src/intrinsics/mod.rs

   1 mod llvm;
   2 mod simd;
   3
   4 pub(crate) use llvm::codegen_llvm_intrinsic_call;
   5
   6 use crate::prelude::*;
   7
   8 macro intrinsic_pat {
   9     (_) => {
  10         _
  11     },
  12     ($name:ident) => {
  13         stringify!($name)
  14     },
  15     ($name:literal) => {
  16         stringify!($name)
  17     },
  18     ($x:ident . $($xs:tt).*) => {
  19         concat!(stringify!($x), ".", intrinsic_pat!($($xs).*))
  20     }
  21 }
  22
  23 macro intrinsic_arg {
  24     (o $fx:expr, $arg:ident) => {
  25         $arg
  26     },
  27     (c $fx:expr, $arg:ident) => {
  28         trans_operand($fx, $arg)
  29     },
  30     (v $fx:expr, $arg:ident) => {
  31         trans_operand($fx, $arg).load_scalar($fx)
  32     }
  33 }
  34
  35 macro intrinsic_substs {
  36     ($substs:expr, $index:expr,) => {},
  37     ($substs:expr, $index:expr, $first:ident $(,$rest:ident)*) => {
  38         let $first = $substs.type_at($index);
  39         intrinsic_substs!($substs, $index+1, $($rest),*);
  40     }
  41 }
  42
  43 macro intrinsic_match {
  44     ($fx:expr, $intrinsic:expr, $substs:expr, $args:expr,
  45     _ => $unknown:block;
  46     $(
  47         $($($name:tt).*)|+ $(if $cond:expr)?, $(<$($subst:ident),*>)? ($($a:ident $arg:ident),*) $content:block;
  48     )*) => {
  49         let _ = $substs; // Silence warning when substs is unused.
  50         match $intrinsic {
  51             $(
  52                 $(intrinsic_pat!($($name).*))|* $(if $cond)? => {
  53                     #[allow(unused_parens, non_snake_case)]
  54                     {
  55                         $(
  56                             intrinsic_substs!($substs, 0, $($subst),*);
  57                         )?
  58                         if let [$($arg),*] = $args {
  59                             let ($($arg,)*) = (
  60                                 $(intrinsic_arg!($a $fx, $arg),)*
  61                             );
  62                             #[warn(unused_parens, non_snake_case)]
  63                             {
  64                                 $content
  65                             }
  66                         } else {
  67                             bug!("wrong number of args for intrinsic {:?}", $intrinsic);
  68                         }
  69                     }
  70                 }
  71             )*
  72             _ => $unknown,
  73         }
  74     }
  75 }
  76
  77 macro call_intrinsic_match {
  78     ($fx:expr, $intrinsic:expr, $substs:expr, $ret:expr, $destination:expr, $args:expr, $(
  79         $name:ident($($arg:ident),*) -> $ty:ident => $func:ident,
  80     )*) => {
  81         match $intrinsic {
  82             $(
  83                 stringify!($name) => {
  84                     assert!($substs.is_noop());
  85                     if let [$(ref $arg),*] = *$args {
  86                         let ($($arg,)*) = (
  87                             $(trans_operand($fx, $arg),)*
  88                         );
  89                         let res = $fx.easy_call(stringify!($func), &[$($arg),*], $fx.tcx.types.$ty);
  90                         $ret.write_cvalue($fx, res);
  91
  92                         if let Some((_, dest)) = $destination {
  93                             let ret_block = $fx.get_block(dest);
  94                             $fx.bcx.ins().jump(ret_block, &[]);
  95                             return;
  96                         } else {
  97                             unreachable!();
  98                         }
  99                     } else {
 100                         bug!("wrong number of args for intrinsic {:?}", $intrinsic);
 101                     }
 102                 }
 103             )*
 104             _ => {}
 105         }
 106     }
 107 }
 108
 109 macro atomic_binop_return_old($fx:expr, $op:ident<$T:ident>($ptr:ident, $src:ident) -> $ret:ident)  {
 110     crate::atomic_shim::lock_global_lock($fx);
 111
 112     let clif_ty = $fx.clif_type($T).unwrap();
 113     let old = $fx.bcx.ins().load(clif_ty, MemFlags::new(), $ptr, 0);
 114     let new = $fx.bcx.ins().$op(old, $src);
 115     $fx.bcx.ins().store(MemFlags::new(), new, $ptr, 0);
 116     $ret.write_cvalue($fx, CValue::by_val(old, $fx.layout_of($T)));
 117
 118     crate::atomic_shim::unlock_global_lock($fx);
 119 }
 120
 121 macro atomic_minmax($fx:expr, $cc:expr, <$T:ident> ($ptr:ident, $src:ident) -> $ret:ident) {
 122     crate::atomic_shim::lock_global_lock($fx);
 123
 124     // Read old
 125     let clif_ty = $fx.clif_type($T).unwrap();
 126     let old = $fx.bcx.ins().load(clif_ty, MemFlags::new(), $ptr, 0);
 127
 128     // Compare
 129     let is_eq = $fx.bcx.ins().icmp(IntCC::SignedGreaterThan, old, $src);
 130     let new = $fx.bcx.ins().select(is_eq, old, $src);
 131
 132     // Write new
 133     $fx.bcx.ins().store(MemFlags::new(), new, $ptr, 0);
 134
 135     let ret_val = CValue::by_val(old, $ret.layout());
 136     $ret.write_cvalue($fx, ret_val);
 137
 138     crate::atomic_shim::unlock_global_lock($fx);
 139 }
 140
 141 fn lane_type_and_count<'tcx>(
 142     tcx: TyCtxt<'tcx>,
 143     layout: TyAndLayout<'tcx>,
 144 ) -> (TyAndLayout<'tcx>, u16) {
 145     assert!(layout.ty.is_simd());
 146     let lane_count = match layout.fields {
 147         rustc_target::abi::FieldsShape::Array { stride: _, count } => u16::try_from(count).unwrap(),
 148         _ => unreachable!("lane_type_and_count({:?})", layout),
 149     };
 150     let lane_layout = layout.field(&ty::layout::LayoutCx {
 151         tcx,
 152         param_env: ParamEnv::reveal_all(),
 153     }, 0).unwrap();
 154     (lane_layout, lane_count)
 155 }
 156
 157 fn clif_vector_type<'tcx>(tcx: TyCtxt<'tcx>, layout: TyAndLayout<'tcx>) -> Option<Type> {
 158     let (element, count) = match &layout.abi {
 159         Abi::Vector { element, count } => (element.clone(), *count),
 160         _ => unreachable!(),
 161     };
 162
 163     match scalar_to_clif_type(tcx, element).by(u16::try_from(count).unwrap()) {
 164         // Cranelift currently only implements icmp for 128bit vectors. While 64bit lanes are
 165         // supported, this needs either the `use_sse41_simd` or `use_sse42_simd` target flag
 166         // to be enabled.
 167         Some(vector_ty) if vector_ty.bits() == 128 && vector_ty.lane_type() != types::I64 => Some(vector_ty),
 168         _ => None,
 169     }
 170 }
 171
 172 fn simd_for_each_lane<'tcx, B: Backend>(
 173     fx: &mut FunctionCx<'_, 'tcx, B>,
 174     val: CValue<'tcx>,
 175     ret: CPlace<'tcx>,
 176     f: impl Fn(
 177         &mut FunctionCx<'_, 'tcx, B>,
 178         TyAndLayout<'tcx>,
 179         TyAndLayout<'tcx>,
 180         Value,
 181     ) -> CValue<'tcx>,
 182 ) {
 183     let layout = val.layout();
 184
 185     let (lane_layout, lane_count) = lane_type_and_count(fx.tcx, layout);
 186     let (ret_lane_layout, ret_lane_count) = lane_type_and_count(fx.tcx, ret.layout());
 187     assert_eq!(lane_count, ret_lane_count);
 188
 189     for lane_idx in 0..lane_count {
 190         let lane_idx = mir::Field::new(lane_idx.try_into().unwrap());
 191         let lane = val.value_field(fx, lane_idx).load_scalar(fx);
 192
 193         let res_lane = f(fx, lane_layout, ret_lane_layout, lane);
 194
 195         ret.place_field(fx, lane_idx).write_cvalue(fx, res_lane);
 196     }
 197 }
 198
 199 fn simd_pair_for_each_lane<'tcx, B: Backend>(
 200     fx: &mut FunctionCx<'_, 'tcx, B>,
 201     x: CValue<'tcx>,
 202     y: CValue<'tcx>,
 203     ret: CPlace<'tcx>,
 204     f: impl Fn(
 205         &mut FunctionCx<'_, 'tcx, B>,
 206         TyAndLayout<'tcx>,
 207         TyAndLayout<'tcx>,
 208         Value,
 209         Value,
 210     ) -> CValue<'tcx>,
 211 ) {
 212     assert_eq!(x.layout(), y.layout());
 213     let layout = x.layout();
 214
 215     let (lane_layout, lane_count) = lane_type_and_count(fx.tcx, layout);
 216     let (ret_lane_layout, ret_lane_count) = lane_type_and_count(fx.tcx, ret.layout());
 217     assert_eq!(lane_count, ret_lane_count);
 218
 219     for lane in 0..lane_count {
 220         let lane = mir::Field::new(lane.try_into().unwrap());
 221         let x_lane = x.value_field(fx, lane).load_scalar(fx);
 222         let y_lane = y.value_field(fx, lane).load_scalar(fx);
 223
 224         let res_lane = f(fx, lane_layout, ret_lane_layout, x_lane, y_lane);
 225
 226         ret.place_field(fx, lane).write_cvalue(fx, res_lane);
 227     }
 228 }
 229
 230 fn bool_to_zero_or_max_uint<'tcx>(
 231     fx: &mut FunctionCx<'_, 'tcx, impl Backend>,
 232     layout: TyAndLayout<'tcx>,
 233     val: Value,
 234 ) -> CValue<'tcx> {
 235     let ty = fx.clif_type(layout.ty).unwrap();
 236
 237     let int_ty = match ty {
 238         types::F32 => types::I32,
 239         types::F64 => types::I64,
 240         ty => ty,
 241     };
 242
 243     let val = fx.bcx.ins().bint(int_ty, val);
 244     let mut res = fx.bcx.ins().ineg(val);
 245
 246     if ty.is_float() {
 247         res = fx.bcx.ins().bitcast(ty, res);
 248     }
 249
 250     CValue::by_val(res, layout)
 251 }
 252
 253 macro simd_cmp {
 254     ($fx:expr, $cc:ident($x:ident, $y:ident) -> $ret:ident) => {
 255         let vector_ty = clif_vector_type($fx.tcx, $x.layout());
 256
 257         if let Some(vector_ty) = vector_ty {
 258             let x = $x.load_scalar($fx);
 259             let y = $y.load_scalar($fx);
 260             let val = $fx.bcx.ins().icmp(IntCC::$cc, x, y);
 261
 262             // HACK This depends on the fact that icmp for vectors represents bools as 0 and !0, not 0 and 1.
 263             let val = $fx.bcx.ins().raw_bitcast(vector_ty, val);
 264
 265             $ret.write_cvalue($fx, CValue::by_val(val, $ret.layout()));
 266         } else {
 267             simd_pair_for_each_lane(
 268                 $fx,
 269                 $x,
 270                 $y,
 271                 $ret,
 272                 |fx, lane_layout, res_lane_layout, x_lane, y_lane| {
 273                     let res_lane = match lane_layout.ty.kind {
 274                         ty::Uint(_) | ty::Int(_) => fx.bcx.ins().icmp(IntCC::$cc, x_lane, y_lane),
 275                         _ => unreachable!("{:?}", lane_layout.ty),
 276                     };
 277                     bool_to_zero_or_max_uint(fx, res_lane_layout, res_lane)
 278                 },
 279             );
 280         }
 281     },
 282     ($fx:expr, $cc_u:ident|$cc_s:ident($x:ident, $y:ident) -> $ret:ident) => {
 283         // FIXME use vector icmp when possible
 284         simd_pair_for_each_lane(
 285             $fx,
 286             $x,
 287             $y,
 288             $ret,
 289             |fx, lane_layout, res_lane_layout, x_lane, y_lane| {
 290                 let res_lane = match lane_layout.ty.kind {
 291                     ty::Uint(_) => fx.bcx.ins().icmp(IntCC::$cc_u, x_lane, y_lane),
 292                     ty::Int(_) => fx.bcx.ins().icmp(IntCC::$cc_s, x_lane, y_lane),
 293                     _ => unreachable!("{:?}", lane_layout.ty),
 294                 };
 295                 bool_to_zero_or_max_uint(fx, res_lane_layout, res_lane)
 296             },
 297         );
 298     },
 299 }
 300
 301 macro simd_int_binop {
 302     ($fx:expr, $op:ident($x:ident, $y:ident) -> $ret:ident) => {
 303         simd_int_binop!($fx, $op|$op($x, $y) -> $ret);
 304     },
 305     ($fx:expr, $op_u:ident|$op_s:ident($x:ident, $y:ident) -> $ret:ident) => {
 306         simd_pair_for_each_lane(
 307             $fx,
 308             $x,
 309             $y,
 310             $ret,
 311             |fx, lane_layout, ret_lane_layout, x_lane, y_lane| {
 312                 let res_lane = match lane_layout.ty.kind {
 313                     ty::Uint(_) => fx.bcx.ins().$op_u(x_lane, y_lane),
 314                     ty::Int(_) => fx.bcx.ins().$op_s(x_lane, y_lane),
 315                     _ => unreachable!("{:?}", lane_layout.ty),
 316                 };
 317                 CValue::by_val(res_lane, ret_lane_layout)
 318             },
 319         );
 320     },
 321 }
 322
 323 macro simd_int_flt_binop {
 324     ($fx:expr, $op:ident|$op_f:ident($x:ident, $y:ident) -> $ret:ident) => {
 325         simd_int_flt_binop!($fx, $op|$op|$op_f($x, $y) -> $ret);
 326     },
 327     ($fx:expr, $op_u:ident|$op_s:ident|$op_f:ident($x:ident, $y:ident) -> $ret:ident) => {
 328         simd_pair_for_each_lane(
 329             $fx,
 330             $x,
 331             $y,
 332             $ret,
 333             |fx, lane_layout, ret_lane_layout, x_lane, y_lane| {
 334                 let res_lane = match lane_layout.ty.kind {
 335                     ty::Uint(_) => fx.bcx.ins().$op_u(x_lane, y_lane),
 336                     ty::Int(_) => fx.bcx.ins().$op_s(x_lane, y_lane),
 337                     ty::Float(_) => fx.bcx.ins().$op_f(x_lane, y_lane),
 338                     _ => unreachable!("{:?}", lane_layout.ty),
 339                 };
 340                 CValue::by_val(res_lane, ret_lane_layout)
 341             },
 342         );
 343     },
 344 }
 345
 346 macro simd_flt_binop($fx:expr, $op:ident($x:ident, $y:ident) -> $ret:ident) {
 347     simd_pair_for_each_lane(
 348         $fx,
 349         $x,
 350         $y,
 351         $ret,
 352         |fx, lane_layout, ret_lane_layout, x_lane, y_lane| {
 353             let res_lane = match lane_layout.ty.kind {
 354                 ty::Float(_) => fx.bcx.ins().$op(x_lane, y_lane),
 355                 _ => unreachable!("{:?}", lane_layout.ty),
 356             };
 357             CValue::by_val(res_lane, ret_lane_layout)
 358         },
 359     );
 360 }
 361
 362 pub(crate) fn codegen_intrinsic_call<'tcx>(
 363     fx: &mut FunctionCx<'_, 'tcx, impl Backend>,
 364     instance: Instance<'tcx>,
 365     args: &[mir::Operand<'tcx>],
 366     destination: Option<(CPlace<'tcx>, BasicBlock)>,
 367     span: Span,
 368 ) {
 369     let def_id = instance.def_id();
 370     let substs = instance.substs;
 371
 372     let intrinsic = fx.tcx.item_name(def_id).as_str();
 373     let intrinsic = &intrinsic[..];
 374
 375     let ret = match destination {
 376         Some((place, _)) => place,
 377         None => {
 378             // Insert non returning intrinsics here
 379             match intrinsic {
 380                 "abort" => {
 381                     trap_panic(fx, "Called intrinsic::abort.");
 382                 }
 383                 "unreachable" => {
 384                     trap_unreachable(fx, "[corruption] Called intrinsic::unreachable.");
 385                 }
 386                 "transmute" => {
 387                     trap_unreachable(
 388                         fx,
 389                         "[corruption] Transmuting to uninhabited type.",
 390                     );
 391                 }
 392                 _ => unimplemented!("unsupported instrinsic {}", intrinsic),
 393             }
 394             return;
 395         }
 396     };
 397
 398     if intrinsic.starts_with("simd_") {
 399         self::simd::codegen_simd_intrinsic_call(fx, instance, args, ret, span);
 400         let ret_block = fx.get_block(destination.expect("SIMD intrinsics don't diverge").1);
 401         fx.bcx.ins().jump(ret_block, &[]);
 402         return;
 403     }
 404
 405     let usize_layout = fx.layout_of(fx.tcx.types.usize);
 406
 407     call_intrinsic_match! {
 408         fx, intrinsic, substs, ret, destination, args,
 409         expf32(flt) -> f32 => expf,
 410         expf64(flt) -> f64 => exp,
 411         exp2f32(flt) -> f32 => exp2f,
 412         exp2f64(flt) -> f64 => exp2,
 413         sqrtf32(flt) -> f32 => sqrtf,
 414         sqrtf64(flt) -> f64 => sqrt,
 415         powif32(a, x) -> f32 => __powisf2, // compiler-builtins
 416         powif64(a, x) -> f64 => __powidf2, // compiler-builtins
 417         powf32(a, x) -> f32 => powf,
 418         powf64(a, x) -> f64 => pow,
 419         logf32(flt) -> f32 => logf,
 420         logf64(flt) -> f64 => log,
 421         log2f32(flt) -> f32 => log2f,
 422         log2f64(flt) -> f64 => log2,
 423         log10f32(flt) -> f32 => log10f,
 424         log10f64(flt) -> f64 => log10,
 425         fabsf32(flt) -> f32 => fabsf,
 426         fabsf64(flt) -> f64 => fabs,
 427         fmaf32(x, y, z) -> f32 => fmaf,
 428         fmaf64(x, y, z) -> f64 => fma,
 429         copysignf32(x, y) -> f32 => copysignf,
 430         copysignf64(x, y) -> f64 => copysign,
 431
 432         // rounding variants
 433         // FIXME use clif insts
 434         floorf32(flt) -> f32 => floorf,
 435         floorf64(flt) -> f64 => floor,
 436         ceilf32(flt) -> f32 => ceilf,
 437         ceilf64(flt) -> f64 => ceil,
 438         truncf32(flt) -> f32 => truncf,
 439         truncf64(flt) -> f64 => trunc,
 440         roundf32(flt) -> f32 => roundf,
 441         roundf64(flt) -> f64 => round,
 442
 443         // trigonometry
 444         sinf32(flt) -> f32 => sinf,
 445         sinf64(flt) -> f64 => sin,
 446         cosf32(flt) -> f32 => cosf,
 447         cosf64(flt) -> f64 => cos,
 448         tanf32(flt) -> f32 => tanf,
 449         tanf64(flt) -> f64 => tan,
 450     }
 451
 452     intrinsic_match! {
 453         fx, intrinsic, substs, args,
 454         _ => {
 455             fx.tcx.sess.span_fatal(span, &format!("unsupported intrinsic {}", intrinsic));
 456         };
 457
 458         assume, (c _a) {};
 459         likely | unlikely, (c a) {
 460             ret.write_cvalue(fx, a);
 461         };
 462         breakpoint, () {
 463             fx.bcx.ins().debugtrap();
 464         };
 465         copy | copy_nonoverlapping, <elem_ty> (v src, v dst, v count) {
 466             let elem_size: u64 = fx.layout_of(elem_ty).size.bytes();
 467             let elem_size = fx
 468                 .bcx
 469                 .ins()
 470                 .iconst(fx.pointer_type, elem_size as i64);
 471             assert_eq!(args.len(), 3);
 472             let byte_amount = fx.bcx.ins().imul(count, elem_size);
 473
 474             if intrinsic.contains("nonoverlapping") {
 475                 // FIXME emit_small_memcpy
 476                 fx.bcx.call_memcpy(fx.module.target_config(), dst, src, byte_amount);
 477             } else {
 478                 // FIXME emit_small_memmove
 479                 fx.bcx.call_memmove(fx.module.target_config(), dst, src, byte_amount);
 480             }
 481         };
 482         // NOTE: the volatile variants have src and dst swapped
 483         volatile_copy_memory | volatile_copy_nonoverlapping_memory, <elem_ty> (v dst, v src, v count) {
 484             let elem_size: u64 = fx.layout_of(elem_ty).size.bytes();
 485             let elem_size = fx
 486                 .bcx
 487                 .ins()
 488                 .iconst(fx.pointer_type, elem_size as i64);
 489             assert_eq!(args.len(), 3);
 490             let byte_amount = fx.bcx.ins().imul(count, elem_size);
 491
 492             // FIXME make the copy actually volatile when using emit_small_mem{cpy,move}
 493             if intrinsic.contains("nonoverlapping") {
 494                 // FIXME emit_small_memcpy
 495                 fx.bcx.call_memcpy(fx.module.target_config(), dst, src, byte_amount);
 496             } else {
 497                 // FIXME emit_small_memmove
 498                 fx.bcx.call_memmove(fx.module.target_config(), dst, src, byte_amount);
 499             }
 500         };
 501         discriminant_value, (c ptr) {
 502             let pointee_layout = fx.layout_of(ptr.layout().ty.builtin_deref(true).unwrap().ty);
 503             let val = CValue::by_ref(Pointer::new(ptr.load_scalar(fx)), pointee_layout);
 504             let discr = crate::discriminant::codegen_get_discriminant(fx, val, ret.layout());
 505             ret.write_cvalue(fx, discr);
 506         };
 507         size_of_val, <T> (c ptr) {
 508             let layout = fx.layout_of(T);
 509             let size = if layout.is_unsized() {
 510                 let (_ptr, info) = ptr.load_scalar_pair(fx);
 511                 let (size, _align) = crate::unsize::size_and_align_of_dst(fx, layout, info);
 512                 size
 513             } else {
 514                 fx
 515                     .bcx
 516                     .ins()
 517                     .iconst(fx.pointer_type, layout.size.bytes() as i64)
 518             };
 519             ret.write_cvalue(fx, CValue::by_val(size, usize_layout));
 520         };
 521         min_align_of_val, <T> (c ptr) {
 522             let layout = fx.layout_of(T);
 523             let align = if layout.is_unsized() {
 524                 let (_ptr, info) = ptr.load_scalar_pair(fx);
 525                 let (_size, align) = crate::unsize::size_and_align_of_dst(fx, layout, info);
 526                 align
 527             } else {
 528                 fx
 529                     .bcx
 530                     .ins()
 531                     .iconst(fx.pointer_type, layout.align.abi.bytes() as i64)
 532             };
 533             ret.write_cvalue(fx, CValue::by_val(align, usize_layout));
 534         };
 535
 536         _ if intrinsic.starts_with("unchecked_") || intrinsic == "exact_div", (c x, c y) {
 537             // FIXME trap on overflow
 538             let bin_op = match intrinsic {
 539                 "unchecked_add" => BinOp::Add,
 540                 "unchecked_sub" => BinOp::Sub,
 541                 "unchecked_div" | "exact_div" => BinOp::Div,
 542                 "unchecked_rem" => BinOp::Rem,
 543                 "unchecked_shl" => BinOp::Shl,
 544                 "unchecked_shr" => BinOp::Shr,
 545                 _ => unreachable!("intrinsic {}", intrinsic),
 546             };
 547             let res = crate::num::trans_int_binop(fx, bin_op, x, y);
 548             ret.write_cvalue(fx, res);
 549         };
 550         _ if intrinsic.ends_with("_with_overflow"), (c x, c y) {
 551             assert_eq!(x.layout().ty, y.layout().ty);
 552             let bin_op = match intrinsic {
 553                 "add_with_overflow" => BinOp::Add,
 554                 "sub_with_overflow" => BinOp::Sub,
 555                 "mul_with_overflow" => BinOp::Mul,
 556                 _ => unreachable!("intrinsic {}", intrinsic),
 557             };
 558
 559             let res = crate::num::trans_checked_int_binop(
 560                 fx,
 561                 bin_op,
 562                 x,
 563                 y,
 564             );
 565             ret.write_cvalue(fx, res);
 566         };
 567         _ if intrinsic.starts_with("wrapping_"), (c x, c y) {
 568             assert_eq!(x.layout().ty, y.layout().ty);
 569             let bin_op = match intrinsic {
 570                 "wrapping_add" => BinOp::Add,
 571                 "wrapping_sub" => BinOp::Sub,
 572                 "wrapping_mul" => BinOp::Mul,
 573                 _ => unreachable!("intrinsic {}", intrinsic),
 574             };
 575             let res = crate::num::trans_int_binop(
 576                 fx,
 577                 bin_op,
 578                 x,
 579                 y,
 580             );
 581             ret.write_cvalue(fx, res);
 582         };
 583         _ if intrinsic.starts_with("saturating_"), <T> (c lhs, c rhs) {
 584             assert_eq!(lhs.layout().ty, rhs.layout().ty);
 585             let bin_op = match intrinsic {
 586                 "saturating_add" => BinOp::Add,
 587                 "saturating_sub" => BinOp::Sub,
 588                 _ => unreachable!("intrinsic {}", intrinsic),
 589             };
 590
 591             let signed = type_sign(T);
 592
 593             let checked_res = crate::num::trans_checked_int_binop(
 594                 fx,
 595                 bin_op,
 596                 lhs,
 597                 rhs,
 598             );
 599
 600             let (val, has_overflow) = checked_res.load_scalar_pair(fx);
 601             let clif_ty = fx.clif_type(T).unwrap();
 602
 603             // `select.i8` is not implemented by Cranelift.
 604             let has_overflow = fx.bcx.ins().uextend(types::I32, has_overflow);
 605
 606             let (min, max) = type_min_max_value(&mut fx.bcx, clif_ty, signed);
 607
 608             let val = match (intrinsic, signed) {
 609                 ("saturating_add", false) => fx.bcx.ins().select(has_overflow, max, val),
 610                 ("saturating_sub", false) => fx.bcx.ins().select(has_overflow, min, val),
 611                 ("saturating_add", true) => {
 612                     let rhs = rhs.load_scalar(fx);
 613                     let rhs_ge_zero = fx.bcx.ins().icmp_imm(IntCC::SignedGreaterThanOrEqual, rhs, 0);
 614                     let sat_val = fx.bcx.ins().select(rhs_ge_zero, max, min);
 615                     fx.bcx.ins().select(has_overflow, sat_val, val)
 616                 }
 617                 ("saturating_sub", true) => {
 618                     let rhs = rhs.load_scalar(fx);
 619                     let rhs_ge_zero = fx.bcx.ins().icmp_imm(IntCC::SignedGreaterThanOrEqual, rhs, 0);
 620                     let sat_val = fx.bcx.ins().select(rhs_ge_zero, min, max);
 621                     fx.bcx.ins().select(has_overflow, sat_val, val)
 622                 }
 623                 _ => unreachable!(),
 624             };
 625
 626             let res = CValue::by_val(val, fx.layout_of(T));
 627
 628             ret.write_cvalue(fx, res);
 629         };
 630         rotate_left, <T>(v x, v y) {
 631             let layout = fx.layout_of(T);
 632             let res = fx.bcx.ins().rotl(x, y);
 633             ret.write_cvalue(fx, CValue::by_val(res, layout));
 634         };
 635         rotate_right, <T>(v x, v y) {
 636             let layout = fx.layout_of(T);
 637             let res = fx.bcx.ins().rotr(x, y);
 638             ret.write_cvalue(fx, CValue::by_val(res, layout));
 639         };
 640
 641         // The only difference between offset and arith_offset is regarding UB. Because Cranelift
 642         // doesn't have UB both are codegen'ed the same way
 643         offset | arith_offset, (c base, v offset) {
 644             let pointee_ty = base.layout().ty.builtin_deref(true).unwrap().ty;
 645             let pointee_size = fx.layout_of(pointee_ty).size.bytes();
 646             let ptr_diff = fx.bcx.ins().imul_imm(offset, pointee_size as i64);
 647             let base_val = base.load_scalar(fx);
 648             let res = fx.bcx.ins().iadd(base_val, ptr_diff);
 649             ret.write_cvalue(fx, CValue::by_val(res, base.layout()));
 650         };
 651
 652         transmute, (c from) {
 653             ret.write_cvalue_transmute(fx, from);
 654         };
 655         write_bytes | volatile_set_memory, (c dst, v val, v count) {
 656             let pointee_ty = dst.layout().ty.builtin_deref(true).unwrap().ty;
 657             let pointee_size = fx.layout_of(pointee_ty).size.bytes();
 658             let count = fx.bcx.ins().imul_imm(count, pointee_size as i64);
 659             let dst_ptr = dst.load_scalar(fx);
 660             // FIXME make the memset actually volatile when switching to emit_small_memset
 661             // FIXME use emit_small_memset
 662             fx.bcx.call_memset(fx.module.target_config(), dst_ptr, val, count);
 663         };
 664         ctlz | ctlz_nonzero, <T> (v arg) {
 665             // FIXME trap on `ctlz_nonzero` with zero arg.
 666             let res = if T == fx.tcx.types.u128 || T == fx.tcx.types.i128 {
 667                 // FIXME verify this algorithm is correct
 668                 let (lsb, msb) = fx.bcx.ins().isplit(arg);
 669                 let lsb_lz = fx.bcx.ins().clz(lsb);
 670                 let msb_lz = fx.bcx.ins().clz(msb);
 671                 let msb_is_zero = fx.bcx.ins().icmp_imm(IntCC::Equal, msb, 0);
 672                 let lsb_lz_plus_64 = fx.bcx.ins().iadd_imm(lsb_lz, 64);
 673                 let res = fx.bcx.ins().select(msb_is_zero, lsb_lz_plus_64, msb_lz);
 674                 fx.bcx.ins().uextend(types::I128, res)
 675             } else {
 676                 fx.bcx.ins().clz(arg)
 677             };
 678             let res = CValue::by_val(res, fx.layout_of(T));
 679             ret.write_cvalue(fx, res);
 680         };
 681         cttz | cttz_nonzero, <T> (v arg) {
 682             // FIXME trap on `cttz_nonzero` with zero arg.
 683             let res = if T == fx.tcx.types.u128 || T == fx.tcx.types.i128 {
 684                 // FIXME verify this algorithm is correct
 685                 let (lsb, msb) = fx.bcx.ins().isplit(arg);
 686                 let lsb_tz = fx.bcx.ins().ctz(lsb);
 687                 let msb_tz = fx.bcx.ins().ctz(msb);
 688                 let lsb_is_zero = fx.bcx.ins().icmp_imm(IntCC::Equal, lsb, 0);
 689                 let msb_tz_plus_64 = fx.bcx.ins().iadd_imm(msb_tz, 64);
 690                 let res = fx.bcx.ins().select(lsb_is_zero, msb_tz_plus_64, lsb_tz);
 691                 fx.bcx.ins().uextend(types::I128, res)
 692             } else {
 693                 fx.bcx.ins().ctz(arg)
 694             };
 695             let res = CValue::by_val(res, fx.layout_of(T));
 696             ret.write_cvalue(fx, res);
 697         };
 698         ctpop, <T> (v arg) {
 699             let res = fx.bcx.ins().popcnt(arg);
 700             let res = CValue::by_val(res, fx.layout_of(T));
 701             ret.write_cvalue(fx, res);
 702         };
 703         bitreverse, <T> (v arg) {
 704             let res = fx.bcx.ins().bitrev(arg);
 705             let res = CValue::by_val(res, fx.layout_of(T));
 706             ret.write_cvalue(fx, res);
 707         };
 708         bswap, <T> (v arg) {
 709             // FIXME(CraneStation/cranelift#794) add bswap instruction to cranelift
 710             fn swap(bcx: &mut FunctionBuilder<'_>, v: Value) -> Value {
 711                 match bcx.func.dfg.value_type(v) {
 712                     types::I8 => v,
 713
 714                     // https://code.woboq.org/gcc/include/bits/byteswap.h.html
 715                     types::I16 => {
 716                         let tmp1 = bcx.ins().ishl_imm(v, 8);
 717                         let n1 = bcx.ins().band_imm(tmp1, 0xFF00);
 718
 719                         let tmp2 = bcx.ins().ushr_imm(v, 8);
 720                         let n2 = bcx.ins().band_imm(tmp2, 0x00FF);
 721
 722                         bcx.ins().bor(n1, n2)
 723                     }
 724                     types::I32 => {
 725                         let tmp1 = bcx.ins().ishl_imm(v, 24);
 726                         let n1 = bcx.ins().band_imm(tmp1, 0xFF00_0000);
 727
 728                         let tmp2 = bcx.ins().ishl_imm(v, 8);
 729                         let n2 = bcx.ins().band_imm(tmp2, 0x00FF_0000);
 730
 731                         let tmp3 = bcx.ins().ushr_imm(v, 8);
 732                         let n3 = bcx.ins().band_imm(tmp3, 0x0000_FF00);
 733
 734                         let tmp4 = bcx.ins().ushr_imm(v, 24);
 735                         let n4 = bcx.ins().band_imm(tmp4, 0x0000_00FF);
 736
 737                         let or_tmp1 = bcx.ins().bor(n1, n2);
 738                         let or_tmp2 = bcx.ins().bor(n3, n4);
 739                         bcx.ins().bor(or_tmp1, or_tmp2)
 740                     }
 741                     types::I64 => {
 742                         let tmp1 = bcx.ins().ishl_imm(v, 56);
 743                         let n1 = bcx.ins().band_imm(tmp1, 0xFF00_0000_0000_0000u64 as i64);
 744
 745                         let tmp2 = bcx.ins().ishl_imm(v, 40);
 746                         let n2 = bcx.ins().band_imm(tmp2, 0x00FF_0000_0000_0000u64 as i64);
 747
 748                         let tmp3 = bcx.ins().ishl_imm(v, 24);
 749                         let n3 = bcx.ins().band_imm(tmp3, 0x0000_FF00_0000_0000u64 as i64);
 750
 751                         let tmp4 = bcx.ins().ishl_imm(v, 8);
 752                         let n4 = bcx.ins().band_imm(tmp4, 0x0000_00FF_0000_0000u64 as i64);
 753
 754                         let tmp5 = bcx.ins().ushr_imm(v, 8);
 755                         let n5 = bcx.ins().band_imm(tmp5, 0x0000_0000_FF00_0000u64 as i64);
 756
 757                         let tmp6 = bcx.ins().ushr_imm(v, 24);
 758                         let n6 = bcx.ins().band_imm(tmp6, 0x0000_0000_00FF_0000u64 as i64);
 759
 760                         let tmp7 = bcx.ins().ushr_imm(v, 40);
 761                         let n7 = bcx.ins().band_imm(tmp7, 0x0000_0000_0000_FF00u64 as i64);
 762
 763                         let tmp8 = bcx.ins().ushr_imm(v, 56);
 764                         let n8 = bcx.ins().band_imm(tmp8, 0x0000_0000_0000_00FFu64 as i64);
 765
 766                         let or_tmp1 = bcx.ins().bor(n1, n2);
 767                         let or_tmp2 = bcx.ins().bor(n3, n4);
 768                         let or_tmp3 = bcx.ins().bor(n5, n6);
 769                         let or_tmp4 = bcx.ins().bor(n7, n8);
 770
 771                         let or_tmp5 = bcx.ins().bor(or_tmp1, or_tmp2);
 772                         let or_tmp6 = bcx.ins().bor(or_tmp3, or_tmp4);
 773                         bcx.ins().bor(or_tmp5, or_tmp6)
 774                     }
 775                     types::I128 => {
 776                         let (lo, hi) = bcx.ins().isplit(v);
 777                         let lo = swap(bcx, lo);
 778                         let hi = swap(bcx, hi);
 779                         bcx.ins().iconcat(hi, lo)
 780                     }
 781                     ty => unreachable!("bswap {}", ty),
 782                 }
 783             };
 784             let res = CValue::by_val(swap(&mut fx.bcx, arg), fx.layout_of(T));
 785             ret.write_cvalue(fx, res);
 786         };
 787         assert_inhabited | assert_zero_valid | assert_uninit_valid, <T> () {
 788             let layout = fx.layout_of(T);
 789             if layout.abi.is_uninhabited() {
 790                 crate::trap::trap_panic(fx, &format!("attempted to instantiate uninhabited type `{}`", T));
 791                 return;
 792             }
 793
 794             if intrinsic == "assert_zero_valid" && !layout.might_permit_raw_init(fx, /*zero:*/ true).unwrap() {
 795                 crate::trap::trap_panic(fx, &format!("attempted to zero-initialize type `{}`, which is invalid", T));
 796                 return;
 797             }
 798
 799             if intrinsic == "assert_uninit_valid" && !layout.might_permit_raw_init(fx, /*zero:*/ false).unwrap() {
 800                 crate::trap::trap_panic(fx, &format!("attempted to leave type `{}` uninitialized, which is invalid", T));
 801                 return;
 802             }
 803         };
 804
 805         volatile_load, (c ptr) {
 806             // Cranelift treats loads as volatile by default
 807             // FIXME ignore during stack2reg optimization
 808             let inner_layout =
 809                 fx.layout_of(ptr.layout().ty.builtin_deref(true).unwrap().ty);
 810             let val = CValue::by_ref(Pointer::new(ptr.load_scalar(fx)), inner_layout);
 811             ret.write_cvalue(fx, val);
 812         };
 813         volatile_store, (v ptr, c val) {
 814             // Cranelift treats stores as volatile by default
 815             // FIXME ignore during stack2reg optimization
 816             let dest = CPlace::for_ptr(Pointer::new(ptr), val.layout());
 817             dest.write_cvalue(fx, val);
 818         };
 819
 820         size_of | pref_align_of | min_align_of | needs_drop | type_id | type_name, () {
 821             let const_val =
 822                 fx.tcx.const_eval_instance(ParamEnv::reveal_all(), instance, None).unwrap();
 823             let val = crate::constant::trans_const_value(
 824                 fx,
 825                 ty::Const::from_value(fx.tcx, const_val, ret.layout().ty),
 826             );
 827             ret.write_cvalue(fx, val);
 828         };
 829
 830         ptr_offset_from, <T> (v ptr, v base) {
 831             let isize_layout = fx.layout_of(fx.tcx.types.isize);
 832
 833             let pointee_size: u64 = fx.layout_of(T).size.bytes();
 834             let diff = fx.bcx.ins().isub(ptr, base);
 835             // FIXME this can be an exact division.
 836             let val = CValue::by_val(fx.bcx.ins().sdiv_imm(diff, pointee_size as i64), isize_layout);
 837             ret.write_cvalue(fx, val);
 838         };
 839
 840         ptr_guaranteed_eq, (c a, c b) {
 841             let val = crate::num::trans_ptr_binop(fx, BinOp::Eq, a, b);
 842             ret.write_cvalue(fx, val);
 843         };
 844
 845         ptr_guaranteed_ne, (c a, c b) {
 846             let val = crate::num::trans_ptr_binop(fx, BinOp::Ne, a, b);
 847             ret.write_cvalue(fx, val);
 848         };
 849
 850         caller_location, () {
 851             let caller_location = fx.get_caller_location(span);
 852             ret.write_cvalue(fx, caller_location);
 853         };
 854
 855         _ if intrinsic.starts_with("atomic_fence"), () {
 856             crate::atomic_shim::lock_global_lock(fx);
 857             crate::atomic_shim::unlock_global_lock(fx);
 858         };
 859         _ if intrinsic.starts_with("atomic_singlethreadfence"), () {
 860             crate::atomic_shim::lock_global_lock(fx);
 861             crate::atomic_shim::unlock_global_lock(fx);
 862         };
 863         _ if intrinsic.starts_with("atomic_load"), (c ptr) {
 864             crate::atomic_shim::lock_global_lock(fx);
 865
 866             let inner_layout =
 867                 fx.layout_of(ptr.layout().ty.builtin_deref(true).unwrap().ty);
 868             let val = CValue::by_ref(Pointer::new(ptr.load_scalar(fx)), inner_layout);
 869             ret.write_cvalue(fx, val);
 870
 871             crate::atomic_shim::unlock_global_lock(fx);
 872         };
 873         _ if intrinsic.starts_with("atomic_store"), (v ptr, c val) {
 874             crate::atomic_shim::lock_global_lock(fx);
 875
 876             let dest = CPlace::for_ptr(Pointer::new(ptr), val.layout());
 877             dest.write_cvalue(fx, val);
 878
 879             crate::atomic_shim::unlock_global_lock(fx);
 880         };
 881         _ if intrinsic.starts_with("atomic_xchg"), <T> (v ptr, c src) {
 882             crate::atomic_shim::lock_global_lock(fx);
 883
 884             // Read old
 885             let clif_ty = fx.clif_type(T).unwrap();
 886             let old = fx.bcx.ins().load(clif_ty, MemFlags::new(), ptr, 0);
 887             ret.write_cvalue(fx, CValue::by_val(old, fx.layout_of(T)));
 888
 889             // Write new
 890             let dest = CPlace::for_ptr(Pointer::new(ptr), src.layout());
 891             dest.write_cvalue(fx, src);
 892
 893             crate::atomic_shim::unlock_global_lock(fx);
 894         };
 895         _ if intrinsic.starts_with("atomic_cxchg"), <T> (v ptr, v test_old, v new) { // both atomic_cxchg_* and atomic_cxchgweak_*
 896             crate::atomic_shim::lock_global_lock(fx);
 897
 898             // Read old
 899             let clif_ty = fx.clif_type(T).unwrap();
 900             let old = fx.bcx.ins().load(clif_ty, MemFlags::new(), ptr, 0);
 901
 902             // Compare
 903             let is_eq = fx.bcx.ins().icmp(IntCC::Equal, old, test_old);
 904             let new = fx.bcx.ins().select(is_eq, new, old); // Keep old if not equal to test_old
 905
 906             // Write new
 907             fx.bcx.ins().store(MemFlags::new(), new, ptr, 0);
 908
 909             let ret_val = CValue::by_val_pair(old, fx.bcx.ins().bint(types::I8, is_eq), ret.layout());
 910             ret.write_cvalue(fx, ret_val);
 911
 912             crate::atomic_shim::unlock_global_lock(fx);
 913         };
 914
 915         _ if intrinsic.starts_with("atomic_xadd"), <T> (v ptr, v amount) {
 916             atomic_binop_return_old! (fx, iadd<T>(ptr, amount) -> ret);
 917         };
 918         _ if intrinsic.starts_with("atomic_xsub"), <T> (v ptr, v amount) {
 919             atomic_binop_return_old! (fx, isub<T>(ptr, amount) -> ret);
 920         };
 921         _ if intrinsic.starts_with("atomic_and"), <T> (v ptr, v src) {
 922             atomic_binop_return_old! (fx, band<T>(ptr, src) -> ret);
 923         };
 924         _ if intrinsic.starts_with("atomic_nand"), <T> (v ptr, v src) {
 925             crate::atomic_shim::lock_global_lock(fx);
 926
 927             let clif_ty = fx.clif_type(T).unwrap();
 928             let old = fx.bcx.ins().load(clif_ty, MemFlags::new(), ptr, 0);
 929             let and = fx.bcx.ins().band(old, src);
 930             let new = fx.bcx.ins().bnot(and);
 931             fx.bcx.ins().store(MemFlags::new(), new, ptr, 0);
 932             ret.write_cvalue(fx, CValue::by_val(old, fx.layout_of(T)));
 933
 934             crate::atomic_shim::unlock_global_lock(fx);
 935         };
 936         _ if intrinsic.starts_with("atomic_or"), <T> (v ptr, v src) {
 937             atomic_binop_return_old! (fx, bor<T>(ptr, src) -> ret);
 938         };
 939         _ if intrinsic.starts_with("atomic_xor"), <T> (v ptr, v src) {
 940             atomic_binop_return_old! (fx, bxor<T>(ptr, src) -> ret);
 941         };
 942
 943         _ if intrinsic.starts_with("atomic_max"), <T> (v ptr, v src) {
 944             atomic_minmax!(fx, IntCC::SignedGreaterThan, <T> (ptr, src) -> ret);
 945         };
 946         _ if intrinsic.starts_with("atomic_umax"), <T> (v ptr, v src) {
 947             atomic_minmax!(fx, IntCC::UnsignedGreaterThan, <T> (ptr, src) -> ret);
 948         };
 949         _ if intrinsic.starts_with("atomic_min"), <T> (v ptr, v src) {
 950             atomic_minmax!(fx, IntCC::SignedLessThan, <T> (ptr, src) -> ret);
 951         };
 952         _ if intrinsic.starts_with("atomic_umin"), <T> (v ptr, v src) {
 953             atomic_minmax!(fx, IntCC::UnsignedLessThan, <T> (ptr, src) -> ret);
 954         };
 955
 956         minnumf32, (v a, v b) {
 957             let val = fx.bcx.ins().fmin(a, b);
 958             let val = CValue::by_val(val, fx.layout_of(fx.tcx.types.f32));
 959             ret.write_cvalue(fx, val);
 960         };
 961         minnumf64, (v a, v b) {
 962             let val = fx.bcx.ins().fmin(a, b);
 963             let val = CValue::by_val(val, fx.layout_of(fx.tcx.types.f64));
 964             ret.write_cvalue(fx, val);
 965         };
 966         maxnumf32, (v a, v b) {
 967             let val = fx.bcx.ins().fmax(a, b);
 968             let val = CValue::by_val(val, fx.layout_of(fx.tcx.types.f32));
 969             ret.write_cvalue(fx, val);
 970         };
 971         maxnumf64, (v a, v b) {
 972             let val = fx.bcx.ins().fmax(a, b);
 973             let val = CValue::by_val(val, fx.layout_of(fx.tcx.types.f64));
 974             ret.write_cvalue(fx, val);
 975         };
 976
 977         try, (v f, v data, v _catch_fn) {
 978             // FIXME once unwinding is supported, change this to actually catch panics
 979             let f_sig = fx.bcx.func.import_signature(Signature {
 980                 call_conv: CallConv::triple_default(fx.triple()),
 981                 params: vec![AbiParam::new(fx.bcx.func.dfg.value_type(data))],
 982                 returns: vec![],
 983             });
 984
 985             fx.bcx.ins().call_indirect(f_sig, f, &[data]);
 986
 987             let ret_val = CValue::const_val(fx, ret.layout(), 0);
 988             ret.write_cvalue(fx, ret_val);
 989         };
 990
 991         fadd_fast | fsub_fast | fmul_fast | fdiv_fast | frem_fast, (c x, c y) {
 992             let res = crate::num::trans_float_binop(fx, match intrinsic {
 993                 "fadd_fast" => BinOp::Add,
 994                 "fsub_fast" => BinOp::Sub,
 995                 "fmul_fast" => BinOp::Mul,
 996                 "fdiv_fast" => BinOp::Div,
 997                 "frem_fast" => BinOp::Rem,
 998                 _ => unreachable!(),
 999             }, x, y);
1000             ret.write_cvalue(fx, res);
1001         };
1002         float_to_int_unchecked, (v f) {
1003             let res = crate::cast::clif_int_or_float_cast(
1004                 fx,
1005                 f,
1006                 false,
1007                 fx.clif_type(ret.layout().ty).unwrap(),
1008                 type_sign(ret.layout().ty),
1009             );
1010             ret.write_cvalue(fx, CValue::by_val(res, ret.layout()));
1011         };
1012     }
1013
1014     if let Some((_, dest)) = destination {
1015         let ret_block = fx.get_block(dest);
1016         fx.bcx.ins().jump(ret_block, &[]);
1017     } else {
1018         trap_unreachable(fx, "[corruption] Diverging intrinsic returned.");
1019     }
1020 }