src/intrinsics.rs

   1 use crate::prelude::*;
   2
   3 use rustc::ty::subst::SubstsRef;
   4
   5 macro_rules! intrinsic_pat {
   6     (_) => {
   7         _
   8     };
   9     ($name:ident) => {
  10         stringify!($name)
  11     }
  12 }
  13
  14 macro_rules! intrinsic_arg {
  15     (c $fx:expr, $arg:ident) => {
  16         $arg
  17     };
  18     (v $fx:expr, $arg:ident) => {
  19         $arg.load_scalar($fx)
  20     };
  21 }
  22
  23 macro_rules! intrinsic_substs {
  24     ($substs:expr, $index:expr,) => {};
  25     ($substs:expr, $index:expr, $first:ident $(,$rest:ident)*) => {
  26         let $first = $substs.type_at($index);
  27         intrinsic_substs!($substs, $index+1, $($rest),*);
  28     };
  29 }
  30
  31 macro_rules! intrinsic_match {
  32     ($fx:expr, $intrinsic:expr, $substs:expr, $args:expr, $(
  33         $($name:tt)|+ $(if $cond:expr)?, $(<$($subst:ident),*>)? ($($a:ident $arg:ident),*) $content:block;
  34     )*) => {
  35         match $intrinsic {
  36             $(
  37                 $(intrinsic_pat!($name))|* $(if $cond)? => {
  38                     #[allow(unused_parens, non_snake_case)]
  39                     {
  40                         $(
  41                             intrinsic_substs!($substs, 0, $($subst),*);
  42                         )?
  43                         if let [$($arg),*] = *$args {
  44                             let ($($arg),*) = (
  45                                 $(intrinsic_arg!($a $fx, $arg)),*
  46                             );
  47                             #[warn(unused_parens, non_snake_case)]
  48                             {
  49                                 $content
  50                             }
  51                         } else {
  52                             bug!("wrong number of args for intrinsic {:?}", $intrinsic);
  53                         }
  54                     }
  55                 }
  56             )*
  57             _ => unimpl!("unsupported intrinsic {}", $intrinsic),
  58         }
  59     };
  60 }
  61
  62 macro_rules! call_intrinsic_match {
  63     ($fx:expr, $intrinsic:expr, $substs:expr, $ret:expr, $destination:expr, $args:expr, $(
  64         $name:ident($($arg:ident),*) -> $ty:ident => $func:ident,
  65     )*) => {
  66         match $intrinsic {
  67             $(
  68                 stringify!($name) => {
  69                     assert!($substs.is_noop());
  70                     if let [$($arg),*] = *$args {
  71                         let res = $fx.easy_call(stringify!($func), &[$($arg),*], $fx.tcx.types.$ty);
  72                         $ret.write_cvalue($fx, res);
  73
  74                         if let Some((_, dest)) = $destination {
  75                             let ret_ebb = $fx.get_ebb(dest);
  76                             $fx.bcx.ins().jump(ret_ebb, &[]);
  77                             return;
  78                         } else {
  79                             unreachable!();
  80                         }
  81                     } else {
  82                         bug!("wrong number of args for intrinsic {:?}", $intrinsic);
  83                     }
  84                 }
  85             )*
  86             _ => {}
  87         }
  88     }
  89 }
  90
  91 macro_rules! atomic_binop_return_old {
  92     ($fx:expr, $op:ident<$T:ident>($ptr:ident, $src:ident) -> $ret:ident) => {
  93         let clif_ty = $fx.clif_type($T).unwrap();
  94         let old = $fx.bcx.ins().load(clif_ty, MemFlags::new(), $ptr, 0);
  95         let new = $fx.bcx.ins().$op(old, $src);
  96         $fx.bcx.ins().store(MemFlags::new(), new, $ptr, 0);
  97         $ret.write_cvalue($fx, CValue::by_val(old, $fx.layout_of($T)));
  98     };
  99 }
 100
 101 macro_rules! atomic_minmax {
 102     ($fx:expr, $cc:expr, <$T:ident> ($ptr:ident, $src:ident) -> $ret:ident) => {
 103         // Read old
 104         let clif_ty = $fx.clif_type($T).unwrap();
 105         let old = $fx.bcx.ins().load(clif_ty, MemFlags::new(), $ptr, 0);
 106
 107         // Compare
 108         let is_eq = $fx.bcx.ins().icmp(IntCC::SignedGreaterThan, old, $src);
 109         let new = crate::common::codegen_select(&mut $fx.bcx, is_eq, old, $src);
 110
 111         // Write new
 112         $fx.bcx.ins().store(MemFlags::new(), new, $ptr, 0);
 113
 114         let ret_val = CValue::by_val(old, $ret.layout());
 115         $ret.write_cvalue($fx, ret_val);
 116     };
 117 }
 118
 119 fn lane_type_and_count<'tcx>(
 120     fx: &FunctionCx<'_, 'tcx, impl Backend>,
 121     layout: TyLayout<'tcx>,
 122     intrinsic: &str,
 123 ) -> (TyLayout<'tcx>, usize) {
 124     assert!(layout.ty.is_simd());
 125     let lane_count = match layout.fields {
 126         layout::FieldPlacement::Array { stride: _, count } => usize::try_from(count).unwrap(),
 127         _ => panic!("Non vector type {:?} passed to or returned from simd_* intrinsic {}", layout.ty, intrinsic),
 128     };
 129     let lane_layout = layout.field(fx, 0);
 130     (lane_layout, lane_count)
 131 }
 132
 133 fn simd_for_each_lane<'tcx, B: Backend>(
 134     fx: &mut FunctionCx<'_, 'tcx, B>,
 135     intrinsic: &str,
 136     x: CValue<'tcx>,
 137     y: CValue<'tcx>,
 138     ret: CPlace<'tcx>,
 139     f: impl Fn(&mut FunctionCx<'_, 'tcx, B>, TyLayout<'tcx>, TyLayout<'tcx>, Value, Value) -> CValue<'tcx>,
 140 ) {
 141     assert_eq!(x.layout(), y.layout());
 142     let layout = x.layout();
 143
 144     let (lane_layout, lane_count) = lane_type_and_count(fx, layout, intrinsic);
 145     let (ret_lane_layout, ret_lane_count) = lane_type_and_count(fx, ret.layout(), intrinsic);
 146     assert_eq!(lane_count, ret_lane_count);
 147
 148     for lane in 0..lane_count {
 149         let lane = mir::Field::new(lane);
 150         let x_lane = x.value_field(fx, lane).load_scalar(fx);
 151         let y_lane = y.value_field(fx, lane).load_scalar(fx);
 152
 153         let res_lane = f(fx, lane_layout, ret_lane_layout, x_lane, y_lane);
 154
 155         ret.place_field(fx, lane).write_cvalue(fx, res_lane);
 156     }
 157 }
 158
 159 fn bool_to_zero_or_max_uint<'tcx>(
 160     fx: &mut FunctionCx<'_, 'tcx, impl Backend>,
 161     layout: TyLayout<'tcx>,
 162     val: Value,
 163 ) -> CValue<'tcx> {
 164     let ty = fx.clif_type(layout.ty).unwrap();
 165
 166     let zero = fx.bcx.ins().iconst(ty, 0);
 167     let max = fx.bcx.ins().iconst(ty, (u64::max_value() >> (64 - ty.bits())) as i64);
 168     let res = crate::common::codegen_select(&mut fx.bcx, val, max, zero);
 169     CValue::by_val(res, layout)
 170 }
 171
 172 macro_rules! simd_cmp {
 173     ($fx:expr, $intrinsic:expr, $cc:ident($x:ident, $y:ident) -> $ret:ident) => {
 174         simd_for_each_lane($fx, $intrinsic, $x, $y, $ret, |fx, _lane_layout, res_lane_layout, x_lane, y_lane| {
 175             let res_lane = fx.bcx.ins().icmp(IntCC::$cc, x_lane, y_lane);
 176             bool_to_zero_or_max_uint(fx, res_lane_layout, res_lane)
 177         });
 178     };
 179     ($fx:expr, $intrinsic:expr, $cc_u:ident|$cc_s:ident($x:ident, $y:ident) -> $ret:ident) => {
 180         simd_for_each_lane($fx, $intrinsic, $x, $y, $ret, |fx, lane_layout, res_lane_layout, x_lane, y_lane| {
 181             let res_lane = match lane_layout.ty.sty {
 182                 ty::Uint(_) => fx.bcx.ins().icmp(IntCC::$cc_u, x_lane, y_lane),
 183                 ty::Int(_) => fx.bcx.ins().icmp(IntCC::$cc_s, x_lane, y_lane),
 184                 _ => unreachable!("{:?}", lane_layout.ty),
 185             };
 186             bool_to_zero_or_max_uint(fx, res_lane_layout, res_lane)
 187         });
 188     };
 189
 190 }
 191
 192 macro_rules! simd_binop {
 193     ($fx:expr, $intrinsic:expr, $op:ident($x:ident, $y:ident) -> $ret:ident) => {
 194         simd_for_each_lane($fx, $intrinsic, $x, $y, $ret, |fx, _lane_layout, ret_lane_layout, x_lane, y_lane| {
 195             let res_lane = fx.bcx.ins().$op(x_lane, y_lane);
 196             CValue::by_val(res_lane, ret_lane_layout)
 197         });
 198     };
 199     ($fx:expr, $intrinsic:expr, $op_u:ident|$op_s:ident($x:ident, $y:ident) -> $ret:ident) => {
 200         simd_for_each_lane($fx, $intrinsic, $x, $y, $ret, |fx, lane_layout, ret_lane_layout, x_lane, y_lane| {
 201             let res_lane = match lane_layout.ty.sty {
 202                 ty::Uint(_) => fx.bcx.ins().$op_u(x_lane, y_lane),
 203                 ty::Int(_) => fx.bcx.ins().$op_s(x_lane, y_lane),
 204                 _ => unreachable!("{:?}", lane_layout.ty),
 205             };
 206             CValue::by_val(res_lane, ret_lane_layout)
 207         });
 208     };
 209 }
 210
 211 pub fn codegen_intrinsic_call<'a, 'tcx: 'a>(
 212     fx: &mut FunctionCx<'a, 'tcx, impl Backend>,
 213     def_id: DefId,
 214     substs: SubstsRef<'tcx>,
 215     args: Vec<CValue<'tcx>>,
 216     destination: Option<(CPlace<'tcx>, BasicBlock)>,
 217 ) {
 218     let intrinsic = fx.tcx.item_name(def_id).as_str();
 219     let intrinsic = &intrinsic[..];
 220
 221     let ret = match destination {
 222         Some((place, _)) => place,
 223         None => {
 224             // Insert non returning intrinsics here
 225             match intrinsic {
 226                 "abort" => {
 227                     trap_panic(fx, "Called intrinsic::abort.");
 228                 }
 229                 "unreachable" => {
 230                     trap_unreachable(fx, "[corruption] Called intrinsic::unreachable.");
 231                 }
 232                 _ => unimplemented!("unsupported instrinsic {}", intrinsic),
 233             }
 234             return;
 235         }
 236     };
 237
 238     let u64_layout = fx.layout_of(fx.tcx.types.u64);
 239     let usize_layout = fx.layout_of(fx.tcx.types.usize);
 240
 241     call_intrinsic_match! {
 242         fx, intrinsic, substs, ret, destination, args,
 243         expf32(flt) -> f32 => expf,
 244         expf64(flt) -> f64 => exp,
 245         exp2f32(flt) -> f32 => exp2f,
 246         exp2f64(flt) -> f64 => exp2,
 247         sqrtf32(flt) -> f32 => sqrtf,
 248         sqrtf64(flt) -> f64 => sqrt,
 249         powif32(a, x) -> f32 => __powisf2, // compiler-builtins
 250         powif64(a, x) -> f64 => __powidf2, // compiler-builtins
 251         logf32(flt) -> f32 => logf,
 252         logf64(flt) -> f64 => log,
 253         fabsf32(flt) -> f32 => fabsf,
 254         fabsf64(flt) -> f64 => fabs,
 255         fmaf32(x, y, z) -> f32 => fmaf,
 256         fmaf64(x, y, z) -> f64 => fma,
 257
 258         // rounding variants
 259         floorf32(flt) -> f32 => floorf,
 260         floorf64(flt) -> f64 => floor,
 261         ceilf32(flt) -> f32 => ceilf,
 262         ceilf64(flt) -> f64 => ceil,
 263         truncf32(flt) -> f32 => truncf,
 264         truncf64(flt) -> f64 => trunc,
 265         roundf32(flt) -> f32 => roundf,
 266         roundf64(flt) -> f64 => round,
 267
 268         // trigonometry
 269         sinf32(flt) -> f32 => sinf,
 270         sinf64(flt) -> f64 => sin,
 271         cosf32(flt) -> f32 => cosf,
 272         cosf64(flt) -> f64 => cos,
 273         tanf32(flt) -> f32 => tanf,
 274         tanf64(flt) -> f64 => tan,
 275     }
 276
 277     intrinsic_match! {
 278         fx, intrinsic, substs, args,
 279
 280         assume, (c _a) {};
 281         likely | unlikely, (c a) {
 282             ret.write_cvalue(fx, a);
 283         };
 284         breakpoint, () {
 285             fx.bcx.ins().debugtrap();
 286         };
 287         copy | copy_nonoverlapping, <elem_ty> (v src, v dst, v count) {
 288             let elem_size: u64 = fx.layout_of(elem_ty).size.bytes();
 289             let elem_size = fx
 290                 .bcx
 291                 .ins()
 292                 .iconst(fx.pointer_type, elem_size as i64);
 293             assert_eq!(args.len(), 3);
 294             let byte_amount = fx.bcx.ins().imul(count, elem_size);
 295
 296             if intrinsic.ends_with("_nonoverlapping") {
 297                 fx.bcx.call_memcpy(fx.module.target_config(), dst, src, byte_amount);
 298             } else {
 299                 fx.bcx.call_memmove(fx.module.target_config(), dst, src, byte_amount);
 300             }
 301         };
 302         discriminant_value, (c val) {
 303             let pointee_layout = fx.layout_of(val.layout().ty.builtin_deref(true).unwrap().ty);
 304             let place = CPlace::for_addr(val.load_scalar(fx), pointee_layout);
 305             let discr = crate::base::trans_get_discriminant(fx, place, ret.layout());
 306             ret.write_cvalue(fx, discr);
 307         };
 308         size_of, <T> () {
 309             let size_of = fx.layout_of(T).size.bytes();
 310             let size_of = CValue::const_val(fx, usize_layout.ty, size_of.into());
 311             ret.write_cvalue(fx, size_of);
 312         };
 313         size_of_val, <T> (c ptr) {
 314             let layout = fx.layout_of(T);
 315             let size = if layout.is_unsized() {
 316                 let (_ptr, info) = ptr.load_scalar_pair(fx);
 317                 let (size, _align) = crate::unsize::size_and_align_of_dst(fx, layout.ty, info);
 318                 size
 319             } else {
 320                 fx
 321                     .bcx
 322                     .ins()
 323                     .iconst(fx.pointer_type, layout.size.bytes() as i64)
 324             };
 325             ret.write_cvalue(fx, CValue::by_val(size, usize_layout));
 326         };
 327         min_align_of, <T> () {
 328             let min_align = fx.layout_of(T).align.abi.bytes();
 329             let min_align = CValue::const_val(fx, usize_layout.ty, min_align.into());
 330             ret.write_cvalue(fx, min_align);
 331         };
 332         min_align_of_val, <T> (c ptr) {
 333             let layout = fx.layout_of(T);
 334             let align = if layout.is_unsized() {
 335                 let (_ptr, info) = ptr.load_scalar_pair(fx);
 336                 let (_size, align) = crate::unsize::size_and_align_of_dst(fx, layout.ty, info);
 337                 align
 338             } else {
 339                 fx
 340                     .bcx
 341                     .ins()
 342                     .iconst(fx.pointer_type, layout.align.abi.bytes() as i64)
 343             };
 344             ret.write_cvalue(fx, CValue::by_val(align, usize_layout));
 345         };
 346         pref_align_of, <T> () {
 347             let pref_align = fx.layout_of(T).align.pref.bytes();
 348             let pref_align = CValue::const_val(fx, usize_layout.ty, pref_align.into());
 349             ret.write_cvalue(fx, pref_align);
 350         };
 351
 352
 353         type_id, <T> () {
 354             let type_id = fx.tcx.type_id_hash(T);
 355             let type_id = CValue::const_val(fx, u64_layout.ty, type_id.into());
 356             ret.write_cvalue(fx, type_id);
 357         };
 358         type_name, <T> () {
 359             let type_name = fx.tcx.type_name(T);
 360             let type_name = crate::constant::trans_const_value(fx, type_name);
 361             ret.write_cvalue(fx, type_name);
 362         };
 363
 364         _ if intrinsic.starts_with("unchecked_") || intrinsic == "exact_div", (c x, c y) {
 365             // FIXME trap on overflow
 366             let bin_op = match intrinsic {
 367                 "unchecked_sub" => BinOp::Sub,
 368                 "unchecked_div" | "exact_div" => BinOp::Div,
 369                 "unchecked_rem" => BinOp::Rem,
 370                 "unchecked_shl" => BinOp::Shl,
 371                 "unchecked_shr" => BinOp::Shr,
 372                 _ => unimplemented!("intrinsic {}", intrinsic),
 373             };
 374             let res = match ret.layout().ty.sty {
 375                 ty::Uint(_) => crate::base::trans_int_binop(
 376                     fx,
 377                     bin_op,
 378                     x,
 379                     y,
 380                     ret.layout().ty,
 381                     false,
 382                 ),
 383                 ty::Int(_) => crate::base::trans_int_binop(
 384                     fx,
 385                     bin_op,
 386                     x,
 387                     y,
 388                     ret.layout().ty,
 389                     true,
 390                 ),
 391                 _ => panic!(),
 392             };
 393             ret.write_cvalue(fx, res);
 394         };
 395         _ if intrinsic.ends_with("_with_overflow"), <T> (c x, c y) {
 396             assert_eq!(x.layout().ty, y.layout().ty);
 397             let bin_op = match intrinsic {
 398                 "add_with_overflow" => BinOp::Add,
 399                 "sub_with_overflow" => BinOp::Sub,
 400                 "mul_with_overflow" => BinOp::Mul,
 401                 _ => unimplemented!("intrinsic {}", intrinsic),
 402             };
 403             let res = match T.sty {
 404                 ty::Uint(_) => crate::base::trans_checked_int_binop(
 405                     fx,
 406                     bin_op,
 407                     x,
 408                     y,
 409                     ret.layout().ty,
 410                     false,
 411                 ),
 412                 ty::Int(_) => crate::base::trans_checked_int_binop(
 413                     fx,
 414                     bin_op,
 415                     x,
 416                     y,
 417                     ret.layout().ty,
 418                     true,
 419                 ),
 420                 _ => panic!(),
 421             };
 422             ret.write_cvalue(fx, res);
 423         };
 424         _ if intrinsic.starts_with("overflowing_"), <T> (c x, c y) {
 425             assert_eq!(x.layout().ty, y.layout().ty);
 426             let bin_op = match intrinsic {
 427                 "overflowing_add" => BinOp::Add,
 428                 "overflowing_sub" => BinOp::Sub,
 429                 "overflowing_mul" => BinOp::Mul,
 430                 _ => unimplemented!("intrinsic {}", intrinsic),
 431             };
 432             let res = match T.sty {
 433                 ty::Uint(_) => crate::base::trans_int_binop(
 434                     fx,
 435                     bin_op,
 436                     x,
 437                     y,
 438                     ret.layout().ty,
 439                     false,
 440                 ),
 441                 ty::Int(_) => crate::base::trans_int_binop(
 442                     fx,
 443                     bin_op,
 444                     x,
 445                     y,
 446                     ret.layout().ty,
 447                     true,
 448                 ),
 449                 _ => panic!(),
 450             };
 451             ret.write_cvalue(fx, res);
 452         };
 453         _ if intrinsic.starts_with("saturating_"), <T> (c x, c y) {
 454             // FIXME implement saturating behavior
 455             assert_eq!(x.layout().ty, y.layout().ty);
 456             let bin_op = match intrinsic {
 457                 "saturating_add" => BinOp::Add,
 458                 "saturating_sub" => BinOp::Sub,
 459                 "saturating_mul" => BinOp::Mul,
 460                 _ => unimplemented!("intrinsic {}", intrinsic),
 461             };
 462             let res = match T.sty {
 463                 ty::Uint(_) => crate::base::trans_int_binop(
 464                     fx,
 465                     bin_op,
 466                     x,
 467                     y,
 468                     ret.layout().ty,
 469                     false,
 470                 ),
 471                 ty::Int(_) => crate::base::trans_int_binop(
 472                     fx,
 473                     bin_op,
 474                     x,
 475                     y,
 476                     ret.layout().ty,
 477                     true,
 478                 ),
 479                 _ => panic!(),
 480             };
 481             ret.write_cvalue(fx, res);
 482         };
 483         rotate_left, <T>(v x, v y) {
 484             let layout = fx.layout_of(T);
 485             let res = fx.bcx.ins().rotl(x, y);
 486             ret.write_cvalue(fx, CValue::by_val(res, layout));
 487         };
 488         rotate_right, <T>(v x, v y) {
 489             let layout = fx.layout_of(T);
 490             let res = fx.bcx.ins().rotr(x, y);
 491             ret.write_cvalue(fx, CValue::by_val(res, layout));
 492         };
 493
 494         // The only difference between offset and arith_offset is regarding UB. Because Cranelift
 495         // doesn't have UB both are codegen'ed the same way
 496         offset | arith_offset, (c base, v offset) {
 497             let pointee_ty = base.layout().ty.builtin_deref(true).unwrap().ty;
 498             let pointee_size = fx.layout_of(pointee_ty).size.bytes();
 499             let ptr_diff = fx.bcx.ins().imul_imm(offset, pointee_size as i64);
 500             let base_val = base.load_scalar(fx);
 501             let res = fx.bcx.ins().iadd(base_val, ptr_diff);
 502             ret.write_cvalue(fx, CValue::by_val(res, args[0].layout()));
 503         };
 504
 505         transmute, <src_ty, dst_ty> (c from) {
 506             assert_eq!(from.layout().ty, src_ty);
 507             let addr = from.force_stack(fx);
 508             let dst_layout = fx.layout_of(dst_ty);
 509             ret.write_cvalue(fx, CValue::by_ref(addr, dst_layout))
 510         };
 511         init, () {
 512             if ret.layout().abi == Abi::Uninhabited {
 513                 crate::trap::trap_panic(fx, "[panic] Called intrinsic::init for uninhabited type.");
 514                 return;
 515             }
 516
 517             match ret {
 518                 CPlace::NoPlace(_layout) => {}
 519                 CPlace::Var(var, layout) => {
 520                     let clif_ty = fx.clif_type(layout.ty).unwrap();
 521                     let val = match clif_ty {
 522                         types::I8 | types::I16 | types::I32 | types::I64 => fx.bcx.ins().iconst(clif_ty, 0),
 523                         types::F32 => {
 524                             let zero = fx.bcx.ins().iconst(types::I32, 0);
 525                             fx.bcx.ins().bitcast(types::F32, zero)
 526                         }
 527                         types::F64 => {
 528                             let zero = fx.bcx.ins().iconst(types::I64, 0);
 529                             fx.bcx.ins().bitcast(types::F64, zero)
 530                         }
 531                         _ => panic!("clif_type returned {}", clif_ty),
 532                     };
 533                     fx.bcx.def_var(mir_var(var), val);
 534                 }
 535                 _ => {
 536                     let addr = ret.to_addr(fx);
 537                     let layout = ret.layout();
 538                     fx.bcx.emit_small_memset(fx.module.target_config(), addr, 0, layout.size.bytes(), 1);
 539                 }
 540             }
 541         };
 542         write_bytes, (c dst, v val, v count) {
 543             let pointee_ty = dst.layout().ty.builtin_deref(true).unwrap().ty;
 544             let pointee_size = fx.layout_of(pointee_ty).size.bytes();
 545             let count = fx.bcx.ins().imul_imm(count, pointee_size as i64);
 546             let dst_ptr = dst.load_scalar(fx);
 547             fx.bcx.call_memset(fx.module.target_config(), dst_ptr, val, count);
 548         };
 549         ctlz | ctlz_nonzero, <T> (v arg) {
 550             let res = if T == fx.tcx.types.u128 || T == fx.tcx.types.i128 {
 551                 // FIXME verify this algorithm is correct
 552                 let (lsb, msb) = fx.bcx.ins().isplit(arg);
 553                 let lsb_lz = fx.bcx.ins().clz(lsb);
 554                 let msb_lz = fx.bcx.ins().clz(msb);
 555                 let msb_is_zero = fx.bcx.ins().icmp_imm(IntCC::Equal, msb, 0);
 556                 let lsb_lz_plus_64 = fx.bcx.ins().iadd_imm(lsb_lz, 64);
 557                 fx.bcx.ins().select(msb_is_zero, lsb_lz_plus_64, msb_lz)
 558             } else {
 559                 fx.bcx.ins().clz(arg)
 560             };
 561             let res = CValue::by_val(res, fx.layout_of(T));
 562             ret.write_cvalue(fx, res);
 563         };
 564         cttz | cttz_nonzero, <T> (v arg) {
 565             let res = if T == fx.tcx.types.u128 || T == fx.tcx.types.i128 {
 566                 // FIXME verify this algorithm is correct
 567                 let (lsb, msb) = fx.bcx.ins().isplit(arg);
 568                 let lsb_tz = fx.bcx.ins().ctz(lsb);
 569                 let msb_tz = fx.bcx.ins().ctz(msb);
 570                 let lsb_is_zero = fx.bcx.ins().icmp_imm(IntCC::Equal, lsb, 0);
 571                 let msb_tz_plus_64 = fx.bcx.ins().iadd_imm(msb_tz, 64);
 572                 fx.bcx.ins().select(lsb_is_zero, msb_tz_plus_64, lsb_tz)
 573             } else {
 574                 fx.bcx.ins().ctz(arg)
 575             };
 576             let res = CValue::by_val(res, fx.layout_of(T));
 577             ret.write_cvalue(fx, res);
 578         };
 579         ctpop, <T> (v arg) {
 580             let res = CValue::by_val(fx.bcx.ins().popcnt(arg), fx.layout_of(T));
 581             ret.write_cvalue(fx, res);
 582         };
 583         bitreverse, <T> (v arg) {
 584             let res = CValue::by_val(fx.bcx.ins().bitrev(arg), fx.layout_of(T));
 585             ret.write_cvalue(fx, res);
 586         };
 587         bswap, <T> (v arg) {
 588             // FIXME(CraneStation/cranelift#794) add bswap instruction to cranelift
 589             fn swap(bcx: &mut FunctionBuilder, v: Value) -> Value {
 590                 match bcx.func.dfg.value_type(v) {
 591                     types::I8 => v,
 592
 593                     // https://code.woboq.org/gcc/include/bits/byteswap.h.html
 594                     types::I16 => {
 595                         let tmp1 = bcx.ins().ishl_imm(v, 8);
 596                         let n1 = bcx.ins().band_imm(tmp1, 0xFF00);
 597
 598                         let tmp2 = bcx.ins().ushr_imm(v, 8);
 599                         let n2 = bcx.ins().band_imm(tmp2, 0x00FF);
 600
 601                         bcx.ins().bor(n1, n2)
 602                     }
 603                     types::I32 => {
 604                         let tmp1 = bcx.ins().ishl_imm(v, 24);
 605                         let n1 = bcx.ins().band_imm(tmp1, 0xFF00_0000);
 606
 607                         let tmp2 = bcx.ins().ishl_imm(v, 8);
 608                         let n2 = bcx.ins().band_imm(tmp2, 0x00FF_0000);
 609
 610                         let tmp3 = bcx.ins().ushr_imm(v, 8);
 611                         let n3 = bcx.ins().band_imm(tmp3, 0x0000_FF00);
 612
 613                         let tmp4 = bcx.ins().ushr_imm(v, 24);
 614                         let n4 = bcx.ins().band_imm(tmp4, 0x0000_00FF);
 615
 616                         let or_tmp1 = bcx.ins().bor(n1, n2);
 617                         let or_tmp2 = bcx.ins().bor(n3, n4);
 618                         bcx.ins().bor(or_tmp1, or_tmp2)
 619                     }
 620                     types::I64 => {
 621                         let tmp1 = bcx.ins().ishl_imm(v, 56);
 622                         let n1 = bcx.ins().band_imm(tmp1, 0xFF00_0000_0000_0000u64 as i64);
 623
 624                         let tmp2 = bcx.ins().ishl_imm(v, 40);
 625                         let n2 = bcx.ins().band_imm(tmp2, 0x00FF_0000_0000_0000u64 as i64);
 626
 627                         let tmp3 = bcx.ins().ishl_imm(v, 24);
 628                         let n3 = bcx.ins().band_imm(tmp3, 0x0000_FF00_0000_0000u64 as i64);
 629
 630                         let tmp4 = bcx.ins().ishl_imm(v, 8);
 631                         let n4 = bcx.ins().band_imm(tmp4, 0x0000_00FF_0000_0000u64 as i64);
 632
 633                         let tmp5 = bcx.ins().ushr_imm(v, 8);
 634                         let n5 = bcx.ins().band_imm(tmp5, 0x0000_0000_FF00_0000u64 as i64);
 635
 636                         let tmp6 = bcx.ins().ushr_imm(v, 24);
 637                         let n6 = bcx.ins().band_imm(tmp6, 0x0000_0000_00FF_0000u64 as i64);
 638
 639                         let tmp7 = bcx.ins().ushr_imm(v, 40);
 640                         let n7 = bcx.ins().band_imm(tmp7, 0x0000_0000_0000_FF00u64 as i64);
 641
 642                         let tmp8 = bcx.ins().ushr_imm(v, 56);
 643                         let n8 = bcx.ins().band_imm(tmp8, 0x0000_0000_0000_00FFu64 as i64);
 644
 645                         let or_tmp1 = bcx.ins().bor(n1, n2);
 646                         let or_tmp2 = bcx.ins().bor(n3, n4);
 647                         let or_tmp3 = bcx.ins().bor(n5, n6);
 648                         let or_tmp4 = bcx.ins().bor(n7, n8);
 649
 650                         let or_tmp5 = bcx.ins().bor(or_tmp1, or_tmp2);
 651                         let or_tmp6 = bcx.ins().bor(or_tmp3, or_tmp4);
 652                         bcx.ins().bor(or_tmp5, or_tmp6)
 653                     }
 654                     types::I128 => {
 655                         let (lo, hi) = bcx.ins().isplit(v);
 656                         let lo = swap(bcx, lo);
 657                         let hi = swap(bcx, hi);
 658                         bcx.ins().iconcat(hi, lo)
 659                     }
 660                     ty => unimplemented!("bswap {}", ty),
 661                 }
 662             };
 663             let res = CValue::by_val(swap(&mut fx.bcx, arg), fx.layout_of(T));
 664             ret.write_cvalue(fx, res);
 665         };
 666         needs_drop, <T> () {
 667             let needs_drop = if T.needs_drop(fx.tcx, ParamEnv::reveal_all()) {
 668                 1
 669             } else {
 670                 0
 671             };
 672             let needs_drop = CValue::const_val(fx, fx.tcx.types.bool, needs_drop);
 673             ret.write_cvalue(fx, needs_drop);
 674         };
 675         panic_if_uninhabited, <T> () {
 676             if fx.layout_of(T).abi.is_uninhabited() {
 677                 crate::trap::trap_panic(fx, "[panic] Called intrinsic::panic_if_uninhabited for uninhabited type.");
 678                 return;
 679             }
 680         };
 681
 682         volatile_load, (c ptr) {
 683             // Cranelift treats loads as volatile by default
 684             let inner_layout =
 685                 fx.layout_of(ptr.layout().ty.builtin_deref(true).unwrap().ty);
 686             let val = CValue::by_ref(ptr.load_scalar(fx), inner_layout);
 687             ret.write_cvalue(fx, val);
 688         };
 689         volatile_store, (v ptr, c val) {
 690             // Cranelift treats stores as volatile by default
 691             let dest = CPlace::for_addr(ptr, val.layout());
 692             dest.write_cvalue(fx, val);
 693         };
 694
 695         _ if intrinsic.starts_with("atomic_fence"), () {};
 696         _ if intrinsic.starts_with("atomic_singlethreadfence"), () {};
 697         _ if intrinsic.starts_with("atomic_load"), (c ptr) {
 698             let inner_layout =
 699                 fx.layout_of(ptr.layout().ty.builtin_deref(true).unwrap().ty);
 700             let val = CValue::by_ref(ptr.load_scalar(fx), inner_layout);
 701             ret.write_cvalue(fx, val);
 702         };
 703         _ if intrinsic.starts_with("atomic_store"), (v ptr, c val) {
 704             let dest = CPlace::for_addr(ptr, val.layout());
 705             dest.write_cvalue(fx, val);
 706         };
 707         _ if intrinsic.starts_with("atomic_xchg"), <T> (v ptr, c src) {
 708             // Read old
 709             let clif_ty = fx.clif_type(T).unwrap();
 710             let old = fx.bcx.ins().load(clif_ty, MemFlags::new(), ptr, 0);
 711             ret.write_cvalue(fx, CValue::by_val(old, fx.layout_of(T)));
 712
 713             // Write new
 714             let dest = CPlace::for_addr(ptr, src.layout());
 715             dest.write_cvalue(fx, src);
 716         };
 717         _ if intrinsic.starts_with("atomic_cxchg"), <T> (v ptr, v test_old, v new) { // both atomic_cxchg_* and atomic_cxchgweak_*
 718             // Read old
 719             let clif_ty = fx.clif_type(T).unwrap();
 720             let old = fx.bcx.ins().load(clif_ty, MemFlags::new(), ptr, 0);
 721
 722             // Compare
 723             let is_eq = fx.bcx.ins().icmp(IntCC::Equal, old, test_old);
 724             let new = crate::common::codegen_select(&mut fx.bcx, is_eq, new, old); // Keep old if not equal to test_old
 725
 726             // Write new
 727             fx.bcx.ins().store(MemFlags::new(), new, ptr, 0);
 728
 729             let ret_val = CValue::by_val_pair(old, fx.bcx.ins().bint(types::I8, is_eq), ret.layout());
 730             ret.write_cvalue(fx, ret_val);
 731         };
 732
 733         _ if intrinsic.starts_with("atomic_xadd"), <T> (v ptr, v amount) {
 734             atomic_binop_return_old! (fx, iadd<T>(ptr, amount) -> ret);
 735         };
 736         _ if intrinsic.starts_with("atomic_xsub"), <T> (v ptr, v amount) {
 737             atomic_binop_return_old! (fx, isub<T>(ptr, amount) -> ret);
 738         };
 739         _ if intrinsic.starts_with("atomic_and"), <T> (v ptr, v src) {
 740             atomic_binop_return_old! (fx, band<T>(ptr, src) -> ret);
 741         };
 742         _ if intrinsic.starts_with("atomic_nand"), <T> (v ptr, v src) {
 743             atomic_binop_return_old! (fx, band_not<T>(ptr, src) -> ret);
 744         };
 745         _ if intrinsic.starts_with("atomic_or"), <T> (v ptr, v src) {
 746             atomic_binop_return_old! (fx, bor<T>(ptr, src) -> ret);
 747         };
 748         _ if intrinsic.starts_with("atomic_xor"), <T> (v ptr, v src) {
 749             atomic_binop_return_old! (fx, bxor<T>(ptr, src) -> ret);
 750         };
 751
 752         _ if intrinsic.starts_with("atomic_max"), <T> (v ptr, v src) {
 753             atomic_minmax!(fx, IntCC::SignedGreaterThan, <T> (ptr, src) -> ret);
 754         };
 755         _ if intrinsic.starts_with("atomic_umax"), <T> (v ptr, v src) {
 756             atomic_minmax!(fx, IntCC::UnsignedGreaterThan, <T> (ptr, src) -> ret);
 757         };
 758         _ if intrinsic.starts_with("atomic_min"), <T> (v ptr, v src) {
 759             atomic_minmax!(fx, IntCC::SignedLessThan, <T> (ptr, src) -> ret);
 760         };
 761         _ if intrinsic.starts_with("atomic_umin"), <T> (v ptr, v src) {
 762             atomic_minmax!(fx, IntCC::UnsignedLessThan, <T> (ptr, src) -> ret);
 763         };
 764
 765         minnumf32, (v a, v b) {
 766             let val = fx.bcx.ins().fmin(a, b);
 767             let val = CValue::by_val(val, fx.layout_of(fx.tcx.types.f32));
 768             ret.write_cvalue(fx, val);
 769         };
 770         minnumf64, (v a, v b) {
 771             let val = fx.bcx.ins().fmin(a, b);
 772             let val = CValue::by_val(val, fx.layout_of(fx.tcx.types.f64));
 773             ret.write_cvalue(fx, val);
 774         };
 775         maxnumf32, (v a, v b) {
 776             let val = fx.bcx.ins().fmax(a, b);
 777             let val = CValue::by_val(val, fx.layout_of(fx.tcx.types.f32));
 778             ret.write_cvalue(fx, val);
 779         };
 780         maxnumf64, (v a, v b) {
 781             let val = fx.bcx.ins().fmax(a, b);
 782             let val = CValue::by_val(val, fx.layout_of(fx.tcx.types.f64));
 783             ret.write_cvalue(fx, val);
 784         };
 785
 786         simd_cast, (c x) {
 787             ret.write_cvalue(fx, x.unchecked_cast_to(ret.layout()));
 788         };
 789
 790         simd_eq, (c x, c y) {
 791             simd_cmp!(fx, intrinsic, Equal(x, y) -> ret);
 792         };
 793         simd_ne, (c x, c y) {
 794             simd_cmp!(fx, intrinsic, NotEqual(x, y) -> ret);
 795         };
 796         simd_lt, (c x, c y) {
 797             simd_cmp!(fx, intrinsic, UnsignedLessThan|SignedLessThan(x, y) -> ret);
 798         };
 799         simd_le, (c x, c y) {
 800             simd_cmp!(fx, intrinsic, UnsignedLessThanOrEqual|SignedLessThanOrEqual(x, y) -> ret);
 801         };
 802         simd_gt, (c x, c y) {
 803             simd_cmp!(fx, intrinsic, UnsignedGreaterThan|SignedGreaterThan(x, y) -> ret);
 804         };
 805         simd_ge, (c x, c y) {
 806             simd_cmp!(fx, intrinsic, UnsignedGreaterThanOrEqual|SignedGreaterThanOrEqual(x, y) -> ret);
 807         };
 808
 809         // simd_shuffle32<T, U>(x: T, y: T, idx: [u32; 32]) -> U
 810         _ if intrinsic.starts_with("simd_shuffle"), (c x, c y, c idx) {
 811             let n: usize = intrinsic["simd_shuffle".len()..].parse().unwrap();
 812
 813             assert_eq!(x.layout(), y.layout());
 814             let layout = x.layout();
 815
 816             let (lane_type, lane_count) = lane_type_and_count(fx, layout, intrinsic);
 817             let (ret_lane_type, ret_lane_count) = lane_type_and_count(fx, ret.layout(), intrinsic);
 818
 819             assert_eq!(lane_type, ret_lane_type);
 820             assert_eq!(n, ret_lane_count);
 821
 822             let total_len = lane_count * 2;
 823
 824             // TODO get shuffle indices
 825             fx.tcx.sess.warn("simd_shuffle* not yet implemented");
 826             crate::trap::trap_unimplemented(fx, "simd_shuffle* not yet implemented");
 827         };
 828
 829         simd_add, (c x, c y) {
 830             simd_binop!(fx, intrinsic, iadd(x, y) -> ret);
 831         };
 832         simd_sub, (c x, c y) {
 833             simd_binop!(fx, intrinsic, isub(x, y) -> ret);
 834         };
 835         simd_mul, (c x, c y) {
 836             simd_binop!(fx, intrinsic, imul(x, y) -> ret);
 837         };
 838         simd_div, (c x, c y) {
 839             simd_binop!(fx, intrinsic, udiv|sdiv(x, y) -> ret);
 840         };
 841         simd_rem, (c x, c y) {
 842             simd_binop!(fx, intrinsic, urem|srem(x, y) -> ret);
 843         };
 844         simd_shl, (c x, c y) {
 845             simd_binop!(fx, intrinsic, ishl(x, y) -> ret);
 846         };
 847         simd_shr, (c x, c y) {
 848             simd_binop!(fx, intrinsic, ushr|sshr(x, y) -> ret);
 849         };
 850         simd_and, (c x, c y) {
 851             simd_binop!(fx, intrinsic, band(x, y) -> ret);
 852         };
 853         simd_or, (c x, c y) {
 854             simd_binop!(fx, intrinsic, bor(x, y) -> ret);
 855         };
 856         simd_xor, (c x, c y) {
 857             simd_binop!(fx, intrinsic, bxor(x, y) -> ret);
 858         };
 859
 860         simd_fmin, (c x, c y) {
 861             simd_binop!(fx, intrinsic, fmin(x, y) -> ret);
 862         };
 863         simd_fmax, (c x, c y) {
 864             simd_binop!(fx, intrinsic, fmax(x, y) -> ret);
 865         };
 866     }
 867
 868     if let Some((_, dest)) = destination {
 869         let ret_ebb = fx.get_ebb(dest);
 870         fx.bcx.ins().jump(ret_ebb, &[]);
 871     } else {
 872         trap_unreachable(fx, "[corruption] Diverging intrinsic returned.");
 873     }
 874 }