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