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     (o $fx:expr, $arg:ident) => {
  16         $arg
  17     };
  18     (c $fx:expr, $arg:ident) => {
  19         trans_operand($fx, $arg)
  20     };
  21     (v $fx:expr, $arg:ident) => {
  22         trans_operand($fx, $arg).load_scalar($fx)
  23     };
  24 }
  25
  26 macro_rules! intrinsic_substs {
  27     ($substs:expr, $index:expr,) => {};
  28     ($substs:expr, $index:expr, $first:ident $(,$rest:ident)*) => {
  29         let $first = $substs.type_at($index);
  30         intrinsic_substs!($substs, $index+1, $($rest),*);
  31     };
  32 }
  33
  34 macro_rules! intrinsic_match {
  35     ($fx:expr, $intrinsic:expr, $substs:expr, $args:expr, $(
  36         $($name:tt)|+ $(if $cond:expr)?, $(<$($subst:ident),*>)? ($($a:ident $arg:ident),*) $content:block;
  37     )*) => {
  38         match $intrinsic {
  39             $(
  40                 $(intrinsic_pat!($name))|* $(if $cond)? => {
  41                     #[allow(unused_parens, non_snake_case)]
  42                     {
  43                         $(
  44                             intrinsic_substs!($substs, 0, $($subst),*);
  45                         )?
  46                         if let [$($arg),*] = $args {
  47                             let ($($arg,)*) = (
  48                                 $(intrinsic_arg!($a $fx, $arg),)*
  49                             );
  50                             #[warn(unused_parens, non_snake_case)]
  51                             {
  52                                 $content
  53                             }
  54                         } else {
  55                             bug!("wrong number of args for intrinsic {:?}", $intrinsic);
  56                         }
  57                     }
  58                 }
  59             )*
  60             _ => unimpl!("unsupported intrinsic {}", $intrinsic),
  61         }
  62     };
  63 }
  64
  65 macro_rules! call_intrinsic_match {
  66     ($fx:expr, $intrinsic:expr, $substs:expr, $ret:expr, $destination:expr, $args:expr, $(
  67         $name:ident($($arg:ident),*) -> $ty:ident => $func:ident,
  68     )*) => {
  69         match $intrinsic {
  70             $(
  71                 stringify!($name) => {
  72                     assert!($substs.is_noop());
  73                     if let [$(ref $arg),*] = *$args {
  74                         let ($($arg,)*) = (
  75                             $(trans_operand($fx, $arg),)*
  76                         );
  77                         let res = $fx.easy_call(stringify!($func), &[$($arg),*], $fx.tcx.types.$ty);
  78                         $ret.write_cvalue($fx, res);
  79
  80                         if let Some((_, dest)) = $destination {
  81                             let ret_ebb = $fx.get_ebb(dest);
  82                             $fx.bcx.ins().jump(ret_ebb, &[]);
  83                             return;
  84                         } else {
  85                             unreachable!();
  86                         }
  87                     } else {
  88                         bug!("wrong number of args for intrinsic {:?}", $intrinsic);
  89                     }
  90                 }
  91             )*
  92             _ => {}
  93         }
  94     }
  95 }
  96
  97 macro_rules! atomic_binop_return_old {
  98     ($fx:expr, $op:ident<$T:ident>($ptr:ident, $src:ident) -> $ret:ident) => {
  99         let clif_ty = $fx.clif_type($T).unwrap();
 100         let old = $fx.bcx.ins().load(clif_ty, MemFlags::new(), $ptr, 0);
 101         let new = $fx.bcx.ins().$op(old, $src);
 102         $fx.bcx.ins().store(MemFlags::new(), new, $ptr, 0);
 103         $ret.write_cvalue($fx, CValue::by_val(old, $fx.layout_of($T)));
 104     };
 105 }
 106
 107 macro_rules! atomic_minmax {
 108     ($fx:expr, $cc:expr, <$T:ident> ($ptr:ident, $src:ident) -> $ret:ident) => {
 109         // Read old
 110         let clif_ty = $fx.clif_type($T).unwrap();
 111         let old = $fx.bcx.ins().load(clif_ty, MemFlags::new(), $ptr, 0);
 112
 113         // Compare
 114         let is_eq = $fx.bcx.ins().icmp(IntCC::SignedGreaterThan, old, $src);
 115         let new = crate::common::codegen_select(&mut $fx.bcx, is_eq, old, $src);
 116
 117         // Write new
 118         $fx.bcx.ins().store(MemFlags::new(), new, $ptr, 0);
 119
 120         let ret_val = CValue::by_val(old, $ret.layout());
 121         $ret.write_cvalue($fx, ret_val);
 122     };
 123 }
 124
 125 fn lane_type_and_count<'tcx>(
 126     fx: &FunctionCx<'_, 'tcx, impl Backend>,
 127     layout: TyLayout<'tcx>,
 128     intrinsic: &str,
 129 ) -> (TyLayout<'tcx>, u32) {
 130     assert!(layout.ty.is_simd());
 131     let lane_count = match layout.fields {
 132         layout::FieldPlacement::Array { stride: _, count } => u32::try_from(count).unwrap(),
 133         _ => panic!("Non vector type {:?} passed to or returned from simd_* intrinsic {}", layout.ty, intrinsic),
 134     };
 135     let lane_layout = layout.field(fx, 0);
 136     (lane_layout, lane_count)
 137 }
 138
 139 fn simd_for_each_lane<'tcx, B: Backend>(
 140     fx: &mut FunctionCx<'_, 'tcx, B>,
 141     intrinsic: &str,
 142     x: CValue<'tcx>,
 143     y: CValue<'tcx>,
 144     ret: CPlace<'tcx>,
 145     f: impl Fn(&mut FunctionCx<'_, 'tcx, B>, TyLayout<'tcx>, TyLayout<'tcx>, Value, Value) -> CValue<'tcx>,
 146 ) {
 147     assert_eq!(x.layout(), y.layout());
 148     let layout = x.layout();
 149
 150     let (lane_layout, lane_count) = lane_type_and_count(fx, layout, intrinsic);
 151     let (ret_lane_layout, ret_lane_count) = lane_type_and_count(fx, ret.layout(), intrinsic);
 152     assert_eq!(lane_count, ret_lane_count);
 153
 154     for lane in 0..lane_count {
 155         let lane = mir::Field::new(lane.try_into().unwrap());
 156         let x_lane = x.value_field(fx, lane).load_scalar(fx);
 157         let y_lane = y.value_field(fx, lane).load_scalar(fx);
 158
 159         let res_lane = f(fx, lane_layout, ret_lane_layout, x_lane, y_lane);
 160
 161         ret.place_field(fx, lane).write_cvalue(fx, res_lane);
 162     }
 163 }
 164
 165 fn bool_to_zero_or_max_uint<'tcx>(
 166     fx: &mut FunctionCx<'_, 'tcx, impl Backend>,
 167     layout: TyLayout<'tcx>,
 168     val: Value,
 169 ) -> CValue<'tcx> {
 170     let ty = fx.clif_type(layout.ty).unwrap();
 171
 172     let zero = fx.bcx.ins().iconst(ty, 0);
 173     let max = fx.bcx.ins().iconst(ty, (u64::max_value() >> (64 - ty.bits())) as i64);
 174     let res = crate::common::codegen_select(&mut fx.bcx, val, max, zero);
 175     CValue::by_val(res, layout)
 176 }
 177
 178 macro_rules! simd_cmp {
 179     ($fx:expr, $intrinsic:expr, $cc:ident($x:ident, $y:ident) -> $ret:ident) => {
 180         simd_for_each_lane($fx, $intrinsic, $x, $y, $ret, |fx, _lane_layout, res_lane_layout, x_lane, y_lane| {
 181             let res_lane = fx.bcx.ins().icmp(IntCC::$cc, x_lane, y_lane);
 182             bool_to_zero_or_max_uint(fx, res_lane_layout, res_lane)
 183         });
 184     };
 185     ($fx:expr, $intrinsic:expr, $cc_u:ident|$cc_s:ident($x:ident, $y:ident) -> $ret:ident) => {
 186         simd_for_each_lane($fx, $intrinsic, $x, $y, $ret, |fx, lane_layout, res_lane_layout, x_lane, y_lane| {
 187             let res_lane = match lane_layout.ty.sty {
 188                 ty::Uint(_) => fx.bcx.ins().icmp(IntCC::$cc_u, x_lane, y_lane),
 189                 ty::Int(_) => fx.bcx.ins().icmp(IntCC::$cc_s, x_lane, y_lane),
 190                 _ => unreachable!("{:?}", lane_layout.ty),
 191             };
 192             bool_to_zero_or_max_uint(fx, res_lane_layout, res_lane)
 193         });
 194     };
 195
 196 }
 197
 198 macro_rules! simd_binop {
 199     ($fx:expr, $intrinsic:expr, $op:ident($x:ident, $y:ident) -> $ret:ident) => {
 200         simd_for_each_lane($fx, $intrinsic, $x, $y, $ret, |fx, _lane_layout, ret_lane_layout, x_lane, y_lane| {
 201             let res_lane = fx.bcx.ins().$op(x_lane, y_lane);
 202             CValue::by_val(res_lane, ret_lane_layout)
 203         });
 204     };
 205     ($fx:expr, $intrinsic:expr, $op_u:ident|$op_s:ident($x:ident, $y:ident) -> $ret:ident) => {
 206         simd_for_each_lane($fx, $intrinsic, $x, $y, $ret, |fx, lane_layout, ret_lane_layout, x_lane, y_lane| {
 207             let res_lane = match lane_layout.ty.sty {
 208                 ty::Uint(_) => fx.bcx.ins().$op_u(x_lane, y_lane),
 209                 ty::Int(_) => fx.bcx.ins().$op_s(x_lane, y_lane),
 210                 _ => unreachable!("{:?}", lane_layout.ty),
 211             };
 212             CValue::by_val(res_lane, ret_lane_layout)
 213         });
 214     };
 215 }
 216
 217 pub fn codegen_intrinsic_call<'a, 'tcx: 'a>(
 218     fx: &mut FunctionCx<'a, 'tcx, impl Backend>,
 219     def_id: DefId,
 220     substs: SubstsRef<'tcx>,
 221     args: &[mir::Operand<'tcx>],
 222     destination: Option<(CPlace<'tcx>, BasicBlock)>,
 223 ) {
 224     let intrinsic = fx.tcx.item_name(def_id).as_str();
 225     let intrinsic = &intrinsic[..];
 226
 227     let ret = match destination {
 228         Some((place, _)) => place,
 229         None => {
 230             // Insert non returning intrinsics here
 231             match intrinsic {
 232                 "abort" => {
 233                     trap_panic(fx, "Called intrinsic::abort.");
 234                 }
 235                 "unreachable" => {
 236                     trap_unreachable(fx, "[corruption] Called intrinsic::unreachable.");
 237                 }
 238                 _ => unimplemented!("unsupported instrinsic {}", intrinsic),
 239             }
 240             return;
 241         }
 242     };
 243
 244     let u64_layout = fx.layout_of(fx.tcx.types.u64);
 245     let usize_layout = fx.layout_of(fx.tcx.types.usize);
 246
 247     call_intrinsic_match! {
 248         fx, intrinsic, substs, ret, destination, args,
 249         expf32(flt) -> f32 => expf,
 250         expf64(flt) -> f64 => exp,
 251         exp2f32(flt) -> f32 => exp2f,
 252         exp2f64(flt) -> f64 => exp2,
 253         sqrtf32(flt) -> f32 => sqrtf,
 254         sqrtf64(flt) -> f64 => sqrt,
 255         powif32(a, x) -> f32 => __powisf2, // compiler-builtins
 256         powif64(a, x) -> f64 => __powidf2, // compiler-builtins
 257         logf32(flt) -> f32 => logf,
 258         logf64(flt) -> f64 => log,
 259         fabsf32(flt) -> f32 => fabsf,
 260         fabsf64(flt) -> f64 => fabs,
 261         fmaf32(x, y, z) -> f32 => fmaf,
 262         fmaf64(x, y, z) -> f64 => fma,
 263
 264         // rounding variants
 265         floorf32(flt) -> f32 => floorf,
 266         floorf64(flt) -> f64 => floor,
 267         ceilf32(flt) -> f32 => ceilf,
 268         ceilf64(flt) -> f64 => ceil,
 269         truncf32(flt) -> f32 => truncf,
 270         truncf64(flt) -> f64 => trunc,
 271         roundf32(flt) -> f32 => roundf,
 272         roundf64(flt) -> f64 => round,
 273
 274         // trigonometry
 275         sinf32(flt) -> f32 => sinf,
 276         sinf64(flt) -> f64 => sin,
 277         cosf32(flt) -> f32 => cosf,
 278         cosf64(flt) -> f64 => cos,
 279         tanf32(flt) -> f32 => tanf,
 280         tanf64(flt) -> f64 => tan,
 281     }
 282
 283     intrinsic_match! {
 284         fx, intrinsic, substs, args,
 285
 286         assume, (c _a) {};
 287         likely | unlikely, (c a) {
 288             ret.write_cvalue(fx, a);
 289         };
 290         breakpoint, () {
 291             fx.bcx.ins().debugtrap();
 292         };
 293         copy | copy_nonoverlapping, <elem_ty> (v src, v dst, v count) {
 294             let elem_size: u64 = fx.layout_of(elem_ty).size.bytes();
 295             let elem_size = fx
 296                 .bcx
 297                 .ins()
 298                 .iconst(fx.pointer_type, elem_size as i64);
 299             assert_eq!(args.len(), 3);
 300             let byte_amount = fx.bcx.ins().imul(count, elem_size);
 301
 302             if intrinsic.ends_with("_nonoverlapping") {
 303                 fx.bcx.call_memcpy(fx.module.target_config(), dst, src, byte_amount);
 304             } else {
 305                 fx.bcx.call_memmove(fx.module.target_config(), dst, src, byte_amount);
 306             }
 307         };
 308         discriminant_value, (c val) {
 309             let pointee_layout = fx.layout_of(val.layout().ty.builtin_deref(true).unwrap().ty);
 310             let place = CPlace::for_addr(val.load_scalar(fx), pointee_layout);
 311             let discr = crate::base::trans_get_discriminant(fx, place, ret.layout());
 312             ret.write_cvalue(fx, discr);
 313         };
 314         size_of, <T> () {
 315             let size_of = fx.layout_of(T).size.bytes();
 316             let size_of = CValue::const_val(fx, usize_layout.ty, size_of.into());
 317             ret.write_cvalue(fx, size_of);
 318         };
 319         size_of_val, <T> (c ptr) {
 320             let layout = fx.layout_of(T);
 321             let size = if layout.is_unsized() {
 322                 let (_ptr, info) = ptr.load_scalar_pair(fx);
 323                 let (size, _align) = crate::unsize::size_and_align_of_dst(fx, layout.ty, info);
 324                 size
 325             } else {
 326                 fx
 327                     .bcx
 328                     .ins()
 329                     .iconst(fx.pointer_type, layout.size.bytes() as i64)
 330             };
 331             ret.write_cvalue(fx, CValue::by_val(size, usize_layout));
 332         };
 333         min_align_of, <T> () {
 334             let min_align = fx.layout_of(T).align.abi.bytes();
 335             let min_align = CValue::const_val(fx, usize_layout.ty, min_align.into());
 336             ret.write_cvalue(fx, min_align);
 337         };
 338         min_align_of_val, <T> (c ptr) {
 339             let layout = fx.layout_of(T);
 340             let align = if layout.is_unsized() {
 341                 let (_ptr, info) = ptr.load_scalar_pair(fx);
 342                 let (_size, align) = crate::unsize::size_and_align_of_dst(fx, layout.ty, info);
 343                 align
 344             } else {
 345                 fx
 346                     .bcx
 347                     .ins()
 348                     .iconst(fx.pointer_type, layout.align.abi.bytes() as i64)
 349             };
 350             ret.write_cvalue(fx, CValue::by_val(align, usize_layout));
 351         };
 352         pref_align_of, <T> () {
 353             let pref_align = fx.layout_of(T).align.pref.bytes();
 354             let pref_align = CValue::const_val(fx, usize_layout.ty, pref_align.into());
 355             ret.write_cvalue(fx, pref_align);
 356         };
 357
 358
 359         type_id, <T> () {
 360             let type_id = fx.tcx.type_id_hash(T);
 361             let type_id = CValue::const_val(fx, u64_layout.ty, type_id.into());
 362             ret.write_cvalue(fx, type_id);
 363         };
 364         type_name, <T> () {
 365             let type_name = fx.tcx.type_name(T);
 366             let type_name = crate::constant::trans_const_value(fx, type_name);
 367             ret.write_cvalue(fx, type_name);
 368         };
 369
 370         _ if intrinsic.starts_with("unchecked_") || intrinsic == "exact_div", (c x, c y) {
 371             // FIXME trap on overflow
 372             let bin_op = match intrinsic {
 373                 "unchecked_sub" => BinOp::Sub,
 374                 "unchecked_div" | "exact_div" => BinOp::Div,
 375                 "unchecked_rem" => BinOp::Rem,
 376                 "unchecked_shl" => BinOp::Shl,
 377                 "unchecked_shr" => BinOp::Shr,
 378                 _ => unimplemented!("intrinsic {}", intrinsic),
 379             };
 380             let res = match ret.layout().ty.sty {
 381                 ty::Uint(_) => crate::base::trans_int_binop(
 382                     fx,
 383                     bin_op,
 384                     x,
 385                     y,
 386                     ret.layout().ty,
 387                     false,
 388                 ),
 389                 ty::Int(_) => crate::base::trans_int_binop(
 390                     fx,
 391                     bin_op,
 392                     x,
 393                     y,
 394                     ret.layout().ty,
 395                     true,
 396                 ),
 397                 _ => panic!(),
 398             };
 399             ret.write_cvalue(fx, res);
 400         };
 401         _ if intrinsic.ends_with("_with_overflow"), <T> (c x, c y) {
 402             assert_eq!(x.layout().ty, y.layout().ty);
 403             let bin_op = match intrinsic {
 404                 "add_with_overflow" => BinOp::Add,
 405                 "sub_with_overflow" => BinOp::Sub,
 406                 "mul_with_overflow" => BinOp::Mul,
 407                 _ => unimplemented!("intrinsic {}", intrinsic),
 408             };
 409             let res = match T.sty {
 410                 ty::Uint(_) => crate::base::trans_checked_int_binop(
 411                     fx,
 412                     bin_op,
 413                     x,
 414                     y,
 415                     ret.layout().ty,
 416                     false,
 417                 ),
 418                 ty::Int(_) => crate::base::trans_checked_int_binop(
 419                     fx,
 420                     bin_op,
 421                     x,
 422                     y,
 423                     ret.layout().ty,
 424                     true,
 425                 ),
 426                 _ => panic!(),
 427             };
 428             ret.write_cvalue(fx, res);
 429         };
 430         _ if intrinsic.starts_with("overflowing_"), <T> (c x, c y) {
 431             assert_eq!(x.layout().ty, y.layout().ty);
 432             let bin_op = match intrinsic {
 433                 "overflowing_add" => BinOp::Add,
 434                 "overflowing_sub" => BinOp::Sub,
 435                 "overflowing_mul" => BinOp::Mul,
 436                 _ => unimplemented!("intrinsic {}", intrinsic),
 437             };
 438             let res = match T.sty {
 439                 ty::Uint(_) => crate::base::trans_int_binop(
 440                     fx,
 441                     bin_op,
 442                     x,
 443                     y,
 444                     ret.layout().ty,
 445                     false,
 446                 ),
 447                 ty::Int(_) => crate::base::trans_int_binop(
 448                     fx,
 449                     bin_op,
 450                     x,
 451                     y,
 452                     ret.layout().ty,
 453                     true,
 454                 ),
 455                 _ => panic!(),
 456             };
 457             ret.write_cvalue(fx, res);
 458         };
 459         _ if intrinsic.starts_with("saturating_"), <T> (c x, c y) {
 460             // FIXME implement saturating behavior
 461             assert_eq!(x.layout().ty, y.layout().ty);
 462             let bin_op = match intrinsic {
 463                 "saturating_add" => BinOp::Add,
 464                 "saturating_sub" => BinOp::Sub,
 465                 "saturating_mul" => BinOp::Mul,
 466                 _ => unimplemented!("intrinsic {}", intrinsic),
 467             };
 468             let res = match T.sty {
 469                 ty::Uint(_) => crate::base::trans_int_binop(
 470                     fx,
 471                     bin_op,
 472                     x,
 473                     y,
 474                     ret.layout().ty,
 475                     false,
 476                 ),
 477                 ty::Int(_) => crate::base::trans_int_binop(
 478                     fx,
 479                     bin_op,
 480                     x,
 481                     y,
 482                     ret.layout().ty,
 483                     true,
 484                 ),
 485                 _ => panic!(),
 486             };
 487             ret.write_cvalue(fx, res);
 488         };
 489         rotate_left, <T>(v x, v y) {
 490             let layout = fx.layout_of(T);
 491             let res = fx.bcx.ins().rotl(x, y);
 492             ret.write_cvalue(fx, CValue::by_val(res, layout));
 493         };
 494         rotate_right, <T>(v x, v y) {
 495             let layout = fx.layout_of(T);
 496             let res = fx.bcx.ins().rotr(x, y);
 497             ret.write_cvalue(fx, CValue::by_val(res, layout));
 498         };
 499
 500         // The only difference between offset and arith_offset is regarding UB. Because Cranelift
 501         // doesn't have UB both are codegen'ed the same way
 502         offset | arith_offset, (c base, v offset) {
 503             let pointee_ty = base.layout().ty.builtin_deref(true).unwrap().ty;
 504             let pointee_size = fx.layout_of(pointee_ty).size.bytes();
 505             let ptr_diff = fx.bcx.ins().imul_imm(offset, pointee_size as i64);
 506             let base_val = base.load_scalar(fx);
 507             let res = fx.bcx.ins().iadd(base_val, ptr_diff);
 508             ret.write_cvalue(fx, CValue::by_val(res, base.layout()));
 509         };
 510
 511         transmute, <src_ty, dst_ty> (c from) {
 512             assert_eq!(from.layout().ty, src_ty);
 513             let addr = from.force_stack(fx);
 514             let dst_layout = fx.layout_of(dst_ty);
 515             ret.write_cvalue(fx, CValue::by_ref(addr, dst_layout))
 516         };
 517         init, () {
 518             if ret.layout().abi == Abi::Uninhabited {
 519                 crate::trap::trap_panic(fx, "[panic] Called intrinsic::init for uninhabited type.");
 520                 return;
 521             }
 522
 523             match ret {
 524                 CPlace::NoPlace(_layout) => {}
 525                 CPlace::Var(var, layout) => {
 526                     let clif_ty = fx.clif_type(layout.ty).unwrap();
 527                     let val = match clif_ty {
 528                         types::I8 | types::I16 | types::I32 | types::I64 => fx.bcx.ins().iconst(clif_ty, 0),
 529                         types::F32 => {
 530                             let zero = fx.bcx.ins().iconst(types::I32, 0);
 531                             fx.bcx.ins().bitcast(types::F32, zero)
 532                         }
 533                         types::F64 => {
 534                             let zero = fx.bcx.ins().iconst(types::I64, 0);
 535                             fx.bcx.ins().bitcast(types::F64, zero)
 536                         }
 537                         _ => panic!("clif_type returned {}", clif_ty),
 538                     };
 539                     fx.bcx.def_var(mir_var(var), val);
 540                 }
 541                 _ => {
 542                     let addr = ret.to_addr(fx);
 543                     let layout = ret.layout();
 544                     fx.bcx.emit_small_memset(fx.module.target_config(), addr, 0, layout.size.bytes(), 1);
 545                 }
 546             }
 547         };
 548         write_bytes, (c dst, v val, v count) {
 549             let pointee_ty = dst.layout().ty.builtin_deref(true).unwrap().ty;
 550             let pointee_size = fx.layout_of(pointee_ty).size.bytes();
 551             let count = fx.bcx.ins().imul_imm(count, pointee_size as i64);
 552             let dst_ptr = dst.load_scalar(fx);
 553             fx.bcx.call_memset(fx.module.target_config(), dst_ptr, val, count);
 554         };
 555         ctlz | ctlz_nonzero, <T> (v arg) {
 556             let res = if T == fx.tcx.types.u128 || T == fx.tcx.types.i128 {
 557                 // FIXME verify this algorithm is correct
 558                 let (lsb, msb) = fx.bcx.ins().isplit(arg);
 559                 let lsb_lz = fx.bcx.ins().clz(lsb);
 560                 let msb_lz = fx.bcx.ins().clz(msb);
 561                 let msb_is_zero = fx.bcx.ins().icmp_imm(IntCC::Equal, msb, 0);
 562                 let lsb_lz_plus_64 = fx.bcx.ins().iadd_imm(lsb_lz, 64);
 563                 fx.bcx.ins().select(msb_is_zero, lsb_lz_plus_64, msb_lz)
 564             } else {
 565                 fx.bcx.ins().clz(arg)
 566             };
 567             let res = CValue::by_val(res, fx.layout_of(T));
 568             ret.write_cvalue(fx, res);
 569         };
 570         cttz | cttz_nonzero, <T> (v arg) {
 571             let res = if T == fx.tcx.types.u128 || T == fx.tcx.types.i128 {
 572                 // FIXME verify this algorithm is correct
 573                 let (lsb, msb) = fx.bcx.ins().isplit(arg);
 574                 let lsb_tz = fx.bcx.ins().ctz(lsb);
 575                 let msb_tz = fx.bcx.ins().ctz(msb);
 576                 let lsb_is_zero = fx.bcx.ins().icmp_imm(IntCC::Equal, lsb, 0);
 577                 let msb_tz_plus_64 = fx.bcx.ins().iadd_imm(msb_tz, 64);
 578                 fx.bcx.ins().select(lsb_is_zero, msb_tz_plus_64, lsb_tz)
 579             } else {
 580                 fx.bcx.ins().ctz(arg)
 581             };
 582             let res = CValue::by_val(res, fx.layout_of(T));
 583             ret.write_cvalue(fx, res);
 584         };
 585         ctpop, <T> (v arg) {
 586             let res = CValue::by_val(fx.bcx.ins().popcnt(arg), fx.layout_of(T));
 587             ret.write_cvalue(fx, res);
 588         };
 589         bitreverse, <T> (v arg) {
 590             let res = CValue::by_val(fx.bcx.ins().bitrev(arg), fx.layout_of(T));
 591             ret.write_cvalue(fx, res);
 592         };
 593         bswap, <T> (v arg) {
 594             // FIXME(CraneStation/cranelift#794) add bswap instruction to cranelift
 595             fn swap(bcx: &mut FunctionBuilder, v: Value) -> Value {
 596                 match bcx.func.dfg.value_type(v) {
 597                     types::I8 => v,
 598
 599                     // https://code.woboq.org/gcc/include/bits/byteswap.h.html
 600                     types::I16 => {
 601                         let tmp1 = bcx.ins().ishl_imm(v, 8);
 602                         let n1 = bcx.ins().band_imm(tmp1, 0xFF00);
 603
 604                         let tmp2 = bcx.ins().ushr_imm(v, 8);
 605                         let n2 = bcx.ins().band_imm(tmp2, 0x00FF);
 606
 607                         bcx.ins().bor(n1, n2)
 608                     }
 609                     types::I32 => {
 610                         let tmp1 = bcx.ins().ishl_imm(v, 24);
 611                         let n1 = bcx.ins().band_imm(tmp1, 0xFF00_0000);
 612
 613                         let tmp2 = bcx.ins().ishl_imm(v, 8);
 614                         let n2 = bcx.ins().band_imm(tmp2, 0x00FF_0000);
 615
 616                         let tmp3 = bcx.ins().ushr_imm(v, 8);
 617                         let n3 = bcx.ins().band_imm(tmp3, 0x0000_FF00);
 618
 619                         let tmp4 = bcx.ins().ushr_imm(v, 24);
 620                         let n4 = bcx.ins().band_imm(tmp4, 0x0000_00FF);
 621
 622                         let or_tmp1 = bcx.ins().bor(n1, n2);
 623                         let or_tmp2 = bcx.ins().bor(n3, n4);
 624                         bcx.ins().bor(or_tmp1, or_tmp2)
 625                     }
 626                     types::I64 => {
 627                         let tmp1 = bcx.ins().ishl_imm(v, 56);
 628                         let n1 = bcx.ins().band_imm(tmp1, 0xFF00_0000_0000_0000u64 as i64);
 629
 630                         let tmp2 = bcx.ins().ishl_imm(v, 40);
 631                         let n2 = bcx.ins().band_imm(tmp2, 0x00FF_0000_0000_0000u64 as i64);
 632
 633                         let tmp3 = bcx.ins().ishl_imm(v, 24);
 634                         let n3 = bcx.ins().band_imm(tmp3, 0x0000_FF00_0000_0000u64 as i64);
 635
 636                         let tmp4 = bcx.ins().ishl_imm(v, 8);
 637                         let n4 = bcx.ins().band_imm(tmp4, 0x0000_00FF_0000_0000u64 as i64);
 638
 639                         let tmp5 = bcx.ins().ushr_imm(v, 8);
 640                         let n5 = bcx.ins().band_imm(tmp5, 0x0000_0000_FF00_0000u64 as i64);
 641
 642                         let tmp6 = bcx.ins().ushr_imm(v, 24);
 643                         let n6 = bcx.ins().band_imm(tmp6, 0x0000_0000_00FF_0000u64 as i64);
 644
 645                         let tmp7 = bcx.ins().ushr_imm(v, 40);
 646                         let n7 = bcx.ins().band_imm(tmp7, 0x0000_0000_0000_FF00u64 as i64);
 647
 648                         let tmp8 = bcx.ins().ushr_imm(v, 56);
 649                         let n8 = bcx.ins().band_imm(tmp8, 0x0000_0000_0000_00FFu64 as i64);
 650
 651                         let or_tmp1 = bcx.ins().bor(n1, n2);
 652                         let or_tmp2 = bcx.ins().bor(n3, n4);
 653                         let or_tmp3 = bcx.ins().bor(n5, n6);
 654                         let or_tmp4 = bcx.ins().bor(n7, n8);
 655
 656                         let or_tmp5 = bcx.ins().bor(or_tmp1, or_tmp2);
 657                         let or_tmp6 = bcx.ins().bor(or_tmp3, or_tmp4);
 658                         bcx.ins().bor(or_tmp5, or_tmp6)
 659                     }
 660                     types::I128 => {
 661                         let (lo, hi) = bcx.ins().isplit(v);
 662                         let lo = swap(bcx, lo);
 663                         let hi = swap(bcx, hi);
 664                         bcx.ins().iconcat(hi, lo)
 665                     }
 666                     ty => unimplemented!("bswap {}", ty),
 667                 }
 668             };
 669             let res = CValue::by_val(swap(&mut fx.bcx, arg), fx.layout_of(T));
 670             ret.write_cvalue(fx, res);
 671         };
 672         needs_drop, <T> () {
 673             let needs_drop = if T.needs_drop(fx.tcx, ParamEnv::reveal_all()) {
 674                 1
 675             } else {
 676                 0
 677             };
 678             let needs_drop = CValue::const_val(fx, fx.tcx.types.bool, needs_drop);
 679             ret.write_cvalue(fx, needs_drop);
 680         };
 681         panic_if_uninhabited, <T> () {
 682             if fx.layout_of(T).abi.is_uninhabited() {
 683                 crate::trap::trap_panic(fx, "[panic] Called intrinsic::panic_if_uninhabited for uninhabited type.");
 684                 return;
 685             }
 686         };
 687
 688         volatile_load, (c ptr) {
 689             // Cranelift treats loads as volatile by default
 690             let inner_layout =
 691                 fx.layout_of(ptr.layout().ty.builtin_deref(true).unwrap().ty);
 692             let val = CValue::by_ref(ptr.load_scalar(fx), inner_layout);
 693             ret.write_cvalue(fx, val);
 694         };
 695         volatile_store, (v ptr, c val) {
 696             // Cranelift treats stores as volatile by default
 697             let dest = CPlace::for_addr(ptr, val.layout());
 698             dest.write_cvalue(fx, val);
 699         };
 700
 701         _ if intrinsic.starts_with("atomic_fence"), () {};
 702         _ if intrinsic.starts_with("atomic_singlethreadfence"), () {};
 703         _ if intrinsic.starts_with("atomic_load"), (c ptr) {
 704             let inner_layout =
 705                 fx.layout_of(ptr.layout().ty.builtin_deref(true).unwrap().ty);
 706             let val = CValue::by_ref(ptr.load_scalar(fx), inner_layout);
 707             ret.write_cvalue(fx, val);
 708         };
 709         _ if intrinsic.starts_with("atomic_store"), (v ptr, c val) {
 710             let dest = CPlace::for_addr(ptr, val.layout());
 711             dest.write_cvalue(fx, val);
 712         };
 713         _ if intrinsic.starts_with("atomic_xchg"), <T> (v ptr, c src) {
 714             // Read old
 715             let clif_ty = fx.clif_type(T).unwrap();
 716             let old = fx.bcx.ins().load(clif_ty, MemFlags::new(), ptr, 0);
 717             ret.write_cvalue(fx, CValue::by_val(old, fx.layout_of(T)));
 718
 719             // Write new
 720             let dest = CPlace::for_addr(ptr, src.layout());
 721             dest.write_cvalue(fx, src);
 722         };
 723         _ if intrinsic.starts_with("atomic_cxchg"), <T> (v ptr, v test_old, v new) { // both atomic_cxchg_* and atomic_cxchgweak_*
 724             // Read old
 725             let clif_ty = fx.clif_type(T).unwrap();
 726             let old = fx.bcx.ins().load(clif_ty, MemFlags::new(), ptr, 0);
 727
 728             // Compare
 729             let is_eq = fx.bcx.ins().icmp(IntCC::Equal, old, test_old);
 730             let new = crate::common::codegen_select(&mut fx.bcx, is_eq, new, old); // Keep old if not equal to test_old
 731
 732             // Write new
 733             fx.bcx.ins().store(MemFlags::new(), new, ptr, 0);
 734
 735             let ret_val = CValue::by_val_pair(old, fx.bcx.ins().bint(types::I8, is_eq), ret.layout());
 736             ret.write_cvalue(fx, ret_val);
 737         };
 738
 739         _ if intrinsic.starts_with("atomic_xadd"), <T> (v ptr, v amount) {
 740             atomic_binop_return_old! (fx, iadd<T>(ptr, amount) -> ret);
 741         };
 742         _ if intrinsic.starts_with("atomic_xsub"), <T> (v ptr, v amount) {
 743             atomic_binop_return_old! (fx, isub<T>(ptr, amount) -> ret);
 744         };
 745         _ if intrinsic.starts_with("atomic_and"), <T> (v ptr, v src) {
 746             atomic_binop_return_old! (fx, band<T>(ptr, src) -> ret);
 747         };
 748         _ if intrinsic.starts_with("atomic_nand"), <T> (v ptr, v src) {
 749             atomic_binop_return_old! (fx, band_not<T>(ptr, src) -> ret);
 750         };
 751         _ if intrinsic.starts_with("atomic_or"), <T> (v ptr, v src) {
 752             atomic_binop_return_old! (fx, bor<T>(ptr, src) -> ret);
 753         };
 754         _ if intrinsic.starts_with("atomic_xor"), <T> (v ptr, v src) {
 755             atomic_binop_return_old! (fx, bxor<T>(ptr, src) -> ret);
 756         };
 757
 758         _ if intrinsic.starts_with("atomic_max"), <T> (v ptr, v src) {
 759             atomic_minmax!(fx, IntCC::SignedGreaterThan, <T> (ptr, src) -> ret);
 760         };
 761         _ if intrinsic.starts_with("atomic_umax"), <T> (v ptr, v src) {
 762             atomic_minmax!(fx, IntCC::UnsignedGreaterThan, <T> (ptr, src) -> ret);
 763         };
 764         _ if intrinsic.starts_with("atomic_min"), <T> (v ptr, v src) {
 765             atomic_minmax!(fx, IntCC::SignedLessThan, <T> (ptr, src) -> ret);
 766         };
 767         _ if intrinsic.starts_with("atomic_umin"), <T> (v ptr, v src) {
 768             atomic_minmax!(fx, IntCC::UnsignedLessThan, <T> (ptr, src) -> ret);
 769         };
 770
 771         minnumf32, (v a, v b) {
 772             let val = fx.bcx.ins().fmin(a, b);
 773             let val = CValue::by_val(val, fx.layout_of(fx.tcx.types.f32));
 774             ret.write_cvalue(fx, val);
 775         };
 776         minnumf64, (v a, v b) {
 777             let val = fx.bcx.ins().fmin(a, b);
 778             let val = CValue::by_val(val, fx.layout_of(fx.tcx.types.f64));
 779             ret.write_cvalue(fx, val);
 780         };
 781         maxnumf32, (v a, v b) {
 782             let val = fx.bcx.ins().fmax(a, b);
 783             let val = CValue::by_val(val, fx.layout_of(fx.tcx.types.f32));
 784             ret.write_cvalue(fx, val);
 785         };
 786         maxnumf64, (v a, v b) {
 787             let val = fx.bcx.ins().fmax(a, b);
 788             let val = CValue::by_val(val, fx.layout_of(fx.tcx.types.f64));
 789             ret.write_cvalue(fx, val);
 790         };
 791
 792         simd_cast, (c x) {
 793             ret.write_cvalue(fx, x.unchecked_cast_to(ret.layout()));
 794         };
 795
 796         simd_eq, (c x, c y) {
 797             simd_cmp!(fx, intrinsic, Equal(x, y) -> ret);
 798         };
 799         simd_ne, (c x, c y) {
 800             simd_cmp!(fx, intrinsic, NotEqual(x, y) -> ret);
 801         };
 802         simd_lt, (c x, c y) {
 803             simd_cmp!(fx, intrinsic, UnsignedLessThan|SignedLessThan(x, y) -> ret);
 804         };
 805         simd_le, (c x, c y) {
 806             simd_cmp!(fx, intrinsic, UnsignedLessThanOrEqual|SignedLessThanOrEqual(x, y) -> ret);
 807         };
 808         simd_gt, (c x, c y) {
 809             simd_cmp!(fx, intrinsic, UnsignedGreaterThan|SignedGreaterThan(x, y) -> ret);
 810         };
 811         simd_ge, (c x, c y) {
 812             simd_cmp!(fx, intrinsic, UnsignedGreaterThanOrEqual|SignedGreaterThanOrEqual(x, y) -> ret);
 813         };
 814
 815         // simd_shuffle32<T, U>(x: T, y: T, idx: [u32; 32]) -> U
 816         _ if intrinsic.starts_with("simd_shuffle"), (c x, c y, o idx) {
 817             let n: u32 = intrinsic["simd_shuffle".len()..].parse().unwrap();
 818
 819             assert_eq!(x.layout(), y.layout());
 820             let layout = x.layout();
 821
 822             let (lane_type, lane_count) = lane_type_and_count(fx, layout, intrinsic);
 823             let (ret_lane_type, ret_lane_count) = lane_type_and_count(fx, ret.layout(), intrinsic);
 824
 825             assert_eq!(lane_type, ret_lane_type);
 826             assert_eq!(n, ret_lane_count);
 827
 828             let total_len = lane_count * 2;
 829
 830             let indexes = {
 831                 use rustc::mir::interpret::*;
 832                 let idx_place = match idx {
 833                     Operand::Copy(idx_place) => {
 834                         idx_place
 835                     }
 836                     _ => panic!("simd_shuffle* idx is not Operand::Copy, but {:?}", idx),
 837                 };
 838
 839                 assert!(idx_place.projection.is_none());
 840                 let static_ = match &idx_place.base {
 841                     PlaceBase::Static(static_) => {
 842                         static_
 843                     }
 844                     PlaceBase::Local(_) => panic!("simd_shuffle* idx is not constant, but a local"),
 845                 };
 846
 847                 let idx_const = match &static_.kind {
 848                     StaticKind::Static(_) => unimplemented!(),
 849                     StaticKind::Promoted(promoted) => {
 850                         fx.tcx.const_eval(ParamEnv::reveal_all().and(GlobalId {
 851                             instance: fx.instance,
 852                             promoted: Some(*promoted),
 853                         })).unwrap()
 854                     }
 855                 };
 856
 857                 let idx_bytes = match idx_const.val {
 858                     ConstValue::ByRef { align: _, offset, alloc } => {
 859                         let ptr = Pointer::new(AllocId(0 /* dummy */), offset);
 860                         let size = Size::from_bytes(4 * u64::from(ret_lane_count) /* size_of([u32; ret_lane_count]) */);
 861                         alloc.get_bytes(fx, ptr, size).unwrap()
 862                     }
 863                     _ => unreachable!("{:?}", idx_const),
 864                 };
 865
 866                 (0..ret_lane_count).map(|i| {
 867                     let i = usize::try_from(i).unwrap();
 868                     let idx = rustc::mir::interpret::read_target_uint(
 869                         fx.tcx.data_layout.endian,
 870                         &idx_bytes[4*i.. 4*i + 4],
 871                     ).expect("read_target_uint");
 872                     u32::try_from(idx).expect("try_from u32")
 873                 }).collect::<Vec<u32>>()
 874             };
 875
 876             for &idx in &indexes {
 877                 assert!(idx < total_len, "idx {} out of range 0..{}", idx, total_len);
 878             }
 879
 880
 881
 882             println!("{:?}", indexes);
 883             unimplemented!();
 884         };
 885
 886         simd_add, (c x, c y) {
 887             simd_binop!(fx, intrinsic, iadd(x, y) -> ret);
 888         };
 889         simd_sub, (c x, c y) {
 890             simd_binop!(fx, intrinsic, isub(x, y) -> ret);
 891         };
 892         simd_mul, (c x, c y) {
 893             simd_binop!(fx, intrinsic, imul(x, y) -> ret);
 894         };
 895         simd_div, (c x, c y) {
 896             simd_binop!(fx, intrinsic, udiv|sdiv(x, y) -> ret);
 897         };
 898         simd_rem, (c x, c y) {
 899             simd_binop!(fx, intrinsic, urem|srem(x, y) -> ret);
 900         };
 901         simd_shl, (c x, c y) {
 902             simd_binop!(fx, intrinsic, ishl(x, y) -> ret);
 903         };
 904         simd_shr, (c x, c y) {
 905             simd_binop!(fx, intrinsic, ushr|sshr(x, y) -> ret);
 906         };
 907         simd_and, (c x, c y) {
 908             simd_binop!(fx, intrinsic, band(x, y) -> ret);
 909         };
 910         simd_or, (c x, c y) {
 911             simd_binop!(fx, intrinsic, bor(x, y) -> ret);
 912         };
 913         simd_xor, (c x, c y) {
 914             simd_binop!(fx, intrinsic, bxor(x, y) -> ret);
 915         };
 916
 917         simd_fmin, (c x, c y) {
 918             simd_binop!(fx, intrinsic, fmin(x, y) -> ret);
 919         };
 920         simd_fmax, (c x, c y) {
 921             simd_binop!(fx, intrinsic, fmax(x, y) -> ret);
 922         };
 923     }
 924
 925     if let Some((_, dest)) = destination {
 926         let ret_ebb = fx.get_ebb(dest);
 927         fx.bcx.ins().jump(ret_ebb, &[]);
 928     } else {
 929         trap_unreachable(fx, "[corruption] Diverging intrinsic returned.");
 930     }
 931 }