src/intrinsics/mod.rs

   1 pub mod llvm;
   2 mod simd;
   3
   4 use crate::prelude::*;
   5
   6 macro intrinsic_pat {
   7     (_) => {
   8         _
   9     },
  10     ($name:ident) => {
  11         stringify!($name)
  12     },
  13     ($name:literal) => {
  14         stringify!($name)
  15     },
  16     ($x:ident . $($xs:tt).*) => {
  17         concat!(stringify!($x), ".", intrinsic_pat!($($xs).*))
  18     }
  19 }
  20
  21 macro intrinsic_arg {
  22     (o $fx:expr, $arg:ident) => {
  23         $arg
  24     },
  25     (c $fx:expr, $arg:ident) => {
  26         trans_operand($fx, $arg)
  27     },
  28     (v $fx:expr, $arg:ident) => {
  29         trans_operand($fx, $arg).load_scalar($fx)
  30     }
  31 }
  32
  33 macro intrinsic_substs {
  34     ($substs:expr, $index:expr,) => {},
  35     ($substs:expr, $index:expr, $first:ident $(,$rest:ident)*) => {
  36         let $first = $substs.type_at($index);
  37         intrinsic_substs!($substs, $index+1, $($rest),*);
  38     }
  39 }
  40
  41 macro intrinsic_match {
  42     ($fx:expr, $intrinsic:expr, $substs:expr, $args:expr,
  43     _ => $unknown:block;
  44     $(
  45         $($($name:tt).*)|+ $(if $cond:expr)?, $(<$($subst:ident),*>)? ($($a:ident $arg:ident),*) $content:block;
  46     )*) => {
  47         match $intrinsic {
  48             $(
  49                 $(intrinsic_pat!($($name).*))|* $(if $cond)? => {
  50                     #[allow(unused_parens, non_snake_case)]
  51                     {
  52                         $(
  53                             intrinsic_substs!($substs, 0, $($subst),*);
  54                         )?
  55                         if let [$($arg),*] = $args {
  56                             let ($($arg,)*) = (
  57                                 $(intrinsic_arg!($a $fx, $arg),)*
  58                             );
  59                             #[warn(unused_parens, non_snake_case)]
  60                             {
  61                                 $content
  62                             }
  63                         } else {
  64                             bug!("wrong number of args for intrinsic {:?}", $intrinsic);
  65                         }
  66                     }
  67                 }
  68             )*
  69             _ => $unknown,
  70         }
  71     }
  72 }
  73
  74 macro call_intrinsic_match {
  75     ($fx:expr, $intrinsic:expr, $substs:expr, $ret:expr, $destination:expr, $args:expr, $(
  76         $name:ident($($arg:ident),*) -> $ty:ident => $func:ident,
  77     )*) => {
  78         match $intrinsic {
  79             $(
  80                 stringify!($name) => {
  81                     assert!($substs.is_noop());
  82                     if let [$(ref $arg),*] = *$args {
  83                         let ($($arg,)*) = (
  84                             $(trans_operand($fx, $arg),)*
  85                         );
  86                         let res = $fx.easy_call(stringify!($func), &[$($arg),*], $fx.tcx.types.$ty);
  87                         $ret.write_cvalue($fx, res);
  88
  89                         if let Some((_, dest)) = $destination {
  90                             let ret_ebb = $fx.get_ebb(dest);
  91                             $fx.bcx.ins().jump(ret_ebb, &[]);
  92                             return;
  93                         } else {
  94                             unreachable!();
  95                         }
  96                     } else {
  97                         bug!("wrong number of args for intrinsic {:?}", $intrinsic);
  98                     }
  99                 }
 100             )*
 101             _ => {}
 102         }
 103     }
 104 }
 105
 106 macro atomic_binop_return_old($fx:expr, $op:ident<$T:ident>($ptr:ident, $src:ident) -> $ret:ident)  {
 107     let clif_ty = $fx.clif_type($T).unwrap();
 108     let old = $fx.bcx.ins().load(clif_ty, MemFlags::new(), $ptr, 0);
 109     let new = $fx.bcx.ins().$op(old, $src);
 110     $fx.bcx.ins().store(MemFlags::new(), new, $ptr, 0);
 111     $ret.write_cvalue($fx, CValue::by_val(old, $fx.layout_of($T)));
 112 }
 113
 114 macro atomic_minmax($fx:expr, $cc:expr, <$T:ident> ($ptr:ident, $src:ident) -> $ret:ident) {
 115     // Read old
 116     let clif_ty = $fx.clif_type($T).unwrap();
 117     let old = $fx.bcx.ins().load(clif_ty, MemFlags::new(), $ptr, 0);
 118
 119     // Compare
 120     let is_eq = codegen_icmp($fx, IntCC::SignedGreaterThan, old, $src);
 121     let new = $fx.bcx.ins().select(is_eq, old, $src);
 122
 123     // Write new
 124     $fx.bcx.ins().store(MemFlags::new(), new, $ptr, 0);
 125
 126     let ret_val = CValue::by_val(old, $ret.layout());
 127     $ret.write_cvalue($fx, ret_val);
 128 }
 129
 130 fn lane_type_and_count<'tcx>(
 131     fx: &FunctionCx<'_, 'tcx, impl Backend>,
 132     layout: TyLayout<'tcx>,
 133     intrinsic: &str,
 134 ) -> (TyLayout<'tcx>, u32) {
 135     assert!(layout.ty.is_simd());
 136     let lane_count = match layout.fields {
 137         layout::FieldPlacement::Array { stride: _, count } => u32::try_from(count).unwrap(),
 138         _ => panic!(
 139             "Non vector type {:?} passed to or returned from simd_* intrinsic {}",
 140             layout.ty, intrinsic
 141         ),
 142     };
 143     let lane_layout = layout.field(fx, 0);
 144     (lane_layout, lane_count)
 145 }
 146
 147 fn simd_for_each_lane<'tcx, B: Backend>(
 148     fx: &mut FunctionCx<'_, 'tcx, B>,
 149     intrinsic: &str,
 150     x: CValue<'tcx>,
 151     y: CValue<'tcx>,
 152     ret: CPlace<'tcx>,
 153     f: impl Fn(
 154         &mut FunctionCx<'_, 'tcx, B>,
 155         TyLayout<'tcx>,
 156         TyLayout<'tcx>,
 157         Value,
 158         Value,
 159     ) -> CValue<'tcx>,
 160 ) {
 161     assert_eq!(x.layout(), y.layout());
 162     let layout = x.layout();
 163
 164     let (lane_layout, lane_count) = lane_type_and_count(fx, layout, intrinsic);
 165     let (ret_lane_layout, ret_lane_count) = lane_type_and_count(fx, ret.layout(), intrinsic);
 166     assert_eq!(lane_count, ret_lane_count);
 167
 168     for lane in 0..lane_count {
 169         let lane = mir::Field::new(lane.try_into().unwrap());
 170         let x_lane = x.value_field(fx, lane).load_scalar(fx);
 171         let y_lane = y.value_field(fx, lane).load_scalar(fx);
 172
 173         let res_lane = f(fx, lane_layout, ret_lane_layout, x_lane, y_lane);
 174
 175         ret.place_field(fx, lane).write_cvalue(fx, res_lane);
 176     }
 177 }
 178
 179 fn bool_to_zero_or_max_uint<'tcx>(
 180     fx: &mut FunctionCx<'_, 'tcx, impl Backend>,
 181     layout: TyLayout<'tcx>,
 182     val: Value,
 183 ) -> CValue<'tcx> {
 184     let ty = fx.clif_type(layout.ty).unwrap();
 185
 186     let int_ty = match ty {
 187         types::F32 => types::I32,
 188         types::F64 => types::I64,
 189         ty => ty,
 190     };
 191
 192     let zero = fx.bcx.ins().iconst(int_ty, 0);
 193     let max = fx
 194         .bcx
 195         .ins()
 196         .iconst(int_ty, (u64::max_value() >> (64 - int_ty.bits())) as i64);
 197     let mut res = fx.bcx.ins().select(val, max, zero);
 198
 199     if ty.is_float() {
 200         res = fx.bcx.ins().bitcast(ty, res);
 201     }
 202
 203     CValue::by_val(res, layout)
 204 }
 205
 206 macro simd_cmp {
 207     ($fx:expr, $intrinsic:expr, $cc:ident($x:ident, $y:ident) -> $ret:ident) => {
 208         simd_for_each_lane(
 209             $fx,
 210             $intrinsic,
 211             $x,
 212             $y,
 213             $ret,
 214             |fx, lane_layout, res_lane_layout, x_lane, y_lane| {
 215                 let res_lane = match lane_layout.ty.kind {
 216                     ty::Uint(_) | ty::Int(_) => codegen_icmp(fx, IntCC::$cc, x_lane, y_lane),
 217                     _ => unreachable!("{:?}", lane_layout.ty),
 218                 };
 219                 bool_to_zero_or_max_uint(fx, res_lane_layout, res_lane)
 220             },
 221         );
 222     },
 223     ($fx:expr, $intrinsic:expr, $cc_u:ident|$cc_s:ident($x:ident, $y:ident) -> $ret:ident) => {
 224         simd_for_each_lane(
 225             $fx,
 226             $intrinsic,
 227             $x,
 228             $y,
 229             $ret,
 230             |fx, lane_layout, res_lane_layout, x_lane, y_lane| {
 231                 let res_lane = match lane_layout.ty.kind {
 232                     ty::Uint(_) => codegen_icmp(fx, IntCC::$cc_u, x_lane, y_lane),
 233                     ty::Int(_) => codegen_icmp(fx, IntCC::$cc_s, x_lane, y_lane),
 234                     _ => unreachable!("{:?}", lane_layout.ty),
 235                 };
 236                 bool_to_zero_or_max_uint(fx, res_lane_layout, res_lane)
 237             },
 238         );
 239     },
 240 }
 241
 242 macro simd_int_binop {
 243     ($fx:expr, $intrinsic:expr, $op:ident($x:ident, $y:ident) -> $ret:ident) => {
 244         simd_for_each_lane(
 245             $fx,
 246             $intrinsic,
 247             $x,
 248             $y,
 249             $ret,
 250             |fx, lane_layout, ret_lane_layout, x_lane, y_lane| {
 251                 let res_lane = match lane_layout.ty.kind {
 252                     ty::Uint(_) | ty::Int(_) => fx.bcx.ins().$op(x_lane, y_lane),
 253                     _ => unreachable!("{:?}", lane_layout.ty),
 254                 };
 255                 CValue::by_val(res_lane, ret_lane_layout)
 256             },
 257         );
 258     },
 259     ($fx:expr, $intrinsic:expr, $op_u:ident|$op_s:ident($x:ident, $y:ident) -> $ret:ident) => {
 260         simd_for_each_lane(
 261             $fx,
 262             $intrinsic,
 263             $x,
 264             $y,
 265             $ret,
 266             |fx, lane_layout, ret_lane_layout, x_lane, y_lane| {
 267                 let res_lane = match lane_layout.ty.kind {
 268                     ty::Uint(_) => fx.bcx.ins().$op_u(x_lane, y_lane),
 269                     ty::Int(_) => fx.bcx.ins().$op_s(x_lane, y_lane),
 270                     _ => unreachable!("{:?}", lane_layout.ty),
 271                 };
 272                 CValue::by_val(res_lane, ret_lane_layout)
 273             },
 274         );
 275     },
 276 }
 277
 278 macro simd_int_flt_binop {
 279     ($fx:expr, $intrinsic:expr, $op:ident|$op_f:ident($x:ident, $y:ident) -> $ret:ident) => {
 280         simd_for_each_lane(
 281             $fx,
 282             $intrinsic,
 283             $x,
 284             $y,
 285             $ret,
 286             |fx, lane_layout, ret_lane_layout, x_lane, y_lane| {
 287                 let res_lane = match lane_layout.ty.kind {
 288                     ty::Uint(_) | ty::Int(_) => fx.bcx.ins().$op(x_lane, y_lane),
 289                     ty::Float(_) => fx.bcx.ins().$op_f(x_lane, y_lane),
 290                     _ => unreachable!("{:?}", lane_layout.ty),
 291                 };
 292                 CValue::by_val(res_lane, ret_lane_layout)
 293             },
 294         );
 295     },
 296     ($fx:expr, $intrinsic:expr, $op_u:ident|$op_s:ident|$op_f:ident($x:ident, $y:ident) -> $ret:ident) => {
 297         simd_for_each_lane(
 298             $fx,
 299             $intrinsic,
 300             $x,
 301             $y,
 302             $ret,
 303             |fx, lane_layout, ret_lane_layout, x_lane, y_lane| {
 304                 let res_lane = match lane_layout.ty.kind {
 305                     ty::Uint(_) => fx.bcx.ins().$op_u(x_lane, y_lane),
 306                     ty::Int(_) => fx.bcx.ins().$op_s(x_lane, y_lane),
 307                     ty::Float(_) => fx.bcx.ins().$op_f(x_lane, y_lane),
 308                     _ => unreachable!("{:?}", lane_layout.ty),
 309                 };
 310                 CValue::by_val(res_lane, ret_lane_layout)
 311             },
 312         );
 313     },
 314 }
 315
 316 macro simd_flt_binop($fx:expr, $intrinsic:expr, $op:ident($x:ident, $y:ident) -> $ret:ident) {
 317     simd_for_each_lane(
 318         $fx,
 319         $intrinsic,
 320         $x,
 321         $y,
 322         $ret,
 323         |fx, lane_layout, ret_lane_layout, x_lane, y_lane| {
 324             let res_lane = match lane_layout.ty.kind {
 325                 ty::Float(_) => fx.bcx.ins().$op(x_lane, y_lane),
 326                 _ => unreachable!("{:?}", lane_layout.ty),
 327             };
 328             CValue::by_val(res_lane, ret_lane_layout)
 329         },
 330     );
 331 }
 332
 333 pub fn codegen_intrinsic_call<'tcx>(
 334     fx: &mut FunctionCx<'_, 'tcx, impl Backend>,
 335     instance: Instance<'tcx>,
 336     args: &[mir::Operand<'tcx>],
 337     destination: Option<(CPlace<'tcx>, BasicBlock)>,
 338     span: Span,
 339 ) {
 340     let def_id = instance.def_id();
 341     let substs = instance.substs;
 342
 343     let intrinsic = fx.tcx.item_name(def_id).as_str();
 344     let intrinsic = &intrinsic[..];
 345
 346     let ret = match destination {
 347         Some((place, _)) => place,
 348         None => {
 349             // Insert non returning intrinsics here
 350             match intrinsic {
 351                 "abort" => {
 352                     trap_panic(fx, "Called intrinsic::abort.");
 353                 }
 354                 "unreachable" => {
 355                     trap_unreachable(fx, "[corruption] Called intrinsic::unreachable.");
 356                 }
 357                 "transmute" => {
 358                     trap_unreachable(
 359                         fx,
 360                         "[corruption] Called intrinsic::transmute with uninhabited argument.",
 361                     );
 362                 }
 363                 _ => unimplemented!("unsupported instrinsic {}", intrinsic),
 364             }
 365             return;
 366         }
 367     };
 368
 369     if intrinsic.starts_with("simd_") {
 370         self::simd::codegen_simd_intrinsic_call(fx, instance, args, ret, span);
 371         let ret_ebb = fx.get_ebb(destination.expect("SIMD intrinsics don't diverge").1);
 372         fx.bcx.ins().jump(ret_ebb, &[]);
 373         return;
 374     }
 375
 376     let usize_layout = fx.layout_of(fx.tcx.types.usize);
 377
 378     call_intrinsic_match! {
 379         fx, intrinsic, substs, ret, destination, args,
 380         expf32(flt) -> f32 => expf,
 381         expf64(flt) -> f64 => exp,
 382         exp2f32(flt) -> f32 => exp2f,
 383         exp2f64(flt) -> f64 => exp2,
 384         sqrtf32(flt) -> f32 => sqrtf,
 385         sqrtf64(flt) -> f64 => sqrt,
 386         powif32(a, x) -> f32 => __powisf2, // compiler-builtins
 387         powif64(a, x) -> f64 => __powidf2, // compiler-builtins
 388         powf32(a, x) -> f32 => powf,
 389         powf64(a, x) -> f64 => pow,
 390         logf32(flt) -> f32 => logf,
 391         logf64(flt) -> f64 => log,
 392         log2f32(flt) -> f32 => log2f,
 393         log2f64(flt) -> f64 => log2,
 394         log10f32(flt) -> f32 => log10f,
 395         log10f64(flt) -> f64 => log10,
 396         fabsf32(flt) -> f32 => fabsf,
 397         fabsf64(flt) -> f64 => fabs,
 398         fmaf32(x, y, z) -> f32 => fmaf,
 399         fmaf64(x, y, z) -> f64 => fma,
 400         copysignf32(x, y) -> f32 => copysignf,
 401         copysignf64(x, y) -> f64 => copysign,
 402
 403         // rounding variants
 404         // FIXME use clif insts
 405         floorf32(flt) -> f32 => floorf,
 406         floorf64(flt) -> f64 => floor,
 407         ceilf32(flt) -> f32 => ceilf,
 408         ceilf64(flt) -> f64 => ceil,
 409         truncf32(flt) -> f32 => truncf,
 410         truncf64(flt) -> f64 => trunc,
 411         roundf32(flt) -> f32 => roundf,
 412         roundf64(flt) -> f64 => round,
 413
 414         // trigonometry
 415         sinf32(flt) -> f32 => sinf,
 416         sinf64(flt) -> f64 => sin,
 417         cosf32(flt) -> f32 => cosf,
 418         cosf64(flt) -> f64 => cos,
 419         tanf32(flt) -> f32 => tanf,
 420         tanf64(flt) -> f64 => tan,
 421     }
 422
 423     intrinsic_match! {
 424         fx, intrinsic, substs, args,
 425         _ => {
 426             unimpl!("unsupported intrinsic {}", intrinsic)
 427         };
 428
 429         assume, (c _a) {};
 430         likely | unlikely, (c a) {
 431             ret.write_cvalue(fx, a);
 432         };
 433         breakpoint, () {
 434             fx.bcx.ins().debugtrap();
 435         };
 436         copy | copy_nonoverlapping, <elem_ty> (v src, v dst, v count) {
 437             let elem_size: u64 = fx.layout_of(elem_ty).size.bytes();
 438             let elem_size = fx
 439                 .bcx
 440                 .ins()
 441                 .iconst(fx.pointer_type, elem_size as i64);
 442             assert_eq!(args.len(), 3);
 443             let byte_amount = fx.bcx.ins().imul(count, elem_size);
 444
 445             if intrinsic.ends_with("_nonoverlapping") {
 446                 fx.bcx.call_memcpy(fx.module.target_config(), dst, src, byte_amount);
 447             } else {
 448                 fx.bcx.call_memmove(fx.module.target_config(), dst, src, byte_amount);
 449             }
 450         };
 451         discriminant_value, (c ptr) {
 452             let pointee_layout = fx.layout_of(ptr.layout().ty.builtin_deref(true).unwrap().ty);
 453             let val = CValue::by_ref(Pointer::new(ptr.load_scalar(fx)), pointee_layout);
 454             let discr = crate::discriminant::codegen_get_discriminant(fx, val, ret.layout());
 455             ret.write_cvalue(fx, discr);
 456         };
 457         size_of_val, <T> (c ptr) {
 458             let layout = fx.layout_of(T);
 459             let size = if layout.is_unsized() {
 460                 let (_ptr, info) = ptr.load_scalar_pair(fx);
 461                 let (size, _align) = crate::unsize::size_and_align_of_dst(fx, layout.ty, info);
 462                 size
 463             } else {
 464                 fx
 465                     .bcx
 466                     .ins()
 467                     .iconst(fx.pointer_type, layout.size.bytes() as i64)
 468             };
 469             ret.write_cvalue(fx, CValue::by_val(size, usize_layout));
 470         };
 471         min_align_of_val, <T> (c ptr) {
 472             let layout = fx.layout_of(T);
 473             let align = if layout.is_unsized() {
 474                 let (_ptr, info) = ptr.load_scalar_pair(fx);
 475                 let (_size, align) = crate::unsize::size_and_align_of_dst(fx, layout.ty, info);
 476                 align
 477             } else {
 478                 fx
 479                     .bcx
 480                     .ins()
 481                     .iconst(fx.pointer_type, layout.align.abi.bytes() as i64)
 482             };
 483             ret.write_cvalue(fx, CValue::by_val(align, usize_layout));
 484         };
 485
 486         _ if intrinsic.starts_with("unchecked_") || intrinsic == "exact_div", (c x, c y) {
 487             // FIXME trap on overflow
 488             let bin_op = match intrinsic {
 489                 "unchecked_sub" => BinOp::Sub,
 490                 "unchecked_div" | "exact_div" => BinOp::Div,
 491                 "unchecked_rem" => BinOp::Rem,
 492                 "unchecked_shl" => BinOp::Shl,
 493                 "unchecked_shr" => BinOp::Shr,
 494                 _ => unimplemented!("intrinsic {}", intrinsic),
 495             };
 496             let res = crate::num::trans_int_binop(fx, bin_op, x, y);
 497             ret.write_cvalue(fx, res);
 498         };
 499         _ if intrinsic.ends_with("_with_overflow"), (c x, c y) {
 500             assert_eq!(x.layout().ty, y.layout().ty);
 501             let bin_op = match intrinsic {
 502                 "add_with_overflow" => BinOp::Add,
 503                 "sub_with_overflow" => BinOp::Sub,
 504                 "mul_with_overflow" => BinOp::Mul,
 505                 _ => unimplemented!("intrinsic {}", intrinsic),
 506             };
 507
 508             let res = crate::num::trans_checked_int_binop(
 509                 fx,
 510                 bin_op,
 511                 x,
 512                 y,
 513             );
 514             ret.write_cvalue(fx, res);
 515         };
 516         _ if intrinsic.starts_with("wrapping_"), (c x, c y) {
 517             assert_eq!(x.layout().ty, y.layout().ty);
 518             let bin_op = match intrinsic {
 519                 "wrapping_add" => BinOp::Add,
 520                 "wrapping_sub" => BinOp::Sub,
 521                 "wrapping_mul" => BinOp::Mul,
 522                 _ => unimplemented!("intrinsic {}", intrinsic),
 523             };
 524             let res = crate::num::trans_int_binop(
 525                 fx,
 526                 bin_op,
 527                 x,
 528                 y,
 529             );
 530             ret.write_cvalue(fx, res);
 531         };
 532         _ if intrinsic.starts_with("saturating_"), <T> (c lhs, c rhs) {
 533             assert_eq!(lhs.layout().ty, rhs.layout().ty);
 534             let bin_op = match intrinsic {
 535                 "saturating_add" => BinOp::Add,
 536                 "saturating_sub" => BinOp::Sub,
 537                 _ => unimplemented!("intrinsic {}", intrinsic),
 538             };
 539
 540             let signed = type_sign(T);
 541
 542             let checked_res = crate::num::trans_checked_int_binop(
 543                 fx,
 544                 bin_op,
 545                 lhs,
 546                 rhs,
 547             );
 548
 549             let (val, has_overflow) = checked_res.load_scalar_pair(fx);
 550             let clif_ty = fx.clif_type(T).unwrap();
 551
 552             // `select.i8` is not implemented by Cranelift.
 553             let has_overflow = fx.bcx.ins().uextend(types::I32, has_overflow);
 554
 555             let (min, max) = type_min_max_value(clif_ty, signed);
 556             let min = fx.bcx.ins().iconst(clif_ty, min);
 557             let max = fx.bcx.ins().iconst(clif_ty, max);
 558
 559             let val = match (intrinsic, signed) {
 560                 ("saturating_add", false) => fx.bcx.ins().select(has_overflow, max, val),
 561                 ("saturating_sub", false) => fx.bcx.ins().select(has_overflow, min, val),
 562                 ("saturating_add", true) => {
 563                     let rhs = rhs.load_scalar(fx);
 564                     let rhs_ge_zero = fx.bcx.ins().icmp_imm(IntCC::SignedGreaterThanOrEqual, rhs, 0);
 565                     let sat_val = fx.bcx.ins().select(rhs_ge_zero, max, min);
 566                     fx.bcx.ins().select(has_overflow, sat_val, val)
 567                 }
 568                 ("saturating_sub", true) => {
 569                     let rhs = rhs.load_scalar(fx);
 570                     let rhs_ge_zero = fx.bcx.ins().icmp_imm(IntCC::SignedGreaterThanOrEqual, rhs, 0);
 571                     let sat_val = fx.bcx.ins().select(rhs_ge_zero, min, max);
 572                     fx.bcx.ins().select(has_overflow, sat_val, val)
 573                 }
 574                 _ => unreachable!(),
 575             };
 576
 577             let res = CValue::by_val(val, fx.layout_of(T));
 578
 579             ret.write_cvalue(fx, res);
 580         };
 581         rotate_left, <T>(v x, v y) {
 582             let layout = fx.layout_of(T);
 583             let res = fx.bcx.ins().rotl(x, y);
 584             ret.write_cvalue(fx, CValue::by_val(res, layout));
 585         };
 586         rotate_right, <T>(v x, v y) {
 587             let layout = fx.layout_of(T);
 588             let res = fx.bcx.ins().rotr(x, y);
 589             ret.write_cvalue(fx, CValue::by_val(res, layout));
 590         };
 591
 592         // The only difference between offset and arith_offset is regarding UB. Because Cranelift
 593         // doesn't have UB both are codegen'ed the same way
 594         offset | arith_offset, (c base, v offset) {
 595             let pointee_ty = base.layout().ty.builtin_deref(true).unwrap().ty;
 596             let pointee_size = fx.layout_of(pointee_ty).size.bytes();
 597             let ptr_diff = fx.bcx.ins().imul_imm(offset, pointee_size as i64);
 598             let base_val = base.load_scalar(fx);
 599             let res = fx.bcx.ins().iadd(base_val, ptr_diff);
 600             ret.write_cvalue(fx, CValue::by_val(res, base.layout()));
 601         };
 602
 603         transmute, <src_ty, dst_ty> (c from) {
 604             assert_eq!(from.layout().ty, src_ty);
 605             let addr = from.force_stack(fx);
 606             let dst_layout = fx.layout_of(dst_ty);
 607             ret.write_cvalue(fx, CValue::by_ref(addr, dst_layout))
 608         };
 609         init, () {
 610             let layout = ret.layout();
 611             if layout.abi == Abi::Uninhabited {
 612                 crate::trap::trap_panic(fx, "[panic] Called intrinsic::init for uninhabited type.");
 613                 return;
 614             }
 615
 616             match *ret.inner() {
 617                 CPlaceInner::NoPlace => {}
 618                 CPlaceInner::Var(var) => {
 619                     let clif_ty = fx.clif_type(layout.ty).unwrap();
 620                     let val = match clif_ty {
 621                         types::I8 | types::I16 | types::I32 | types::I64 => fx.bcx.ins().iconst(clif_ty, 0),
 622                         types::I128 => {
 623                             let zero = fx.bcx.ins().iconst(types::I64, 0);
 624                             fx.bcx.ins().iconcat(zero, zero)
 625                         }
 626                         types::F32 => {
 627                             let zero = fx.bcx.ins().iconst(types::I32, 0);
 628                             fx.bcx.ins().bitcast(types::F32, zero)
 629                         }
 630                         types::F64 => {
 631                             let zero = fx.bcx.ins().iconst(types::I64, 0);
 632                             fx.bcx.ins().bitcast(types::F64, zero)
 633                         }
 634                         _ => panic!("clif_type returned {}", clif_ty),
 635                     };
 636                     fx.bcx.set_val_label(val, cranelift::codegen::ir::ValueLabel::from_u32(var.as_u32()));
 637                     fx.bcx.def_var(mir_var(var), val);
 638                 }
 639                 _ => {
 640                     let addr = ret.to_ptr(fx).get_addr(fx);
 641                     let layout = ret.layout();
 642                     fx.bcx.emit_small_memset(fx.module.target_config(), addr, 0, layout.size.bytes(), 1);
 643                 }
 644             }
 645         };
 646         uninit, () {
 647             let layout = ret.layout();
 648             if layout.abi == Abi::Uninhabited {
 649                 crate::trap::trap_panic(fx, "[panic] Called intrinsic::uninit for uninhabited type.");
 650                 return;
 651             }
 652             match *ret.inner() {
 653                 CPlaceInner::NoPlace => {},
 654                 CPlaceInner::Var(var) => {
 655                     let clif_ty = fx.clif_type(layout.ty).unwrap();
 656                     let val = match clif_ty {
 657                         types::I8 | types::I16 | types::I32 | types::I64 => fx.bcx.ins().iconst(clif_ty, 42),
 658                         types::I128 => {
 659                             let zero = fx.bcx.ins().iconst(types::I64, 0);
 660                             let fourty_two = fx.bcx.ins().iconst(types::I64, 42);
 661                             fx.bcx.ins().iconcat(fourty_two, zero)
 662                         }
 663                         types::F32 => {
 664                             let zero = fx.bcx.ins().iconst(types::I32, 0xdeadbeef);
 665                             fx.bcx.ins().bitcast(types::F32, zero)
 666                         }
 667                         types::F64 => {
 668                             let zero = fx.bcx.ins().iconst(types::I64, 0xcafebabedeadbeefu64 as i64);
 669                             fx.bcx.ins().bitcast(types::F64, zero)
 670                         }
 671                         _ => panic!("clif_type returned {}", clif_ty),
 672                     };
 673                     fx.bcx.set_val_label(val, cranelift::codegen::ir::ValueLabel::from_u32(var.as_u32()));
 674                     fx.bcx.def_var(mir_var(var), val);
 675                 }
 676                 CPlaceInner::Addr(_, _) => {
 677                     // Don't write to `ret`, as the destination memory is already uninitialized.
 678                 }
 679             }
 680         };
 681         write_bytes, (c dst, v val, v count) {
 682             let pointee_ty = dst.layout().ty.builtin_deref(true).unwrap().ty;
 683             let pointee_size = fx.layout_of(pointee_ty).size.bytes();
 684             let count = fx.bcx.ins().imul_imm(count, pointee_size as i64);
 685             let dst_ptr = dst.load_scalar(fx);
 686             fx.bcx.call_memset(fx.module.target_config(), dst_ptr, val, count);
 687         };
 688         ctlz | ctlz_nonzero, <T> (v arg) {
 689             // FIXME trap on `ctlz_nonzero` with zero arg.
 690             let res = if T == fx.tcx.types.u128 || T == fx.tcx.types.i128 {
 691                 // FIXME verify this algorithm is correct
 692                 let (lsb, msb) = fx.bcx.ins().isplit(arg);
 693                 let lsb_lz = fx.bcx.ins().clz(lsb);
 694                 let msb_lz = fx.bcx.ins().clz(msb);
 695                 let msb_is_zero = fx.bcx.ins().icmp_imm(IntCC::Equal, msb, 0);
 696                 let lsb_lz_plus_64 = fx.bcx.ins().iadd_imm(lsb_lz, 64);
 697                 let res = fx.bcx.ins().select(msb_is_zero, lsb_lz_plus_64, msb_lz);
 698                 fx.bcx.ins().uextend(types::I128, res)
 699             } else {
 700                 fx.bcx.ins().clz(arg)
 701             };
 702             let res = CValue::by_val(res, fx.layout_of(T));
 703             ret.write_cvalue(fx, res);
 704         };
 705         cttz | cttz_nonzero, <T> (v arg) {
 706             // FIXME trap on `cttz_nonzero` with zero arg.
 707             let res = if T == fx.tcx.types.u128 || T == fx.tcx.types.i128 {
 708                 // FIXME verify this algorithm is correct
 709                 let (lsb, msb) = fx.bcx.ins().isplit(arg);
 710                 let lsb_tz = fx.bcx.ins().ctz(lsb);
 711                 let msb_tz = fx.bcx.ins().ctz(msb);
 712                 let lsb_is_zero = fx.bcx.ins().icmp_imm(IntCC::Equal, lsb, 0);
 713                 let msb_tz_plus_64 = fx.bcx.ins().iadd_imm(msb_tz, 64);
 714                 let res = fx.bcx.ins().select(lsb_is_zero, msb_tz_plus_64, lsb_tz);
 715                 fx.bcx.ins().uextend(types::I128, res)
 716             } else {
 717                 fx.bcx.ins().ctz(arg)
 718             };
 719             let res = CValue::by_val(res, fx.layout_of(T));
 720             ret.write_cvalue(fx, res);
 721         };
 722         ctpop, <T> (v arg) {
 723             let res = fx.bcx.ins().popcnt(arg);
 724             let res = CValue::by_val(res, fx.layout_of(T));
 725             ret.write_cvalue(fx, res);
 726         };
 727         bitreverse, <T> (v arg) {
 728             let res = fx.bcx.ins().bitrev(arg);
 729             let res = CValue::by_val(res, fx.layout_of(T));
 730             ret.write_cvalue(fx, res);
 731         };
 732         bswap, <T> (v arg) {
 733             // FIXME(CraneStation/cranelift#794) add bswap instruction to cranelift
 734             fn swap(bcx: &mut FunctionBuilder, v: Value) -> Value {
 735                 match bcx.func.dfg.value_type(v) {
 736                     types::I8 => v,
 737
 738                     // https://code.woboq.org/gcc/include/bits/byteswap.h.html
 739                     types::I16 => {
 740                         let tmp1 = bcx.ins().ishl_imm(v, 8);
 741                         let n1 = bcx.ins().band_imm(tmp1, 0xFF00);
 742
 743                         let tmp2 = bcx.ins().ushr_imm(v, 8);
 744                         let n2 = bcx.ins().band_imm(tmp2, 0x00FF);
 745
 746                         bcx.ins().bor(n1, n2)
 747                     }
 748                     types::I32 => {
 749                         let tmp1 = bcx.ins().ishl_imm(v, 24);
 750                         let n1 = bcx.ins().band_imm(tmp1, 0xFF00_0000);
 751
 752                         let tmp2 = bcx.ins().ishl_imm(v, 8);
 753                         let n2 = bcx.ins().band_imm(tmp2, 0x00FF_0000);
 754
 755                         let tmp3 = bcx.ins().ushr_imm(v, 8);
 756                         let n3 = bcx.ins().band_imm(tmp3, 0x0000_FF00);
 757
 758                         let tmp4 = bcx.ins().ushr_imm(v, 24);
 759                         let n4 = bcx.ins().band_imm(tmp4, 0x0000_00FF);
 760
 761                         let or_tmp1 = bcx.ins().bor(n1, n2);
 762                         let or_tmp2 = bcx.ins().bor(n3, n4);
 763                         bcx.ins().bor(or_tmp1, or_tmp2)
 764                     }
 765                     types::I64 => {
 766                         let tmp1 = bcx.ins().ishl_imm(v, 56);
 767                         let n1 = bcx.ins().band_imm(tmp1, 0xFF00_0000_0000_0000u64 as i64);
 768
 769                         let tmp2 = bcx.ins().ishl_imm(v, 40);
 770                         let n2 = bcx.ins().band_imm(tmp2, 0x00FF_0000_0000_0000u64 as i64);
 771
 772                         let tmp3 = bcx.ins().ishl_imm(v, 24);
 773                         let n3 = bcx.ins().band_imm(tmp3, 0x0000_FF00_0000_0000u64 as i64);
 774
 775                         let tmp4 = bcx.ins().ishl_imm(v, 8);
 776                         let n4 = bcx.ins().band_imm(tmp4, 0x0000_00FF_0000_0000u64 as i64);
 777
 778                         let tmp5 = bcx.ins().ushr_imm(v, 8);
 779                         let n5 = bcx.ins().band_imm(tmp5, 0x0000_0000_FF00_0000u64 as i64);
 780
 781                         let tmp6 = bcx.ins().ushr_imm(v, 24);
 782                         let n6 = bcx.ins().band_imm(tmp6, 0x0000_0000_00FF_0000u64 as i64);
 783
 784                         let tmp7 = bcx.ins().ushr_imm(v, 40);
 785                         let n7 = bcx.ins().band_imm(tmp7, 0x0000_0000_0000_FF00u64 as i64);
 786
 787                         let tmp8 = bcx.ins().ushr_imm(v, 56);
 788                         let n8 = bcx.ins().band_imm(tmp8, 0x0000_0000_0000_00FFu64 as i64);
 789
 790                         let or_tmp1 = bcx.ins().bor(n1, n2);
 791                         let or_tmp2 = bcx.ins().bor(n3, n4);
 792                         let or_tmp3 = bcx.ins().bor(n5, n6);
 793                         let or_tmp4 = bcx.ins().bor(n7, n8);
 794
 795                         let or_tmp5 = bcx.ins().bor(or_tmp1, or_tmp2);
 796                         let or_tmp6 = bcx.ins().bor(or_tmp3, or_tmp4);
 797                         bcx.ins().bor(or_tmp5, or_tmp6)
 798                     }
 799                     types::I128 => {
 800                         let (lo, hi) = bcx.ins().isplit(v);
 801                         let lo = swap(bcx, lo);
 802                         let hi = swap(bcx, hi);
 803                         bcx.ins().iconcat(hi, lo)
 804                     }
 805                     ty => unimplemented!("bswap {}", ty),
 806                 }
 807             };
 808             let res = CValue::by_val(swap(&mut fx.bcx, arg), fx.layout_of(T));
 809             ret.write_cvalue(fx, res);
 810         };
 811         panic_if_uninhabited, <T> () {
 812             if fx.layout_of(T).abi.is_uninhabited() {
 813                 crate::trap::trap_panic(fx, "[panic] Called intrinsic::panic_if_uninhabited for uninhabited type.");
 814                 return;
 815             }
 816         };
 817
 818         volatile_load, (c ptr) {
 819             // Cranelift treats loads as volatile by default
 820             let inner_layout =
 821                 fx.layout_of(ptr.layout().ty.builtin_deref(true).unwrap().ty);
 822             let val = CValue::by_ref(Pointer::new(ptr.load_scalar(fx)), inner_layout);
 823             ret.write_cvalue(fx, val);
 824         };
 825         volatile_store, (v ptr, c val) {
 826             // Cranelift treats stores as volatile by default
 827             let dest = CPlace::for_ptr(Pointer::new(ptr), val.layout());
 828             dest.write_cvalue(fx, val);
 829         };
 830
 831         size_of | pref_align_of | min_align_of | needs_drop | type_id | type_name, () {
 832             let const_val =
 833                 fx.tcx.const_eval_instance(ParamEnv::reveal_all(), instance, None).unwrap();
 834             let val = crate::constant::trans_const_value(fx, const_val);
 835             ret.write_cvalue(fx, val);
 836         };
 837
 838         ptr_offset_from, <T> (v ptr, v base) {
 839             let isize_layout = fx.layout_of(fx.tcx.types.isize);
 840
 841             let pointee_size: u64 = fx.layout_of(T).size.bytes();
 842             let diff = fx.bcx.ins().isub(ptr, base);
 843             // FIXME this can be an exact division.
 844             let val = CValue::by_val(fx.bcx.ins().udiv_imm(diff, pointee_size as i64), isize_layout);
 845             ret.write_cvalue(fx, val);
 846         };
 847
 848         caller_location, () {
 849             let caller_location = fx.get_caller_location(span);
 850             ret.write_cvalue(fx, caller_location);
 851         };
 852
 853         _ if intrinsic.starts_with("atomic_fence"), () {};
 854         _ if intrinsic.starts_with("atomic_singlethreadfence"), () {};
 855         _ if intrinsic.starts_with("atomic_load"), (c ptr) {
 856             let inner_layout =
 857                 fx.layout_of(ptr.layout().ty.builtin_deref(true).unwrap().ty);
 858             let val = CValue::by_ref(Pointer::new(ptr.load_scalar(fx)), inner_layout);
 859             ret.write_cvalue(fx, val);
 860         };
 861         _ if intrinsic.starts_with("atomic_store"), (v ptr, c val) {
 862             let dest = CPlace::for_ptr(Pointer::new(ptr), val.layout());
 863             dest.write_cvalue(fx, val);
 864         };
 865         _ if intrinsic.starts_with("atomic_xchg"), <T> (v ptr, c src) {
 866             // Read old
 867             let clif_ty = fx.clif_type(T).unwrap();
 868             let old = fx.bcx.ins().load(clif_ty, MemFlags::new(), ptr, 0);
 869             ret.write_cvalue(fx, CValue::by_val(old, fx.layout_of(T)));
 870
 871             // Write new
 872             let dest = CPlace::for_ptr(Pointer::new(ptr), src.layout());
 873             dest.write_cvalue(fx, src);
 874         };
 875         _ if intrinsic.starts_with("atomic_cxchg"), <T> (v ptr, v test_old, v new) { // both atomic_cxchg_* and atomic_cxchgweak_*
 876             // Read old
 877             let clif_ty = fx.clif_type(T).unwrap();
 878             let old = fx.bcx.ins().load(clif_ty, MemFlags::new(), ptr, 0);
 879
 880             // Compare
 881             let is_eq = codegen_icmp(fx, IntCC::Equal, old, test_old);
 882             let new = fx.bcx.ins().select(is_eq, new, old); // Keep old if not equal to test_old
 883
 884             // Write new
 885             fx.bcx.ins().store(MemFlags::new(), new, ptr, 0);
 886
 887             let ret_val = CValue::by_val_pair(old, fx.bcx.ins().bint(types::I8, is_eq), ret.layout());
 888             ret.write_cvalue(fx, ret_val);
 889         };
 890
 891         _ if intrinsic.starts_with("atomic_xadd"), <T> (v ptr, v amount) {
 892             atomic_binop_return_old! (fx, iadd<T>(ptr, amount) -> ret);
 893         };
 894         _ if intrinsic.starts_with("atomic_xsub"), <T> (v ptr, v amount) {
 895             atomic_binop_return_old! (fx, isub<T>(ptr, amount) -> ret);
 896         };
 897         _ if intrinsic.starts_with("atomic_and"), <T> (v ptr, v src) {
 898             atomic_binop_return_old! (fx, band<T>(ptr, src) -> ret);
 899         };
 900         _ if intrinsic.starts_with("atomic_nand"), <T> (v ptr, v src) {
 901             let clif_ty = fx.clif_type(T).unwrap();
 902             let old = fx.bcx.ins().load(clif_ty, MemFlags::new(), ptr, 0);
 903             let and = fx.bcx.ins().band(old, src);
 904             let new = fx.bcx.ins().bnot(and);
 905             fx.bcx.ins().store(MemFlags::new(), new, ptr, 0);
 906             ret.write_cvalue(fx, CValue::by_val(old, fx.layout_of(T)));
 907         };
 908         _ if intrinsic.starts_with("atomic_or"), <T> (v ptr, v src) {
 909             atomic_binop_return_old! (fx, bor<T>(ptr, src) -> ret);
 910         };
 911         _ if intrinsic.starts_with("atomic_xor"), <T> (v ptr, v src) {
 912             atomic_binop_return_old! (fx, bxor<T>(ptr, src) -> ret);
 913         };
 914
 915         _ if intrinsic.starts_with("atomic_max"), <T> (v ptr, v src) {
 916             atomic_minmax!(fx, IntCC::SignedGreaterThan, <T> (ptr, src) -> ret);
 917         };
 918         _ if intrinsic.starts_with("atomic_umax"), <T> (v ptr, v src) {
 919             atomic_minmax!(fx, IntCC::UnsignedGreaterThan, <T> (ptr, src) -> ret);
 920         };
 921         _ if intrinsic.starts_with("atomic_min"), <T> (v ptr, v src) {
 922             atomic_minmax!(fx, IntCC::SignedLessThan, <T> (ptr, src) -> ret);
 923         };
 924         _ if intrinsic.starts_with("atomic_umin"), <T> (v ptr, v src) {
 925             atomic_minmax!(fx, IntCC::UnsignedLessThan, <T> (ptr, src) -> ret);
 926         };
 927
 928         minnumf32, (v a, v b) {
 929             let val = fx.bcx.ins().fmin(a, b);
 930             let val = CValue::by_val(val, fx.layout_of(fx.tcx.types.f32));
 931             ret.write_cvalue(fx, val);
 932         };
 933         minnumf64, (v a, v b) {
 934             let val = fx.bcx.ins().fmin(a, b);
 935             let val = CValue::by_val(val, fx.layout_of(fx.tcx.types.f64));
 936             ret.write_cvalue(fx, val);
 937         };
 938         maxnumf32, (v a, v b) {
 939             let val = fx.bcx.ins().fmax(a, b);
 940             let val = CValue::by_val(val, fx.layout_of(fx.tcx.types.f32));
 941             ret.write_cvalue(fx, val);
 942         };
 943         maxnumf64, (v a, v b) {
 944             let val = fx.bcx.ins().fmax(a, b);
 945             let val = CValue::by_val(val, fx.layout_of(fx.tcx.types.f64));
 946             ret.write_cvalue(fx, val);
 947         };
 948
 949         try, (v f, v data, v _local_ptr) {
 950             // FIXME once unwinding is supported, change this to actually catch panics
 951             let f_sig = fx.bcx.func.import_signature(Signature {
 952                 call_conv: CallConv::triple_default(fx.triple()),
 953                 params: vec![AbiParam::new(fx.bcx.func.dfg.value_type(data))],
 954                 returns: vec![],
 955             });
 956
 957             fx.bcx.ins().call_indirect(f_sig, f, &[data]);
 958
 959             let ret_val = CValue::const_val(fx, ret.layout().ty, 0);
 960             ret.write_cvalue(fx, ret_val);
 961         };
 962     }
 963
 964     if let Some((_, dest)) = destination {
 965         let ret_ebb = fx.get_ebb(dest);
 966         fx.bcx.ins().jump(ret_ebb, &[]);
 967     } else {
 968         trap_unreachable(fx, "[corruption] Diverging intrinsic returned.");
 969     }
 970 }