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