src/shims/intrinsics.rs

   1 use std::iter;
   2
   3 use log::trace;
   4
   5 use rustc_apfloat::{Float, Round};
   6 use rustc_middle::ty::layout::{HasParamEnv, IntegerExt, LayoutOf};
   7 use rustc_middle::{mir, mir::BinOp, ty, ty::FloatTy};
   8 use rustc_target::abi::{Align, Endian, HasDataLayout, Integer, Size};
   9
  10 use crate::*;
  11 use helpers::check_arg_count;
  12
  13 pub enum AtomicOp {
  14     MirOp(mir::BinOp, bool),
  15     Max,
  16     Min,
  17 }
  18
  19 impl<'mir, 'tcx: 'mir> EvalContextExt<'mir, 'tcx> for crate::MiriEvalContext<'mir, 'tcx> {}
  20 pub trait EvalContextExt<'mir, 'tcx: 'mir>: crate::MiriEvalContextExt<'mir, 'tcx> {
  21     fn call_intrinsic(
  22         &mut self,
  23         instance: ty::Instance<'tcx>,
  24         args: &[OpTy<'tcx, Tag>],
  25         ret: Option<(&PlaceTy<'tcx, Tag>, mir::BasicBlock)>,
  26         _unwind: StackPopUnwind,
  27     ) -> InterpResult<'tcx> {
  28         let this = self.eval_context_mut();
  29
  30         if this.emulate_intrinsic(instance, args, ret)? {
  31             return Ok(());
  32         }
  33
  34         // All supported intrinsics have a return place.
  35         let intrinsic_name = this.tcx.item_name(instance.def_id());
  36         let intrinsic_name = intrinsic_name.as_str();
  37         let (dest, ret) = match ret {
  38             None => throw_unsup_format!("unimplemented (diverging) intrinsic: {}", intrinsic_name),
  39             Some(p) => p,
  40         };
  41
  42         // Then handle terminating intrinsics.
  43         match intrinsic_name {
  44             // Miri overwriting CTFE intrinsics.
  45             "ptr_guaranteed_eq" => {
  46                 let &[ref left, ref right] = check_arg_count(args)?;
  47                 let left = this.read_immediate(left)?;
  48                 let right = this.read_immediate(right)?;
  49                 this.binop_ignore_overflow(mir::BinOp::Eq, &left, &right, dest)?;
  50             }
  51             "ptr_guaranteed_ne" => {
  52                 let &[ref left, ref right] = check_arg_count(args)?;
  53                 let left = this.read_immediate(left)?;
  54                 let right = this.read_immediate(right)?;
  55                 this.binop_ignore_overflow(mir::BinOp::Ne, &left, &right, dest)?;
  56             }
  57             "const_allocate" => {
  58                 // For now, for compatibility with the run-time implementation of this, we just return null.
  59                 // See <https://github.com/rust-lang/rust/issues/93935>.
  60                 this.write_null(dest)?;
  61             }
  62             "const_deallocate" => {
  63                 // complete NOP
  64             }
  65
  66             // Raw memory accesses
  67             "volatile_load" => {
  68                 let &[ref place] = check_arg_count(args)?;
  69                 let place = this.deref_operand(place)?;
  70                 this.copy_op(&place.into(), dest)?;
  71             }
  72             "volatile_store" => {
  73                 let &[ref place, ref dest] = check_arg_count(args)?;
  74                 let place = this.deref_operand(place)?;
  75                 this.copy_op(dest, &place.into())?;
  76             }
  77
  78             "write_bytes" | "volatile_set_memory" => {
  79                 let &[ref ptr, ref val_byte, ref count] = check_arg_count(args)?;
  80                 let ty = instance.substs.type_at(0);
  81                 let ty_layout = this.layout_of(ty)?;
  82                 let val_byte = this.read_scalar(val_byte)?.to_u8()?;
  83                 let ptr = this.read_pointer(ptr)?;
  84                 let count = this.read_scalar(count)?.to_machine_usize(this)?;
  85                 // `checked_mul` enforces a too small bound (the correct one would probably be machine_isize_max),
  86                 // but no actual allocation can be big enough for the difference to be noticeable.
  87                 let byte_count = ty_layout.size.checked_mul(count, this).ok_or_else(|| {
  88                     err_ub_format!("overflow computing total size of `{}`", intrinsic_name)
  89                 })?;
  90                 this.write_bytes_ptr(
  91                     ptr,
  92                     iter::repeat(val_byte).take(byte_count.bytes() as usize),
  93                 )?;
  94             }
  95
  96             // Floating-point operations
  97             "fabsf32" => {
  98                 let &[ref f] = check_arg_count(args)?;
  99                 let f = this.read_scalar(f)?.to_f32()?;
 100                 // Can be implemented in soft-floats.
 101                 this.write_scalar(Scalar::from_f32(f.abs()), dest)?;
 102             }
 103             "fabsf64" => {
 104                 let &[ref f] = check_arg_count(args)?;
 105                 let f = this.read_scalar(f)?.to_f64()?;
 106                 // Can be implemented in soft-floats.
 107                 this.write_scalar(Scalar::from_f64(f.abs()), dest)?;
 108             }
 109             #[rustfmt::skip]
 110             | "sinf32"
 111             | "cosf32"
 112             | "sqrtf32"
 113             | "expf32"
 114             | "exp2f32"
 115             | "logf32"
 116             | "log10f32"
 117             | "log2f32"
 118             | "floorf32"
 119             | "ceilf32"
 120             | "truncf32"
 121             | "roundf32"
 122             => {
 123                 let &[ref f] = check_arg_count(args)?;
 124                 // FIXME: Using host floats.
 125                 let f = f32::from_bits(this.read_scalar(f)?.to_u32()?);
 126                 let f = match intrinsic_name {
 127                     "sinf32" => f.sin(),
 128                     "cosf32" => f.cos(),
 129                     "sqrtf32" => f.sqrt(),
 130                     "expf32" => f.exp(),
 131                     "exp2f32" => f.exp2(),
 132                     "logf32" => f.ln(),
 133                     "log10f32" => f.log10(),
 134                     "log2f32" => f.log2(),
 135                     "floorf32" => f.floor(),
 136                     "ceilf32" => f.ceil(),
 137                     "truncf32" => f.trunc(),
 138                     "roundf32" => f.round(),
 139                     _ => bug!(),
 140                 };
 141                 this.write_scalar(Scalar::from_u32(f.to_bits()), dest)?;
 142             }
 143
 144             #[rustfmt::skip]
 145             | "sinf64"
 146             | "cosf64"
 147             | "sqrtf64"
 148             | "expf64"
 149             | "exp2f64"
 150             | "logf64"
 151             | "log10f64"
 152             | "log2f64"
 153             | "floorf64"
 154             | "ceilf64"
 155             | "truncf64"
 156             | "roundf64"
 157             => {
 158                 let &[ref f] = check_arg_count(args)?;
 159                 // FIXME: Using host floats.
 160                 let f = f64::from_bits(this.read_scalar(f)?.to_u64()?);
 161                 let f = match intrinsic_name {
 162                     "sinf64" => f.sin(),
 163                     "cosf64" => f.cos(),
 164                     "sqrtf64" => f.sqrt(),
 165                     "expf64" => f.exp(),
 166                     "exp2f64" => f.exp2(),
 167                     "logf64" => f.ln(),
 168                     "log10f64" => f.log10(),
 169                     "log2f64" => f.log2(),
 170                     "floorf64" => f.floor(),
 171                     "ceilf64" => f.ceil(),
 172                     "truncf64" => f.trunc(),
 173                     "roundf64" => f.round(),
 174                     _ => bug!(),
 175                 };
 176                 this.write_scalar(Scalar::from_u64(f.to_bits()), dest)?;
 177             }
 178
 179             #[rustfmt::skip]
 180             | "fadd_fast"
 181             | "fsub_fast"
 182             | "fmul_fast"
 183             | "fdiv_fast"
 184             | "frem_fast"
 185             => {
 186                 let &[ref a, ref b] = check_arg_count(args)?;
 187                 let a = this.read_immediate(a)?;
 188                 let b = this.read_immediate(b)?;
 189                 let op = match intrinsic_name {
 190                     "fadd_fast" => mir::BinOp::Add,
 191                     "fsub_fast" => mir::BinOp::Sub,
 192                     "fmul_fast" => mir::BinOp::Mul,
 193                     "fdiv_fast" => mir::BinOp::Div,
 194                     "frem_fast" => mir::BinOp::Rem,
 195                     _ => bug!(),
 196                 };
 197                 let float_finite = |x: ImmTy<'tcx, _>| -> InterpResult<'tcx, bool> {
 198                     Ok(match x.layout.ty.kind() {
 199                         ty::Float(FloatTy::F32) => x.to_scalar()?.to_f32()?.is_finite(),
 200                         ty::Float(FloatTy::F64) => x.to_scalar()?.to_f64()?.is_finite(),
 201                         _ => bug!(
 202                             "`{}` called with non-float input type {:?}",
 203                             intrinsic_name,
 204                             x.layout.ty
 205                         ),
 206                     })
 207                 };
 208                 match (float_finite(a)?, float_finite(b)?) {
 209                     (false, false) => throw_ub_format!(
 210                         "`{}` intrinsic called with non-finite value as both parameters",
 211                         intrinsic_name,
 212                     ),
 213                     (false, _) => throw_ub_format!(
 214                         "`{}` intrinsic called with non-finite value as first parameter",
 215                         intrinsic_name,
 216                     ),
 217                     (_, false) => throw_ub_format!(
 218                         "`{}` intrinsic called with non-finite value as second parameter",
 219                         intrinsic_name,
 220                     ),
 221                     _ => {}
 222                 }
 223                 this.binop_ignore_overflow(op, &a, &b, dest)?;
 224             }
 225
 226             #[rustfmt::skip]
 227             | "minnumf32"
 228             | "maxnumf32"
 229             | "copysignf32"
 230             => {
 231                 let &[ref a, ref b] = check_arg_count(args)?;
 232                 let a = this.read_scalar(a)?.to_f32()?;
 233                 let b = this.read_scalar(b)?.to_f32()?;
 234                 let res = match intrinsic_name {
 235                     "minnumf32" => a.min(b),
 236                     "maxnumf32" => a.max(b),
 237                     "copysignf32" => a.copy_sign(b),
 238                     _ => bug!(),
 239                 };
 240                 this.write_scalar(Scalar::from_f32(res), dest)?;
 241             }
 242
 243             #[rustfmt::skip]
 244             | "minnumf64"
 245             | "maxnumf64"
 246             | "copysignf64"
 247             => {
 248                 let &[ref a, ref b] = check_arg_count(args)?;
 249                 let a = this.read_scalar(a)?.to_f64()?;
 250                 let b = this.read_scalar(b)?.to_f64()?;
 251                 let res = match intrinsic_name {
 252                     "minnumf64" => a.min(b),
 253                     "maxnumf64" => a.max(b),
 254                     "copysignf64" => a.copy_sign(b),
 255                     _ => bug!(),
 256                 };
 257                 this.write_scalar(Scalar::from_f64(res), dest)?;
 258             }
 259
 260             "powf32" => {
 261                 let &[ref f, ref f2] = check_arg_count(args)?;
 262                 // FIXME: Using host floats.
 263                 let f = f32::from_bits(this.read_scalar(f)?.to_u32()?);
 264                 let f2 = f32::from_bits(this.read_scalar(f2)?.to_u32()?);
 265                 this.write_scalar(Scalar::from_u32(f.powf(f2).to_bits()), dest)?;
 266             }
 267
 268             "powf64" => {
 269                 let &[ref f, ref f2] = check_arg_count(args)?;
 270                 // FIXME: Using host floats.
 271                 let f = f64::from_bits(this.read_scalar(f)?.to_u64()?);
 272                 let f2 = f64::from_bits(this.read_scalar(f2)?.to_u64()?);
 273                 this.write_scalar(Scalar::from_u64(f.powf(f2).to_bits()), dest)?;
 274             }
 275
 276             "fmaf32" => {
 277                 let &[ref a, ref b, ref c] = check_arg_count(args)?;
 278                 let a = this.read_scalar(a)?.to_f32()?;
 279                 let b = this.read_scalar(b)?.to_f32()?;
 280                 let c = this.read_scalar(c)?.to_f32()?;
 281                 let res = a.mul_add(b, c).value;
 282                 this.write_scalar(Scalar::from_f32(res), dest)?;
 283             }
 284
 285             "fmaf64" => {
 286                 let &[ref a, ref b, ref c] = check_arg_count(args)?;
 287                 let a = this.read_scalar(a)?.to_f64()?;
 288                 let b = this.read_scalar(b)?.to_f64()?;
 289                 let c = this.read_scalar(c)?.to_f64()?;
 290                 let res = a.mul_add(b, c).value;
 291                 this.write_scalar(Scalar::from_f64(res), dest)?;
 292             }
 293
 294             "powif32" => {
 295                 let &[ref f, ref i] = check_arg_count(args)?;
 296                 // FIXME: Using host floats.
 297                 let f = f32::from_bits(this.read_scalar(f)?.to_u32()?);
 298                 let i = this.read_scalar(i)?.to_i32()?;
 299                 this.write_scalar(Scalar::from_u32(f.powi(i).to_bits()), dest)?;
 300             }
 301
 302             "powif64" => {
 303                 let &[ref f, ref i] = check_arg_count(args)?;
 304                 // FIXME: Using host floats.
 305                 let f = f64::from_bits(this.read_scalar(f)?.to_u64()?);
 306                 let i = this.read_scalar(i)?.to_i32()?;
 307                 this.write_scalar(Scalar::from_u64(f.powi(i).to_bits()), dest)?;
 308             }
 309
 310             "float_to_int_unchecked" => {
 311                 let &[ref val] = check_arg_count(args)?;
 312                 let val = this.read_immediate(val)?;
 313
 314                 let res = match val.layout.ty.kind() {
 315                     ty::Float(FloatTy::F32) =>
 316                         this.float_to_int_unchecked(val.to_scalar()?.to_f32()?, dest.layout.ty)?,
 317                     ty::Float(FloatTy::F64) =>
 318                         this.float_to_int_unchecked(val.to_scalar()?.to_f64()?, dest.layout.ty)?,
 319                     _ =>
 320                         bug!(
 321                             "`float_to_int_unchecked` called with non-float input type {:?}",
 322                             val.layout.ty
 323                         ),
 324                 };
 325
 326                 this.write_scalar(res, dest)?;
 327             }
 328
 329             // SIMD operations
 330             #[rustfmt::skip]
 331             | "simd_neg"
 332             | "simd_fabs"
 333             | "simd_ceil"
 334             | "simd_floor"
 335             | "simd_round"
 336             | "simd_trunc"
 337             | "simd_fsqrt" => {
 338                 let &[ref op] = check_arg_count(args)?;
 339                 let (op, op_len) = this.operand_to_simd(op)?;
 340                 let (dest, dest_len) = this.place_to_simd(dest)?;
 341
 342                 assert_eq!(dest_len, op_len);
 343
 344                 #[derive(Copy, Clone)]
 345                 enum HostFloatOp {
 346                     Ceil,
 347                     Floor,
 348                     Round,
 349                     Trunc,
 350                     Sqrt,
 351                 }
 352                 #[derive(Copy, Clone)]
 353                 enum Op {
 354                     MirOp(mir::UnOp),
 355                     Abs,
 356                     HostOp(HostFloatOp),
 357                 }
 358                 let which = match intrinsic_name {
 359                     "simd_neg" => Op::MirOp(mir::UnOp::Neg),
 360                     "simd_fabs" => Op::Abs,
 361                     "simd_ceil" => Op::HostOp(HostFloatOp::Ceil),
 362                     "simd_floor" => Op::HostOp(HostFloatOp::Floor),
 363                     "simd_round" => Op::HostOp(HostFloatOp::Round),
 364                     "simd_trunc" => Op::HostOp(HostFloatOp::Trunc),
 365                     "simd_fsqrt" => Op::HostOp(HostFloatOp::Sqrt),
 366                     _ => unreachable!(),
 367                 };
 368
 369                 for i in 0..dest_len {
 370                     let op = this.read_immediate(&this.mplace_index(&op, i)?.into())?;
 371                     let dest = this.mplace_index(&dest, i)?;
 372                     let val = match which {
 373                         Op::MirOp(mir_op) => this.unary_op(mir_op, &op)?.to_scalar()?,
 374                         Op::Abs => {
 375                             // Works for f32 and f64.
 376                             let ty::Float(float_ty) = op.layout.ty.kind() else {
 377                                 bug!("{} operand is not a float", intrinsic_name)
 378                             };
 379                             let op = op.to_scalar()?;
 380                             match float_ty {
 381                                 FloatTy::F32 => Scalar::from_f32(op.to_f32()?.abs()),
 382                                 FloatTy::F64 => Scalar::from_f64(op.to_f64()?.abs()),
 383                             }
 384                         }
 385                         Op::HostOp(host_op) => {
 386                             let ty::Float(float_ty) = op.layout.ty.kind() else {
 387                                 bug!("{} operand is not a float", intrinsic_name)
 388                             };
 389                             // FIXME using host floats
 390                             match float_ty {
 391                                 FloatTy::F32 => {
 392                                     let f = f32::from_bits(op.to_scalar()?.to_u32()?);
 393                                     let res = match host_op {
 394                                         HostFloatOp::Ceil => f.ceil(),
 395                                         HostFloatOp::Floor => f.floor(),
 396                                         HostFloatOp::Round => f.round(),
 397                                         HostFloatOp::Trunc => f.trunc(),
 398                                         HostFloatOp::Sqrt => f.sqrt(),
 399                                     };
 400                                     Scalar::from_u32(res.to_bits())
 401                                 }
 402                                 FloatTy::F64 => {
 403                                     let f = f64::from_bits(op.to_scalar()?.to_u64()?);
 404                                     let res = match host_op {
 405                                         HostFloatOp::Ceil => f.ceil(),
 406                                         HostFloatOp::Floor => f.floor(),
 407                                         HostFloatOp::Round => f.round(),
 408                                         HostFloatOp::Trunc => f.trunc(),
 409                                         HostFloatOp::Sqrt => f.sqrt(),
 410                                     };
 411                                     Scalar::from_u64(res.to_bits())
 412                                 }
 413                             }
 414
 415                         }
 416                     };
 417                     this.write_scalar(val, &dest.into())?;
 418                 }
 419             }
 420             #[rustfmt::skip]
 421             | "simd_add"
 422             | "simd_sub"
 423             | "simd_mul"
 424             | "simd_div"
 425             | "simd_rem"
 426             | "simd_shl"
 427             | "simd_shr"
 428             | "simd_and"
 429             | "simd_or"
 430             | "simd_xor"
 431             | "simd_eq"
 432             | "simd_ne"
 433             | "simd_lt"
 434             | "simd_le"
 435             | "simd_gt"
 436             | "simd_ge"
 437             | "simd_fmax"
 438             | "simd_fmin"
 439             | "simd_saturating_add"
 440             | "simd_saturating_sub"
 441             | "simd_arith_offset" => {
 442                 use mir::BinOp;
 443
 444                 let &[ref left, ref right] = check_arg_count(args)?;
 445                 let (left, left_len) = this.operand_to_simd(left)?;
 446                 let (right, right_len) = this.operand_to_simd(right)?;
 447                 let (dest, dest_len) = this.place_to_simd(dest)?;
 448
 449                 assert_eq!(dest_len, left_len);
 450                 assert_eq!(dest_len, right_len);
 451
 452                 enum Op {
 453                     MirOp(BinOp),
 454                     SaturatingOp(BinOp),
 455                     FMax,
 456                     FMin,
 457                     WrappingOffset,
 458                 }
 459                 let which = match intrinsic_name {
 460                     "simd_add" => Op::MirOp(BinOp::Add),
 461                     "simd_sub" => Op::MirOp(BinOp::Sub),
 462                     "simd_mul" => Op::MirOp(BinOp::Mul),
 463                     "simd_div" => Op::MirOp(BinOp::Div),
 464                     "simd_rem" => Op::MirOp(BinOp::Rem),
 465                     "simd_shl" => Op::MirOp(BinOp::Shl),
 466                     "simd_shr" => Op::MirOp(BinOp::Shr),
 467                     "simd_and" => Op::MirOp(BinOp::BitAnd),
 468                     "simd_or" => Op::MirOp(BinOp::BitOr),
 469                     "simd_xor" => Op::MirOp(BinOp::BitXor),
 470                     "simd_eq" => Op::MirOp(BinOp::Eq),
 471                     "simd_ne" => Op::MirOp(BinOp::Ne),
 472                     "simd_lt" => Op::MirOp(BinOp::Lt),
 473                     "simd_le" => Op::MirOp(BinOp::Le),
 474                     "simd_gt" => Op::MirOp(BinOp::Gt),
 475                     "simd_ge" => Op::MirOp(BinOp::Ge),
 476                     "simd_fmax" => Op::FMax,
 477                     "simd_fmin" => Op::FMin,
 478                     "simd_saturating_add" => Op::SaturatingOp(BinOp::Add),
 479                     "simd_saturating_sub" => Op::SaturatingOp(BinOp::Sub),
 480                     "simd_arith_offset" => Op::WrappingOffset,
 481                     _ => unreachable!(),
 482                 };
 483
 484                 for i in 0..dest_len {
 485                     let left = this.read_immediate(&this.mplace_index(&left, i)?.into())?;
 486                     let right = this.read_immediate(&this.mplace_index(&right, i)?.into())?;
 487                     let dest = this.mplace_index(&dest, i)?;
 488                     let val = match which {
 489                         Op::MirOp(mir_op) => {
 490                             let (val, overflowed, ty) = this.overflowing_binary_op(mir_op, &left, &right)?;
 491                             if matches!(mir_op, BinOp::Shl | BinOp::Shr) {
 492                                 // Shifts have extra UB as SIMD operations that the MIR binop does not have.
 493                                 // See <https://github.com/rust-lang/rust/issues/91237>.
 494                                 if overflowed {
 495                                     let r_val = right.to_scalar()?.to_bits(right.layout.size)?;
 496                                     throw_ub_format!("overflowing shift by {} in `{}` in SIMD lane {}", r_val, intrinsic_name, i);
 497                                 }
 498                             }
 499                             if matches!(mir_op, BinOp::Eq | BinOp::Ne | BinOp::Lt | BinOp::Le | BinOp::Gt | BinOp::Ge) {
 500                                 // Special handling for boolean-returning operations
 501                                 assert_eq!(ty, this.tcx.types.bool);
 502                                 let val = val.to_bool().unwrap();
 503                                 bool_to_simd_element(val, dest.layout.size)
 504                             } else {
 505                                 assert_ne!(ty, this.tcx.types.bool);
 506                                 assert_eq!(ty, dest.layout.ty);
 507                                 val
 508                             }
 509                         }
 510                         Op::SaturatingOp(mir_op) => {
 511                             this.saturating_arith(mir_op, &left, &right)?
 512                         }
 513                         Op::WrappingOffset => {
 514                             let ptr = this.scalar_to_ptr(left.to_scalar()?)?;
 515                             let offset_count = right.to_scalar()?.to_machine_isize(this)?;
 516                             let pointee_ty = left.layout.ty.builtin_deref(true).unwrap().ty;
 517
 518                             let pointee_size = i64::try_from(this.layout_of(pointee_ty)?.size.bytes()).unwrap();
 519                             let offset_bytes = offset_count.wrapping_mul(pointee_size);
 520                             let offset_ptr = ptr.wrapping_signed_offset(offset_bytes, this);
 521                             Scalar::from_maybe_pointer(offset_ptr, this)
 522                         }
 523                         Op::FMax => {
 524                             fmax_op(&left, &right)?
 525                         }
 526                         Op::FMin => {
 527                             fmin_op(&left, &right)?
 528                         }
 529                     };
 530                     this.write_scalar(val, &dest.into())?;
 531                 }
 532             }
 533             "simd_fma" => {
 534                 let &[ref a, ref b, ref c] = check_arg_count(args)?;
 535                 let (a, a_len) = this.operand_to_simd(a)?;
 536                 let (b, b_len) = this.operand_to_simd(b)?;
 537                 let (c, c_len) = this.operand_to_simd(c)?;
 538                 let (dest, dest_len) = this.place_to_simd(dest)?;
 539
 540                 assert_eq!(dest_len, a_len);
 541                 assert_eq!(dest_len, b_len);
 542                 assert_eq!(dest_len, c_len);
 543
 544                 for i in 0..dest_len {
 545                     let a = this.read_immediate(&this.mplace_index(&a, i)?.into())?.to_scalar()?;
 546                     let b = this.read_immediate(&this.mplace_index(&b, i)?.into())?.to_scalar()?;
 547                     let c = this.read_immediate(&this.mplace_index(&c, i)?.into())?.to_scalar()?;
 548                     let dest = this.mplace_index(&dest, i)?;
 549
 550                     // Works for f32 and f64.
 551                     let ty::Float(float_ty) = dest.layout.ty.kind() else {
 552                         bug!("{} operand is not a float", intrinsic_name)
 553                     };
 554                     let val = match float_ty {
 555                         FloatTy::F32 =>
 556                             Scalar::from_f32(a.to_f32()?.mul_add(b.to_f32()?, c.to_f32()?).value),
 557                         FloatTy::F64 =>
 558                             Scalar::from_f64(a.to_f64()?.mul_add(b.to_f64()?, c.to_f64()?).value),
 559                     };
 560                     this.write_scalar(val, &dest.into())?;
 561                 }
 562             }
 563             #[rustfmt::skip]
 564             | "simd_reduce_and"
 565             | "simd_reduce_or"
 566             | "simd_reduce_xor"
 567             | "simd_reduce_any"
 568             | "simd_reduce_all"
 569             | "simd_reduce_max"
 570             | "simd_reduce_min" => {
 571                 use mir::BinOp;
 572
 573                 let &[ref op] = check_arg_count(args)?;
 574                 let (op, op_len) = this.operand_to_simd(op)?;
 575
 576                 let imm_from_bool =
 577                     |b| ImmTy::from_scalar(Scalar::from_bool(b), this.machine.layouts.bool);
 578
 579                 enum Op {
 580                     MirOp(BinOp),
 581                     MirOpBool(BinOp),
 582                     Max,
 583                     Min,
 584                 }
 585                 let which = match intrinsic_name {
 586                     "simd_reduce_and" => Op::MirOp(BinOp::BitAnd),
 587                     "simd_reduce_or" => Op::MirOp(BinOp::BitOr),
 588                     "simd_reduce_xor" => Op::MirOp(BinOp::BitXor),
 589                     "simd_reduce_any" => Op::MirOpBool(BinOp::BitOr),
 590                     "simd_reduce_all" => Op::MirOpBool(BinOp::BitAnd),
 591                     "simd_reduce_max" => Op::Max,
 592                     "simd_reduce_min" => Op::Min,
 593                     _ => unreachable!(),
 594                 };
 595
 596                 // Initialize with first lane, then proceed with the rest.
 597                 let mut res = this.read_immediate(&this.mplace_index(&op, 0)?.into())?;
 598                 if matches!(which, Op::MirOpBool(_)) {
 599                     // Convert to `bool` scalar.
 600                     res = imm_from_bool(simd_element_to_bool(res)?);
 601                 }
 602                 for i in 1..op_len {
 603                     let op = this.read_immediate(&this.mplace_index(&op, i)?.into())?;
 604                     res = match which {
 605                         Op::MirOp(mir_op) => {
 606                             this.binary_op(mir_op, &res, &op)?
 607                         }
 608                         Op::MirOpBool(mir_op) => {
 609                             let op = imm_from_bool(simd_element_to_bool(op)?);
 610                             this.binary_op(mir_op, &res, &op)?
 611                         }
 612                         Op::Max => {
 613                             if matches!(res.layout.ty.kind(), ty::Float(_)) {
 614                                 ImmTy::from_scalar(fmax_op(&res, &op)?, res.layout)
 615                             } else {
 616                                 // Just boring integers, so NaNs to worry about
 617                                 if this.binary_op(BinOp::Ge, &res, &op)?.to_scalar()?.to_bool()? {
 618                                     res
 619                                 } else {
 620                                     op
 621                                 }
 622                             }
 623                         }
 624                         Op::Min => {
 625                             if matches!(res.layout.ty.kind(), ty::Float(_)) {
 626                                 ImmTy::from_scalar(fmin_op(&res, &op)?, res.layout)
 627                             } else {
 628                                 // Just boring integers, so NaNs to worry about
 629                                 if this.binary_op(BinOp::Le, &res, &op)?.to_scalar()?.to_bool()? {
 630                                     res
 631                                 } else {
 632                                     op
 633                                 }
 634                             }
 635                         }
 636                     };
 637                 }
 638                 this.write_immediate(*res, dest)?;
 639             }
 640             #[rustfmt::skip]
 641             | "simd_reduce_add_ordered"
 642             | "simd_reduce_mul_ordered" => {
 643                 use mir::BinOp;
 644
 645                 let &[ref op, ref init] = check_arg_count(args)?;
 646                 let (op, op_len) = this.operand_to_simd(op)?;
 647                 let init = this.read_immediate(init)?;
 648
 649                 let mir_op = match intrinsic_name {
 650                     "simd_reduce_add_ordered" => BinOp::Add,
 651                     "simd_reduce_mul_ordered" => BinOp::Mul,
 652                     _ => unreachable!(),
 653                 };
 654
 655                 let mut res = init;
 656                 for i in 0..op_len {
 657                     let op = this.read_immediate(&this.mplace_index(&op, i)?.into())?;
 658                     res = this.binary_op(mir_op, &res, &op)?;
 659                 }
 660                 this.write_immediate(*res, dest)?;
 661             }
 662             "simd_select" => {
 663                 let &[ref mask, ref yes, ref no] = check_arg_count(args)?;
 664                 let (mask, mask_len) = this.operand_to_simd(mask)?;
 665                 let (yes, yes_len) = this.operand_to_simd(yes)?;
 666                 let (no, no_len) = this.operand_to_simd(no)?;
 667                 let (dest, dest_len) = this.place_to_simd(dest)?;
 668
 669                 assert_eq!(dest_len, mask_len);
 670                 assert_eq!(dest_len, yes_len);
 671                 assert_eq!(dest_len, no_len);
 672
 673                 for i in 0..dest_len {
 674                     let mask = this.read_immediate(&this.mplace_index(&mask, i)?.into())?;
 675                     let yes = this.read_immediate(&this.mplace_index(&yes, i)?.into())?;
 676                     let no = this.read_immediate(&this.mplace_index(&no, i)?.into())?;
 677                     let dest = this.mplace_index(&dest, i)?;
 678
 679                     let val = if simd_element_to_bool(mask)? { yes } else { no };
 680                     this.write_immediate(*val, &dest.into())?;
 681                 }
 682             }
 683             "simd_select_bitmask" => {
 684                 let &[ref mask, ref yes, ref no] = check_arg_count(args)?;
 685                 let (yes, yes_len) = this.operand_to_simd(yes)?;
 686                 let (no, no_len) = this.operand_to_simd(no)?;
 687                 let (dest, dest_len) = this.place_to_simd(dest)?;
 688                 let bitmask_len = dest_len.max(8);
 689
 690                 assert!(mask.layout.ty.is_integral());
 691                 assert!(bitmask_len <= 64);
 692                 assert_eq!(bitmask_len, mask.layout.size.bits());
 693                 assert_eq!(dest_len, yes_len);
 694                 assert_eq!(dest_len, no_len);
 695
 696                 let mask: u64 = this
 697                     .read_scalar(mask)?
 698                     .check_init()?
 699                     .to_bits(mask.layout.size)?
 700                     .try_into()
 701                     .unwrap();
 702                 for i in 0..dest_len {
 703                     let mask =
 704                         mask & (1 << simd_bitmask_index(i, dest_len, this.data_layout().endian));
 705                     let yes = this.read_immediate(&this.mplace_index(&yes, i)?.into())?;
 706                     let no = this.read_immediate(&this.mplace_index(&no, i)?.into())?;
 707                     let dest = this.mplace_index(&dest, i)?;
 708
 709                     let val = if mask != 0 { yes } else { no };
 710                     this.write_immediate(*val, &dest.into())?;
 711                 }
 712                 for i in dest_len..bitmask_len {
 713                     // If the mask is "padded", ensure that padding is all-zero.
 714                     let mask = mask & (1 << i);
 715                     if mask != 0 {
 716                         throw_ub_format!(
 717                             "a SIMD bitmask less than 8 bits long must be filled with 0s for the remaining bits"
 718                         );
 719                     }
 720                 }
 721             }
 722             #[rustfmt::skip]
 723             "simd_cast" | "simd_as" => {
 724                 let &[ref op] = check_arg_count(args)?;
 725                 let (op, op_len) = this.operand_to_simd(op)?;
 726                 let (dest, dest_len) = this.place_to_simd(dest)?;
 727
 728                 assert_eq!(dest_len, op_len);
 729
 730                 let safe_cast = intrinsic_name == "simd_as";
 731
 732                 for i in 0..dest_len {
 733                     let op = this.read_immediate(&this.mplace_index(&op, i)?.into())?;
 734                     let dest = this.mplace_index(&dest, i)?;
 735
 736                     let val = match (op.layout.ty.kind(), dest.layout.ty.kind()) {
 737                         // Int-to-(int|float): always safe
 738                         (ty::Int(_) | ty::Uint(_), ty::Int(_) | ty::Uint(_) | ty::Float(_)) =>
 739                             this.misc_cast(&op, dest.layout.ty)?,
 740                         // Float-to-float: always safe
 741                         (ty::Float(_), ty::Float(_)) =>
 742                             this.misc_cast(&op, dest.layout.ty)?,
 743                         // Float-to-int in safe mode
 744                         (ty::Float(_), ty::Int(_) | ty::Uint(_)) if safe_cast =>
 745                             this.misc_cast(&op, dest.layout.ty)?,
 746                         // Float-to-int in unchecked mode
 747                         (ty::Float(FloatTy::F32), ty::Int(_) | ty::Uint(_)) if !safe_cast =>
 748                             this.float_to_int_unchecked(op.to_scalar()?.to_f32()?, dest.layout.ty)?.into(),
 749                         (ty::Float(FloatTy::F64), ty::Int(_) | ty::Uint(_)) if !safe_cast =>
 750                             this.float_to_int_unchecked(op.to_scalar()?.to_f64()?, dest.layout.ty)?.into(),
 751                         _ =>
 752                             throw_unsup_format!(
 753                                 "Unsupported SIMD cast from element type {} to {}",
 754                                 op.layout.ty,
 755                                 dest.layout.ty
 756                             ),
 757                     };
 758                     this.write_immediate(val, &dest.into())?;
 759                 }
 760             }
 761             "simd_shuffle" => {
 762                 let &[ref left, ref right, ref index] = check_arg_count(args)?;
 763                 let (left, left_len) = this.operand_to_simd(left)?;
 764                 let (right, right_len) = this.operand_to_simd(right)?;
 765                 let (dest, dest_len) = this.place_to_simd(dest)?;
 766
 767                 // `index` is an array, not a SIMD type
 768                 let ty::Array(_, index_len) = index.layout.ty.kind() else {
 769                     bug!("simd_shuffle index argument has non-array type {}", index.layout.ty)
 770                 };
 771                 let index_len = index_len.eval_usize(*this.tcx, this.param_env());
 772
 773                 assert_eq!(left_len, right_len);
 774                 assert_eq!(index_len, dest_len);
 775
 776                 for i in 0..dest_len {
 777                     let src_index: u64 = this
 778                         .read_immediate(&this.operand_index(&index, i)?.into())?
 779                         .to_scalar()?
 780                         .to_u32()?
 781                         .into();
 782                     let dest = this.mplace_index(&dest, i)?;
 783
 784                     let val = if src_index < left_len {
 785                         this.read_immediate(&this.mplace_index(&left, src_index)?.into())?
 786                     } else if src_index < left_len.checked_add(right_len).unwrap() {
 787                         this.read_immediate(
 788                             &this.mplace_index(&right, src_index - left_len)?.into(),
 789                         )?
 790                     } else {
 791                         bug!(
 792                             "simd_shuffle index {} is out of bounds for 2 vectors of size {}",
 793                             src_index,
 794                             left_len
 795                         );
 796                     };
 797                     this.write_immediate(*val, &dest.into())?;
 798                 }
 799             }
 800             "simd_gather" => {
 801                 let &[ref passthru, ref ptrs, ref mask] = check_arg_count(args)?;
 802                 let (passthru, passthru_len) = this.operand_to_simd(passthru)?;
 803                 let (ptrs, ptrs_len) = this.operand_to_simd(ptrs)?;
 804                 let (mask, mask_len) = this.operand_to_simd(mask)?;
 805                 let (dest, dest_len) = this.place_to_simd(dest)?;
 806
 807                 assert_eq!(dest_len, passthru_len);
 808                 assert_eq!(dest_len, ptrs_len);
 809                 assert_eq!(dest_len, mask_len);
 810
 811                 for i in 0..dest_len {
 812                     let passthru = this.read_immediate(&this.mplace_index(&passthru, i)?.into())?;
 813                     let ptr = this.read_immediate(&this.mplace_index(&ptrs, i)?.into())?;
 814                     let mask = this.read_immediate(&this.mplace_index(&mask, i)?.into())?;
 815                     let dest = this.mplace_index(&dest, i)?;
 816
 817                     let val = if simd_element_to_bool(mask)? {
 818                         let place = this.deref_operand(&ptr.into())?;
 819                         this.read_immediate(&place.into())?
 820                     } else {
 821                         passthru
 822                     };
 823                     this.write_immediate(*val, &dest.into())?;
 824                 }
 825             }
 826             "simd_scatter" => {
 827                 let &[ref value, ref ptrs, ref mask] = check_arg_count(args)?;
 828                 let (value, value_len) = this.operand_to_simd(value)?;
 829                 let (ptrs, ptrs_len) = this.operand_to_simd(ptrs)?;
 830                 let (mask, mask_len) = this.operand_to_simd(mask)?;
 831
 832                 assert_eq!(ptrs_len, value_len);
 833                 assert_eq!(ptrs_len, mask_len);
 834
 835                 for i in 0..ptrs_len {
 836                     let value = this.read_immediate(&this.mplace_index(&value, i)?.into())?;
 837                     let ptr = this.read_immediate(&this.mplace_index(&ptrs, i)?.into())?;
 838                     let mask = this.read_immediate(&this.mplace_index(&mask, i)?.into())?;
 839
 840                     if simd_element_to_bool(mask)? {
 841                         let place = this.deref_operand(&ptr.into())?;
 842                         this.write_immediate(*value, &place.into())?;
 843                     }
 844                 }
 845             }
 846             "simd_bitmask" => {
 847                 let &[ref op] = check_arg_count(args)?;
 848                 let (op, op_len) = this.operand_to_simd(op)?;
 849                 let bitmask_len = op_len.max(8);
 850
 851                 assert!(dest.layout.ty.is_integral());
 852                 assert!(bitmask_len <= 64);
 853                 assert_eq!(bitmask_len, dest.layout.size.bits());
 854
 855                 let mut res = 0u64;
 856                 for i in 0..op_len {
 857                     let op = this.read_immediate(&this.mplace_index(&op, i)?.into())?;
 858                     if simd_element_to_bool(op)? {
 859                         res |= 1 << simd_bitmask_index(i, op_len, this.data_layout().endian);
 860                     }
 861                 }
 862                 this.write_int(res, dest)?;
 863             }
 864
 865             // Atomic operations
 866             "atomic_load" => this.atomic_load(args, dest, AtomicReadOp::SeqCst)?,
 867             "atomic_load_relaxed" => this.atomic_load(args, dest, AtomicReadOp::Relaxed)?,
 868             "atomic_load_acq" => this.atomic_load(args, dest, AtomicReadOp::Acquire)?,
 869
 870             "atomic_store" => this.atomic_store(args, AtomicWriteOp::SeqCst)?,
 871             "atomic_store_relaxed" => this.atomic_store(args, AtomicWriteOp::Relaxed)?,
 872             "atomic_store_rel" => this.atomic_store(args, AtomicWriteOp::Release)?,
 873
 874             "atomic_fence_acq" => this.atomic_fence(args, AtomicFenceOp::Acquire)?,
 875             "atomic_fence_rel" => this.atomic_fence(args, AtomicFenceOp::Release)?,
 876             "atomic_fence_acqrel" => this.atomic_fence(args, AtomicFenceOp::AcqRel)?,
 877             "atomic_fence" => this.atomic_fence(args, AtomicFenceOp::SeqCst)?,
 878
 879             "atomic_singlethreadfence_acq" => this.compiler_fence(args, AtomicFenceOp::Acquire)?,
 880             "atomic_singlethreadfence_rel" => this.compiler_fence(args, AtomicFenceOp::Release)?,
 881             "atomic_singlethreadfence_acqrel" =>
 882                 this.compiler_fence(args, AtomicFenceOp::AcqRel)?,
 883             "atomic_singlethreadfence" => this.compiler_fence(args, AtomicFenceOp::SeqCst)?,
 884
 885             "atomic_xchg" => this.atomic_exchange(args, dest, AtomicRwOp::SeqCst)?,
 886             "atomic_xchg_acq" => this.atomic_exchange(args, dest, AtomicRwOp::Acquire)?,
 887             "atomic_xchg_rel" => this.atomic_exchange(args, dest, AtomicRwOp::Release)?,
 888             "atomic_xchg_acqrel" => this.atomic_exchange(args, dest, AtomicRwOp::AcqRel)?,
 889             "atomic_xchg_relaxed" => this.atomic_exchange(args, dest, AtomicRwOp::Relaxed)?,
 890
 891             #[rustfmt::skip]
 892             "atomic_cxchg" =>
 893                 this.atomic_compare_exchange(args, dest, AtomicRwOp::SeqCst, AtomicReadOp::SeqCst)?,
 894             #[rustfmt::skip]
 895             "atomic_cxchg_acq" =>
 896                 this.atomic_compare_exchange(args, dest, AtomicRwOp::Acquire, AtomicReadOp::Acquire)?,
 897             #[rustfmt::skip]
 898             "atomic_cxchg_rel" =>
 899                 this.atomic_compare_exchange(args, dest, AtomicRwOp::Release, AtomicReadOp::Relaxed)?,
 900             #[rustfmt::skip]
 901             "atomic_cxchg_acqrel" =>
 902                 this.atomic_compare_exchange(args, dest, AtomicRwOp::AcqRel, AtomicReadOp::Acquire)?,
 903             #[rustfmt::skip]
 904             "atomic_cxchg_relaxed" =>
 905                 this.atomic_compare_exchange(args, dest, AtomicRwOp::Relaxed, AtomicReadOp::Relaxed)?,
 906             #[rustfmt::skip]
 907             "atomic_cxchg_acq_failrelaxed" =>
 908                 this.atomic_compare_exchange(args, dest, AtomicRwOp::Acquire, AtomicReadOp::Relaxed)?,
 909             #[rustfmt::skip]
 910             "atomic_cxchg_acqrel_failrelaxed" =>
 911                 this.atomic_compare_exchange(args, dest, AtomicRwOp::AcqRel, AtomicReadOp::Relaxed)?,
 912             #[rustfmt::skip]
 913             "atomic_cxchg_failrelaxed" =>
 914                 this.atomic_compare_exchange(args, dest, AtomicRwOp::SeqCst, AtomicReadOp::Relaxed)?,
 915             #[rustfmt::skip]
 916             "atomic_cxchg_failacq" =>
 917                 this.atomic_compare_exchange(args, dest, AtomicRwOp::SeqCst, AtomicReadOp::Acquire)?,
 918
 919             #[rustfmt::skip]
 920             "atomic_cxchgweak" =>
 921                 this.atomic_compare_exchange_weak(args, dest, AtomicRwOp::SeqCst, AtomicReadOp::SeqCst)?,
 922             #[rustfmt::skip]
 923             "atomic_cxchgweak_acq" =>
 924                 this.atomic_compare_exchange_weak(args, dest, AtomicRwOp::Acquire, AtomicReadOp::Acquire)?,
 925             #[rustfmt::skip]
 926             "atomic_cxchgweak_rel" =>
 927                 this.atomic_compare_exchange_weak(args, dest, AtomicRwOp::Release, AtomicReadOp::Relaxed)?,
 928             #[rustfmt::skip]
 929             "atomic_cxchgweak_acqrel" =>
 930                 this.atomic_compare_exchange_weak(args, dest, AtomicRwOp::AcqRel, AtomicReadOp::Acquire)?,
 931             #[rustfmt::skip]
 932             "atomic_cxchgweak_relaxed" =>
 933                 this.atomic_compare_exchange_weak(args, dest, AtomicRwOp::Relaxed, AtomicReadOp::Relaxed)?,
 934             #[rustfmt::skip]
 935             "atomic_cxchgweak_acq_failrelaxed" =>
 936                 this.atomic_compare_exchange_weak(args, dest, AtomicRwOp::Acquire, AtomicReadOp::Relaxed)?,
 937             #[rustfmt::skip]
 938             "atomic_cxchgweak_acqrel_failrelaxed" =>
 939                 this.atomic_compare_exchange_weak(args, dest, AtomicRwOp::AcqRel, AtomicReadOp::Relaxed)?,
 940             #[rustfmt::skip]
 941             "atomic_cxchgweak_failrelaxed" =>
 942                 this.atomic_compare_exchange_weak(args, dest, AtomicRwOp::SeqCst, AtomicReadOp::Relaxed)?,
 943             #[rustfmt::skip]
 944             "atomic_cxchgweak_failacq" =>
 945                 this.atomic_compare_exchange_weak(args, dest, AtomicRwOp::SeqCst, AtomicReadOp::Acquire)?,
 946
 947             #[rustfmt::skip]
 948             "atomic_or" =>
 949                 this.atomic_op(args, dest, AtomicOp::MirOp(BinOp::BitOr, false), AtomicRwOp::SeqCst)?,
 950             #[rustfmt::skip]
 951             "atomic_or_acq" =>
 952                 this.atomic_op(args, dest, AtomicOp::MirOp(BinOp::BitOr, false), AtomicRwOp::Acquire)?,
 953             #[rustfmt::skip]
 954             "atomic_or_rel" =>
 955                 this.atomic_op(args, dest, AtomicOp::MirOp(BinOp::BitOr, false), AtomicRwOp::Release)?,
 956             #[rustfmt::skip]
 957             "atomic_or_acqrel" =>
 958                 this.atomic_op(args, dest, AtomicOp::MirOp(BinOp::BitOr, false), AtomicRwOp::AcqRel)?,
 959             #[rustfmt::skip]
 960             "atomic_or_relaxed" =>
 961                 this.atomic_op(args, dest, AtomicOp::MirOp(BinOp::BitOr, false), AtomicRwOp::Relaxed)?,
 962             #[rustfmt::skip]
 963             "atomic_xor" =>
 964                 this.atomic_op(args, dest, AtomicOp::MirOp(BinOp::BitXor, false), AtomicRwOp::SeqCst)?,
 965             #[rustfmt::skip]
 966             "atomic_xor_acq" =>
 967                 this.atomic_op(args, dest, AtomicOp::MirOp(BinOp::BitXor, false), AtomicRwOp::Acquire)?,
 968             #[rustfmt::skip]
 969             "atomic_xor_rel" =>
 970                 this.atomic_op(args, dest, AtomicOp::MirOp(BinOp::BitXor, false), AtomicRwOp::Release)?,
 971             #[rustfmt::skip]
 972             "atomic_xor_acqrel" =>
 973                 this.atomic_op(args, dest, AtomicOp::MirOp(BinOp::BitXor, false), AtomicRwOp::AcqRel)?,
 974             #[rustfmt::skip]
 975             "atomic_xor_relaxed" =>
 976                 this.atomic_op(args, dest, AtomicOp::MirOp(BinOp::BitXor, false), AtomicRwOp::Relaxed)?,
 977             #[rustfmt::skip]
 978             "atomic_and" =>
 979                 this.atomic_op(args, dest, AtomicOp::MirOp(BinOp::BitAnd, false), AtomicRwOp::SeqCst)?,
 980             #[rustfmt::skip]
 981             "atomic_and_acq" =>
 982                 this.atomic_op(args, dest, AtomicOp::MirOp(BinOp::BitAnd, false), AtomicRwOp::Acquire)?,
 983             #[rustfmt::skip]
 984             "atomic_and_rel" =>
 985                 this.atomic_op(args, dest, AtomicOp::MirOp(BinOp::BitAnd, false), AtomicRwOp::Release)?,
 986             #[rustfmt::skip]
 987             "atomic_and_acqrel" =>
 988                 this.atomic_op(args, dest, AtomicOp::MirOp(BinOp::BitAnd, false), AtomicRwOp::AcqRel)?,
 989             #[rustfmt::skip]
 990             "atomic_and_relaxed" =>
 991                 this.atomic_op(args, dest, AtomicOp::MirOp(BinOp::BitAnd, false), AtomicRwOp::Relaxed)?,
 992             #[rustfmt::skip]
 993             "atomic_nand" =>
 994                 this.atomic_op(args, dest, AtomicOp::MirOp(BinOp::BitAnd, true), AtomicRwOp::SeqCst)?,
 995             #[rustfmt::skip]
 996             "atomic_nand_acq" =>
 997                 this.atomic_op(args, dest, AtomicOp::MirOp(BinOp::BitAnd, true), AtomicRwOp::Acquire)?,
 998             #[rustfmt::skip]
 999             "atomic_nand_rel" =>
1000                 this.atomic_op(args, dest, AtomicOp::MirOp(BinOp::BitAnd, true), AtomicRwOp::Release)?,
1001             #[rustfmt::skip]
1002             "atomic_nand_acqrel" =>
1003                 this.atomic_op(args, dest, AtomicOp::MirOp(BinOp::BitAnd, true), AtomicRwOp::AcqRel)?,
1004             #[rustfmt::skip]
1005             "atomic_nand_relaxed" =>
1006                 this.atomic_op(args, dest, AtomicOp::MirOp(BinOp::BitAnd, true), AtomicRwOp::Relaxed)?,
1007             #[rustfmt::skip]
1008             "atomic_xadd" =>
1009                 this.atomic_op(args, dest, AtomicOp::MirOp(BinOp::Add, false), AtomicRwOp::SeqCst)?,
1010             #[rustfmt::skip]
1011             "atomic_xadd_acq" =>
1012                 this.atomic_op(args, dest, AtomicOp::MirOp(BinOp::Add, false), AtomicRwOp::Acquire)?,
1013             #[rustfmt::skip]
1014             "atomic_xadd_rel" =>
1015                 this.atomic_op(args, dest, AtomicOp::MirOp(BinOp::Add, false), AtomicRwOp::Release)?,
1016             #[rustfmt::skip]
1017             "atomic_xadd_acqrel" =>
1018                 this.atomic_op(args, dest, AtomicOp::MirOp(BinOp::Add, false), AtomicRwOp::AcqRel)?,
1019             #[rustfmt::skip]
1020             "atomic_xadd_relaxed" =>
1021                 this.atomic_op(args, dest, AtomicOp::MirOp(BinOp::Add, false), AtomicRwOp::Relaxed)?,
1022             #[rustfmt::skip]
1023             "atomic_xsub" =>
1024                 this.atomic_op(args, dest, AtomicOp::MirOp(BinOp::Sub, false), AtomicRwOp::SeqCst)?,
1025             #[rustfmt::skip]
1026             "atomic_xsub_acq" =>
1027                 this.atomic_op(args, dest, AtomicOp::MirOp(BinOp::Sub, false), AtomicRwOp::Acquire)?,
1028             #[rustfmt::skip]
1029             "atomic_xsub_rel" =>
1030                 this.atomic_op(args, dest, AtomicOp::MirOp(BinOp::Sub, false), AtomicRwOp::Release)?,
1031             #[rustfmt::skip]
1032             "atomic_xsub_acqrel" =>
1033                 this.atomic_op(args, dest, AtomicOp::MirOp(BinOp::Sub, false), AtomicRwOp::AcqRel)?,
1034             #[rustfmt::skip]
1035             "atomic_xsub_relaxed" =>
1036                 this.atomic_op(args, dest, AtomicOp::MirOp(BinOp::Sub, false), AtomicRwOp::Relaxed)?,
1037             "atomic_min" => this.atomic_op(args, dest, AtomicOp::Min, AtomicRwOp::SeqCst)?,
1038             "atomic_min_acq" => this.atomic_op(args, dest, AtomicOp::Min, AtomicRwOp::Acquire)?,
1039             "atomic_min_rel" => this.atomic_op(args, dest, AtomicOp::Min, AtomicRwOp::Release)?,
1040             "atomic_min_acqrel" => this.atomic_op(args, dest, AtomicOp::Min, AtomicRwOp::AcqRel)?,
1041             "atomic_min_relaxed" =>
1042                 this.atomic_op(args, dest, AtomicOp::Min, AtomicRwOp::Relaxed)?,
1043             "atomic_max" => this.atomic_op(args, dest, AtomicOp::Max, AtomicRwOp::SeqCst)?,
1044             "atomic_max_acq" => this.atomic_op(args, dest, AtomicOp::Max, AtomicRwOp::Acquire)?,
1045             "atomic_max_rel" => this.atomic_op(args, dest, AtomicOp::Max, AtomicRwOp::Release)?,
1046             "atomic_max_acqrel" => this.atomic_op(args, dest, AtomicOp::Max, AtomicRwOp::AcqRel)?,
1047             "atomic_max_relaxed" =>
1048                 this.atomic_op(args, dest, AtomicOp::Max, AtomicRwOp::Relaxed)?,
1049             "atomic_umin" => this.atomic_op(args, dest, AtomicOp::Min, AtomicRwOp::SeqCst)?,
1050             "atomic_umin_acq" => this.atomic_op(args, dest, AtomicOp::Min, AtomicRwOp::Acquire)?,
1051             "atomic_umin_rel" => this.atomic_op(args, dest, AtomicOp::Min, AtomicRwOp::Release)?,
1052             "atomic_umin_acqrel" =>
1053                 this.atomic_op(args, dest, AtomicOp::Min, AtomicRwOp::AcqRel)?,
1054             "atomic_umin_relaxed" =>
1055                 this.atomic_op(args, dest, AtomicOp::Min, AtomicRwOp::Relaxed)?,
1056             "atomic_umax" => this.atomic_op(args, dest, AtomicOp::Max, AtomicRwOp::SeqCst)?,
1057             "atomic_umax_acq" => this.atomic_op(args, dest, AtomicOp::Max, AtomicRwOp::Acquire)?,
1058             "atomic_umax_rel" => this.atomic_op(args, dest, AtomicOp::Max, AtomicRwOp::Release)?,
1059             "atomic_umax_acqrel" =>
1060                 this.atomic_op(args, dest, AtomicOp::Max, AtomicRwOp::AcqRel)?,
1061             "atomic_umax_relaxed" =>
1062                 this.atomic_op(args, dest, AtomicOp::Max, AtomicRwOp::Relaxed)?,
1063
1064             // Other
1065             "exact_div" => {
1066                 let &[ref num, ref denom] = check_arg_count(args)?;
1067                 this.exact_div(&this.read_immediate(num)?, &this.read_immediate(denom)?, dest)?;
1068             }
1069
1070             "try" => return this.handle_try(args, dest, ret),
1071
1072             "breakpoint" => {
1073                 let &[] = check_arg_count(args)?;
1074                 // normally this would raise a SIGTRAP, which aborts if no debugger is connected
1075                 throw_machine_stop!(TerminationInfo::Abort("Trace/breakpoint trap".to_string()))
1076             }
1077
1078             name => throw_unsup_format!("unimplemented intrinsic: {}", name),
1079         }
1080
1081         trace!("{:?}", this.dump_place(**dest));
1082         this.go_to_block(ret);
1083         Ok(())
1084     }
1085
1086     fn atomic_load(
1087         &mut self,
1088         args: &[OpTy<'tcx, Tag>],
1089         dest: &PlaceTy<'tcx, Tag>,
1090         atomic: AtomicReadOp,
1091     ) -> InterpResult<'tcx> {
1092         let this = self.eval_context_mut();
1093
1094         let &[ref place] = check_arg_count(args)?;
1095         let place = this.deref_operand(place)?;
1096
1097         // make sure it fits into a scalar; otherwise it cannot be atomic
1098         let val = this.read_scalar_atomic(&place, atomic)?;
1099
1100         // Check alignment requirements. Atomics must always be aligned to their size,
1101         // even if the type they wrap would be less aligned (e.g. AtomicU64 on 32bit must
1102         // be 8-aligned).
1103         let align = Align::from_bytes(place.layout.size.bytes()).unwrap();
1104         this.check_ptr_access_align(
1105             place.ptr,
1106             place.layout.size,
1107             align,
1108             CheckInAllocMsg::MemoryAccessTest,
1109         )?;
1110         // Perform regular access.
1111         this.write_scalar(val, dest)?;
1112         Ok(())
1113     }
1114
1115     fn atomic_store(
1116         &mut self,
1117         args: &[OpTy<'tcx, Tag>],
1118         atomic: AtomicWriteOp,
1119     ) -> InterpResult<'tcx> {
1120         let this = self.eval_context_mut();
1121
1122         let &[ref place, ref val] = check_arg_count(args)?;
1123         let place = this.deref_operand(place)?;
1124         let val = this.read_scalar(val)?; // make sure it fits into a scalar; otherwise it cannot be atomic
1125
1126         // Check alignment requirements. Atomics must always be aligned to their size,
1127         // even if the type they wrap would be less aligned (e.g. AtomicU64 on 32bit must
1128         // be 8-aligned).
1129         let align = Align::from_bytes(place.layout.size.bytes()).unwrap();
1130         this.check_ptr_access_align(
1131             place.ptr,
1132             place.layout.size,
1133             align,
1134             CheckInAllocMsg::MemoryAccessTest,
1135         )?;
1136
1137         // Perform atomic store
1138         this.write_scalar_atomic(val, &place, atomic)?;
1139         Ok(())
1140     }
1141
1142     fn compiler_fence(
1143         &mut self,
1144         args: &[OpTy<'tcx, Tag>],
1145         atomic: AtomicFenceOp,
1146     ) -> InterpResult<'tcx> {
1147         let &[] = check_arg_count(args)?;
1148         let _ = atomic;
1149         //FIXME: compiler fences are currently ignored
1150         Ok(())
1151     }
1152
1153     fn atomic_fence(
1154         &mut self,
1155         args: &[OpTy<'tcx, Tag>],
1156         atomic: AtomicFenceOp,
1157     ) -> InterpResult<'tcx> {
1158         let this = self.eval_context_mut();
1159         let &[] = check_arg_count(args)?;
1160         this.validate_atomic_fence(atomic)?;
1161         Ok(())
1162     }
1163
1164     fn atomic_op(
1165         &mut self,
1166         args: &[OpTy<'tcx, Tag>],
1167         dest: &PlaceTy<'tcx, Tag>,
1168         atomic_op: AtomicOp,
1169         atomic: AtomicRwOp,
1170     ) -> InterpResult<'tcx> {
1171         let this = self.eval_context_mut();
1172
1173         let &[ref place, ref rhs] = check_arg_count(args)?;
1174         let place = this.deref_operand(place)?;
1175
1176         if !place.layout.ty.is_integral() {
1177             bug!("Atomic arithmetic operations only work on integer types");
1178         }
1179         let rhs = this.read_immediate(rhs)?;
1180
1181         // Check alignment requirements. Atomics must always be aligned to their size,
1182         // even if the type they wrap would be less aligned (e.g. AtomicU64 on 32bit must
1183         // be 8-aligned).
1184         let align = Align::from_bytes(place.layout.size.bytes()).unwrap();
1185         this.check_ptr_access_align(
1186             place.ptr,
1187             place.layout.size,
1188             align,
1189             CheckInAllocMsg::MemoryAccessTest,
1190         )?;
1191
1192         match atomic_op {
1193             AtomicOp::Min => {
1194                 let old = this.atomic_min_max_scalar(&place, rhs, true, atomic)?;
1195                 this.write_immediate(*old, &dest)?; // old value is returned
1196                 Ok(())
1197             }
1198             AtomicOp::Max => {
1199                 let old = this.atomic_min_max_scalar(&place, rhs, false, atomic)?;
1200                 this.write_immediate(*old, &dest)?; // old value is returned
1201                 Ok(())
1202             }
1203             AtomicOp::MirOp(op, neg) => {
1204                 let old = this.atomic_op_immediate(&place, &rhs, op, neg, atomic)?;
1205                 this.write_immediate(*old, dest)?; // old value is returned
1206                 Ok(())
1207             }
1208         }
1209     }
1210
1211     fn atomic_exchange(
1212         &mut self,
1213         args: &[OpTy<'tcx, Tag>],
1214         dest: &PlaceTy<'tcx, Tag>,
1215         atomic: AtomicRwOp,
1216     ) -> InterpResult<'tcx> {
1217         let this = self.eval_context_mut();
1218
1219         let &[ref place, ref new] = check_arg_count(args)?;
1220         let place = this.deref_operand(place)?;
1221         let new = this.read_scalar(new)?;
1222
1223         // Check alignment requirements. Atomics must always be aligned to their size,
1224         // even if the type they wrap would be less aligned (e.g. AtomicU64 on 32bit must
1225         // be 8-aligned).
1226         let align = Align::from_bytes(place.layout.size.bytes()).unwrap();
1227         this.check_ptr_access_align(
1228             place.ptr,
1229             place.layout.size,
1230             align,
1231             CheckInAllocMsg::MemoryAccessTest,
1232         )?;
1233
1234         let old = this.atomic_exchange_scalar(&place, new, atomic)?;
1235         this.write_scalar(old, dest)?; // old value is returned
1236         Ok(())
1237     }
1238
1239     fn atomic_compare_exchange_impl(
1240         &mut self,
1241         args: &[OpTy<'tcx, Tag>],
1242         dest: &PlaceTy<'tcx, Tag>,
1243         success: AtomicRwOp,
1244         fail: AtomicReadOp,
1245         can_fail_spuriously: bool,
1246     ) -> InterpResult<'tcx> {
1247         let this = self.eval_context_mut();
1248
1249         let &[ref place, ref expect_old, ref new] = check_arg_count(args)?;
1250         let place = this.deref_operand(place)?;
1251         let expect_old = this.read_immediate(expect_old)?; // read as immediate for the sake of `binary_op()`
1252         let new = this.read_scalar(new)?;
1253
1254         // Check alignment requirements. Atomics must always be aligned to their size,
1255         // even if the type they wrap would be less aligned (e.g. AtomicU64 on 32bit must
1256         // be 8-aligned).
1257         let align = Align::from_bytes(place.layout.size.bytes()).unwrap();
1258         this.check_ptr_access_align(
1259             place.ptr,
1260             place.layout.size,
1261             align,
1262             CheckInAllocMsg::MemoryAccessTest,
1263         )?;
1264
1265         let old = this.atomic_compare_exchange_scalar(
1266             &place,
1267             &expect_old,
1268             new,
1269             success,
1270             fail,
1271             can_fail_spuriously,
1272         )?;
1273
1274         // Return old value.
1275         this.write_immediate(old, dest)?;
1276         Ok(())
1277     }
1278
1279     fn atomic_compare_exchange(
1280         &mut self,
1281         args: &[OpTy<'tcx, Tag>],
1282         dest: &PlaceTy<'tcx, Tag>,
1283         success: AtomicRwOp,
1284         fail: AtomicReadOp,
1285     ) -> InterpResult<'tcx> {
1286         self.atomic_compare_exchange_impl(args, dest, success, fail, false)
1287     }
1288
1289     fn atomic_compare_exchange_weak(
1290         &mut self,
1291         args: &[OpTy<'tcx, Tag>],
1292         dest: &PlaceTy<'tcx, Tag>,
1293         success: AtomicRwOp,
1294         fail: AtomicReadOp,
1295     ) -> InterpResult<'tcx> {
1296         self.atomic_compare_exchange_impl(args, dest, success, fail, true)
1297     }
1298
1299     fn float_to_int_unchecked<F>(
1300         &self,
1301         f: F,
1302         dest_ty: ty::Ty<'tcx>,
1303     ) -> InterpResult<'tcx, Scalar<Tag>>
1304     where
1305         F: Float + Into<Scalar<Tag>>,
1306     {
1307         let this = self.eval_context_ref();
1308
1309         // Step 1: cut off the fractional part of `f`. The result of this is
1310         // guaranteed to be precisely representable in IEEE floats.
1311         let f = f.round_to_integral(Round::TowardZero).value;
1312
1313         // Step 2: Cast the truncated float to the target integer type and see if we lose any information in this step.
1314         Ok(match dest_ty.kind() {
1315             // Unsigned
1316             ty::Uint(t) => {
1317                 let size = Integer::from_uint_ty(this, *t).size();
1318                 let res = f.to_u128(size.bits_usize());
1319                 if res.status.is_empty() {
1320                     // No status flags means there was no further rounding or other loss of precision.
1321                     Scalar::from_uint(res.value, size)
1322                 } else {
1323                     // `f` was not representable in this integer type.
1324                     throw_ub_format!(
1325                         "`float_to_int_unchecked` intrinsic called on {} which cannot be represented in target type `{:?}`",
1326                         f,
1327                         dest_ty,
1328                     );
1329                 }
1330             }
1331             // Signed
1332             ty::Int(t) => {
1333                 let size = Integer::from_int_ty(this, *t).size();
1334                 let res = f.to_i128(size.bits_usize());
1335                 if res.status.is_empty() {
1336                     // No status flags means there was no further rounding or other loss of precision.
1337                     Scalar::from_int(res.value, size)
1338                 } else {
1339                     // `f` was not representable in this integer type.
1340                     throw_ub_format!(
1341                         "`float_to_int_unchecked` intrinsic called on {} which cannot be represented in target type `{:?}`",
1342                         f,
1343                         dest_ty,
1344                     );
1345                 }
1346             }
1347             // Nothing else
1348             _ => bug!("`float_to_int_unchecked` called with non-int output type {:?}", dest_ty),
1349         })
1350     }
1351 }
1352
1353 fn fmax_op<'tcx>(
1354     left: &ImmTy<'tcx, Tag>,
1355     right: &ImmTy<'tcx, Tag>,
1356 ) -> InterpResult<'tcx, Scalar<Tag>> {
1357     assert_eq!(left.layout.ty, right.layout.ty);
1358     let ty::Float(float_ty) = left.layout.ty.kind() else {
1359         bug!("fmax operand is not a float")
1360     };
1361     let left = left.to_scalar()?;
1362     let right = right.to_scalar()?;
1363     Ok(match float_ty {
1364         FloatTy::F32 => Scalar::from_f32(left.to_f32()?.max(right.to_f32()?)),
1365         FloatTy::F64 => Scalar::from_f64(left.to_f64()?.max(right.to_f64()?)),
1366     })
1367 }
1368
1369 fn fmin_op<'tcx>(
1370     left: &ImmTy<'tcx, Tag>,
1371     right: &ImmTy<'tcx, Tag>,
1372 ) -> InterpResult<'tcx, Scalar<Tag>> {
1373     assert_eq!(left.layout.ty, right.layout.ty);
1374     let ty::Float(float_ty) = left.layout.ty.kind() else {
1375         bug!("fmin operand is not a float")
1376     };
1377     let left = left.to_scalar()?;
1378     let right = right.to_scalar()?;
1379     Ok(match float_ty {
1380         FloatTy::F32 => Scalar::from_f32(left.to_f32()?.min(right.to_f32()?)),
1381         FloatTy::F64 => Scalar::from_f64(left.to_f64()?.min(right.to_f64()?)),
1382     })
1383 }
1384
1385 fn bool_to_simd_element(b: bool, size: Size) -> Scalar<Tag> {
1386     // SIMD uses all-1 as pattern for "true"
1387     let val = if b { -1 } else { 0 };
1388     Scalar::from_int(val, size)
1389 }
1390
1391 fn simd_element_to_bool<'tcx>(elem: ImmTy<'tcx, Tag>) -> InterpResult<'tcx, bool> {
1392     let val = elem.to_scalar()?.to_int(elem.layout.size)?;
1393     Ok(match val {
1394         0 => false,
1395         -1 => true,
1396         _ => throw_ub_format!("each element of a SIMD mask must be all-0-bits or all-1-bits"),
1397     })
1398 }
1399
1400 fn simd_bitmask_index(idx: u64, vec_len: u64, endianess: Endian) -> u64 {
1401     assert!(idx < vec_len);
1402     match endianess {
1403         Endian::Little => idx,
1404         Endian::Big => vec_len - 1 - idx, // reverse order of bits
1405     }
1406 }