src/librustc_codegen_llvm/intrinsic.rs

   1 use crate::abi::{Abi, FnAbi, LlvmType, PassMode};
   2 use crate::builder::Builder;
   3 use crate::context::CodegenCx;
   4 use crate::llvm;
   5 use crate::type_::Type;
   6 use crate::type_of::LayoutLlvmExt;
   7 use crate::va_arg::emit_va_arg;
   8 use crate::value::Value;
   9
  10 use rustc_ast::ast;
  11 use rustc_codegen_ssa::base::{compare_simd_types, to_immediate, wants_msvc_seh};
  12 use rustc_codegen_ssa::common::span_invalid_monomorphization_error;
  13 use rustc_codegen_ssa::common::TypeKind;
  14 use rustc_codegen_ssa::glue;
  15 use rustc_codegen_ssa::mir::operand::{OperandRef, OperandValue};
  16 use rustc_codegen_ssa::mir::place::PlaceRef;
  17 use rustc_codegen_ssa::traits::*;
  18 use rustc_codegen_ssa::MemFlags;
  19 use rustc_hir as hir;
  20 use rustc_middle::ty::layout::{FnAbiExt, HasTyCtxt};
  21 use rustc_middle::ty::{self, Ty};
  22 use rustc_middle::{bug, span_bug};
  23 use rustc_span::Span;
  24 use rustc_target::abi::{self, HasDataLayout, LayoutOf, Primitive};
  25
  26 use std::cmp::Ordering;
  27 use std::iter;
  28
  29 fn get_simple_intrinsic(cx: &CodegenCx<'ll, '_>, name: &str) -> Option<&'ll Value> {
  30     let llvm_name = match name {
  31         "sqrtf32" => "llvm.sqrt.f32",
  32         "sqrtf64" => "llvm.sqrt.f64",
  33         "powif32" => "llvm.powi.f32",
  34         "powif64" => "llvm.powi.f64",
  35         "sinf32" => "llvm.sin.f32",
  36         "sinf64" => "llvm.sin.f64",
  37         "cosf32" => "llvm.cos.f32",
  38         "cosf64" => "llvm.cos.f64",
  39         "powf32" => "llvm.pow.f32",
  40         "powf64" => "llvm.pow.f64",
  41         "expf32" => "llvm.exp.f32",
  42         "expf64" => "llvm.exp.f64",
  43         "exp2f32" => "llvm.exp2.f32",
  44         "exp2f64" => "llvm.exp2.f64",
  45         "logf32" => "llvm.log.f32",
  46         "logf64" => "llvm.log.f64",
  47         "log10f32" => "llvm.log10.f32",
  48         "log10f64" => "llvm.log10.f64",
  49         "log2f32" => "llvm.log2.f32",
  50         "log2f64" => "llvm.log2.f64",
  51         "fmaf32" => "llvm.fma.f32",
  52         "fmaf64" => "llvm.fma.f64",
  53         "fabsf32" => "llvm.fabs.f32",
  54         "fabsf64" => "llvm.fabs.f64",
  55         "minnumf32" => "llvm.minnum.f32",
  56         "minnumf64" => "llvm.minnum.f64",
  57         "maxnumf32" => "llvm.maxnum.f32",
  58         "maxnumf64" => "llvm.maxnum.f64",
  59         "copysignf32" => "llvm.copysign.f32",
  60         "copysignf64" => "llvm.copysign.f64",
  61         "floorf32" => "llvm.floor.f32",
  62         "floorf64" => "llvm.floor.f64",
  63         "ceilf32" => "llvm.ceil.f32",
  64         "ceilf64" => "llvm.ceil.f64",
  65         "truncf32" => "llvm.trunc.f32",
  66         "truncf64" => "llvm.trunc.f64",
  67         "rintf32" => "llvm.rint.f32",
  68         "rintf64" => "llvm.rint.f64",
  69         "nearbyintf32" => "llvm.nearbyint.f32",
  70         "nearbyintf64" => "llvm.nearbyint.f64",
  71         "roundf32" => "llvm.round.f32",
  72         "roundf64" => "llvm.round.f64",
  73         "assume" => "llvm.assume",
  74         "abort" => "llvm.trap",
  75         _ => return None,
  76     };
  77     Some(cx.get_intrinsic(&llvm_name))
  78 }
  79
  80 impl IntrinsicCallMethods<'tcx> for Builder<'a, 'll, 'tcx> {
  81     fn codegen_intrinsic_call(
  82         &mut self,
  83         instance: ty::Instance<'tcx>,
  84         fn_abi: &FnAbi<'tcx, Ty<'tcx>>,
  85         args: &[OperandRef<'tcx, &'ll Value>],
  86         llresult: &'ll Value,
  87         span: Span,
  88     ) {
  89         let tcx = self.tcx;
  90         let callee_ty = instance.monomorphic_ty(tcx);
  91
  92         let (def_id, substs) = match callee_ty.kind {
  93             ty::FnDef(def_id, substs) => (def_id, substs),
  94             _ => bug!("expected fn item type, found {}", callee_ty),
  95         };
  96
  97         let sig = callee_ty.fn_sig(tcx);
  98         let sig = tcx.normalize_erasing_late_bound_regions(ty::ParamEnv::reveal_all(), &sig);
  99         let arg_tys = sig.inputs();
 100         let ret_ty = sig.output();
 101         let name = &*tcx.item_name(def_id).as_str();
 102
 103         let llret_ty = self.layout_of(ret_ty).llvm_type(self);
 104         let result = PlaceRef::new_sized(llresult, fn_abi.ret.layout);
 105
 106         let simple = get_simple_intrinsic(self, name);
 107         let llval = match name {
 108             _ if simple.is_some() => self.call(
 109                 simple.unwrap(),
 110                 &args.iter().map(|arg| arg.immediate()).collect::<Vec<_>>(),
 111                 None,
 112             ),
 113             "unreachable" => {
 114                 return;
 115             }
 116             "likely" => {
 117                 let expect = self.get_intrinsic(&("llvm.expect.i1"));
 118                 self.call(expect, &[args[0].immediate(), self.const_bool(true)], None)
 119             }
 120             "unlikely" => {
 121                 let expect = self.get_intrinsic(&("llvm.expect.i1"));
 122                 self.call(expect, &[args[0].immediate(), self.const_bool(false)], None)
 123             }
 124             "try" => {
 125                 try_intrinsic(
 126                     self,
 127                     args[0].immediate(),
 128                     args[1].immediate(),
 129                     args[2].immediate(),
 130                     llresult,
 131                 );
 132                 return;
 133             }
 134             "breakpoint" => {
 135                 let llfn = self.get_intrinsic(&("llvm.debugtrap"));
 136                 self.call(llfn, &[], None)
 137             }
 138             "va_start" => self.va_start(args[0].immediate()),
 139             "va_end" => self.va_end(args[0].immediate()),
 140             "va_copy" => {
 141                 let intrinsic = self.cx().get_intrinsic(&("llvm.va_copy"));
 142                 self.call(intrinsic, &[args[0].immediate(), args[1].immediate()], None)
 143             }
 144             "va_arg" => {
 145                 match fn_abi.ret.layout.abi {
 146                     abi::Abi::Scalar(ref scalar) => {
 147                         match scalar.value {
 148                             Primitive::Int(..) => {
 149                                 if self.cx().size_of(ret_ty).bytes() < 4 {
 150                                     // `va_arg` should not be called on a integer type
 151                                     // less than 4 bytes in length. If it is, promote
 152                                     // the integer to a `i32` and truncate the result
 153                                     // back to the smaller type.
 154                                     let promoted_result = emit_va_arg(self, args[0], tcx.types.i32);
 155                                     self.trunc(promoted_result, llret_ty)
 156                                 } else {
 157                                     emit_va_arg(self, args[0], ret_ty)
 158                                 }
 159                             }
 160                             Primitive::F64 | Primitive::Pointer => {
 161                                 emit_va_arg(self, args[0], ret_ty)
 162                             }
 163                             // `va_arg` should never be used with the return type f32.
 164                             Primitive::F32 => bug!("the va_arg intrinsic does not work with `f32`"),
 165                         }
 166                     }
 167                     _ => bug!("the va_arg intrinsic does not work with non-scalar types"),
 168                 }
 169             }
 170             "size_of_val" => {
 171                 let tp_ty = substs.type_at(0);
 172                 if let OperandValue::Pair(_, meta) = args[0].val {
 173                     let (llsize, _) = glue::size_and_align_of_dst(self, tp_ty, Some(meta));
 174                     llsize
 175                 } else {
 176                     self.const_usize(self.size_of(tp_ty).bytes())
 177                 }
 178             }
 179             "min_align_of_val" => {
 180                 let tp_ty = substs.type_at(0);
 181                 if let OperandValue::Pair(_, meta) = args[0].val {
 182                     let (_, llalign) = glue::size_and_align_of_dst(self, tp_ty, Some(meta));
 183                     llalign
 184                 } else {
 185                     self.const_usize(self.align_of(tp_ty).bytes())
 186                 }
 187             }
 188             "size_of" | "pref_align_of" | "min_align_of" | "needs_drop" | "type_id"
 189             | "type_name" => {
 190                 let ty_name = self
 191                     .tcx
 192                     .const_eval_instance(ty::ParamEnv::reveal_all(), instance, None)
 193                     .unwrap();
 194                 OperandRef::from_const(self, ty_name, ret_ty).immediate_or_packed_pair(self)
 195             }
 196             // Effectively no-op
 197             "forget" => {
 198                 return;
 199             }
 200             "offset" => {
 201                 let ptr = args[0].immediate();
 202                 let offset = args[1].immediate();
 203                 self.inbounds_gep(ptr, &[offset])
 204             }
 205             "arith_offset" => {
 206                 let ptr = args[0].immediate();
 207                 let offset = args[1].immediate();
 208                 self.gep(ptr, &[offset])
 209             }
 210
 211             "copy_nonoverlapping" => {
 212                 copy_intrinsic(
 213                     self,
 214                     false,
 215                     false,
 216                     substs.type_at(0),
 217                     args[1].immediate(),
 218                     args[0].immediate(),
 219                     args[2].immediate(),
 220                 );
 221                 return;
 222             }
 223             "copy" => {
 224                 copy_intrinsic(
 225                     self,
 226                     true,
 227                     false,
 228                     substs.type_at(0),
 229                     args[1].immediate(),
 230                     args[0].immediate(),
 231                     args[2].immediate(),
 232                 );
 233                 return;
 234             }
 235             "write_bytes" => {
 236                 memset_intrinsic(
 237                     self,
 238                     false,
 239                     substs.type_at(0),
 240                     args[0].immediate(),
 241                     args[1].immediate(),
 242                     args[2].immediate(),
 243                 );
 244                 return;
 245             }
 246
 247             "volatile_copy_nonoverlapping_memory" => {
 248                 copy_intrinsic(
 249                     self,
 250                     false,
 251                     true,
 252                     substs.type_at(0),
 253                     args[0].immediate(),
 254                     args[1].immediate(),
 255                     args[2].immediate(),
 256                 );
 257                 return;
 258             }
 259             "volatile_copy_memory" => {
 260                 copy_intrinsic(
 261                     self,
 262                     true,
 263                     true,
 264                     substs.type_at(0),
 265                     args[0].immediate(),
 266                     args[1].immediate(),
 267                     args[2].immediate(),
 268                 );
 269                 return;
 270             }
 271             "volatile_set_memory" => {
 272                 memset_intrinsic(
 273                     self,
 274                     true,
 275                     substs.type_at(0),
 276                     args[0].immediate(),
 277                     args[1].immediate(),
 278                     args[2].immediate(),
 279                 );
 280                 return;
 281             }
 282             "volatile_load" | "unaligned_volatile_load" => {
 283                 let tp_ty = substs.type_at(0);
 284                 let mut ptr = args[0].immediate();
 285                 if let PassMode::Cast(ty) = fn_abi.ret.mode {
 286                     ptr = self.pointercast(ptr, self.type_ptr_to(ty.llvm_type(self)));
 287                 }
 288                 let load = self.volatile_load(ptr);
 289                 let align = if name == "unaligned_volatile_load" {
 290                     1
 291                 } else {
 292                     self.align_of(tp_ty).bytes() as u32
 293                 };
 294                 unsafe {
 295                     llvm::LLVMSetAlignment(load, align);
 296                 }
 297                 to_immediate(self, load, self.layout_of(tp_ty))
 298             }
 299             "volatile_store" => {
 300                 let dst = args[0].deref(self.cx());
 301                 args[1].val.volatile_store(self, dst);
 302                 return;
 303             }
 304             "unaligned_volatile_store" => {
 305                 let dst = args[0].deref(self.cx());
 306                 args[1].val.unaligned_volatile_store(self, dst);
 307                 return;
 308             }
 309             "prefetch_read_data"
 310             | "prefetch_write_data"
 311             | "prefetch_read_instruction"
 312             | "prefetch_write_instruction" => {
 313                 let expect = self.get_intrinsic(&("llvm.prefetch"));
 314                 let (rw, cache_type) = match name {
 315                     "prefetch_read_data" => (0, 1),
 316                     "prefetch_write_data" => (1, 1),
 317                     "prefetch_read_instruction" => (0, 0),
 318                     "prefetch_write_instruction" => (1, 0),
 319                     _ => bug!(),
 320                 };
 321                 self.call(
 322                     expect,
 323                     &[
 324                         args[0].immediate(),
 325                         self.const_i32(rw),
 326                         args[1].immediate(),
 327                         self.const_i32(cache_type),
 328                     ],
 329                     None,
 330                 )
 331             }
 332             "ctlz" | "ctlz_nonzero" | "cttz" | "cttz_nonzero" | "ctpop" | "bswap"
 333             | "bitreverse" | "add_with_overflow" | "sub_with_overflow" | "mul_with_overflow"
 334             | "wrapping_add" | "wrapping_sub" | "wrapping_mul" | "unchecked_div"
 335             | "unchecked_rem" | "unchecked_shl" | "unchecked_shr" | "unchecked_add"
 336             | "unchecked_sub" | "unchecked_mul" | "exact_div" | "rotate_left" | "rotate_right"
 337             | "saturating_add" | "saturating_sub" => {
 338                 let ty = arg_tys[0];
 339                 match int_type_width_signed(ty, self) {
 340                     Some((width, signed)) => match name {
 341                         "ctlz" | "cttz" => {
 342                             let y = self.const_bool(false);
 343                             let llfn = self.get_intrinsic(&format!("llvm.{}.i{}", name, width));
 344                             self.call(llfn, &[args[0].immediate(), y], None)
 345                         }
 346                         "ctlz_nonzero" | "cttz_nonzero" => {
 347                             let y = self.const_bool(true);
 348                             let llvm_name = &format!("llvm.{}.i{}", &name[..4], width);
 349                             let llfn = self.get_intrinsic(llvm_name);
 350                             self.call(llfn, &[args[0].immediate(), y], None)
 351                         }
 352                         "ctpop" => self.call(
 353                             self.get_intrinsic(&format!("llvm.ctpop.i{}", width)),
 354                             &[args[0].immediate()],
 355                             None,
 356                         ),
 357                         "bswap" => {
 358                             if width == 8 {
 359                                 args[0].immediate() // byte swap a u8/i8 is just a no-op
 360                             } else {
 361                                 self.call(
 362                                     self.get_intrinsic(&format!("llvm.bswap.i{}", width)),
 363                                     &[args[0].immediate()],
 364                                     None,
 365                                 )
 366                             }
 367                         }
 368                         "bitreverse" => self.call(
 369                             self.get_intrinsic(&format!("llvm.bitreverse.i{}", width)),
 370                             &[args[0].immediate()],
 371                             None,
 372                         ),
 373                         "add_with_overflow" | "sub_with_overflow" | "mul_with_overflow" => {
 374                             let intrinsic = format!(
 375                                 "llvm.{}{}.with.overflow.i{}",
 376                                 if signed { 's' } else { 'u' },
 377                                 &name[..3],
 378                                 width
 379                             );
 380                             let llfn = self.get_intrinsic(&intrinsic);
 381
 382                             // Convert `i1` to a `bool`, and write it to the out parameter
 383                             let pair =
 384                                 self.call(llfn, &[args[0].immediate(), args[1].immediate()], None);
 385                             let val = self.extract_value(pair, 0);
 386                             let overflow = self.extract_value(pair, 1);
 387                             let overflow = self.zext(overflow, self.type_bool());
 388
 389                             let dest = result.project_field(self, 0);
 390                             self.store(val, dest.llval, dest.align);
 391                             let dest = result.project_field(self, 1);
 392                             self.store(overflow, dest.llval, dest.align);
 393
 394                             return;
 395                         }
 396                         "wrapping_add" => self.add(args[0].immediate(), args[1].immediate()),
 397                         "wrapping_sub" => self.sub(args[0].immediate(), args[1].immediate()),
 398                         "wrapping_mul" => self.mul(args[0].immediate(), args[1].immediate()),
 399                         "exact_div" => {
 400                             if signed {
 401                                 self.exactsdiv(args[0].immediate(), args[1].immediate())
 402                             } else {
 403                                 self.exactudiv(args[0].immediate(), args[1].immediate())
 404                             }
 405                         }
 406                         "unchecked_div" => {
 407                             if signed {
 408                                 self.sdiv(args[0].immediate(), args[1].immediate())
 409                             } else {
 410                                 self.udiv(args[0].immediate(), args[1].immediate())
 411                             }
 412                         }
 413                         "unchecked_rem" => {
 414                             if signed {
 415                                 self.srem(args[0].immediate(), args[1].immediate())
 416                             } else {
 417                                 self.urem(args[0].immediate(), args[1].immediate())
 418                             }
 419                         }
 420                         "unchecked_shl" => self.shl(args[0].immediate(), args[1].immediate()),
 421                         "unchecked_shr" => {
 422                             if signed {
 423                                 self.ashr(args[0].immediate(), args[1].immediate())
 424                             } else {
 425                                 self.lshr(args[0].immediate(), args[1].immediate())
 426                             }
 427                         }
 428                         "unchecked_add" => {
 429                             if signed {
 430                                 self.unchecked_sadd(args[0].immediate(), args[1].immediate())
 431                             } else {
 432                                 self.unchecked_uadd(args[0].immediate(), args[1].immediate())
 433                             }
 434                         }
 435                         "unchecked_sub" => {
 436                             if signed {
 437                                 self.unchecked_ssub(args[0].immediate(), args[1].immediate())
 438                             } else {
 439                                 self.unchecked_usub(args[0].immediate(), args[1].immediate())
 440                             }
 441                         }
 442                         "unchecked_mul" => {
 443                             if signed {
 444                                 self.unchecked_smul(args[0].immediate(), args[1].immediate())
 445                             } else {
 446                                 self.unchecked_umul(args[0].immediate(), args[1].immediate())
 447                             }
 448                         }
 449                         "rotate_left" | "rotate_right" => {
 450                             let is_left = name == "rotate_left";
 451                             let val = args[0].immediate();
 452                             let raw_shift = args[1].immediate();
 453                             // rotate = funnel shift with first two args the same
 454                             let llvm_name =
 455                                 &format!("llvm.fsh{}.i{}", if is_left { 'l' } else { 'r' }, width);
 456                             let llfn = self.get_intrinsic(llvm_name);
 457                             self.call(llfn, &[val, val, raw_shift], None)
 458                         }
 459                         "saturating_add" | "saturating_sub" => {
 460                             let is_add = name == "saturating_add";
 461                             let lhs = args[0].immediate();
 462                             let rhs = args[1].immediate();
 463                             let llvm_name = &format!(
 464                                 "llvm.{}{}.sat.i{}",
 465                                 if signed { 's' } else { 'u' },
 466                                 if is_add { "add" } else { "sub" },
 467                                 width
 468                             );
 469                             let llfn = self.get_intrinsic(llvm_name);
 470                             self.call(llfn, &[lhs, rhs], None)
 471                         }
 472                         _ => bug!(),
 473                     },
 474                     None => {
 475                         span_invalid_monomorphization_error(
 476                             tcx.sess,
 477                             span,
 478                             &format!(
 479                                 "invalid monomorphization of `{}` intrinsic: \
 480                                       expected basic integer type, found `{}`",
 481                                 name, ty
 482                             ),
 483                         );
 484                         return;
 485                     }
 486                 }
 487             }
 488             "fadd_fast" | "fsub_fast" | "fmul_fast" | "fdiv_fast" | "frem_fast" => {
 489                 match float_type_width(arg_tys[0]) {
 490                     Some(_width) => match name {
 491                         "fadd_fast" => self.fadd_fast(args[0].immediate(), args[1].immediate()),
 492                         "fsub_fast" => self.fsub_fast(args[0].immediate(), args[1].immediate()),
 493                         "fmul_fast" => self.fmul_fast(args[0].immediate(), args[1].immediate()),
 494                         "fdiv_fast" => self.fdiv_fast(args[0].immediate(), args[1].immediate()),
 495                         "frem_fast" => self.frem_fast(args[0].immediate(), args[1].immediate()),
 496                         _ => bug!(),
 497                     },
 498                     None => {
 499                         span_invalid_monomorphization_error(
 500                             tcx.sess,
 501                             span,
 502                             &format!(
 503                                 "invalid monomorphization of `{}` intrinsic: \
 504                                       expected basic float type, found `{}`",
 505                                 name, arg_tys[0]
 506                             ),
 507                         );
 508                         return;
 509                     }
 510                 }
 511             }
 512
 513             "float_to_int_unchecked" => {
 514                 if float_type_width(arg_tys[0]).is_none() {
 515                     span_invalid_monomorphization_error(
 516                         tcx.sess,
 517                         span,
 518                         &format!(
 519                             "invalid monomorphization of `float_to_int_unchecked` \
 520                                   intrinsic: expected basic float type, \
 521                                   found `{}`",
 522                             arg_tys[0]
 523                         ),
 524                     );
 525                     return;
 526                 }
 527                 match int_type_width_signed(ret_ty, self.cx) {
 528                     Some((width, signed)) => {
 529                         if signed {
 530                             self.fptosi(args[0].immediate(), self.cx.type_ix(width))
 531                         } else {
 532                             self.fptoui(args[0].immediate(), self.cx.type_ix(width))
 533                         }
 534                     }
 535                     None => {
 536                         span_invalid_monomorphization_error(
 537                             tcx.sess,
 538                             span,
 539                             &format!(
 540                                 "invalid monomorphization of `float_to_int_unchecked` \
 541                                       intrinsic:  expected basic integer type, \
 542                                       found `{}`",
 543                                 ret_ty
 544                             ),
 545                         );
 546                         return;
 547                     }
 548                 }
 549             }
 550
 551             "discriminant_value" => args[0].deref(self.cx()).codegen_get_discr(self, ret_ty),
 552
 553             name if name.starts_with("simd_") => {
 554                 match generic_simd_intrinsic(self, name, callee_ty, args, ret_ty, llret_ty, span) {
 555                     Ok(llval) => llval,
 556                     Err(()) => return,
 557                 }
 558             }
 559             // This requires that atomic intrinsics follow a specific naming pattern:
 560             // "atomic_<operation>[_<ordering>]", and no ordering means SeqCst
 561             name if name.starts_with("atomic_") => {
 562                 use rustc_codegen_ssa::common::AtomicOrdering::*;
 563                 use rustc_codegen_ssa::common::{AtomicRmwBinOp, SynchronizationScope};
 564
 565                 let split: Vec<&str> = name.split('_').collect();
 566
 567                 let is_cxchg = split[1] == "cxchg" || split[1] == "cxchgweak";
 568                 let (order, failorder) = match split.len() {
 569                     2 => (SequentiallyConsistent, SequentiallyConsistent),
 570                     3 => match split[2] {
 571                         "unordered" => (Unordered, Unordered),
 572                         "relaxed" => (Monotonic, Monotonic),
 573                         "acq" => (Acquire, Acquire),
 574                         "rel" => (Release, Monotonic),
 575                         "acqrel" => (AcquireRelease, Acquire),
 576                         "failrelaxed" if is_cxchg => (SequentiallyConsistent, Monotonic),
 577                         "failacq" if is_cxchg => (SequentiallyConsistent, Acquire),
 578                         _ => self.sess().fatal("unknown ordering in atomic intrinsic"),
 579                     },
 580                     4 => match (split[2], split[3]) {
 581                         ("acq", "failrelaxed") if is_cxchg => (Acquire, Monotonic),
 582                         ("acqrel", "failrelaxed") if is_cxchg => (AcquireRelease, Monotonic),
 583                         _ => self.sess().fatal("unknown ordering in atomic intrinsic"),
 584                     },
 585                     _ => self.sess().fatal("Atomic intrinsic not in correct format"),
 586                 };
 587
 588                 let invalid_monomorphization = |ty| {
 589                     span_invalid_monomorphization_error(
 590                         tcx.sess,
 591                         span,
 592                         &format!(
 593                             "invalid monomorphization of `{}` intrinsic: \
 594                                   expected basic integer type, found `{}`",
 595                             name, ty
 596                         ),
 597                     );
 598                 };
 599
 600                 match split[1] {
 601                     "cxchg" | "cxchgweak" => {
 602                         let ty = substs.type_at(0);
 603                         if int_type_width_signed(ty, self).is_some() {
 604                             let weak = split[1] == "cxchgweak";
 605                             let pair = self.atomic_cmpxchg(
 606                                 args[0].immediate(),
 607                                 args[1].immediate(),
 608                                 args[2].immediate(),
 609                                 order,
 610                                 failorder,
 611                                 weak,
 612                             );
 613                             let val = self.extract_value(pair, 0);
 614                             let success = self.extract_value(pair, 1);
 615                             let success = self.zext(success, self.type_bool());
 616
 617                             let dest = result.project_field(self, 0);
 618                             self.store(val, dest.llval, dest.align);
 619                             let dest = result.project_field(self, 1);
 620                             self.store(success, dest.llval, dest.align);
 621                             return;
 622                         } else {
 623                             return invalid_monomorphization(ty);
 624                         }
 625                     }
 626
 627                     "load" => {
 628                         let ty = substs.type_at(0);
 629                         if int_type_width_signed(ty, self).is_some() {
 630                             let size = self.size_of(ty);
 631                             self.atomic_load(args[0].immediate(), order, size)
 632                         } else {
 633                             return invalid_monomorphization(ty);
 634                         }
 635                     }
 636
 637                     "store" => {
 638                         let ty = substs.type_at(0);
 639                         if int_type_width_signed(ty, self).is_some() {
 640                             let size = self.size_of(ty);
 641                             self.atomic_store(
 642                                 args[1].immediate(),
 643                                 args[0].immediate(),
 644                                 order,
 645                                 size,
 646                             );
 647                             return;
 648                         } else {
 649                             return invalid_monomorphization(ty);
 650                         }
 651                     }
 652
 653                     "fence" => {
 654                         self.atomic_fence(order, SynchronizationScope::CrossThread);
 655                         return;
 656                     }
 657
 658                     "singlethreadfence" => {
 659                         self.atomic_fence(order, SynchronizationScope::SingleThread);
 660                         return;
 661                     }
 662
 663                     // These are all AtomicRMW ops
 664                     op => {
 665                         let atom_op = match op {
 666                             "xchg" => AtomicRmwBinOp::AtomicXchg,
 667                             "xadd" => AtomicRmwBinOp::AtomicAdd,
 668                             "xsub" => AtomicRmwBinOp::AtomicSub,
 669                             "and" => AtomicRmwBinOp::AtomicAnd,
 670                             "nand" => AtomicRmwBinOp::AtomicNand,
 671                             "or" => AtomicRmwBinOp::AtomicOr,
 672                             "xor" => AtomicRmwBinOp::AtomicXor,
 673                             "max" => AtomicRmwBinOp::AtomicMax,
 674                             "min" => AtomicRmwBinOp::AtomicMin,
 675                             "umax" => AtomicRmwBinOp::AtomicUMax,
 676                             "umin" => AtomicRmwBinOp::AtomicUMin,
 677                             _ => self.sess().fatal("unknown atomic operation"),
 678                         };
 679
 680                         let ty = substs.type_at(0);
 681                         if int_type_width_signed(ty, self).is_some() {
 682                             self.atomic_rmw(
 683                                 atom_op,
 684                                 args[0].immediate(),
 685                                 args[1].immediate(),
 686                                 order,
 687                             )
 688                         } else {
 689                             return invalid_monomorphization(ty);
 690                         }
 691                     }
 692                 }
 693             }
 694
 695             "nontemporal_store" => {
 696                 let dst = args[0].deref(self.cx());
 697                 args[1].val.nontemporal_store(self, dst);
 698                 return;
 699             }
 700
 701             "ptr_offset_from" => {
 702                 let ty = substs.type_at(0);
 703                 let pointee_size = self.size_of(ty);
 704
 705                 // This is the same sequence that Clang emits for pointer subtraction.
 706                 // It can be neither `nsw` nor `nuw` because the input is treated as
 707                 // unsigned but then the output is treated as signed, so neither works.
 708                 let a = args[0].immediate();
 709                 let b = args[1].immediate();
 710                 let a = self.ptrtoint(a, self.type_isize());
 711                 let b = self.ptrtoint(b, self.type_isize());
 712                 let d = self.sub(a, b);
 713                 let pointee_size = self.const_usize(pointee_size.bytes());
 714                 // this is where the signed magic happens (notice the `s` in `exactsdiv`)
 715                 self.exactsdiv(d, pointee_size)
 716             }
 717
 718             _ => bug!("unknown intrinsic '{}'", name),
 719         };
 720
 721         if !fn_abi.ret.is_ignore() {
 722             if let PassMode::Cast(ty) = fn_abi.ret.mode {
 723                 let ptr_llty = self.type_ptr_to(ty.llvm_type(self));
 724                 let ptr = self.pointercast(result.llval, ptr_llty);
 725                 self.store(llval, ptr, result.align);
 726             } else {
 727                 OperandRef::from_immediate_or_packed_pair(self, llval, result.layout)
 728                     .val
 729                     .store(self, result);
 730             }
 731         }
 732     }
 733
 734     fn abort(&mut self) {
 735         let fnname = self.get_intrinsic(&("llvm.trap"));
 736         self.call(fnname, &[], None);
 737     }
 738
 739     fn assume(&mut self, val: Self::Value) {
 740         let assume_intrinsic = self.get_intrinsic("llvm.assume");
 741         self.call(assume_intrinsic, &[val], None);
 742     }
 743
 744     fn expect(&mut self, cond: Self::Value, expected: bool) -> Self::Value {
 745         let expect = self.get_intrinsic(&"llvm.expect.i1");
 746         self.call(expect, &[cond, self.const_bool(expected)], None)
 747     }
 748
 749     fn sideeffect(&mut self) {
 750         if self.tcx.sess.opts.debugging_opts.insert_sideeffect {
 751             let fnname = self.get_intrinsic(&("llvm.sideeffect"));
 752             self.call(fnname, &[], None);
 753         }
 754     }
 755
 756     fn va_start(&mut self, va_list: &'ll Value) -> &'ll Value {
 757         let intrinsic = self.cx().get_intrinsic("llvm.va_start");
 758         self.call(intrinsic, &[va_list], None)
 759     }
 760
 761     fn va_end(&mut self, va_list: &'ll Value) -> &'ll Value {
 762         let intrinsic = self.cx().get_intrinsic("llvm.va_end");
 763         self.call(intrinsic, &[va_list], None)
 764     }
 765 }
 766
 767 fn copy_intrinsic(
 768     bx: &mut Builder<'a, 'll, 'tcx>,
 769     allow_overlap: bool,
 770     volatile: bool,
 771     ty: Ty<'tcx>,
 772     dst: &'ll Value,
 773     src: &'ll Value,
 774     count: &'ll Value,
 775 ) {
 776     let (size, align) = bx.size_and_align_of(ty);
 777     let size = bx.mul(bx.const_usize(size.bytes()), count);
 778     let flags = if volatile { MemFlags::VOLATILE } else { MemFlags::empty() };
 779     if allow_overlap {
 780         bx.memmove(dst, align, src, align, size, flags);
 781     } else {
 782         bx.memcpy(dst, align, src, align, size, flags);
 783     }
 784 }
 785
 786 fn memset_intrinsic(
 787     bx: &mut Builder<'a, 'll, 'tcx>,
 788     volatile: bool,
 789     ty: Ty<'tcx>,
 790     dst: &'ll Value,
 791     val: &'ll Value,
 792     count: &'ll Value,
 793 ) {
 794     let (size, align) = bx.size_and_align_of(ty);
 795     let size = bx.mul(bx.const_usize(size.bytes()), count);
 796     let flags = if volatile { MemFlags::VOLATILE } else { MemFlags::empty() };
 797     bx.memset(dst, val, size, align, flags);
 798 }
 799
 800 fn try_intrinsic(
 801     bx: &mut Builder<'a, 'll, 'tcx>,
 802     try_func: &'ll Value,
 803     data: &'ll Value,
 804     catch_func: &'ll Value,
 805     dest: &'ll Value,
 806 ) {
 807     if bx.sess().no_landing_pads() {
 808         bx.call(try_func, &[data], None);
 809         // Return 0 unconditionally from the intrinsic call;
 810         // we can never unwind.
 811         let ret_align = bx.tcx().data_layout.i32_align.abi;
 812         bx.store(bx.const_i32(0), dest, ret_align);
 813     } else if wants_msvc_seh(bx.sess()) {
 814         codegen_msvc_try(bx, try_func, data, catch_func, dest);
 815     } else {
 816         codegen_gnu_try(bx, try_func, data, catch_func, dest);
 817     }
 818 }
 819
 820 // MSVC's definition of the `rust_try` function.
 821 //
 822 // This implementation uses the new exception handling instructions in LLVM
 823 // which have support in LLVM for SEH on MSVC targets. Although these
 824 // instructions are meant to work for all targets, as of the time of this
 825 // writing, however, LLVM does not recommend the usage of these new instructions
 826 // as the old ones are still more optimized.
 827 fn codegen_msvc_try(
 828     bx: &mut Builder<'a, 'll, 'tcx>,
 829     try_func: &'ll Value,
 830     data: &'ll Value,
 831     catch_func: &'ll Value,
 832     dest: &'ll Value,
 833 ) {
 834     let llfn = get_rust_try_fn(bx, &mut |mut bx| {
 835         bx.set_personality_fn(bx.eh_personality());
 836         bx.sideeffect();
 837
 838         let mut normal = bx.build_sibling_block("normal");
 839         let mut catchswitch = bx.build_sibling_block("catchswitch");
 840         let mut catchpad = bx.build_sibling_block("catchpad");
 841         let mut caught = bx.build_sibling_block("caught");
 842
 843         let try_func = llvm::get_param(bx.llfn(), 0);
 844         let data = llvm::get_param(bx.llfn(), 1);
 845         let catch_func = llvm::get_param(bx.llfn(), 2);
 846
 847         // We're generating an IR snippet that looks like:
 848         //
 849         //   declare i32 @rust_try(%try_func, %data, %catch_func) {
 850         //      %slot = alloca u8*
 851         //      invoke %try_func(%data) to label %normal unwind label %catchswitch
 852         //
 853         //   normal:
 854         //      ret i32 0
 855         //
 856         //   catchswitch:
 857         //      %cs = catchswitch within none [%catchpad] unwind to caller
 858         //
 859         //   catchpad:
 860         //      %tok = catchpad within %cs [%type_descriptor, 0, %slot]
 861         //      %ptr = load %slot
 862         //      call %catch_func(%data, %ptr)
 863         //      catchret from %tok to label %caught
 864         //
 865         //   caught:
 866         //      ret i32 1
 867         //   }
 868         //
 869         // This structure follows the basic usage of throw/try/catch in LLVM.
 870         // For example, compile this C++ snippet to see what LLVM generates:
 871         //
 872         //      #include <stdint.h>
 873         //
 874         //      struct rust_panic {
 875         //          rust_panic(const rust_panic&);
 876         //          ~rust_panic();
 877         //
 878         //          uint64_t x[2];
 879         //      };
 880         //
 881         //      int __rust_try(
 882         //          void (*try_func)(void*),
 883         //          void *data,
 884         //          void (*catch_func)(void*, void*) noexcept
 885         //      ) {
 886         //          try {
 887         //              try_func(data);
 888         //              return 0;
 889         //          } catch(rust_panic& a) {
 890         //              catch_func(data, &a);
 891         //              return 1;
 892         //          }
 893         //      }
 894         //
 895         // More information can be found in libstd's seh.rs implementation.
 896         let ptr_align = bx.tcx().data_layout.pointer_align.abi;
 897         let slot = bx.alloca(bx.type_i8p(), ptr_align);
 898         bx.invoke(try_func, &[data], normal.llbb(), catchswitch.llbb(), None);
 899
 900         normal.ret(bx.const_i32(0));
 901
 902         let cs = catchswitch.catch_switch(None, None, 1);
 903         catchswitch.add_handler(cs, catchpad.llbb());
 904
 905         // We can't use the TypeDescriptor defined in libpanic_unwind because it
 906         // might be in another DLL and the SEH encoding only supports specifying
 907         // a TypeDescriptor from the current module.
 908         //
 909         // However this isn't an issue since the MSVC runtime uses string
 910         // comparison on the type name to match TypeDescriptors rather than
 911         // pointer equality.
 912         //
 913         // So instead we generate a new TypeDescriptor in each module that uses
 914         // `try` and let the linker merge duplicate definitions in the same
 915         // module.
 916         //
 917         // When modifying, make sure that the type_name string exactly matches
 918         // the one used in src/libpanic_unwind/seh.rs.
 919         let type_info_vtable = bx.declare_global("??_7type_info@@6B@", bx.type_i8p());
 920         let type_name = bx.const_bytes(b"rust_panic\0");
 921         let type_info =
 922             bx.const_struct(&[type_info_vtable, bx.const_null(bx.type_i8p()), type_name], false);
 923         let tydesc = bx.declare_global("__rust_panic_type_info", bx.val_ty(type_info));
 924         unsafe {
 925             llvm::LLVMRustSetLinkage(tydesc, llvm::Linkage::LinkOnceODRLinkage);
 926             llvm::SetUniqueComdat(bx.llmod, tydesc);
 927             llvm::LLVMSetInitializer(tydesc, type_info);
 928         }
 929
 930         // The flag value of 8 indicates that we are catching the exception by
 931         // reference instead of by value. We can't use catch by value because
 932         // that requires copying the exception object, which we don't support
 933         // since our exception object effectively contains a Box.
 934         //
 935         // Source: MicrosoftCXXABI::getAddrOfCXXCatchHandlerType in clang
 936         let flags = bx.const_i32(8);
 937         let funclet = catchpad.catch_pad(cs, &[tydesc, flags, slot]);
 938         let ptr = catchpad.load(slot, ptr_align);
 939         catchpad.call(catch_func, &[data, ptr], Some(&funclet));
 940
 941         catchpad.catch_ret(&funclet, caught.llbb());
 942
 943         caught.ret(bx.const_i32(1));
 944     });
 945
 946     // Note that no invoke is used here because by definition this function
 947     // can't panic (that's what it's catching).
 948     let ret = bx.call(llfn, &[try_func, data, catch_func], None);
 949     let i32_align = bx.tcx().data_layout.i32_align.abi;
 950     bx.store(ret, dest, i32_align);
 951 }
 952
 953 // Definition of the standard `try` function for Rust using the GNU-like model
 954 // of exceptions (e.g., the normal semantics of LLVM's `landingpad` and `invoke`
 955 // instructions).
 956 //
 957 // This codegen is a little surprising because we always call a shim
 958 // function instead of inlining the call to `invoke` manually here. This is done
 959 // because in LLVM we're only allowed to have one personality per function
 960 // definition. The call to the `try` intrinsic is being inlined into the
 961 // function calling it, and that function may already have other personality
 962 // functions in play. By calling a shim we're guaranteed that our shim will have
 963 // the right personality function.
 964 fn codegen_gnu_try(
 965     bx: &mut Builder<'a, 'll, 'tcx>,
 966     try_func: &'ll Value,
 967     data: &'ll Value,
 968     catch_func: &'ll Value,
 969     dest: &'ll Value,
 970 ) {
 971     let llfn = get_rust_try_fn(bx, &mut |mut bx| {
 972         // Codegens the shims described above:
 973         //
 974         //   bx:
 975         //      invoke %try_func(%data) normal %normal unwind %catch
 976         //
 977         //   normal:
 978         //      ret 0
 979         //
 980         //   catch:
 981         //      (%ptr, _) = landingpad
 982         //      call %catch_func(%data, %ptr)
 983         //      ret 1
 984
 985         bx.sideeffect();
 986
 987         let mut then = bx.build_sibling_block("then");
 988         let mut catch = bx.build_sibling_block("catch");
 989
 990         let try_func = llvm::get_param(bx.llfn(), 0);
 991         let data = llvm::get_param(bx.llfn(), 1);
 992         let catch_func = llvm::get_param(bx.llfn(), 2);
 993         bx.invoke(try_func, &[data], then.llbb(), catch.llbb(), None);
 994         then.ret(bx.const_i32(0));
 995
 996         // Type indicator for the exception being thrown.
 997         //
 998         // The first value in this tuple is a pointer to the exception object
 999         // being thrown.  The second value is a "selector" indicating which of
1000         // the landing pad clauses the exception's type had been matched to.
1001         // rust_try ignores the selector.
1002         let lpad_ty = bx.type_struct(&[bx.type_i8p(), bx.type_i32()], false);
1003         let vals = catch.landing_pad(lpad_ty, bx.eh_personality(), 1);
1004         let tydesc = match bx.tcx().lang_items().eh_catch_typeinfo() {
1005             Some(tydesc) => {
1006                 let tydesc = bx.get_static(tydesc);
1007                 bx.bitcast(tydesc, bx.type_i8p())
1008             }
1009             None => bx.const_null(bx.type_i8p()),
1010         };
1011         catch.add_clause(vals, tydesc);
1012         let ptr = catch.extract_value(vals, 0);
1013         catch.call(catch_func, &[data, ptr], None);
1014         catch.ret(bx.const_i32(1));
1015     });
1016
1017     // Note that no invoke is used here because by definition this function
1018     // can't panic (that's what it's catching).
1019     let ret = bx.call(llfn, &[try_func, data, catch_func], None);
1020     let i32_align = bx.tcx().data_layout.i32_align.abi;
1021     bx.store(ret, dest, i32_align);
1022 }
1023
1024 // Helper function to give a Block to a closure to codegen a shim function.
1025 // This is currently primarily used for the `try` intrinsic functions above.
1026 fn gen_fn<'ll, 'tcx>(
1027     cx: &CodegenCx<'ll, 'tcx>,
1028     name: &str,
1029     inputs: Vec<Ty<'tcx>>,
1030     output: Ty<'tcx>,
1031     codegen: &mut dyn FnMut(Builder<'_, 'll, 'tcx>),
1032 ) -> &'ll Value {
1033     let rust_fn_sig = ty::Binder::bind(cx.tcx.mk_fn_sig(
1034         inputs.into_iter(),
1035         output,
1036         false,
1037         hir::Unsafety::Unsafe,
1038         Abi::Rust,
1039     ));
1040     let fn_abi = FnAbi::of_fn_ptr(cx, rust_fn_sig, &[]);
1041     let llfn = cx.declare_fn(name, &fn_abi);
1042     cx.set_frame_pointer_elimination(llfn);
1043     cx.apply_target_cpu_attr(llfn);
1044     // FIXME(eddyb) find a nicer way to do this.
1045     unsafe { llvm::LLVMRustSetLinkage(llfn, llvm::Linkage::InternalLinkage) };
1046     let bx = Builder::new_block(cx, llfn, "entry-block");
1047     codegen(bx);
1048     llfn
1049 }
1050
1051 // Helper function used to get a handle to the `__rust_try` function used to
1052 // catch exceptions.
1053 //
1054 // This function is only generated once and is then cached.
1055 fn get_rust_try_fn<'ll, 'tcx>(
1056     cx: &CodegenCx<'ll, 'tcx>,
1057     codegen: &mut dyn FnMut(Builder<'_, 'll, 'tcx>),
1058 ) -> &'ll Value {
1059     if let Some(llfn) = cx.rust_try_fn.get() {
1060         return llfn;
1061     }
1062
1063     // Define the type up front for the signature of the rust_try function.
1064     let tcx = cx.tcx;
1065     let i8p = tcx.mk_mut_ptr(tcx.types.i8);
1066     let try_fn_ty = tcx.mk_fn_ptr(ty::Binder::bind(tcx.mk_fn_sig(
1067         iter::once(i8p),
1068         tcx.mk_unit(),
1069         false,
1070         hir::Unsafety::Unsafe,
1071         Abi::Rust,
1072     )));
1073     let catch_fn_ty = tcx.mk_fn_ptr(ty::Binder::bind(tcx.mk_fn_sig(
1074         [i8p, i8p].iter().cloned(),
1075         tcx.mk_unit(),
1076         false,
1077         hir::Unsafety::Unsafe,
1078         Abi::Rust,
1079     )));
1080     let output = tcx.types.i32;
1081     let rust_try = gen_fn(cx, "__rust_try", vec![try_fn_ty, i8p, catch_fn_ty], output, codegen);
1082     cx.rust_try_fn.set(Some(rust_try));
1083     rust_try
1084 }
1085
1086 fn generic_simd_intrinsic(
1087     bx: &mut Builder<'a, 'll, 'tcx>,
1088     name: &str,
1089     callee_ty: Ty<'tcx>,
1090     args: &[OperandRef<'tcx, &'ll Value>],
1091     ret_ty: Ty<'tcx>,
1092     llret_ty: &'ll Type,
1093     span: Span,
1094 ) -> Result<&'ll Value, ()> {
1095     // macros for error handling:
1096     macro_rules! emit_error {
1097         ($msg: tt) => {
1098             emit_error!($msg, )
1099         };
1100         ($msg: tt, $($fmt: tt)*) => {
1101             span_invalid_monomorphization_error(
1102                 bx.sess(), span,
1103                 &format!(concat!("invalid monomorphization of `{}` intrinsic: ", $msg),
1104                          name, $($fmt)*));
1105         }
1106     }
1107
1108     macro_rules! return_error {
1109         ($($fmt: tt)*) => {
1110             {
1111                 emit_error!($($fmt)*);
1112                 return Err(());
1113             }
1114         }
1115     }
1116
1117     macro_rules! require {
1118         ($cond: expr, $($fmt: tt)*) => {
1119             if !$cond {
1120                 return_error!($($fmt)*);
1121             }
1122         };
1123     }
1124
1125     macro_rules! require_simd {
1126         ($ty: expr, $position: expr) => {
1127             require!($ty.is_simd(), "expected SIMD {} type, found non-SIMD `{}`", $position, $ty)
1128         };
1129     }
1130
1131     let tcx = bx.tcx();
1132     let sig = tcx
1133         .normalize_erasing_late_bound_regions(ty::ParamEnv::reveal_all(), &callee_ty.fn_sig(tcx));
1134     let arg_tys = sig.inputs();
1135
1136     if name == "simd_select_bitmask" {
1137         let in_ty = arg_tys[0];
1138         let m_len = match in_ty.kind {
1139             // Note that this `.unwrap()` crashes for isize/usize, that's sort
1140             // of intentional as there's not currently a use case for that.
1141             ty::Int(i) => i.bit_width().unwrap(),
1142             ty::Uint(i) => i.bit_width().unwrap(),
1143             _ => return_error!("`{}` is not an integral type", in_ty),
1144         };
1145         require_simd!(arg_tys[1], "argument");
1146         let v_len = arg_tys[1].simd_size(tcx);
1147         require!(
1148             m_len == v_len,
1149             "mismatched lengths: mask length `{}` != other vector length `{}`",
1150             m_len,
1151             v_len
1152         );
1153         let i1 = bx.type_i1();
1154         let i1xn = bx.type_vector(i1, m_len);
1155         let m_i1s = bx.bitcast(args[0].immediate(), i1xn);
1156         return Ok(bx.select(m_i1s, args[1].immediate(), args[2].immediate()));
1157     }
1158
1159     // every intrinsic below takes a SIMD vector as its first argument
1160     require_simd!(arg_tys[0], "input");
1161     let in_ty = arg_tys[0];
1162     let in_elem = arg_tys[0].simd_type(tcx);
1163     let in_len = arg_tys[0].simd_size(tcx);
1164
1165     let comparison = match name {
1166         "simd_eq" => Some(hir::BinOpKind::Eq),
1167         "simd_ne" => Some(hir::BinOpKind::Ne),
1168         "simd_lt" => Some(hir::BinOpKind::Lt),
1169         "simd_le" => Some(hir::BinOpKind::Le),
1170         "simd_gt" => Some(hir::BinOpKind::Gt),
1171         "simd_ge" => Some(hir::BinOpKind::Ge),
1172         _ => None,
1173     };
1174
1175     if let Some(cmp_op) = comparison {
1176         require_simd!(ret_ty, "return");
1177
1178         let out_len = ret_ty.simd_size(tcx);
1179         require!(
1180             in_len == out_len,
1181             "expected return type with length {} (same as input type `{}`), \
1182                   found `{}` with length {}",
1183             in_len,
1184             in_ty,
1185             ret_ty,
1186             out_len
1187         );
1188         require!(
1189             bx.type_kind(bx.element_type(llret_ty)) == TypeKind::Integer,
1190             "expected return type with integer elements, found `{}` with non-integer `{}`",
1191             ret_ty,
1192             ret_ty.simd_type(tcx)
1193         );
1194
1195         return Ok(compare_simd_types(
1196             bx,
1197             args[0].immediate(),
1198             args[1].immediate(),
1199             in_elem,
1200             llret_ty,
1201             cmp_op,
1202         ));
1203     }
1204
1205     if name.starts_with("simd_shuffle") {
1206         let n: u64 = name["simd_shuffle".len()..].parse().unwrap_or_else(|_| {
1207             span_bug!(span, "bad `simd_shuffle` instruction only caught in codegen?")
1208         });
1209
1210         require_simd!(ret_ty, "return");
1211
1212         let out_len = ret_ty.simd_size(tcx);
1213         require!(
1214             out_len == n,
1215             "expected return type of length {}, found `{}` with length {}",
1216             n,
1217             ret_ty,
1218             out_len
1219         );
1220         require!(
1221             in_elem == ret_ty.simd_type(tcx),
1222             "expected return element type `{}` (element of input `{}`), \
1223                   found `{}` with element type `{}`",
1224             in_elem,
1225             in_ty,
1226             ret_ty,
1227             ret_ty.simd_type(tcx)
1228         );
1229
1230         let total_len = u128::from(in_len) * 2;
1231
1232         let vector = args[2].immediate();
1233
1234         let indices: Option<Vec<_>> = (0..n)
1235             .map(|i| {
1236                 let arg_idx = i;
1237                 let val = bx.const_get_elt(vector, i as u64);
1238                 match bx.const_to_opt_u128(val, true) {
1239                     None => {
1240                         emit_error!("shuffle index #{} is not a constant", arg_idx);
1241                         None
1242                     }
1243                     Some(idx) if idx >= total_len => {
1244                         emit_error!(
1245                             "shuffle index #{} is out of bounds (limit {})",
1246                             arg_idx,
1247                             total_len
1248                         );
1249                         None
1250                     }
1251                     Some(idx) => Some(bx.const_i32(idx as i32)),
1252                 }
1253             })
1254             .collect();
1255         let indices = match indices {
1256             Some(i) => i,
1257             None => return Ok(bx.const_null(llret_ty)),
1258         };
1259
1260         return Ok(bx.shuffle_vector(
1261             args[0].immediate(),
1262             args[1].immediate(),
1263             bx.const_vector(&indices),
1264         ));
1265     }
1266
1267     if name == "simd_insert" {
1268         require!(
1269             in_elem == arg_tys[2],
1270             "expected inserted type `{}` (element of input `{}`), found `{}`",
1271             in_elem,
1272             in_ty,
1273             arg_tys[2]
1274         );
1275         return Ok(bx.insert_element(
1276             args[0].immediate(),
1277             args[2].immediate(),
1278             args[1].immediate(),
1279         ));
1280     }
1281     if name == "simd_extract" {
1282         require!(
1283             ret_ty == in_elem,
1284             "expected return type `{}` (element of input `{}`), found `{}`",
1285             in_elem,
1286             in_ty,
1287             ret_ty
1288         );
1289         return Ok(bx.extract_element(args[0].immediate(), args[1].immediate()));
1290     }
1291
1292     if name == "simd_select" {
1293         let m_elem_ty = in_elem;
1294         let m_len = in_len;
1295         require_simd!(arg_tys[1], "argument");
1296         let v_len = arg_tys[1].simd_size(tcx);
1297         require!(
1298             m_len == v_len,
1299             "mismatched lengths: mask length `{}` != other vector length `{}`",
1300             m_len,
1301             v_len
1302         );
1303         match m_elem_ty.kind {
1304             ty::Int(_) => {}
1305             _ => return_error!("mask element type is `{}`, expected `i_`", m_elem_ty),
1306         }
1307         // truncate the mask to a vector of i1s
1308         let i1 = bx.type_i1();
1309         let i1xn = bx.type_vector(i1, m_len as u64);
1310         let m_i1s = bx.trunc(args[0].immediate(), i1xn);
1311         return Ok(bx.select(m_i1s, args[1].immediate(), args[2].immediate()));
1312     }
1313
1314     if name == "simd_bitmask" {
1315         // The `fn simd_bitmask(vector) -> unsigned integer` intrinsic takes a
1316         // vector mask and returns an unsigned integer containing the most
1317         // significant bit (MSB) of each lane.
1318
1319         // If the vector has less than 8 lanes, an u8 is returned with zeroed
1320         // trailing bits.
1321         let expected_int_bits = in_len.max(8);
1322         match ret_ty.kind {
1323             ty::Uint(i) if i.bit_width() == Some(expected_int_bits) => (),
1324             _ => return_error!("bitmask `{}`, expected `u{}`", ret_ty, expected_int_bits),
1325         }
1326
1327         // Integer vector <i{in_bitwidth} x in_len>:
1328         let (i_xn, in_elem_bitwidth) = match in_elem.kind {
1329             ty::Int(i) => {
1330                 (args[0].immediate(), i.bit_width().unwrap_or(bx.data_layout().pointer_size.bits()))
1331             }
1332             ty::Uint(i) => {
1333                 (args[0].immediate(), i.bit_width().unwrap_or(bx.data_layout().pointer_size.bits()))
1334             }
1335             _ => return_error!(
1336                 "vector argument `{}`'s element type `{}`, expected integer element type",
1337                 in_ty,
1338                 in_elem
1339             ),
1340         };
1341
1342         // Shift the MSB to the right by "in_elem_bitwidth - 1" into the first bit position.
1343         let shift_indices =
1344             vec![
1345                 bx.cx.const_int(bx.type_ix(in_elem_bitwidth), (in_elem_bitwidth - 1) as _);
1346                 in_len as _
1347             ];
1348         let i_xn_msb = bx.lshr(i_xn, bx.const_vector(shift_indices.as_slice()));
1349         // Truncate vector to an <i1 x N>
1350         let i1xn = bx.trunc(i_xn_msb, bx.type_vector(bx.type_i1(), in_len));
1351         // Bitcast <i1 x N> to iN:
1352         let i_ = bx.bitcast(i1xn, bx.type_ix(in_len));
1353         // Zero-extend iN to the bitmask type:
1354         return Ok(bx.zext(i_, bx.type_ix(expected_int_bits)));
1355     }
1356
1357     fn simd_simple_float_intrinsic(
1358         name: &str,
1359         in_elem: &::rustc_middle::ty::TyS<'_>,
1360         in_ty: &::rustc_middle::ty::TyS<'_>,
1361         in_len: u64,
1362         bx: &mut Builder<'a, 'll, 'tcx>,
1363         span: Span,
1364         args: &[OperandRef<'tcx, &'ll Value>],
1365     ) -> Result<&'ll Value, ()> {
1366         macro_rules! emit_error {
1367             ($msg: tt) => {
1368                 emit_error!($msg, )
1369             };
1370             ($msg: tt, $($fmt: tt)*) => {
1371                 span_invalid_monomorphization_error(
1372                     bx.sess(), span,
1373                     &format!(concat!("invalid monomorphization of `{}` intrinsic: ", $msg),
1374                              name, $($fmt)*));
1375             }
1376         }
1377         macro_rules! return_error {
1378             ($($fmt: tt)*) => {
1379                 {
1380                     emit_error!($($fmt)*);
1381                     return Err(());
1382                 }
1383             }
1384         }
1385         let ety = match in_elem.kind {
1386             ty::Float(f) if f.bit_width() == 32 => {
1387                 if in_len < 2 || in_len > 16 {
1388                     return_error!(
1389                         "unsupported floating-point vector `{}` with length `{}` \
1390                          out-of-range [2, 16]",
1391                         in_ty,
1392                         in_len
1393                     );
1394                 }
1395                 "f32"
1396             }
1397             ty::Float(f) if f.bit_width() == 64 => {
1398                 if in_len < 2 || in_len > 8 {
1399                     return_error!(
1400                         "unsupported floating-point vector `{}` with length `{}` \
1401                                    out-of-range [2, 8]",
1402                         in_ty,
1403                         in_len
1404                     );
1405                 }
1406                 "f64"
1407             }
1408             ty::Float(f) => {
1409                 return_error!(
1410                     "unsupported element type `{}` of floating-point vector `{}`",
1411                     f.name_str(),
1412                     in_ty
1413                 );
1414             }
1415             _ => {
1416                 return_error!("`{}` is not a floating-point type", in_ty);
1417             }
1418         };
1419
1420         let llvm_name = &format!("llvm.{0}.v{1}{2}", name, in_len, ety);
1421         let intrinsic = bx.get_intrinsic(&llvm_name);
1422         let c =
1423             bx.call(intrinsic, &args.iter().map(|arg| arg.immediate()).collect::<Vec<_>>(), None);
1424         unsafe { llvm::LLVMRustSetHasUnsafeAlgebra(c) };
1425         Ok(c)
1426     }
1427
1428     match name {
1429         "simd_fsqrt" => {
1430             return simd_simple_float_intrinsic("sqrt", in_elem, in_ty, in_len, bx, span, args);
1431         }
1432         "simd_fsin" => {
1433             return simd_simple_float_intrinsic("sin", in_elem, in_ty, in_len, bx, span, args);
1434         }
1435         "simd_fcos" => {
1436             return simd_simple_float_intrinsic("cos", in_elem, in_ty, in_len, bx, span, args);
1437         }
1438         "simd_fabs" => {
1439             return simd_simple_float_intrinsic("fabs", in_elem, in_ty, in_len, bx, span, args);
1440         }
1441         "simd_floor" => {
1442             return simd_simple_float_intrinsic("floor", in_elem, in_ty, in_len, bx, span, args);
1443         }
1444         "simd_ceil" => {
1445             return simd_simple_float_intrinsic("ceil", in_elem, in_ty, in_len, bx, span, args);
1446         }
1447         "simd_fexp" => {
1448             return simd_simple_float_intrinsic("exp", in_elem, in_ty, in_len, bx, span, args);
1449         }
1450         "simd_fexp2" => {
1451             return simd_simple_float_intrinsic("exp2", in_elem, in_ty, in_len, bx, span, args);
1452         }
1453         "simd_flog10" => {
1454             return simd_simple_float_intrinsic("log10", in_elem, in_ty, in_len, bx, span, args);
1455         }
1456         "simd_flog2" => {
1457             return simd_simple_float_intrinsic("log2", in_elem, in_ty, in_len, bx, span, args);
1458         }
1459         "simd_flog" => {
1460             return simd_simple_float_intrinsic("log", in_elem, in_ty, in_len, bx, span, args);
1461         }
1462         "simd_fpowi" => {
1463             return simd_simple_float_intrinsic("powi", in_elem, in_ty, in_len, bx, span, args);
1464         }
1465         "simd_fpow" => {
1466             return simd_simple_float_intrinsic("pow", in_elem, in_ty, in_len, bx, span, args);
1467         }
1468         "simd_fma" => {
1469             return simd_simple_float_intrinsic("fma", in_elem, in_ty, in_len, bx, span, args);
1470         }
1471         _ => { /* fallthrough */ }
1472     }
1473
1474     // FIXME: use:
1475     //  https://github.com/llvm-mirror/llvm/blob/master/include/llvm/IR/Function.h#L182
1476     //  https://github.com/llvm-mirror/llvm/blob/master/include/llvm/IR/Intrinsics.h#L81
1477     fn llvm_vector_str(elem_ty: Ty<'_>, vec_len: u64, no_pointers: usize) -> String {
1478         let p0s: String = "p0".repeat(no_pointers);
1479         match elem_ty.kind {
1480             ty::Int(v) => format!("v{}{}i{}", vec_len, p0s, v.bit_width().unwrap()),
1481             ty::Uint(v) => format!("v{}{}i{}", vec_len, p0s, v.bit_width().unwrap()),
1482             ty::Float(v) => format!("v{}{}f{}", vec_len, p0s, v.bit_width()),
1483             _ => unreachable!(),
1484         }
1485     }
1486
1487     fn llvm_vector_ty(
1488         cx: &CodegenCx<'ll, '_>,
1489         elem_ty: Ty<'_>,
1490         vec_len: u64,
1491         mut no_pointers: usize,
1492     ) -> &'ll Type {
1493         // FIXME: use cx.layout_of(ty).llvm_type() ?
1494         let mut elem_ty = match elem_ty.kind {
1495             ty::Int(v) => cx.type_int_from_ty(v),
1496             ty::Uint(v) => cx.type_uint_from_ty(v),
1497             ty::Float(v) => cx.type_float_from_ty(v),
1498             _ => unreachable!(),
1499         };
1500         while no_pointers > 0 {
1501             elem_ty = cx.type_ptr_to(elem_ty);
1502             no_pointers -= 1;
1503         }
1504         cx.type_vector(elem_ty, vec_len)
1505     }
1506
1507     if name == "simd_gather" {
1508         // simd_gather(values: <N x T>, pointers: <N x *_ T>,
1509         //             mask: <N x i{M}>) -> <N x T>
1510         // * N: number of elements in the input vectors
1511         // * T: type of the element to load
1512         // * M: any integer width is supported, will be truncated to i1
1513
1514         // All types must be simd vector types
1515         require_simd!(in_ty, "first");
1516         require_simd!(arg_tys[1], "second");
1517         require_simd!(arg_tys[2], "third");
1518         require_simd!(ret_ty, "return");
1519
1520         // Of the same length:
1521         require!(
1522             in_len == arg_tys[1].simd_size(tcx),
1523             "expected {} argument with length {} (same as input type `{}`), \
1524                   found `{}` with length {}",
1525             "second",
1526             in_len,
1527             in_ty,
1528             arg_tys[1],
1529             arg_tys[1].simd_size(tcx)
1530         );
1531         require!(
1532             in_len == arg_tys[2].simd_size(tcx),
1533             "expected {} argument with length {} (same as input type `{}`), \
1534                   found `{}` with length {}",
1535             "third",
1536             in_len,
1537             in_ty,
1538             arg_tys[2],
1539             arg_tys[2].simd_size(tcx)
1540         );
1541
1542         // The return type must match the first argument type
1543         require!(ret_ty == in_ty, "expected return type `{}`, found `{}`", in_ty, ret_ty);
1544
1545         // This counts how many pointers
1546         fn ptr_count(t: Ty<'_>) -> usize {
1547             match t.kind {
1548                 ty::RawPtr(p) => 1 + ptr_count(p.ty),
1549                 _ => 0,
1550             }
1551         }
1552
1553         // Non-ptr type
1554         fn non_ptr(t: Ty<'_>) -> Ty<'_> {
1555             match t.kind {
1556                 ty::RawPtr(p) => non_ptr(p.ty),
1557                 _ => t,
1558             }
1559         }
1560
1561         // The second argument must be a simd vector with an element type that's a pointer
1562         // to the element type of the first argument
1563         let (pointer_count, underlying_ty) = match arg_tys[1].simd_type(tcx).kind {
1564             ty::RawPtr(p) if p.ty == in_elem => {
1565                 (ptr_count(arg_tys[1].simd_type(tcx)), non_ptr(arg_tys[1].simd_type(tcx)))
1566             }
1567             _ => {
1568                 require!(
1569                     false,
1570                     "expected element type `{}` of second argument `{}` \
1571                                  to be a pointer to the element type `{}` of the first \
1572                                  argument `{}`, found `{}` != `*_ {}`",
1573                     arg_tys[1].simd_type(tcx),
1574                     arg_tys[1],
1575                     in_elem,
1576                     in_ty,
1577                     arg_tys[1].simd_type(tcx),
1578                     in_elem
1579                 );
1580                 unreachable!();
1581             }
1582         };
1583         assert!(pointer_count > 0);
1584         assert_eq!(pointer_count - 1, ptr_count(arg_tys[0].simd_type(tcx)));
1585         assert_eq!(underlying_ty, non_ptr(arg_tys[0].simd_type(tcx)));
1586
1587         // The element type of the third argument must be a signed integer type of any width:
1588         match arg_tys[2].simd_type(tcx).kind {
1589             ty::Int(_) => (),
1590             _ => {
1591                 require!(
1592                     false,
1593                     "expected element type `{}` of third argument `{}` \
1594                                  to be a signed integer type",
1595                     arg_tys[2].simd_type(tcx),
1596                     arg_tys[2]
1597                 );
1598             }
1599         }
1600
1601         // Alignment of T, must be a constant integer value:
1602         let alignment_ty = bx.type_i32();
1603         let alignment = bx.const_i32(bx.align_of(in_elem).bytes() as i32);
1604
1605         // Truncate the mask vector to a vector of i1s:
1606         let (mask, mask_ty) = {
1607             let i1 = bx.type_i1();
1608             let i1xn = bx.type_vector(i1, in_len);
1609             (bx.trunc(args[2].immediate(), i1xn), i1xn)
1610         };
1611
1612         // Type of the vector of pointers:
1613         let llvm_pointer_vec_ty = llvm_vector_ty(bx, underlying_ty, in_len, pointer_count);
1614         let llvm_pointer_vec_str = llvm_vector_str(underlying_ty, in_len, pointer_count);
1615
1616         // Type of the vector of elements:
1617         let llvm_elem_vec_ty = llvm_vector_ty(bx, underlying_ty, in_len, pointer_count - 1);
1618         let llvm_elem_vec_str = llvm_vector_str(underlying_ty, in_len, pointer_count - 1);
1619
1620         let llvm_intrinsic =
1621             format!("llvm.masked.gather.{}.{}", llvm_elem_vec_str, llvm_pointer_vec_str);
1622         let f = bx.declare_cfn(
1623             &llvm_intrinsic,
1624             bx.type_func(
1625                 &[llvm_pointer_vec_ty, alignment_ty, mask_ty, llvm_elem_vec_ty],
1626                 llvm_elem_vec_ty,
1627             ),
1628         );
1629         llvm::SetUnnamedAddress(f, llvm::UnnamedAddr::No);
1630         let v = bx.call(f, &[args[1].immediate(), alignment, mask, args[0].immediate()], None);
1631         return Ok(v);
1632     }
1633
1634     if name == "simd_scatter" {
1635         // simd_scatter(values: <N x T>, pointers: <N x *mut T>,
1636         //             mask: <N x i{M}>) -> ()
1637         // * N: number of elements in the input vectors
1638         // * T: type of the element to load
1639         // * M: any integer width is supported, will be truncated to i1
1640
1641         // All types must be simd vector types
1642         require_simd!(in_ty, "first");
1643         require_simd!(arg_tys[1], "second");
1644         require_simd!(arg_tys[2], "third");
1645
1646         // Of the same length:
1647         require!(
1648             in_len == arg_tys[1].simd_size(tcx),
1649             "expected {} argument with length {} (same as input type `{}`), \
1650                   found `{}` with length {}",
1651             "second",
1652             in_len,
1653             in_ty,
1654             arg_tys[1],
1655             arg_tys[1].simd_size(tcx)
1656         );
1657         require!(
1658             in_len == arg_tys[2].simd_size(tcx),
1659             "expected {} argument with length {} (same as input type `{}`), \
1660                   found `{}` with length {}",
1661             "third",
1662             in_len,
1663             in_ty,
1664             arg_tys[2],
1665             arg_tys[2].simd_size(tcx)
1666         );
1667
1668         // This counts how many pointers
1669         fn ptr_count(t: Ty<'_>) -> usize {
1670             match t.kind {
1671                 ty::RawPtr(p) => 1 + ptr_count(p.ty),
1672                 _ => 0,
1673             }
1674         }
1675
1676         // Non-ptr type
1677         fn non_ptr(t: Ty<'_>) -> Ty<'_> {
1678             match t.kind {
1679                 ty::RawPtr(p) => non_ptr(p.ty),
1680                 _ => t,
1681             }
1682         }
1683
1684         // The second argument must be a simd vector with an element type that's a pointer
1685         // to the element type of the first argument
1686         let (pointer_count, underlying_ty) = match arg_tys[1].simd_type(tcx).kind {
1687             ty::RawPtr(p) if p.ty == in_elem && p.mutbl == hir::Mutability::Mut => {
1688                 (ptr_count(arg_tys[1].simd_type(tcx)), non_ptr(arg_tys[1].simd_type(tcx)))
1689             }
1690             _ => {
1691                 require!(
1692                     false,
1693                     "expected element type `{}` of second argument `{}` \
1694                                  to be a pointer to the element type `{}` of the first \
1695                                  argument `{}`, found `{}` != `*mut {}`",
1696                     arg_tys[1].simd_type(tcx),
1697                     arg_tys[1],
1698                     in_elem,
1699                     in_ty,
1700                     arg_tys[1].simd_type(tcx),
1701                     in_elem
1702                 );
1703                 unreachable!();
1704             }
1705         };
1706         assert!(pointer_count > 0);
1707         assert_eq!(pointer_count - 1, ptr_count(arg_tys[0].simd_type(tcx)));
1708         assert_eq!(underlying_ty, non_ptr(arg_tys[0].simd_type(tcx)));
1709
1710         // The element type of the third argument must be a signed integer type of any width:
1711         match arg_tys[2].simd_type(tcx).kind {
1712             ty::Int(_) => (),
1713             _ => {
1714                 require!(
1715                     false,
1716                     "expected element type `{}` of third argument `{}` \
1717                                  to be a signed integer type",
1718                     arg_tys[2].simd_type(tcx),
1719                     arg_tys[2]
1720                 );
1721             }
1722         }
1723
1724         // Alignment of T, must be a constant integer value:
1725         let alignment_ty = bx.type_i32();
1726         let alignment = bx.const_i32(bx.align_of(in_elem).bytes() as i32);
1727
1728         // Truncate the mask vector to a vector of i1s:
1729         let (mask, mask_ty) = {
1730             let i1 = bx.type_i1();
1731             let i1xn = bx.type_vector(i1, in_len);
1732             (bx.trunc(args[2].immediate(), i1xn), i1xn)
1733         };
1734
1735         let ret_t = bx.type_void();
1736
1737         // Type of the vector of pointers:
1738         let llvm_pointer_vec_ty = llvm_vector_ty(bx, underlying_ty, in_len, pointer_count);
1739         let llvm_pointer_vec_str = llvm_vector_str(underlying_ty, in_len, pointer_count);
1740
1741         // Type of the vector of elements:
1742         let llvm_elem_vec_ty = llvm_vector_ty(bx, underlying_ty, in_len, pointer_count - 1);
1743         let llvm_elem_vec_str = llvm_vector_str(underlying_ty, in_len, pointer_count - 1);
1744
1745         let llvm_intrinsic =
1746             format!("llvm.masked.scatter.{}.{}", llvm_elem_vec_str, llvm_pointer_vec_str);
1747         let f = bx.declare_cfn(
1748             &llvm_intrinsic,
1749             bx.type_func(&[llvm_elem_vec_ty, llvm_pointer_vec_ty, alignment_ty, mask_ty], ret_t),
1750         );
1751         llvm::SetUnnamedAddress(f, llvm::UnnamedAddr::No);
1752         let v = bx.call(f, &[args[0].immediate(), args[1].immediate(), alignment, mask], None);
1753         return Ok(v);
1754     }
1755
1756     macro_rules! arith_red {
1757         ($name:tt : $integer_reduce:ident, $float_reduce:ident, $ordered:expr) => {
1758             if name == $name {
1759                 require!(
1760                     ret_ty == in_elem,
1761                     "expected return type `{}` (element of input `{}`), found `{}`",
1762                     in_elem,
1763                     in_ty,
1764                     ret_ty
1765                 );
1766                 return match in_elem.kind {
1767                     ty::Int(_) | ty::Uint(_) => {
1768                         let r = bx.$integer_reduce(args[0].immediate());
1769                         if $ordered {
1770                             // if overflow occurs, the result is the
1771                             // mathematical result modulo 2^n:
1772                             if name.contains("mul") {
1773                                 Ok(bx.mul(args[1].immediate(), r))
1774                             } else {
1775                                 Ok(bx.add(args[1].immediate(), r))
1776                             }
1777                         } else {
1778                             Ok(bx.$integer_reduce(args[0].immediate()))
1779                         }
1780                     }
1781                     ty::Float(f) => {
1782                         let acc = if $ordered {
1783                             // ordered arithmetic reductions take an accumulator
1784                             args[1].immediate()
1785                         } else {
1786                             // unordered arithmetic reductions use the identity accumulator
1787                             let identity_acc = if $name.contains("mul") { 1.0 } else { 0.0 };
1788                             match f.bit_width() {
1789                                 32 => bx.const_real(bx.type_f32(), identity_acc),
1790                                 64 => bx.const_real(bx.type_f64(), identity_acc),
1791                                 v => return_error!(
1792                                     r#"
1793 unsupported {} from `{}` with element `{}` of size `{}` to `{}`"#,
1794                                     $name,
1795                                     in_ty,
1796                                     in_elem,
1797                                     v,
1798                                     ret_ty
1799                                 ),
1800                             }
1801                         };
1802                         Ok(bx.$float_reduce(acc, args[0].immediate()))
1803                     }
1804                     _ => return_error!(
1805                         "unsupported {} from `{}` with element `{}` to `{}`",
1806                         $name,
1807                         in_ty,
1808                         in_elem,
1809                         ret_ty
1810                     ),
1811                 };
1812             }
1813         };
1814     }
1815
1816     arith_red!("simd_reduce_add_ordered": vector_reduce_add, vector_reduce_fadd, true);
1817     arith_red!("simd_reduce_mul_ordered": vector_reduce_mul, vector_reduce_fmul, true);
1818     arith_red!("simd_reduce_add_unordered": vector_reduce_add, vector_reduce_fadd_fast, false);
1819     arith_red!("simd_reduce_mul_unordered": vector_reduce_mul, vector_reduce_fmul_fast, false);
1820
1821     macro_rules! minmax_red {
1822         ($name:tt: $int_red:ident, $float_red:ident) => {
1823             if name == $name {
1824                 require!(
1825                     ret_ty == in_elem,
1826                     "expected return type `{}` (element of input `{}`), found `{}`",
1827                     in_elem,
1828                     in_ty,
1829                     ret_ty
1830                 );
1831                 return match in_elem.kind {
1832                     ty::Int(_i) => Ok(bx.$int_red(args[0].immediate(), true)),
1833                     ty::Uint(_u) => Ok(bx.$int_red(args[0].immediate(), false)),
1834                     ty::Float(_f) => Ok(bx.$float_red(args[0].immediate())),
1835                     _ => return_error!(
1836                         "unsupported {} from `{}` with element `{}` to `{}`",
1837                         $name,
1838                         in_ty,
1839                         in_elem,
1840                         ret_ty
1841                     ),
1842                 };
1843             }
1844         };
1845     }
1846
1847     minmax_red!("simd_reduce_min": vector_reduce_min, vector_reduce_fmin);
1848     minmax_red!("simd_reduce_max": vector_reduce_max, vector_reduce_fmax);
1849
1850     minmax_red!("simd_reduce_min_nanless": vector_reduce_min, vector_reduce_fmin_fast);
1851     minmax_red!("simd_reduce_max_nanless": vector_reduce_max, vector_reduce_fmax_fast);
1852
1853     macro_rules! bitwise_red {
1854         ($name:tt : $red:ident, $boolean:expr) => {
1855             if name == $name {
1856                 let input = if !$boolean {
1857                     require!(
1858                         ret_ty == in_elem,
1859                         "expected return type `{}` (element of input `{}`), found `{}`",
1860                         in_elem,
1861                         in_ty,
1862                         ret_ty
1863                     );
1864                     args[0].immediate()
1865                 } else {
1866                     match in_elem.kind {
1867                         ty::Int(_) | ty::Uint(_) => {}
1868                         _ => return_error!(
1869                             "unsupported {} from `{}` with element `{}` to `{}`",
1870                             $name,
1871                             in_ty,
1872                             in_elem,
1873                             ret_ty
1874                         ),
1875                     }
1876
1877                     // boolean reductions operate on vectors of i1s:
1878                     let i1 = bx.type_i1();
1879                     let i1xn = bx.type_vector(i1, in_len as u64);
1880                     bx.trunc(args[0].immediate(), i1xn)
1881                 };
1882                 return match in_elem.kind {
1883                     ty::Int(_) | ty::Uint(_) => {
1884                         let r = bx.$red(input);
1885                         Ok(if !$boolean { r } else { bx.zext(r, bx.type_bool()) })
1886                     }
1887                     _ => return_error!(
1888                         "unsupported {} from `{}` with element `{}` to `{}`",
1889                         $name,
1890                         in_ty,
1891                         in_elem,
1892                         ret_ty
1893                     ),
1894                 };
1895             }
1896         };
1897     }
1898
1899     bitwise_red!("simd_reduce_and": vector_reduce_and, false);
1900     bitwise_red!("simd_reduce_or": vector_reduce_or, false);
1901     bitwise_red!("simd_reduce_xor": vector_reduce_xor, false);
1902     bitwise_red!("simd_reduce_all": vector_reduce_and, true);
1903     bitwise_red!("simd_reduce_any": vector_reduce_or, true);
1904
1905     if name == "simd_cast" {
1906         require_simd!(ret_ty, "return");
1907         let out_len = ret_ty.simd_size(tcx);
1908         require!(
1909             in_len == out_len,
1910             "expected return type with length {} (same as input type `{}`), \
1911                   found `{}` with length {}",
1912             in_len,
1913             in_ty,
1914             ret_ty,
1915             out_len
1916         );
1917         // casting cares about nominal type, not just structural type
1918         let out_elem = ret_ty.simd_type(tcx);
1919
1920         if in_elem == out_elem {
1921             return Ok(args[0].immediate());
1922         }
1923
1924         enum Style {
1925             Float,
1926             Int(/* is signed? */ bool),
1927             Unsupported,
1928         }
1929
1930         let (in_style, in_width) = match in_elem.kind {
1931             // vectors of pointer-sized integers should've been
1932             // disallowed before here, so this unwrap is safe.
1933             ty::Int(i) => (Style::Int(true), i.bit_width().unwrap()),
1934             ty::Uint(u) => (Style::Int(false), u.bit_width().unwrap()),
1935             ty::Float(f) => (Style::Float, f.bit_width()),
1936             _ => (Style::Unsupported, 0),
1937         };
1938         let (out_style, out_width) = match out_elem.kind {
1939             ty::Int(i) => (Style::Int(true), i.bit_width().unwrap()),
1940             ty::Uint(u) => (Style::Int(false), u.bit_width().unwrap()),
1941             ty::Float(f) => (Style::Float, f.bit_width()),
1942             _ => (Style::Unsupported, 0),
1943         };
1944
1945         match (in_style, out_style) {
1946             (Style::Int(in_is_signed), Style::Int(_)) => {
1947                 return Ok(match in_width.cmp(&out_width) {
1948                     Ordering::Greater => bx.trunc(args[0].immediate(), llret_ty),
1949                     Ordering::Equal => args[0].immediate(),
1950                     Ordering::Less => {
1951                         if in_is_signed {
1952                             bx.sext(args[0].immediate(), llret_ty)
1953                         } else {
1954                             bx.zext(args[0].immediate(), llret_ty)
1955                         }
1956                     }
1957                 });
1958             }
1959             (Style::Int(in_is_signed), Style::Float) => {
1960                 return Ok(if in_is_signed {
1961                     bx.sitofp(args[0].immediate(), llret_ty)
1962                 } else {
1963                     bx.uitofp(args[0].immediate(), llret_ty)
1964                 });
1965             }
1966             (Style::Float, Style::Int(out_is_signed)) => {
1967                 return Ok(if out_is_signed {
1968                     bx.fptosi(args[0].immediate(), llret_ty)
1969                 } else {
1970                     bx.fptoui(args[0].immediate(), llret_ty)
1971                 });
1972             }
1973             (Style::Float, Style::Float) => {
1974                 return Ok(match in_width.cmp(&out_width) {
1975                     Ordering::Greater => bx.fptrunc(args[0].immediate(), llret_ty),
1976                     Ordering::Equal => args[0].immediate(),
1977                     Ordering::Less => bx.fpext(args[0].immediate(), llret_ty),
1978                 });
1979             }
1980             _ => { /* Unsupported. Fallthrough. */ }
1981         }
1982         require!(
1983             false,
1984             "unsupported cast from `{}` with element `{}` to `{}` with element `{}`",
1985             in_ty,
1986             in_elem,
1987             ret_ty,
1988             out_elem
1989         );
1990     }
1991     macro_rules! arith {
1992         ($($name: ident: $($($p: ident),* => $call: ident),*;)*) => {
1993             $(if name == stringify!($name) {
1994                 match in_elem.kind {
1995                     $($(ty::$p(_))|* => {
1996                         return Ok(bx.$call(args[0].immediate(), args[1].immediate()))
1997                     })*
1998                     _ => {},
1999                 }
2000                 require!(false,
2001                          "unsupported operation on `{}` with element `{}`",
2002                          in_ty,
2003                          in_elem)
2004             })*
2005         }
2006     }
2007     arith! {
2008         simd_add: Uint, Int => add, Float => fadd;
2009         simd_sub: Uint, Int => sub, Float => fsub;
2010         simd_mul: Uint, Int => mul, Float => fmul;
2011         simd_div: Uint => udiv, Int => sdiv, Float => fdiv;
2012         simd_rem: Uint => urem, Int => srem, Float => frem;
2013         simd_shl: Uint, Int => shl;
2014         simd_shr: Uint => lshr, Int => ashr;
2015         simd_and: Uint, Int => and;
2016         simd_or: Uint, Int => or;
2017         simd_xor: Uint, Int => xor;
2018         simd_fmax: Float => maxnum;
2019         simd_fmin: Float => minnum;
2020
2021     }
2022
2023     if name == "simd_saturating_add" || name == "simd_saturating_sub" {
2024         let lhs = args[0].immediate();
2025         let rhs = args[1].immediate();
2026         let is_add = name == "simd_saturating_add";
2027         let ptr_bits = bx.tcx().data_layout.pointer_size.bits() as _;
2028         let (signed, elem_width, elem_ty) = match in_elem.kind {
2029             ty::Int(i) => (true, i.bit_width().unwrap_or(ptr_bits), bx.cx.type_int_from_ty(i)),
2030             ty::Uint(i) => (false, i.bit_width().unwrap_or(ptr_bits), bx.cx.type_uint_from_ty(i)),
2031             _ => {
2032                 return_error!(
2033                     "expected element type `{}` of vector type `{}` \
2034                      to be a signed or unsigned integer type",
2035                     arg_tys[0].simd_type(tcx),
2036                     arg_tys[0]
2037                 );
2038             }
2039         };
2040         let llvm_intrinsic = &format!(
2041             "llvm.{}{}.sat.v{}i{}",
2042             if signed { 's' } else { 'u' },
2043             if is_add { "add" } else { "sub" },
2044             in_len,
2045             elem_width
2046         );
2047         let vec_ty = bx.cx.type_vector(elem_ty, in_len as u64);
2048
2049         let f = bx.declare_cfn(&llvm_intrinsic, bx.type_func(&[vec_ty, vec_ty], vec_ty));
2050         llvm::SetUnnamedAddress(f, llvm::UnnamedAddr::No);
2051         let v = bx.call(f, &[lhs, rhs], None);
2052         return Ok(v);
2053     }
2054
2055     span_bug!(span, "unknown SIMD intrinsic");
2056 }
2057
2058 // Returns the width of an int Ty, and if it's signed or not
2059 // Returns None if the type is not an integer
2060 // FIXME: there’s multiple of this functions, investigate using some of the already existing
2061 // stuffs.
2062 fn int_type_width_signed(ty: Ty<'_>, cx: &CodegenCx<'_, '_>) -> Option<(u64, bool)> {
2063     match ty.kind {
2064         ty::Int(t) => Some((
2065             match t {
2066                 ast::IntTy::Isize => u64::from(cx.tcx.sess.target.ptr_width),
2067                 ast::IntTy::I8 => 8,
2068                 ast::IntTy::I16 => 16,
2069                 ast::IntTy::I32 => 32,
2070                 ast::IntTy::I64 => 64,
2071                 ast::IntTy::I128 => 128,
2072             },
2073             true,
2074         )),
2075         ty::Uint(t) => Some((
2076             match t {
2077                 ast::UintTy::Usize => u64::from(cx.tcx.sess.target.ptr_width),
2078                 ast::UintTy::U8 => 8,
2079                 ast::UintTy::U16 => 16,
2080                 ast::UintTy::U32 => 32,
2081                 ast::UintTy::U64 => 64,
2082                 ast::UintTy::U128 => 128,
2083             },
2084             false,
2085         )),
2086         _ => None,
2087     }
2088 }
2089
2090 // Returns the width of a float Ty
2091 // Returns None if the type is not a float
2092 fn float_type_width(ty: Ty<'_>) -> Option<u64> {
2093     match ty.kind {
2094         ty::Float(t) => Some(t.bit_width()),
2095         _ => None,
2096     }
2097 }