewCall = Builder.CreateLifetimeStart(Ptr); |
| 5580 | else |
| 5581 | NewCall = Builder.CreateLifetimeEnd(Ptr); |
| 5582 | break; |
| 5583 | } |
| 5584 | |
| 5585 | // Otherwise remove the lifetime marker. |
| 5586 | CI->eraseFromParent(); |
| 5587 | return; |
| 5588 | } |
| 5589 | |
| 5590 | case Intrinsic::x86_avx512_vpdpbusd_128: |
| 5591 | case Intrinsic::x86_avx512_vpdpbusd_256: |
| 5592 | case Intrinsic::x86_avx512_vpdpbusd_512: |
| 5593 | case Intrinsic::x86_avx512_vpdpbusds_128: |
| 5594 | case Intrinsic::x86_avx512_vpdpbusds_256: |
| 5595 | case Intrinsic::x86_avx512_vpdpbusds_512: |
| 5596 | case Intrinsic::x86_avx2_vpdpbssd_128: |
| 5597 | case Intrinsic::x86_avx2_vpdpbssd_256: |
| 5598 | case Intrinsic::x86_avx10_vpdpbssd_512: |
| 5599 | case Intrinsic::x86_avx2_vpdpbssds_128: |
| 5600 | case Intrinsic::x86_avx2_vpdpbssds_256: |
| 5601 | case Intrinsic::x86_avx10_vpdpbssds_512: |
| 5602 | case Intrinsic::x86_avx2_vpdpbsud_128: |
| 5603 | case Intrinsic::x86_avx2_vpdpbsud_256: |
| 5604 | case Intrinsic::x86_avx10_vpdpbsud_512: |
| 5605 | case Intrinsic::x86_avx2_vpdpbsuds_128: |
| 5606 | case Intrinsic::x86_avx2_vpdpbsuds_256: |
| 5607 | case Intrinsic::x86_avx10_vpdpbsuds_512: |
| 5608 | case Intrinsic::x86_avx2_vpdpbuud_128: |
| 5609 | case Intrinsic::x86_avx2_vpdpbuud_256: |
| 5610 | case Intrinsic::x86_avx10_vpdpbuud_512: |
| 5611 | case Intrinsic::x86_avx2_vpdpbuuds_128: |
| 5612 | case Intrinsic::x86_avx2_vpdpbuuds_256: |
| 5613 | case Intrinsic::x86_avx10_vpdpbuuds_512: { |
| 5614 | unsigned NumElts = CI->getType()->getPrimitiveSizeInBits() / 8; |
| 5615 | Value *Args[] = {CI->getArgOperand(i: 0), CI->getArgOperand(i: 1), |
| 5616 | CI->getArgOperand(i: 2)}; |
| 5617 | Type *NewArgType = VectorType::get(ElementType: Builder.getInt8Ty(), NumElements: NumElts, Scalable: false); |
| 5618 | Args[1] = Builder.CreateBitCast(V: Args[1], DestTy: NewArgType); |
| 5619 | Args[2] = Builder.CreateBitCast(V: Args[2], DestTy: NewArgType); |
| 5620 | |
| 5621 | NewCall = Builder.CreateCall(Callee: NewFn, Args); |
| 5622 | break; |
| 5623 | } |
| 5624 | case Intrinsic::x86_avx512_vpdpwssd_128: |
| 5625 | case Intrinsic::x86_avx512_vpdpwssd_256: |
| 5626 | case Intrinsic::x86_avx512_vpdpwssd_512: |
| 5627 | case Intrinsic::x86_avx512_vpdpwssds_128: |
| 5628 | case Intrinsic::x86_avx512_vpdpwssds_256: |
| 5629 | case Intrinsic::x86_avx512_vpdpwssds_512: |
| 5630 | case Intrinsic::x86_avx2_vpdpwsud_128: |
| 5631 | case Intrinsic::x86_avx2_vpdpwsud_256: |
| 5632 | case Intrinsic::x86_avx10_vpdpwsud_512: |
| 5633 | case Intrinsic::x86_avx2_vpdpwsuds_128: |
| 5634 | case Intrinsic::x86_avx2_vpdpwsuds_256: |
| 5635 | case Intrinsic::x86_avx10_vpdpwsuds_512: |
| 5636 | case Intrinsic::x86_avx2_vpdpwusd_128: |
| 5637 | case Intrinsic::x86_avx2_vpdpwusd_256: |
| 5638 | case Intrinsic::x86_avx10_vpdpwusd_512: |
| 5639 | case Intrinsic::x86_avx2_vpdpwusds_128: |
| 5640 | case Intrinsic::x86_avx2_vpdpwusds_256: |
| 5641 | case Intrinsic::x86_avx10_vpdpwusds_512: |
| 5642 | case Intrinsic::x86_avx2_vpdpwuud_128: |
| 5643 | case Intrinsic::x86_avx2_vpdpwuud_256: |
| 5644 | case Intrinsic::x86_avx10_vpdpwuud_512: |
| 5645 | case Intrinsic::x86_avx2_vpdpwuuds_128: |
| 5646 | case Intrinsic::x86_avx2_vpdpwuuds_256: |
| 5647 | case Intrinsic::x86_avx10_vpdpwuuds_512: |
| 5648 | unsigned NumElts = CI->getType()->getPrimitiveSizeInBits() / 16; |
| 5649 | Value *Args[] = {CI->getArgOperand(i: 0), CI->getArgOperand(i: 1), |
| 5650 | CI->getArgOperand(i: 2)}; |
| 5651 | Type *NewArgType = VectorType::get(ElementType: Builder.getInt16Ty(), NumElements: NumElts, Scalable: false); |
| 5652 | Args[1] = Builder.CreateBitCast(V: Args[1], DestTy: NewArgType); |
| 5653 | Args[2] = Builder.CreateBitCast(V: Args[2], DestTy: NewArgType); |
| 5654 | |
| 5655 | NewCall = Builder.CreateCall(Callee: NewFn, Args); |
| 5656 | break; |
| 5657 | } |
| 5658 | assert(NewCall && "Should have either set this variable or returned through " |
| 5659 | "the default case" ); |
| 5660 | NewCall->takeName(V: CI); |
| 5661 | CI->replaceAllUsesWith(V: NewCall); |
| 5662 | CI->eraseFromParent(); |
| 5663 | } |
| 5664 | |
| 5665 | void llvm::UpgradeCallsToIntrinsic(Function *F) { |
| 5666 | assert(F && "Illegal attempt to upgrade a non-existent intrinsic." ); |
| 5667 | |
| 5668 | // Check if this function should be upgraded and get the replacement function |
| 5669 | // if there is one. |
| 5670 | Function *NewFn; |
| 5671 | if (UpgradeIntrinsicFunction(F, NewFn)) { |
| 5672 | // Replace all users of the old function with the new function or new |
| 5673 | // instructions. This is not a range loop because the call is deleted. |
| 5674 | for (User *U : make_early_inc_range(Range: F->users())) |
| 5675 | if (CallBase *CB = dyn_cast<CallBase>(Val: U)) |
| 5676 | UpgradeIntrinsicCall(CI: CB, NewFn); |
| 5677 | |
| 5678 | // Remove old function, no longer used, from the module. |
| 5679 | if (F != NewFn) |
| 5680 | F->eraseFromParent(); |
| 5681 | } |
| 5682 | } |
| 5683 | |
| 5684 | MDNode *llvm::UpgradeTBAANode(MDNode &MD) { |
| 5685 | const unsigned NumOperands = MD.getNumOperands(); |
| 5686 | if (NumOperands == 0) |
| 5687 | return &MD; // Invalid, punt to a verifier error. |
| 5688 | |
| 5689 | // Check if the tag uses struct-path aware TBAA format. |
| 5690 | if (isa<MDNode>(Val: MD.getOperand(I: 0)) && NumOperands >= 3) |
| 5691 | return &MD; |
| 5692 | |
| 5693 | auto &Context = MD.getContext(); |
| 5694 | if (NumOperands == 3) { |
| 5695 | Metadata *Elts[] = {MD.getOperand(I: 0), MD.getOperand(I: 1)}; |
| 5696 | MDNode *ScalarType = MDNode::get(Context, MDs: Elts); |
| 5697 | // Create a MDNode <ScalarType, ScalarType, offset 0, const> |
| 5698 | Metadata *Elts2[] = {ScalarType, ScalarType, |
| 5699 | ConstantAsMetadata::get( |
| 5700 | C: Constant::getNullValue(Ty: Type::getInt64Ty(C&: Context))), |
| 5701 | MD.getOperand(I: 2)}; |
| 5702 | return MDNode::get(Context, MDs: Elts2); |
| 5703 | } |
| 5704 | // Create a MDNode <MD, MD, offset 0> |
| 5705 | Metadata *Elts[] = {&MD, &MD, ConstantAsMetadata::get(C: Constant::getNullValue( |
| 5706 | Ty: Type::getInt64Ty(C&: Context)))}; |
| 5707 | return MDNode::get(Context, MDs: Elts); |
| 5708 | } |
| 5709 | |
| 5710 | Instruction *llvm::UpgradeBitCastInst(unsigned Opc, Value *V, Type *DestTy, |
| 5711 | Instruction *&Temp) { |
| 5712 | if (Opc != Instruction::BitCast) |
| 5713 | return nullptr; |
| 5714 | |
| 5715 | Temp = nullptr; |
| 5716 | Type *SrcTy = V->getType(); |
| 5717 | if (SrcTy->isPtrOrPtrVectorTy() && DestTy->isPtrOrPtrVectorTy() && |
| 5718 | SrcTy->getPointerAddressSpace() != DestTy->getPointerAddressSpace()) { |
| 5719 | LLVMContext &Context = V->getContext(); |
| 5720 | |
| 5721 | // We have no information about target data layout, so we assume that |
| 5722 | // the maximum pointer size is 64bit. |
| 5723 | Type *MidTy = Type::getInt64Ty(C&: Context); |
| 5724 | Temp = CastInst::Create(Instruction::PtrToInt, S: V, Ty: MidTy); |
| 5725 | |
| 5726 | return CastInst::Create(Instruction::IntToPtr, S: Temp, Ty: DestTy); |
| 5727 | } |
| 5728 | |
| 5729 | return nullptr; |
| 5730 | } |
| 5731 | |
| 5732 | Constant *llvm::UpgradeBitCastExpr(unsigned Opc, Constant *C, Type *DestTy) { |
| 5733 | if (Opc != Instruction::BitCast) |
| 5734 | return nullptr; |
| 5735 | |
| 5736 | Type *SrcTy = C->getType(); |
| 5737 | if (SrcTy->isPtrOrPtrVectorTy() && DestTy->isPtrOrPtrVectorTy() && |
| 5738 | SrcTy->getPointerAddressSpace() != DestTy->getPointerAddressSpace()) { |
| 5739 | LLVMContext &Context = C->getContext(); |
| 5740 | |
| 5741 | // We have no information about target data layout, so we assume that |
| 5742 | // the maximum pointer size is 64bit. |
| 5743 | Type *MidTy = Type::getInt64Ty(C&: Context); |
| 5744 | |
| 5745 | return ConstantExpr::getIntToPtr(C: ConstantExpr::getPtrToInt(C, Ty: MidTy), |
| 5746 | Ty: DestTy); |
| 5747 | } |
| 5748 | |
| 5749 | return nullptr; |
| 5750 | } |
| 5751 | |
| 5752 | /// Check the debug info version number, if it is out-dated, drop the debug |
| 5753 | /// info. Return true if module is modified. |
| 5754 | bool llvm::UpgradeDebugInfo(Module &M) { |
| 5755 | if (DisableAutoUpgradeDebugInfo) |
| 5756 | return false; |
| 5757 | |
| 5758 | llvm::TimeTraceScope timeScope("Upgrade debug info" ); |
| 5759 | // We need to get metadata before the module is verified (i.e., getModuleFlag |
| 5760 | // makes assumptions that we haven't verified yet). Carefully extract the flag |
| 5761 | // from the metadata. |
| 5762 | unsigned Version = 0; |
| 5763 | if (NamedMDNode *ModFlags = M.getModuleFlagsMetadata()) { |
| 5764 | auto OpIt = find_if(Range: ModFlags->operands(), P: [](const MDNode *Flag) { |
| 5765 | if (Flag->getNumOperands() < 3) |
| 5766 | return false; |
| 5767 | if (MDString *K = dyn_cast_or_null<MDString>(Val: Flag->getOperand(I: 1))) |
| 5768 | return K->getString() == "Debug Info Version" ; |
| 5769 | return false; |
| 5770 | }); |
| 5771 | if (OpIt != ModFlags->op_end()) { |
| 5772 | const MDOperand &ValOp = (*OpIt)->getOperand(I: 2); |
| 5773 | if (auto *CI = mdconst::dyn_extract_or_null<ConstantInt>(MD: ValOp)) |
| 5774 | Version = CI->getZExtValue(); |
| 5775 | } |
| 5776 | } |
| 5777 | |
| 5778 | if (Version == DEBUG_METADATA_VERSION) { |
| 5779 | bool BrokenDebugInfo = false; |
| 5780 | if (verifyModule(M, OS: &llvm::errs(), BrokenDebugInfo: &BrokenDebugInfo)) |
| 5781 | report_fatal_error(reason: "Broken module found, compilation aborted!" ); |
| 5782 | if (!BrokenDebugInfo) |
| 5783 | // Everything is ok. |
| 5784 | return false; |
| 5785 | else { |
| 5786 | // Diagnose malformed debug info. |
| 5787 | DiagnosticInfoIgnoringInvalidDebugMetadata Diag(M); |
| 5788 | M.getContext().diagnose(DI: Diag); |
| 5789 | } |
| 5790 | } |
| 5791 | bool Modified = StripDebugInfo(M); |
| 5792 | if (Modified && Version != DEBUG_METADATA_VERSION) { |
| 5793 | // Diagnose a version mismatch. |
| 5794 | DiagnosticInfoDebugMetadataVersion DiagVersion(M, Version); |
| 5795 | M.getContext().diagnose(DI: DiagVersion); |
| 5796 | } |
| 5797 | return Modified; |
| 5798 | } |
| 5799 | |
| 5800 | static void upgradeNVVMFnVectorAttr(const StringRef Attr, const char DimC, |
| 5801 | GlobalValue *GV, const Metadata *V) { |
| 5802 | Function *F = cast<Function>(Val: GV); |
| 5803 | |
| 5804 | constexpr StringLiteral DefaultValue = "1" ; |
| 5805 | StringRef Vect3[3] = {DefaultValue, DefaultValue, DefaultValue}; |
| 5806 | unsigned Length = 0; |
| 5807 | |
| 5808 | if (F->hasFnAttribute(Kind: Attr)) { |
| 5809 | // We expect the existing attribute to have the form "x[,y[,z]]". Here we |
| 5810 | // parse these elements placing them into Vect3 |
| 5811 | StringRef S = F->getFnAttribute(Kind: Attr).getValueAsString(); |
| 5812 | for (; Length < 3 && !S.empty(); Length++) { |
| 5813 | auto [Part, Rest] = S.split(Separator: ','); |
| 5814 | Vect3[Length] = Part.trim(); |
| 5815 | S = Rest; |
| 5816 | } |
| 5817 | } |
| 5818 | |
| 5819 | const unsigned Dim = DimC - 'x'; |
| 5820 | assert(Dim < 3 && "Unexpected dim char" ); |
| 5821 | |
| 5822 | const uint64_t VInt = mdconst::extract<ConstantInt>(MD&: V)->getZExtValue(); |
| 5823 | |
| 5824 | // local variable required for StringRef in Vect3 to point to. |
| 5825 | const std::string VStr = llvm::utostr(X: VInt); |
| 5826 | Vect3[Dim] = VStr; |
| 5827 | Length = std::max(a: Length, b: Dim + 1); |
| 5828 | |
| 5829 | const std::string NewAttr = llvm::join(R: ArrayRef(Vect3, Length), Separator: "," ); |
| 5830 | F->addFnAttr(Kind: Attr, Val: NewAttr); |
| 5831 | } |
| 5832 | |
| 5833 | static inline bool isXYZ(StringRef S) { |
| 5834 | return S == "x" || S == "y" || S == "z" ; |
| 5835 | } |
| 5836 | |
| 5837 | bool static upgradeSingleNVVMAnnotation(GlobalValue *GV, StringRef K, |
| 5838 | const Metadata *V) { |
| 5839 | if (K == "kernel" ) { |
| 5840 | if (!mdconst::extract<ConstantInt>(MD&: V)->isZero()) |
| 5841 | cast<Function>(Val: GV)->setCallingConv(CallingConv::PTX_Kernel); |
| 5842 | return true; |
| 5843 | } |
| 5844 | if (K == "align" ) { |
| 5845 | // V is a bitfeild specifying two 16-bit values. The alignment value is |
| 5846 | // specfied in low 16-bits, The index is specified in the high bits. For the |
| 5847 | // index, 0 indicates the return value while higher values correspond to |
| 5848 | // each parameter (idx = param + 1). |
| 5849 | const uint64_t AlignIdxValuePair = |
| 5850 | mdconst::extract<ConstantInt>(MD&: V)->getZExtValue(); |
| 5851 | const unsigned Idx = (AlignIdxValuePair >> 16); |
| 5852 | const Align StackAlign = Align(AlignIdxValuePair & 0xFFFF); |
| 5853 | cast<Function>(Val: GV)->addAttributeAtIndex( |
| 5854 | i: Idx, Attr: Attribute::getWithStackAlignment(Context&: GV->getContext(), Alignment: StackAlign)); |
| 5855 | return true; |
| 5856 | } |
| 5857 | if (K == "maxclusterrank" || K == "cluster_max_blocks" ) { |
| 5858 | const auto CV = mdconst::extract<ConstantInt>(MD&: V)->getZExtValue(); |
| 5859 | cast<Function>(Val: GV)->addFnAttr(Kind: "nvvm.maxclusterrank" , Val: llvm::utostr(X: CV)); |
| 5860 | return true; |
| 5861 | } |
| 5862 | if (K == "minctasm" ) { |
| 5863 | const auto CV = mdconst::extract<ConstantInt>(MD&: V)->getZExtValue(); |
| 5864 | cast<Function>(Val: GV)->addFnAttr(Kind: "nvvm.minctasm" , Val: llvm::utostr(X: CV)); |
| 5865 | return true; |
| 5866 | } |
| 5867 | if (K == "maxnreg" ) { |
| 5868 | const auto CV = mdconst::extract<ConstantInt>(MD&: V)->getZExtValue(); |
| 5869 | cast<Function>(Val: GV)->addFnAttr(Kind: "nvvm.maxnreg" , Val: llvm::utostr(X: CV)); |
| 5870 | return true; |
| 5871 | } |
| 5872 | if (K.consume_front(Prefix: "maxntid" ) && isXYZ(S: K)) { |
| 5873 | upgradeNVVMFnVectorAttr(Attr: "nvvm.maxntid" , DimC: K[0], GV, V); |
| 5874 | return true; |
| 5875 | } |
| 5876 | if (K.consume_front(Prefix: "reqntid" ) && isXYZ(S: K)) { |
| 5877 | upgradeNVVMFnVectorAttr(Attr: "nvvm.reqntid" , DimC: K[0], GV, V); |
| 5878 | return true; |
| 5879 | } |
| 5880 | if (K.consume_front(Prefix: "cluster_dim_" ) && isXYZ(S: K)) { |
| 5881 | upgradeNVVMFnVectorAttr(Attr: "nvvm.cluster_dim" , DimC: K[0], GV, V); |
| 5882 | return true; |
| 5883 | } |
| 5884 | if (K == "grid_constant" ) { |
| 5885 | const auto Attr = Attribute::get(Context&: GV->getContext(), Kind: "nvvm.grid_constant" ); |
| 5886 | for (const auto &Op : cast<MDNode>(Val: V)->operands()) { |
| 5887 | // For some reason, the index is 1-based in the metadata. Good thing we're |
| 5888 | // able to auto-upgrade it! |
| 5889 | const auto Index = mdconst::extract<ConstantInt>(MD: Op)->getZExtValue() - 1; |
| 5890 | cast<Function>(Val: GV)->addParamAttr(ArgNo: Index, Attr); |
| 5891 | } |
| 5892 | return true; |
| 5893 | } |
| 5894 | |
| 5895 | return false; |
| 5896 | } |
| 5897 | |
| 5898 | void llvm::UpgradeNVVMAnnotations(Module &M) { |
| 5899 | NamedMDNode *NamedMD = M.getNamedMetadata(Name: "nvvm.annotations" ); |
| 5900 | if (!NamedMD) |
| 5901 | return; |
| 5902 | |
| 5903 | SmallVector<MDNode *, 8> NewNodes; |
| 5904 | SmallPtrSet<const MDNode *, 8> SeenNodes; |
| 5905 | for (MDNode *MD : NamedMD->operands()) { |
| 5906 | if (!SeenNodes.insert(Ptr: MD).second) |
| 5907 | continue; |
| 5908 | |
| 5909 | auto *GV = mdconst::dyn_extract_or_null<GlobalValue>(MD: MD->getOperand(I: 0)); |
| 5910 | if (!GV) |
| 5911 | continue; |
| 5912 | |
| 5913 | assert((MD->getNumOperands() % 2) == 1 && "Invalid number of operands" ); |
| 5914 | |
| 5915 | SmallVector<Metadata *, 8> NewOperands{MD->getOperand(I: 0)}; |
| 5916 | // Each nvvm.annotations metadata entry will be of the following form: |
| 5917 | // !{ ptr @gv, !"key1", value1, !"key2", value2, ... } |
| 5918 | // start index = 1, to skip the global variable key |
| 5919 | // increment = 2, to skip the value for each property-value pairs |
| 5920 | for (unsigned j = 1, je = MD->getNumOperands(); j < je; j += 2) { |
| 5921 | MDString *K = cast<MDString>(Val: MD->getOperand(I: j)); |
| 5922 | const MDOperand &V = MD->getOperand(I: j + 1); |
| 5923 | bool Upgraded = upgradeSingleNVVMAnnotation(GV, K: K->getString(), V); |
| 5924 | if (!Upgraded) |
| 5925 | NewOperands.append(IL: {K, V}); |
| 5926 | } |
| 5927 | |
| 5928 | if (NewOperands.size() > 1) |
| 5929 | NewNodes.push_back(Elt: MDNode::get(Context&: M.getContext(), MDs: NewOperands)); |
| 5930 | } |
| 5931 | |
| 5932 | NamedMD->clearOperands(); |
| 5933 | for (MDNode *N : NewNodes) |
| 5934 | NamedMD->addOperand(M: N); |
| 5935 | } |
| 5936 | |
| 5937 | /// This checks for objc retain release marker which should be upgraded. It |
| 5938 | /// returns true if module is modified. |
| 5939 | static bool upgradeRetainReleaseMarker(Module &M) { |
| 5940 | bool Changed = false; |
| 5941 | const char *MarkerKey = "clang.arc.retainAutoreleasedReturnValueMarker" ; |
| 5942 | NamedMDNode *ModRetainReleaseMarker = M.getNamedMetadata(Name: MarkerKey); |
| 5943 | if (ModRetainReleaseMarker) { |
| 5944 | MDNode *Op = ModRetainReleaseMarker->getOperand(i: 0); |
| 5945 | if (Op) { |
| 5946 | MDString *ID = dyn_cast_or_null<MDString>(Val: Op->getOperand(I: 0)); |
| 5947 | if (ID) { |
| 5948 | SmallVector<StringRef, 4> ValueComp; |
| 5949 | ID->getString().split(A&: ValueComp, Separator: "#" ); |
| 5950 | if (ValueComp.size() == 2) { |
| 5951 | std::string NewValue = ValueComp[0].str() + ";" + ValueComp[1].str(); |
| 5952 | ID = MDString::get(Context&: M.getContext(), Str: NewValue); |
| 5953 | } |
| 5954 | M.addModuleFlag(Behavior: Module::Error, Key: MarkerKey, Val: ID); |
| 5955 | M.eraseNamedMetadata(NMD: ModRetainReleaseMarker); |
| 5956 | Changed = true; |
| 5957 | } |
| 5958 | } |
| 5959 | } |
| 5960 | return Changed; |
| 5961 | } |
| 5962 | |
| 5963 | void llvm::UpgradeARCRuntime(Module &M) { |
| 5964 | // This lambda converts normal function calls to ARC runtime functions to |
| 5965 | // intrinsic calls. |
| 5966 | auto UpgradeToIntrinsic = [&](const char *OldFunc, |
| 5967 | llvm::Intrinsic::ID IntrinsicFunc) { |
| 5968 | Function *Fn = M.getFunction(Name: OldFunc); |
| 5969 | |
| 5970 | if (!Fn) |
| 5971 | return; |
| 5972 | |
| 5973 | Function *NewFn = |
| 5974 | llvm::Intrinsic::getOrInsertDeclaration(M: &M, id: IntrinsicFunc); |
| 5975 | |
| 5976 | for (User *U : make_early_inc_range(Range: Fn->users())) { |
| 5977 | CallInst *CI = dyn_cast<CallInst>(Val: U); |
| 5978 | if (!CI || CI->getCalledFunction() != Fn) |
| 5979 | continue; |
| 5980 | |
| 5981 | IRBuilder<> Builder(CI->getParent(), CI->getIterator()); |
| 5982 | FunctionType *NewFuncTy = NewFn->getFunctionType(); |
| 5983 | SmallVector<Value *, 2> Args; |
| 5984 | |
| 5985 | // Don't upgrade the intrinsic if it's not valid to bitcast the return |
| 5986 | // value to the return type of the old function. |
| 5987 | if (NewFuncTy->getReturnType() != CI->getType() && |
| 5988 | !CastInst::castIsValid(op: Instruction::BitCast, S: CI, |
| 5989 | DstTy: NewFuncTy->getReturnType())) |
| 5990 | continue; |
| 5991 | |
| 5992 | bool InvalidCast = false; |
| 5993 | |
| 5994 | for (unsigned I = 0, E = CI->arg_size(); I != E; ++I) { |
| 5995 | Value *Arg = CI->getArgOperand(i: I); |
| 5996 | |
| 5997 | // Bitcast argument to the parameter type of the new function if it's |
| 5998 | // not a variadic argument. |
| 5999 | if (I < NewFuncTy->getNumParams()) { |
| 6000 | // Don't upgrade the intrinsic if it's not valid to bitcast the argument |
| 6001 | // to the parameter type of the new function. |
| 6002 | if (!CastInst::castIsValid(op: Instruction::BitCast, S: Arg, |
| 6003 | DstTy: NewFuncTy->getParamType(i: I))) { |
| 6004 | InvalidCast = true; |
| 6005 | break; |
| 6006 | } |
| 6007 | Arg = Builder.CreateBitCast(V: Arg, DestTy: NewFuncTy->getParamType(i: I)); |
| 6008 | } |
| 6009 | Args.push_back(Elt: Arg); |
| 6010 | } |
| 6011 | |
| 6012 | if (InvalidCast) |
| 6013 | continue; |
| 6014 | |
| 6015 | // Create a call instruction that calls the new function. |
| 6016 | CallInst *NewCall = Builder.CreateCall(FTy: NewFuncTy, Callee: NewFn, Args); |
| 6017 | NewCall->setTailCallKind(cast<CallInst>(Val: CI)->getTailCallKind()); |
| 6018 | NewCall->takeName(V: CI); |
| 6019 | |
| 6020 | // Bitcast the return value back to the type of the old call. |
| 6021 | Value *NewRetVal = Builder.CreateBitCast(V: NewCall, DestTy: CI->getType()); |
| 6022 | |
| 6023 | if (!CI->use_empty()) |
| 6024 | CI->replaceAllUsesWith(V: NewRetVal); |
| 6025 | CI->eraseFromParent(); |
| 6026 | } |
| 6027 | |
| 6028 | if (Fn->use_empty()) |
| 6029 | Fn->eraseFromParent(); |
| 6030 | }; |
| 6031 | |
| 6032 | // Unconditionally convert a call to "clang.arc.use" to a call to |
| 6033 | // "llvm.objc.clang.arc.use". |
| 6034 | UpgradeToIntrinsic("clang.arc.use" , llvm::Intrinsic::objc_clang_arc_use); |
| 6035 | |
| 6036 | // Upgrade the retain release marker. If there is no need to upgrade |
| 6037 | // the marker, that means either the module is already new enough to contain |
| 6038 | // new intrinsics or it is not ARC. There is no need to upgrade runtime call. |
| 6039 | if (!upgradeRetainReleaseMarker(M)) |
| 6040 | return; |
| 6041 | |
| 6042 | std::pair<const char *, llvm::Intrinsic::ID> RuntimeFuncs[] = { |
| 6043 | {"objc_autorelease" , llvm::Intrinsic::objc_autorelease}, |
| 6044 | {"objc_autoreleasePoolPop" , llvm::Intrinsic::objc_autoreleasePoolPop}, |
| 6045 | {"objc_autoreleasePoolPush" , llvm::Intrinsic::objc_autoreleasePoolPush}, |
| 6046 | {"objc_autoreleaseReturnValue" , |
| 6047 | llvm::Intrinsic::objc_autoreleaseReturnValue}, |
| 6048 | {"objc_copyWeak" , llvm::Intrinsic::objc_copyWeak}, |
| 6049 | {"objc_destroyWeak" , llvm::Intrinsic::objc_destroyWeak}, |
| 6050 | {"objc_initWeak" , llvm::Intrinsic::objc_initWeak}, |
| 6051 | {"objc_loadWeak" , llvm::Intrinsic::objc_loadWeak}, |
| 6052 | {"objc_loadWeakRetained" , llvm::Intrinsic::objc_loadWeakRetained}, |
| 6053 | {"objc_moveWeak" , llvm::Intrinsic::objc_moveWeak}, |
| 6054 | {"objc_release" , llvm::Intrinsic::objc_release}, |
| 6055 | {"objc_retain" , llvm::Intrinsic::objc_retain}, |
| 6056 | {"objc_retainAutorelease" , llvm::Intrinsic::objc_retainAutorelease}, |
| 6057 | {"objc_retainAutoreleaseReturnValue" , |
| 6058 | llvm::Intrinsic::objc_retainAutoreleaseReturnValue}, |
| 6059 | {"objc_retainAutoreleasedReturnValue" , |
| 6060 | llvm::Intrinsic::objc_retainAutoreleasedReturnValue}, |
| 6061 | {"objc_retainBlock" , llvm::Intrinsic::objc_retainBlock}, |
| 6062 | {"objc_storeStrong" , llvm::Intrinsic::objc_storeStrong}, |
| 6063 | {"objc_storeWeak" , llvm::Intrinsic::objc_storeWeak}, |
| 6064 | {"objc_unsafeClaimAutoreleasedReturnValue" , |
| 6065 | llvm::Intrinsic::objc_unsafeClaimAutoreleasedReturnValue}, |
| 6066 | {"objc_retainedObject" , llvm::Intrinsic::objc_retainedObject}, |
| 6067 | {"objc_unretainedObject" , llvm::Intrinsic::objc_unretainedObject}, |
| 6068 | {"objc_unretainedPointer" , llvm::Intrinsic::objc_unretainedPointer}, |
| 6069 | {"objc_retain_autorelease" , llvm::Intrinsic::objc_retain_autorelease}, |
| 6070 | {"objc_sync_enter" , llvm::Intrinsic::objc_sync_enter}, |
| 6071 | {"objc_sync_exit" , llvm::Intrinsic::objc_sync_exit}, |
| 6072 | {"objc_arc_annotation_topdown_bbstart" , |
| 6073 | llvm::Intrinsic::objc_arc_annotation_topdown_bbstart}, |
| 6074 | {"objc_arc_annotation_topdown_bbend" , |
| 6075 | llvm::Intrinsic::objc_arc_annotation_topdown_bbend}, |
| 6076 | {"objc_arc_annotation_bottomup_bbstart" , |
| 6077 | llvm::Intrinsic::objc_arc_annotation_bottomup_bbstart}, |
| 6078 | {"objc_arc_annotation_bottomup_bbend" , |
| 6079 | llvm::Intrinsic::objc_arc_annotation_bottomup_bbend}}; |
| 6080 | |
| 6081 | for (auto &I : RuntimeFuncs) |
| 6082 | UpgradeToIntrinsic(I.first, I.second); |
| 6083 | } |
| 6084 | |
| 6085 | bool llvm::UpgradeModuleFlags(Module &M) { |
| 6086 | NamedMDNode *ModFlags = M.getModuleFlagsMetadata(); |
| 6087 | if (!ModFlags) |
| 6088 | return false; |
| 6089 | |
| 6090 | bool HasObjCFlag = false, HasClassProperties = false, Changed = false; |
| 6091 | bool HasSwiftVersionFlag = false; |
| 6092 | uint8_t SwiftMajorVersion, SwiftMinorVersion; |
| 6093 | uint32_t SwiftABIVersion; |
| 6094 | auto Int8Ty = Type::getInt8Ty(C&: M.getContext()); |
| 6095 | auto Int32Ty = Type::getInt32Ty(C&: M.getContext()); |
| 6096 | |
| 6097 | for (unsigned I = 0, E = ModFlags->getNumOperands(); I != E; ++I) { |
| 6098 | MDNode *Op = ModFlags->getOperand(i: I); |
| 6099 | if (Op->getNumOperands() != 3) |
| 6100 | continue; |
| 6101 | MDString *ID = dyn_cast_or_null<MDString>(Val: Op->getOperand(I: 1)); |
| 6102 | if (!ID) |
| 6103 | continue; |
| 6104 | auto SetBehavior = [&](Module::ModFlagBehavior B) { |
| 6105 | Metadata *Ops[3] = {ConstantAsMetadata::get(C: ConstantInt::get( |
| 6106 | Ty: Type::getInt32Ty(C&: M.getContext()), V: B)), |
| 6107 | MDString::get(Context&: M.getContext(), Str: ID->getString()), |
| 6108 | Op->getOperand(I: 2)}; |
| 6109 | ModFlags->setOperand(I, New: MDNode::get(Context&: M.getContext(), MDs: Ops)); |
| 6110 | Changed = true; |
| 6111 | }; |
| 6112 | |
| 6113 | if (ID->getString() == "Objective-C Image Info Version" ) |
| 6114 | HasObjCFlag = true; |
| 6115 | if (ID->getString() == "Objective-C Class Properties" ) |
| 6116 | HasClassProperties = true; |
| 6117 | // Upgrade PIC from Error/Max to Min. |
| 6118 | if (ID->getString() == "PIC Level" ) { |
| 6119 | if (auto *Behavior = |
| 6120 | mdconst::dyn_extract_or_null<ConstantInt>(MD: Op->getOperand(I: 0))) { |
| 6121 | uint64_t V = Behavior->getLimitedValue(); |
| 6122 | if (V == Module::Error || V == Module::Max) |
| 6123 | SetBehavior(Module::Min); |
| 6124 | } |
| 6125 | } |
| 6126 | // Upgrade "PIE Level" from Error to Max. |
| 6127 | if (ID->getString() == "PIE Level" ) |
| 6128 | if (auto *Behavior = |
| 6129 | mdconst::dyn_extract_or_null<ConstantInt>(MD: Op->getOperand(I: 0))) |
| 6130 | if (Behavior->getLimitedValue() == Module::Error) |
| 6131 | SetBehavior(Module::Max); |
| 6132 | |
| 6133 | // Upgrade branch protection and return address signing module flags. The |
| 6134 | // module flag behavior for these fields were Error and now they are Min. |
| 6135 | if (ID->getString() == "branch-target-enforcement" || |
| 6136 | ID->getString().starts_with(Prefix: "sign-return-address" )) { |
| 6137 | if (auto *Behavior = |
| 6138 | mdconst::dyn_extract_or_null<ConstantInt>(MD: Op->getOperand(I: 0))) { |
| 6139 | if (Behavior->getLimitedValue() == Module::Error) { |
| 6140 | Type *Int32Ty = Type::getInt32Ty(C&: M.getContext()); |
| 6141 | Metadata *Ops[3] = { |
| 6142 | ConstantAsMetadata::get(C: ConstantInt::get(Ty: Int32Ty, V: Module::Min)), |
| 6143 | Op->getOperand(I: 1), Op->getOperand(I: 2)}; |
| 6144 | ModFlags->setOperand(I, New: MDNode::get(Context&: M.getContext(), MDs: Ops)); |
| 6145 | Changed = true; |
| 6146 | } |
| 6147 | } |
| 6148 | } |
| 6149 | |
| 6150 | // Upgrade Objective-C Image Info Section. Removed the whitespce in the |
| 6151 | // section name so that llvm-lto will not complain about mismatching |
| 6152 | // module flags that is functionally the same. |
| 6153 | if (ID->getString() == "Objective-C Image Info Section" ) { |
| 6154 | if (auto *Value = dyn_cast_or_null<MDString>(Val: Op->getOperand(I: 2))) { |
| 6155 | SmallVector<StringRef, 4> ValueComp; |
| 6156 | Value->getString().split(A&: ValueComp, Separator: " " ); |
| 6157 | if (ValueComp.size() != 1) { |
| 6158 | std::string NewValue; |
| 6159 | for (auto &S : ValueComp) |
| 6160 | NewValue += S.str(); |
| 6161 | Metadata *Ops[3] = {Op->getOperand(I: 0), Op->getOperand(I: 1), |
| 6162 | MDString::get(Context&: M.getContext(), Str: NewValue)}; |
| 6163 | ModFlags->setOperand(I, New: MDNode::get(Context&: M.getContext(), MDs: Ops)); |
| 6164 | Changed = true; |
| 6165 | } |
| 6166 | } |
| 6167 | } |
| 6168 | |
| 6169 | // IRUpgrader turns a i32 type "Objective-C Garbage Collection" into i8 value. |
| 6170 | // If the higher bits are set, it adds new module flag for swift info. |
| 6171 | if (ID->getString() == "Objective-C Garbage Collection" ) { |
| 6172 | auto Md = dyn_cast<ConstantAsMetadata>(Val: Op->getOperand(I: 2)); |
| 6173 | if (Md) { |
| 6174 | assert(Md->getValue() && "Expected non-empty metadata" ); |
| 6175 | auto Type = Md->getValue()->getType(); |
| 6176 | if (Type == Int8Ty) |
| 6177 | continue; |
| 6178 | unsigned Val = Md->getValue()->getUniqueInteger().getZExtValue(); |
| 6179 | if ((Val & 0xff) != Val) { |
| 6180 | HasSwiftVersionFlag = true; |
| 6181 | SwiftABIVersion = (Val & 0xff00) >> 8; |
| 6182 | SwiftMajorVersion = (Val & 0xff000000) >> 24; |
| 6183 | SwiftMinorVersion = (Val & 0xff0000) >> 16; |
| 6184 | } |
| 6185 | Metadata *Ops[3] = { |
| 6186 | ConstantAsMetadata::get(C: ConstantInt::get(Ty: Int32Ty,V: Module::Error)), |
| 6187 | Op->getOperand(I: 1), |
| 6188 | ConstantAsMetadata::get(C: ConstantInt::get(Ty: Int8Ty,V: Val & 0xff))}; |
| 6189 | ModFlags->setOperand(I, New: MDNode::get(Context&: M.getContext(), MDs: Ops)); |
| 6190 | Changed = true; |
| 6191 | } |
| 6192 | } |
| 6193 | |
| 6194 | if (ID->getString() == "amdgpu_code_object_version" ) { |
| 6195 | Metadata *Ops[3] = { |
| 6196 | Op->getOperand(I: 0), |
| 6197 | MDString::get(Context&: M.getContext(), Str: "amdhsa_code_object_version" ), |
| 6198 | Op->getOperand(I: 2)}; |
| 6199 | ModFlags->setOperand(I, New: MDNode::get(Context&: M.getContext(), MDs: Ops)); |
| 6200 | Changed = true; |
| 6201 | } |
| 6202 | } |
| 6203 | |
| 6204 | // "Objective-C Class Properties" is recently added for Objective-C. We |
| 6205 | // upgrade ObjC bitcodes to contain a "Objective-C Class Properties" module |
| 6206 | // flag of value 0, so we can correclty downgrade this flag when trying to |
| 6207 | // link an ObjC bitcode without this module flag with an ObjC bitcode with |
| 6208 | // this module flag. |
| 6209 | if (HasObjCFlag && !HasClassProperties) { |
| 6210 | M.addModuleFlag(Behavior: llvm::Module::Override, Key: "Objective-C Class Properties" , |
| 6211 | Val: (uint32_t)0); |
| 6212 | Changed = true; |
| 6213 | } |
| 6214 | |
| 6215 | if (HasSwiftVersionFlag) { |
| 6216 | M.addModuleFlag(Behavior: Module::Error, Key: "Swift ABI Version" , |
| 6217 | Val: SwiftABIVersion); |
| 6218 | M.addModuleFlag(Behavior: Module::Error, Key: "Swift Major Version" , |
| 6219 | Val: ConstantInt::get(Ty: Int8Ty, V: SwiftMajorVersion)); |
| 6220 | M.addModuleFlag(Behavior: Module::Error, Key: "Swift Minor Version" , |
| 6221 | Val: ConstantInt::get(Ty: Int8Ty, V: SwiftMinorVersion)); |
| 6222 | Changed = true; |
| 6223 | } |
| 6224 | |
| 6225 | return Changed; |
| 6226 | } |
| 6227 | |
| 6228 | void llvm::UpgradeSectionAttributes(Module &M) { |
| 6229 | auto TrimSpaces = [](StringRef Section) -> std::string { |
| 6230 | SmallVector<StringRef, 5> Components; |
| 6231 | Section.split(A&: Components, Separator: ','); |
| 6232 | |
| 6233 | SmallString<32> Buffer; |
| 6234 | raw_svector_ostream OS(Buffer); |
| 6235 | |
| 6236 | for (auto Component : Components) |
| 6237 | OS << ',' << Component.trim(); |
| 6238 | |
| 6239 | return std::string(OS.str().substr(Start: 1)); |
| 6240 | }; |
| 6241 | |
| 6242 | for (auto &GV : M.globals()) { |
| 6243 | if (!GV.hasSection()) |
| 6244 | continue; |
| 6245 | |
| 6246 | StringRef Section = GV.getSection(); |
| 6247 | |
| 6248 | if (!Section.starts_with(Prefix: "__DATA, __objc_catlist" )) |
| 6249 | continue; |
| 6250 | |
| 6251 | // __DATA, __objc_catlist, regular, no_dead_strip |
| 6252 | // __DATA,__objc_catlist,regular,no_dead_strip |
| 6253 | GV.setSection(TrimSpaces(Section)); |
| 6254 | } |
| 6255 | } |
| 6256 | |
| 6257 | namespace { |
| 6258 | // Prior to LLVM 10.0, the strictfp attribute could be used on individual |
| 6259 | // callsites within a function that did not also have the strictfp attribute. |
| 6260 | // Since 10.0, if strict FP semantics are needed within a function, the |
| 6261 | // function must have the strictfp attribute and all calls within the function |
| 6262 | // must also have the strictfp attribute. This latter restriction is |
| 6263 | // necessary to prevent unwanted libcall simplification when a function is |
| 6264 | // being cloned (such as for inlining). |
| 6265 | // |
| 6266 | // The "dangling" strictfp attribute usage was only used to prevent constant |
| 6267 | // folding and other libcall simplification. The nobuiltin attribute on the |
| 6268 | // callsite has the same effect. |
| 6269 | struct StrictFPUpgradeVisitor : public InstVisitor<StrictFPUpgradeVisitor> { |
| 6270 | StrictFPUpgradeVisitor() = default; |
| 6271 | |
| 6272 | void visitCallBase(CallBase &Call) { |
| 6273 | if (!Call.isStrictFP()) |
| 6274 | return; |
| 6275 | if (isa<ConstrainedFPIntrinsic>(Val: &Call)) |
| 6276 | return; |
| 6277 | // If we get here, the caller doesn't have the strictfp attribute |
| 6278 | // but this callsite does. Replace the strictfp attribute with nobuiltin. |
| 6279 | Call.removeFnAttr(Kind: Attribute::StrictFP); |
| 6280 | Call.addFnAttr(Kind: Attribute::NoBuiltin); |
| 6281 | } |
| 6282 | }; |
| 6283 | |
| 6284 | /// Replace "amdgpu-unsafe-fp-atomics" metadata with atomicrmw metadata |
| 6285 | struct AMDGPUUnsafeFPAtomicsUpgradeVisitor |
| 6286 | : public InstVisitor<AMDGPUUnsafeFPAtomicsUpgradeVisitor> { |
| 6287 | AMDGPUUnsafeFPAtomicsUpgradeVisitor() = default; |
| 6288 | |
| 6289 | void visitAtomicRMWInst(AtomicRMWInst &RMW) { |
| 6290 | if (!RMW.isFloatingPointOperation()) |
| 6291 | return; |
| 6292 | |
| 6293 | MDNode *Empty = MDNode::get(Context&: RMW.getContext(), MDs: {}); |
| 6294 | RMW.setMetadata(Kind: "amdgpu.no.fine.grained.host.memory" , Node: Empty); |
| 6295 | RMW.setMetadata(Kind: "amdgpu.no.remote.memory.access" , Node: Empty); |
| 6296 | RMW.setMetadata(Kind: "amdgpu.ignore.denormal.mode" , Node: Empty); |
| 6297 | } |
| 6298 | }; |
| 6299 | } // namespace |
| 6300 | |
| 6301 | void llvm::UpgradeFunctionAttributes(Function &F) { |
| 6302 | // If a function definition doesn't have the strictfp attribute, |
| 6303 | // convert any callsite strictfp attributes to nobuiltin. |
| 6304 | if (!F.isDeclaration() && !F.hasFnAttribute(Kind: Attribute::StrictFP)) { |
| 6305 | StrictFPUpgradeVisitor SFPV; |
| 6306 | SFPV.visit(F); |
| 6307 | } |
| 6308 | |
| 6309 | // Remove all incompatibile attributes from function. |
| 6310 | F.removeRetAttrs(Attrs: AttributeFuncs::typeIncompatible( |
| 6311 | Ty: F.getReturnType(), AS: F.getAttributes().getRetAttrs())); |
| 6312 | for (auto &Arg : F.args()) |
| 6313 | Arg.removeAttrs( |
| 6314 | AM: AttributeFuncs::typeIncompatible(Ty: Arg.getType(), AS: Arg.getAttributes())); |
| 6315 | |
| 6316 | // Older versions of LLVM treated an "implicit-section-name" attribute |
| 6317 | // similarly to directly setting the section on a Function. |
| 6318 | if (Attribute A = F.getFnAttribute(Kind: "implicit-section-name" ); |
| 6319 | A.isValid() && A.isStringAttribute()) { |
| 6320 | F.setSection(A.getValueAsString()); |
| 6321 | F.removeFnAttr(Kind: "implicit-section-name" ); |
| 6322 | } |
| 6323 | |
| 6324 | if (!F.empty()) { |
| 6325 | // For some reason this is called twice, and the first time is before any |
| 6326 | // instructions are loaded into the body. |
| 6327 | |
| 6328 | if (Attribute A = F.getFnAttribute(Kind: "amdgpu-unsafe-fp-atomics" ); |
| 6329 | A.isValid()) { |
| 6330 | |
| 6331 | if (A.getValueAsBool()) { |
| 6332 | AMDGPUUnsafeFPAtomicsUpgradeVisitor Visitor; |
| 6333 | Visitor.visit(F); |
| 6334 | } |
| 6335 | |
| 6336 | // We will leave behind dead attribute uses on external declarations, but |
| 6337 | // clang never added these to declarations anyway. |
| 6338 | F.removeFnAttr(Kind: "amdgpu-unsafe-fp-atomics" ); |
| 6339 | } |
| 6340 | } |
| 6341 | } |
| 6342 | |
| 6343 | // Check if the function attribute is not present and set it. |
| 6344 | static void setFunctionAttrIfNotSet(Function &F, StringRef FnAttrName, |
| 6345 | StringRef Value) { |
| 6346 | if (!F.hasFnAttribute(Kind: FnAttrName)) |
| 6347 | F.addFnAttr(Kind: FnAttrName, Val: Value); |
| 6348 | } |
| 6349 | |
| 6350 | // Check if the function attribute is not present and set it if needed. |
| 6351 | // If the attribute is "false" then removes it. |
| 6352 | // If the attribute is "true" resets it to a valueless attribute. |
| 6353 | static void ConvertFunctionAttr(Function &F, bool Set, StringRef FnAttrName) { |
| 6354 | if (!F.hasFnAttribute(Kind: FnAttrName)) { |
| 6355 | if (Set) |
| 6356 | F.addFnAttr(Kind: FnAttrName); |
| 6357 | } else { |
| 6358 | auto A = F.getFnAttribute(Kind: FnAttrName); |
| 6359 | if ("false" == A.getValueAsString()) |
| 6360 | F.removeFnAttr(Kind: FnAttrName); |
| 6361 | else if ("true" == A.getValueAsString()) { |
| 6362 | F.removeFnAttr(Kind: FnAttrName); |
| 6363 | F.addFnAttr(Kind: FnAttrName); |
| 6364 | } |
| 6365 | } |
| 6366 | } |
| 6367 | |
| 6368 | void llvm::copyModuleAttrToFunctions(Module &M) { |
| 6369 | Triple T(M.getTargetTriple()); |
| 6370 | if (!T.isThumb() && !T.isARM() && !T.isAArch64()) |
| 6371 | return; |
| 6372 | |
| 6373 | uint64_t BTEValue = 0; |
| 6374 | uint64_t BPPLRValue = 0; |
| 6375 | uint64_t GCSValue = 0; |
| 6376 | uint64_t SRAValue = 0; |
| 6377 | uint64_t SRAALLValue = 0; |
| 6378 | uint64_t SRABKeyValue = 0; |
| 6379 | |
| 6380 | NamedMDNode *ModFlags = M.getModuleFlagsMetadata(); |
| 6381 | if (ModFlags) { |
| 6382 | for (unsigned I = 0, E = ModFlags->getNumOperands(); I != E; ++I) { |
| 6383 | MDNode *Op = ModFlags->getOperand(i: I); |
| 6384 | if (Op->getNumOperands() != 3) |
| 6385 | continue; |
| 6386 | |
| 6387 | MDString *ID = dyn_cast_or_null<MDString>(Val: Op->getOperand(I: 1)); |
| 6388 | auto *CI = mdconst::dyn_extract<ConstantInt>(MD: Op->getOperand(I: 2)); |
| 6389 | if (!ID || !CI) |
| 6390 | continue; |
| 6391 | |
| 6392 | StringRef IDStr = ID->getString(); |
| 6393 | uint64_t *ValPtr = IDStr == "branch-target-enforcement" ? &BTEValue |
| 6394 | : IDStr == "branch-protection-pauth-lr" ? &BPPLRValue |
| 6395 | : IDStr == "guarded-control-stack" ? &GCSValue |
| 6396 | : IDStr == "sign-return-address" ? &SRAValue |
| 6397 | : IDStr == "sign-return-address-all" ? &SRAALLValue |
| 6398 | : IDStr == "sign-return-address-with-bkey" |
| 6399 | ? &SRABKeyValue |
| 6400 | : nullptr; |
| 6401 | if (!ValPtr) |
| 6402 | continue; |
| 6403 | |
| 6404 | *ValPtr = CI->getZExtValue(); |
| 6405 | if (*ValPtr == 2) |
| 6406 | return; |
| 6407 | } |
| 6408 | } |
| 6409 | |
| 6410 | bool BTE = BTEValue == 1; |
| 6411 | bool BPPLR = BPPLRValue == 1; |
| 6412 | bool GCS = GCSValue == 1; |
| 6413 | bool SRA = SRAValue == 1; |
| 6414 | |
| 6415 | StringRef SignTypeValue = "non-leaf" ; |
| 6416 | if (SRA && SRAALLValue == 1) |
| 6417 | SignTypeValue = "all" ; |
| 6418 | |
| 6419 | StringRef SignKeyValue = "a_key" ; |
| 6420 | if (SRA && SRABKeyValue == 1) |
| 6421 | SignKeyValue = "b_key" ; |
| 6422 | |
| 6423 | for (Function &F : M.getFunctionList()) { |
| 6424 | if (F.isDeclaration()) |
| 6425 | continue; |
| 6426 | |
| 6427 | if (SRA) { |
| 6428 | setFunctionAttrIfNotSet(F, FnAttrName: "sign-return-address" , Value: SignTypeValue); |
| 6429 | setFunctionAttrIfNotSet(F, FnAttrName: "sign-return-address-key" , Value: SignKeyValue); |
| 6430 | } else { |
| 6431 | if (auto A = F.getFnAttribute(Kind: "sign-return-address" ); |
| 6432 | A.isValid() && "none" == A.getValueAsString()) { |
| 6433 | F.removeFnAttr(Kind: "sign-return-address" ); |
| 6434 | F.removeFnAttr(Kind: "sign-return-address-key" ); |
| 6435 | } |
| 6436 | } |
| 6437 | ConvertFunctionAttr(F, Set: BTE, FnAttrName: "branch-target-enforcement" ); |
| 6438 | ConvertFunctionAttr(F, Set: BPPLR, FnAttrName: "branch-protection-pauth-lr" ); |
| 6439 | ConvertFunctionAttr(F, Set: GCS, FnAttrName: "guarded-control-stack" ); |
| 6440 | } |
| 6441 | |
| 6442 | if (BTE) |
| 6443 | M.setModuleFlag(Behavior: llvm::Module::Min, Key: "branch-target-enforcement" , Val: 2); |
| 6444 | if (BPPLR) |
| 6445 | M.setModuleFlag(Behavior: llvm::Module::Min, Key: "branch-protection-pauth-lr" , Val: 2); |
| 6446 | if (GCS) |
| 6447 | M.setModuleFlag(Behavior: llvm::Module::Min, Key: "guarded-control-stack" , Val: 2); |
| 6448 | if (SRA) { |
| 6449 | M.setModuleFlag(Behavior: llvm::Module::Min, Key: "sign-return-address" , Val: 2); |
| 6450 | if (SRAALLValue == 1) |
| 6451 | M.setModuleFlag(Behavior: llvm::Module::Min, Key: "sign-return-address-all" , Val: 2); |
| 6452 | if (SRABKeyValue == 1) |
| 6453 | M.setModuleFlag(Behavior: llvm::Module::Min, Key: "sign-return-address-with-bkey" , Val: 2); |
| 6454 | } |
| 6455 | } |
| 6456 | |
| 6457 | static bool isOldLoopArgument(Metadata *MD) { |
| 6458 | auto *T = dyn_cast_or_null<MDTuple>(Val: MD); |
| 6459 | if (!T) |
| 6460 | return false; |
| 6461 | if (T->getNumOperands() < 1) |
| 6462 | return false; |
| 6463 | auto *S = dyn_cast_or_null<MDString>(Val: T->getOperand(I: 0)); |
| 6464 | if (!S) |
| 6465 | return false; |
| 6466 | return S->getString().starts_with(Prefix: "llvm.vectorizer." ); |
| 6467 | } |
| 6468 | |
| 6469 | static MDString *upgradeLoopTag(LLVMContext &C, StringRef OldTag) { |
| 6470 | StringRef OldPrefix = "llvm.vectorizer." ; |
| 6471 | assert(OldTag.starts_with(OldPrefix) && "Expected old prefix" ); |
| 6472 | |
| 6473 | if (OldTag == "llvm.vectorizer.unroll" ) |
| 6474 | return MDString::get(Context&: C, Str: "llvm.loop.interleave.count" ); |
| 6475 | |
| 6476 | return MDString::get( |
| 6477 | Context&: C, Str: (Twine("llvm.loop.vectorize." ) + OldTag.drop_front(N: OldPrefix.size())) |
| 6478 | .str()); |
| 6479 | } |
| 6480 | |
| 6481 | static Metadata *upgradeLoopArgument(Metadata *MD) { |
| 6482 | auto *T = dyn_cast_or_null<MDTuple>(Val: MD); |
| 6483 | if (!T) |
| 6484 | return MD; |
| 6485 | if (T->getNumOperands() < 1) |
| 6486 | return MD; |
| 6487 | auto *OldTag = dyn_cast_or_null<MDString>(Val: T->getOperand(I: 0)); |
| 6488 | if (!OldTag) |
| 6489 | return MD; |
| 6490 | if (!OldTag->getString().starts_with(Prefix: "llvm.vectorizer." )) |
| 6491 | return MD; |
| 6492 | |
| 6493 | // This has an old tag. Upgrade it. |
| 6494 | SmallVector<Metadata *, 8> Ops; |
| 6495 | Ops.reserve(N: T->getNumOperands()); |
| 6496 | Ops.push_back(Elt: upgradeLoopTag(C&: T->getContext(), OldTag: OldTag->getString())); |
| 6497 | for (unsigned I = 1, E = T->getNumOperands(); I != E; ++I) |
| 6498 | Ops.push_back(Elt: T->getOperand(I)); |
| 6499 | |
| 6500 | return MDTuple::get(Context&: T->getContext(), MDs: Ops); |
| 6501 | } |
| 6502 | |
| 6503 | MDNode *llvm::upgradeInstructionLoopAttachment(MDNode &N) { |
| 6504 | auto *T = dyn_cast<MDTuple>(Val: &N); |
| 6505 | if (!T) |
| 6506 | return &N; |
| 6507 | |
| 6508 | if (none_of(Range: T->operands(), P: isOldLoopArgument)) |
| 6509 | return &N; |
| 6510 | |
| 6511 | SmallVector<Metadata *, 8> Ops; |
| 6512 | Ops.reserve(N: T->getNumOperands()); |
| 6513 | for (Metadata *MD : T->operands()) |
| 6514 | Ops.push_back(Elt: upgradeLoopArgument(MD)); |
| 6515 | |
| 6516 | return MDTuple::get(Context&: T->getContext(), MDs: Ops); |
| 6517 | } |
| 6518 | |
| 6519 | std::string llvm::UpgradeDataLayoutString(StringRef DL, StringRef TT) { |
| 6520 | Triple T(TT); |
| 6521 | // The only data layout upgrades needed for pre-GCN, SPIR or SPIRV are setting |
| 6522 | // the address space of globals to 1. This does not apply to SPIRV Logical. |
| 6523 | if ((T.isSPIR() || (T.isSPIRV() && !T.isSPIRVLogical())) && |
| 6524 | !DL.contains(Other: "-G" ) && !DL.starts_with(Prefix: "G" )) { |
| 6525 | return DL.empty() ? std::string("G1" ) : (DL + "-G1" ).str(); |
| 6526 | } |
| 6527 | |
| 6528 | if (T.isLoongArch64() || T.isRISCV64()) { |
| 6529 | // Make i32 a native type for 64-bit LoongArch and RISC-V. |
| 6530 | auto I = DL.find(Str: "-n64-" ); |
| 6531 | if (I != StringRef::npos) |
| 6532 | return (DL.take_front(N: I) + "-n32:64-" + DL.drop_front(N: I + 5)).str(); |
| 6533 | return DL.str(); |
| 6534 | } |
| 6535 | |
| 6536 | // AMDGPU data layout upgrades. |
| 6537 | std::string Res = DL.str(); |
| 6538 | if (T.isAMDGPU()) { |
| 6539 | // Define address spaces for constants. |
| 6540 | if (!DL.contains(Other: "-G" ) && !DL.starts_with(Prefix: "G" )) |
| 6541 | Res.append(s: Res.empty() ? "G1" : "-G1" ); |
| 6542 | |
| 6543 | // AMDGCN data layout upgrades. |
| 6544 | if (T.isAMDGCN()) { |
| 6545 | |
| 6546 | // Add missing non-integral declarations. |
| 6547 | // This goes before adding new address spaces to prevent incoherent string |
| 6548 | // values. |
| 6549 | if (!DL.contains(Other: "-ni" ) && !DL.starts_with(Prefix: "ni" )) |
| 6550 | Res.append(s: "-ni:7:8:9" ); |
| 6551 | // Update ni:7 to ni:7:8:9. |
| 6552 | if (DL.ends_with(Suffix: "ni:7" )) |
| 6553 | Res.append(s: ":8:9" ); |
| 6554 | if (DL.ends_with(Suffix: "ni:7:8" )) |
| 6555 | Res.append(s: ":9" ); |
| 6556 | |
| 6557 | // Add sizing for address spaces 7 and 8 (fat raw buffers and buffer |
| 6558 | // resources) An empty data layout has already been upgraded to G1 by now. |
| 6559 | if (!DL.contains(Other: "-p7" ) && !DL.starts_with(Prefix: "p7" )) |
| 6560 | Res.append(s: "-p7:160:256:256:32" ); |
| 6561 | if (!DL.contains(Other: "-p8" ) && !DL.starts_with(Prefix: "p8" )) |
| 6562 | Res.append(s: "-p8:128:128:128:48" ); |
| 6563 | constexpr StringRef OldP8("-p8:128:128-" ); |
| 6564 | if (DL.contains(Other: OldP8)) |
| 6565 | Res.replace(pos: Res.find(svt: OldP8), n1: OldP8.size(), s: "-p8:128:128:128:48-" ); |
| 6566 | if (!DL.contains(Other: "-p9" ) && !DL.starts_with(Prefix: "p9" )) |
| 6567 | Res.append(s: "-p9:192:256:256:32" ); |
| 6568 | } |
| 6569 | |
| 6570 | // Upgrade the ELF mangling mode. |
| 6571 | if (!DL.contains(Other: "m:e" )) |
| 6572 | Res = Res.empty() ? "m:e" : "m:e-" + Res; |
| 6573 | |
| 6574 | return Res; |
| 6575 | } |
| 6576 | |
| 6577 | if (T.isSystemZ() && !DL.empty()) { |
| 6578 | // Make sure the stack alignment is present. |
| 6579 | if (!DL.contains(Other: "-S64" )) |
| 6580 | return "E-S64" + DL.drop_front(N: 1).str(); |
| 6581 | return DL.str(); |
| 6582 | } |
| 6583 | |
| 6584 | auto AddPtr32Ptr64AddrSpaces = [&DL, &Res]() { |
| 6585 | // If the datalayout matches the expected format, add pointer size address |
| 6586 | // spaces to the datalayout. |
| 6587 | StringRef AddrSpaces{"-p270:32:32-p271:32:32-p272:64:64" }; |
| 6588 | if (!DL.contains(Other: AddrSpaces)) { |
| 6589 | SmallVector<StringRef, 4> Groups; |
| 6590 | Regex R("^([Ee]-m:[a-z](-p:32:32)?)(-.*)$" ); |
| 6591 | if (R.match(String: Res, Matches: &Groups)) |
| 6592 | Res = (Groups[1] + AddrSpaces + Groups[3]).str(); |
| 6593 | } |
| 6594 | }; |
| 6595 | |
| 6596 | // AArch64 data layout upgrades. |
| 6597 | if (T.isAArch64()) { |
| 6598 | // Add "-Fn32" |
| 6599 | if (!DL.empty() && !DL.contains(Other: "-Fn32" )) |
| 6600 | Res.append(s: "-Fn32" ); |
| 6601 | AddPtr32Ptr64AddrSpaces(); |
| 6602 | return Res; |
| 6603 | } |
| 6604 | |
| 6605 | if (T.isSPARC() || (T.isMIPS64() && !DL.contains(Other: "m:m" )) || T.isPPC64() || |
| 6606 | T.isWasm()) { |
| 6607 | // Mips64 with o32 ABI did not add "-i128:128". |
| 6608 | // Add "-i128:128" |
| 6609 | std::string I64 = "-i64:64" ; |
| 6610 | std::string I128 = "-i128:128" ; |
| 6611 | if (!StringRef(Res).contains(Other: I128)) { |
| 6612 | size_t Pos = Res.find(str: I64); |
| 6613 | if (Pos != size_t(-1)) |
| 6614 | Res.insert(pos1: Pos + I64.size(), str: I128); |
| 6615 | } |
| 6616 | } |
| 6617 | |
| 6618 | if (T.isPPC() && T.isOSAIX() && !DL.contains(Other: "f64:32:64" ) && !DL.empty()) { |
| 6619 | size_t Pos = Res.find(s: "-S128" ); |
| 6620 | if (Pos == StringRef::npos) |
| 6621 | Pos = Res.size(); |
| 6622 | Res.insert(pos: Pos, s: "-f64:32:64" ); |
| 6623 | } |
| 6624 | |
| 6625 | if (!T.isX86()) |
| 6626 | return Res; |
| 6627 | |
| 6628 | AddPtr32Ptr64AddrSpaces(); |
| 6629 | |
| 6630 | // i128 values need to be 16-byte-aligned. LLVM already called into libgcc |
| 6631 | // for i128 operations prior to this being reflected in the data layout, and |
| 6632 | // clang mostly produced LLVM IR that already aligned i128 to 16 byte |
| 6633 | // boundaries, so although this is a breaking change, the upgrade is expected |
| 6634 | // to fix more IR than it breaks. |
| 6635 | // Intel MCU is an exception and uses 4-byte-alignment. |
| 6636 | if (!T.isOSIAMCU()) { |
| 6637 | std::string I128 = "-i128:128" ; |
| 6638 | if (StringRef Ref = Res; !Ref.contains(Other: I128)) { |
| 6639 | SmallVector<StringRef, 4> Groups; |
| 6640 | Regex R("^(e(-[mpi][^-]*)*)((-[^mpi][^-]*)*)$" ); |
| 6641 | if (R.match(String: Res, Matches: &Groups)) |
| 6642 | Res = (Groups[1] + I128 + Groups[3]).str(); |
| 6643 | } |
| 6644 | } |
| 6645 | |
| 6646 | // For 32-bit MSVC targets, raise the alignment of f80 values to 16 bytes. |
| 6647 | // Raising the alignment is safe because Clang did not produce f80 values in |
| 6648 | // the MSVC environment before this upgrade was added. |
| 6649 | if (T.isWindowsMSVCEnvironment() && !T.isArch64Bit()) { |
| 6650 | StringRef Ref = Res; |
| 6651 | auto I = Ref.find(Str: "-f80:32-" ); |
| 6652 | if (I != StringRef::npos) |
| 6653 | Res = (Ref.take_front(N: I) + "-f80:128-" + Ref.drop_front(N: I + 8)).str(); |
| 6654 | } |
| 6655 | |
| 6656 | return Res; |
| 6657 | } |
| 6658 | |
| 6659 | void llvm::UpgradeAttributes(AttrBuilder &B) { |
| 6660 | StringRef FramePointer; |
| 6661 | Attribute A = B.getAttribute(Kind: "no-frame-pointer-elim" ); |
| 6662 | if (A.isValid()) { |
| 6663 | // The value can be "true" or "false". |
| 6664 | FramePointer = A.getValueAsString() == "true" ? "all" : "none" ; |
| 6665 | B.removeAttribute(A: "no-frame-pointer-elim" ); |
| 6666 | } |
| 6667 | if (B.contains(A: "no-frame-pointer-elim-non-leaf" )) { |
| 6668 | // The value is ignored. "no-frame-pointer-elim"="true" takes priority. |
| 6669 | if (FramePointer != "all" ) |
| 6670 | FramePointer = "non-leaf" ; |
| 6671 | B.removeAttribute(A: "no-frame-pointer-elim-non-leaf" ); |
| 6672 | } |
| 6673 | if (!FramePointer.empty()) |
| 6674 | B.addAttribute(A: "frame-pointer" , V: FramePointer); |
| 6675 | |
| 6676 | A = B.getAttribute(Kind: "null-pointer-is-valid" ); |
| 6677 | if (A.isValid()) { |
| 6678 | // The value can be "true" or "false". |
| 6679 | bool NullPointerIsValid = A.getValueAsString() == "true" ; |
| 6680 | B.removeAttribute(A: "null-pointer-is-valid" ); |
| 6681 | if (NullPointerIsValid) |
| 6682 | B.addAttribute(Val: Attribute::NullPointerIsValid); |
| 6683 | } |
| 6684 | } |
| 6685 | |
| 6686 | void llvm::UpgradeOperandBundles(std::vector<OperandBundleDef> &Bundles) { |
| 6687 | // clang.arc.attachedcall bundles are now required to have an operand. |
| 6688 | // If they don't, it's okay to drop them entirely: when there is an operand, |
| 6689 | // the "attachedcall" is meaningful and required, but without an operand, |
| 6690 | // it's just a marker NOP. Dropping it merely prevents an optimization. |
| 6691 | erase_if(C&: Bundles, P: [&](OperandBundleDef &OBD) { |
| 6692 | return OBD.getTag() == "clang.arc.attachedcall" && |
| 6693 | OBD.inputs().empty(); |
| 6694 | }); |
| 6695 | } |
| 6696 | |