src/intrinsics/mod.rs

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