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::llvm_util;
   6 use crate::type_::Type;
   7 use crate::type_of::LayoutLlvmExt;
   8 use crate::va_arg::emit_va_arg;
   9 use crate::value::Value;
  10 use rustc::ty::layout::{self, FnAbiExt, HasTyCtxt, LayoutOf, Primitive};
  11 use rustc::ty::{self, Ty};
  12 use rustc::{bug, span_bug};
  13 use rustc_codegen_ssa::base::{compare_simd_types, to_immediate, wants_msvc_seh};
  14 use rustc_codegen_ssa::common::{IntPredicate, TypeKind};
  15 use rustc_codegen_ssa::glue;
  16 use rustc_codegen_ssa::mir::operand::{OperandRef, OperandValue};
  17 use rustc_codegen_ssa::mir::place::PlaceRef;
  18 use rustc_codegen_ssa::MemFlags;
  19 use rustc_hir as hir;
  20 use rustc_target::abi::HasDataLayout;
  21 use syntax::ast;
  22
  23 use rustc_codegen_ssa::common::span_invalid_monomorphization_error;
  24 use rustc_codegen_ssa::traits::*;
  25
  26 use rustc_span::Span;
  27
  28 use std::cmp::Ordering;
  29 use std::{i128, iter, u128};
  30
  31 fn get_simple_intrinsic(cx: &CodegenCx<'ll, '_>, name: &str) -> Option<&'ll Value> {
  32     let llvm_name = match name {
  33         "sqrtf32" => "llvm.sqrt.f32",
  34         "sqrtf64" => "llvm.sqrt.f64",
  35         "powif32" => "llvm.powi.f32",
  36         "powif64" => "llvm.powi.f64",
  37         "sinf32" => "llvm.sin.f32",
  38         "sinf64" => "llvm.sin.f64",
  39         "cosf32" => "llvm.cos.f32",
  40         "cosf64" => "llvm.cos.f64",
  41         "powf32" => "llvm.pow.f32",
  42         "powf64" => "llvm.pow.f64",
  43         "expf32" => "llvm.exp.f32",
  44         "expf64" => "llvm.exp.f64",
  45         "exp2f32" => "llvm.exp2.f32",
  46         "exp2f64" => "llvm.exp2.f64",
  47         "logf32" => "llvm.log.f32",
  48         "logf64" => "llvm.log.f64",
  49         "log10f32" => "llvm.log10.f32",
  50         "log10f64" => "llvm.log10.f64",
  51         "log2f32" => "llvm.log2.f32",
  52         "log2f64" => "llvm.log2.f64",
  53         "fmaf32" => "llvm.fma.f32",
  54         "fmaf64" => "llvm.fma.f64",
  55         "fabsf32" => "llvm.fabs.f32",
  56         "fabsf64" => "llvm.fabs.f64",
  57         "minnumf32" => "llvm.minnum.f32",
  58         "minnumf64" => "llvm.minnum.f64",
  59         "maxnumf32" => "llvm.maxnum.f32",
  60         "maxnumf64" => "llvm.maxnum.f64",
  61         "copysignf32" => "llvm.copysign.f32",
  62         "copysignf64" => "llvm.copysign.f64",
  63         "floorf32" => "llvm.floor.f32",
  64         "floorf64" => "llvm.floor.f64",
  65         "ceilf32" => "llvm.ceil.f32",
  66         "ceilf64" => "llvm.ceil.f64",
  67         "truncf32" => "llvm.trunc.f32",
  68         "truncf64" => "llvm.trunc.f64",
  69         "rintf32" => "llvm.rint.f32",
  70         "rintf64" => "llvm.rint.f64",
  71         "nearbyintf32" => "llvm.nearbyint.f32",
  72         "nearbyintf64" => "llvm.nearbyint.f64",
  73         "roundf32" => "llvm.round.f32",
  74         "roundf64" => "llvm.round.f64",
  75         "assume" => "llvm.assume",
  76         "abort" => "llvm.trap",
  77         _ => return None,
  78     };
  79     Some(cx.get_intrinsic(&llvm_name))
  80 }
  81
  82 impl IntrinsicCallMethods<'tcx> for Builder<'a, 'll, 'tcx> {
  83     fn codegen_intrinsic_call(
  84         &mut self,
  85         instance: ty::Instance<'tcx>,
  86         fn_abi: &FnAbi<'tcx, Ty<'tcx>>,
  87         args: &[OperandRef<'tcx, &'ll Value>],
  88         llresult: &'ll Value,
  89         span: Span,
  90     ) {
  91         let tcx = self.tcx;
  92         let callee_ty = instance.monomorphic_ty(tcx);
  93
  94         let (def_id, substs) = match callee_ty.kind {
  95             ty::FnDef(def_id, substs) => (def_id, substs),
  96             _ => bug!("expected fn item type, found {}", callee_ty),
  97         };
  98
  99         let sig = callee_ty.fn_sig(tcx);
 100         let sig = tcx.normalize_erasing_late_bound_regions(ty::ParamEnv::reveal_all(), &sig);
 101         let arg_tys = sig.inputs();
 102         let ret_ty = sig.output();
 103         let name = &*tcx.item_name(def_id).as_str();
 104
 105         let llret_ty = self.layout_of(ret_ty).llvm_type(self);
 106         let result = PlaceRef::new_sized(llresult, fn_abi.ret.layout);
 107
 108         let simple = get_simple_intrinsic(self, name);
 109         let llval = match name {
 110             _ if simple.is_some() => self.call(
 111                 simple.unwrap(),
 112                 &args.iter().map(|arg| arg.immediate()).collect::<Vec<_>>(),
 113                 None,
 114             ),
 115             "unreachable" => {
 116                 return;
 117             }
 118             "likely" => {
 119                 let expect = self.get_intrinsic(&("llvm.expect.i1"));
 120                 self.call(expect, &[args[0].immediate(), self.const_bool(true)], None)
 121             }
 122             "unlikely" => {
 123                 let expect = self.get_intrinsic(&("llvm.expect.i1"));
 124                 self.call(expect, &[args[0].immediate(), self.const_bool(false)], None)
 125             }
 126             "try" => {
 127                 try_intrinsic(
 128                     self,
 129                     args[0].immediate(),
 130                     args[1].immediate(),
 131                     args[2].immediate(),
 132                     llresult,
 133                 );
 134                 return;
 135             }
 136             "breakpoint" => {
 137                 let llfn = self.get_intrinsic(&("llvm.debugtrap"));
 138                 self.call(llfn, &[], None)
 139             }
 140             "va_start" => self.va_start(args[0].immediate()),
 141             "va_end" => self.va_end(args[0].immediate()),
 142             "va_copy" => {
 143                 let intrinsic = self.cx().get_intrinsic(&("llvm.va_copy"));
 144                 self.call(intrinsic, &[args[0].immediate(), args[1].immediate()], None)
 145             }
 146             "va_arg" => {
 147                 match fn_abi.ret.layout.abi {
 148                     layout::Abi::Scalar(ref scalar) => {
 149                         match scalar.value {
 150                             Primitive::Int(..) => {
 151                                 if self.cx().size_of(ret_ty).bytes() < 4 {
 152                                     // `va_arg` should not be called on a integer type
 153                                     // less than 4 bytes in length. If it is, promote
 154                                     // the integer to a `i32` and truncate the result
 155                                     // back to the smaller type.
 156                                     let promoted_result = emit_va_arg(self, args[0], tcx.types.i32);
 157                                     self.trunc(promoted_result, llret_ty)
 158                                 } else {
 159                                     emit_va_arg(self, args[0], ret_ty)
 160                                 }
 161                             }
 162                             Primitive::F64 | Primitive::Pointer => {
 163                                 emit_va_arg(self, args[0], ret_ty)
 164                             }
 165                             // `va_arg` should never be used with the return type f32.
 166                             Primitive::F32 => bug!("the va_arg intrinsic does not work with `f32`"),
 167                         }
 168                     }
 169                     _ => bug!("the va_arg intrinsic does not work with non-scalar types"),
 170                 }
 171             }
 172             "size_of_val" => {
 173                 let tp_ty = substs.type_at(0);
 174                 if let OperandValue::Pair(_, meta) = args[0].val {
 175                     let (llsize, _) = glue::size_and_align_of_dst(self, tp_ty, Some(meta));
 176                     llsize
 177                 } else {
 178                     self.const_usize(self.size_of(tp_ty).bytes())
 179                 }
 180             }
 181             "min_align_of_val" => {
 182                 let tp_ty = substs.type_at(0);
 183                 if let OperandValue::Pair(_, meta) = args[0].val {
 184                     let (_, llalign) = glue::size_and_align_of_dst(self, tp_ty, Some(meta));
 185                     llalign
 186                 } else {
 187                     self.const_usize(self.align_of(tp_ty).bytes())
 188                 }
 189             }
 190             "size_of" | "pref_align_of" | "min_align_of" | "needs_drop" | "type_id"
 191             | "type_name" => {
 192                 let ty_name = self
 193                     .tcx
 194                     .const_eval_instance(ty::ParamEnv::reveal_all(), instance, None)
 195                     .unwrap();
 196                 OperandRef::from_const(self, ty_name).immediate_or_packed_pair(self)
 197             }
 198             "init" => {
 199                 let ty = substs.type_at(0);
 200                 if !self.layout_of(ty).is_zst() {
 201                     // Just zero out the stack slot.
 202                     // If we store a zero constant, LLVM will drown in vreg allocation for large
 203                     // data structures, and the generated code will be awful. (A telltale sign of
 204                     // this is large quantities of `mov [byte ptr foo],0` in the generated code.)
 205                     memset_intrinsic(
 206                         self,
 207                         false,
 208                         ty,
 209                         llresult,
 210                         self.const_u8(0),
 211                         self.const_usize(1),
 212                     );
 213                 }
 214                 return;
 215             }
 216             // Effectively no-ops
 217             "uninit" | "forget" => {
 218                 return;
 219             }
 220             "offset" => {
 221                 let ptr = args[0].immediate();
 222                 let offset = args[1].immediate();
 223                 self.inbounds_gep(ptr, &[offset])
 224             }
 225             "arith_offset" => {
 226                 let ptr = args[0].immediate();
 227                 let offset = args[1].immediate();
 228                 self.gep(ptr, &[offset])
 229             }
 230
 231             "copy_nonoverlapping" => {
 232                 copy_intrinsic(
 233                     self,
 234                     false,
 235                     false,
 236                     substs.type_at(0),
 237                     args[1].immediate(),
 238                     args[0].immediate(),
 239                     args[2].immediate(),
 240                 );
 241                 return;
 242             }
 243             "copy" => {
 244                 copy_intrinsic(
 245                     self,
 246                     true,
 247                     false,
 248                     substs.type_at(0),
 249                     args[1].immediate(),
 250                     args[0].immediate(),
 251                     args[2].immediate(),
 252                 );
 253                 return;
 254             }
 255             "write_bytes" => {
 256                 memset_intrinsic(
 257                     self,
 258                     false,
 259                     substs.type_at(0),
 260                     args[0].immediate(),
 261                     args[1].immediate(),
 262                     args[2].immediate(),
 263                 );
 264                 return;
 265             }
 266
 267             "volatile_copy_nonoverlapping_memory" => {
 268                 copy_intrinsic(
 269                     self,
 270                     false,
 271                     true,
 272                     substs.type_at(0),
 273                     args[0].immediate(),
 274                     args[1].immediate(),
 275                     args[2].immediate(),
 276                 );
 277                 return;
 278             }
 279             "volatile_copy_memory" => {
 280                 copy_intrinsic(
 281                     self,
 282                     true,
 283                     true,
 284                     substs.type_at(0),
 285                     args[0].immediate(),
 286                     args[1].immediate(),
 287                     args[2].immediate(),
 288                 );
 289                 return;
 290             }
 291             "volatile_set_memory" => {
 292                 memset_intrinsic(
 293                     self,
 294                     true,
 295                     substs.type_at(0),
 296                     args[0].immediate(),
 297                     args[1].immediate(),
 298                     args[2].immediate(),
 299                 );
 300                 return;
 301             }
 302             "volatile_load" | "unaligned_volatile_load" => {
 303                 let tp_ty = substs.type_at(0);
 304                 let mut ptr = args[0].immediate();
 305                 if let PassMode::Cast(ty) = fn_abi.ret.mode {
 306                     ptr = self.pointercast(ptr, self.type_ptr_to(ty.llvm_type(self)));
 307                 }
 308                 let load = self.volatile_load(ptr);
 309                 let align = if name == "unaligned_volatile_load" {
 310                     1
 311                 } else {
 312                     self.align_of(tp_ty).bytes() as u32
 313                 };
 314                 unsafe {
 315                     llvm::LLVMSetAlignment(load, align);
 316                 }
 317                 to_immediate(self, load, self.layout_of(tp_ty))
 318             }
 319             "volatile_store" => {
 320                 let dst = args[0].deref(self.cx());
 321                 args[1].val.volatile_store(self, dst);
 322                 return;
 323             }
 324             "unaligned_volatile_store" => {
 325                 let dst = args[0].deref(self.cx());
 326                 args[1].val.unaligned_volatile_store(self, dst);
 327                 return;
 328             }
 329             "prefetch_read_data"
 330             | "prefetch_write_data"
 331             | "prefetch_read_instruction"
 332             | "prefetch_write_instruction" => {
 333                 let expect = self.get_intrinsic(&("llvm.prefetch"));
 334                 let (rw, cache_type) = match name {
 335                     "prefetch_read_data" => (0, 1),
 336                     "prefetch_write_data" => (1, 1),
 337                     "prefetch_read_instruction" => (0, 0),
 338                     "prefetch_write_instruction" => (1, 0),
 339                     _ => bug!(),
 340                 };
 341                 self.call(
 342                     expect,
 343                     &[
 344                         args[0].immediate(),
 345                         self.const_i32(rw),
 346                         args[1].immediate(),
 347                         self.const_i32(cache_type),
 348                     ],
 349                     None,
 350                 )
 351             }
 352             "ctlz" | "ctlz_nonzero" | "cttz" | "cttz_nonzero" | "ctpop" | "bswap"
 353             | "bitreverse" | "add_with_overflow" | "sub_with_overflow" | "mul_with_overflow"
 354             | "wrapping_add" | "wrapping_sub" | "wrapping_mul" | "unchecked_div"
 355             | "unchecked_rem" | "unchecked_shl" | "unchecked_shr" | "unchecked_add"
 356             | "unchecked_sub" | "unchecked_mul" | "exact_div" | "rotate_left" | "rotate_right"
 357             | "saturating_add" | "saturating_sub" => {
 358                 let ty = arg_tys[0];
 359                 match int_type_width_signed(ty, self) {
 360                     Some((width, signed)) => match name {
 361                         "ctlz" | "cttz" => {
 362                             let y = self.const_bool(false);
 363                             let llfn = self.get_intrinsic(&format!("llvm.{}.i{}", name, width));
 364                             self.call(llfn, &[args[0].immediate(), y], None)
 365                         }
 366                         "ctlz_nonzero" | "cttz_nonzero" => {
 367                             let y = self.const_bool(true);
 368                             let llvm_name = &format!("llvm.{}.i{}", &name[..4], width);
 369                             let llfn = self.get_intrinsic(llvm_name);
 370                             self.call(llfn, &[args[0].immediate(), y], None)
 371                         }
 372                         "ctpop" => self.call(
 373                             self.get_intrinsic(&format!("llvm.ctpop.i{}", width)),
 374                             &[args[0].immediate()],
 375                             None,
 376                         ),
 377                         "bswap" => {
 378                             if width == 8 {
 379                                 args[0].immediate() // byte swap a u8/i8 is just a no-op
 380                             } else {
 381                                 self.call(
 382                                     self.get_intrinsic(&format!("llvm.bswap.i{}", width)),
 383                                     &[args[0].immediate()],
 384                                     None,
 385                                 )
 386                             }
 387                         }
 388                         "bitreverse" => self.call(
 389                             self.get_intrinsic(&format!("llvm.bitreverse.i{}", width)),
 390                             &[args[0].immediate()],
 391                             None,
 392                         ),
 393                         "add_with_overflow" | "sub_with_overflow" | "mul_with_overflow" => {
 394                             let intrinsic = format!(
 395                                 "llvm.{}{}.with.overflow.i{}",
 396                                 if signed { 's' } else { 'u' },
 397                                 &name[..3],
 398                                 width
 399                             );
 400                             let llfn = self.get_intrinsic(&intrinsic);
 401
 402                             // Convert `i1` to a `bool`, and write it to the out parameter
 403                             let pair =
 404                                 self.call(llfn, &[args[0].immediate(), args[1].immediate()], None);
 405                             let val = self.extract_value(pair, 0);
 406                             let overflow = self.extract_value(pair, 1);
 407                             let overflow = self.zext(overflow, self.type_bool());
 408
 409                             let dest = result.project_field(self, 0);
 410                             self.store(val, dest.llval, dest.align);
 411                             let dest = result.project_field(self, 1);
 412                             self.store(overflow, dest.llval, dest.align);
 413
 414                             return;
 415                         }
 416                         "wrapping_add" => self.add(args[0].immediate(), args[1].immediate()),
 417                         "wrapping_sub" => self.sub(args[0].immediate(), args[1].immediate()),
 418                         "wrapping_mul" => self.mul(args[0].immediate(), args[1].immediate()),
 419                         "exact_div" => {
 420                             if signed {
 421                                 self.exactsdiv(args[0].immediate(), args[1].immediate())
 422                             } else {
 423                                 self.exactudiv(args[0].immediate(), args[1].immediate())
 424                             }
 425                         }
 426                         "unchecked_div" => {
 427                             if signed {
 428                                 self.sdiv(args[0].immediate(), args[1].immediate())
 429                             } else {
 430                                 self.udiv(args[0].immediate(), args[1].immediate())
 431                             }
 432                         }
 433                         "unchecked_rem" => {
 434                             if signed {
 435                                 self.srem(args[0].immediate(), args[1].immediate())
 436                             } else {
 437                                 self.urem(args[0].immediate(), args[1].immediate())
 438                             }
 439                         }
 440                         "unchecked_shl" => self.shl(args[0].immediate(), args[1].immediate()),
 441                         "unchecked_shr" => {
 442                             if signed {
 443                                 self.ashr(args[0].immediate(), args[1].immediate())
 444                             } else {
 445                                 self.lshr(args[0].immediate(), args[1].immediate())
 446                             }
 447                         }
 448                         "unchecked_add" => {
 449                             if signed {
 450                                 self.unchecked_sadd(args[0].immediate(), args[1].immediate())
 451                             } else {
 452                                 self.unchecked_uadd(args[0].immediate(), args[1].immediate())
 453                             }
 454                         }
 455                         "unchecked_sub" => {
 456                             if signed {
 457                                 self.unchecked_ssub(args[0].immediate(), args[1].immediate())
 458                             } else {
 459                                 self.unchecked_usub(args[0].immediate(), args[1].immediate())
 460                             }
 461                         }
 462                         "unchecked_mul" => {
 463                             if signed {
 464                                 self.unchecked_smul(args[0].immediate(), args[1].immediate())
 465                             } else {
 466                                 self.unchecked_umul(args[0].immediate(), args[1].immediate())
 467                             }
 468                         }
 469                         "rotate_left" | "rotate_right" => {
 470                             let is_left = name == "rotate_left";
 471                             let val = args[0].immediate();
 472                             let raw_shift = args[1].immediate();
 473                             // rotate = funnel shift with first two args the same
 474                             let llvm_name =
 475                                 &format!("llvm.fsh{}.i{}", if is_left { 'l' } else { 'r' }, width);
 476                             let llfn = self.get_intrinsic(llvm_name);
 477                             self.call(llfn, &[val, val, raw_shift], None)
 478                         }
 479                         "saturating_add" | "saturating_sub" => {
 480                             let is_add = name == "saturating_add";
 481                             let lhs = args[0].immediate();
 482                             let rhs = args[1].immediate();
 483                             if llvm_util::get_major_version() >= 8 {
 484                                 let llvm_name = &format!(
 485                                     "llvm.{}{}.sat.i{}",
 486                                     if signed { 's' } else { 'u' },
 487                                     if is_add { "add" } else { "sub" },
 488                                     width
 489                                 );
 490                                 let llfn = self.get_intrinsic(llvm_name);
 491                                 self.call(llfn, &[lhs, rhs], None)
 492                             } else {
 493                                 let llvm_name = &format!(
 494                                     "llvm.{}{}.with.overflow.i{}",
 495                                     if signed { 's' } else { 'u' },
 496                                     if is_add { "add" } else { "sub" },
 497                                     width
 498                                 );
 499                                 let llfn = self.get_intrinsic(llvm_name);
 500                                 let pair = self.call(llfn, &[lhs, rhs], None);
 501                                 let val = self.extract_value(pair, 0);
 502                                 let overflow = self.extract_value(pair, 1);
 503                                 let llty = self.type_ix(width);
 504
 505                                 let limit = if signed {
 506                                     let limit_lo = self
 507                                         .const_uint_big(llty, (i128::MIN >> (128 - width)) as u128);
 508                                     let limit_hi = self
 509                                         .const_uint_big(llty, (i128::MAX >> (128 - width)) as u128);
 510                                     let neg = self.icmp(
 511                                         IntPredicate::IntSLT,
 512                                         val,
 513                                         self.const_uint(llty, 0),
 514                                     );
 515                                     self.select(neg, limit_hi, limit_lo)
 516                                 } else if is_add {
 517                                     self.const_uint_big(llty, u128::MAX >> (128 - width))
 518                                 } else {
 519                                     self.const_uint(llty, 0)
 520                                 };
 521                                 self.select(overflow, limit, val)
 522                             }
 523                         }
 524                         _ => bug!(),
 525                     },
 526                     None => {
 527                         span_invalid_monomorphization_error(
 528                             tcx.sess,
 529                             span,
 530                             &format!(
 531                                 "invalid monomorphization of `{}` intrinsic: \
 532                                       expected basic integer type, found `{}`",
 533                                 name, ty
 534                             ),
 535                         );
 536                         return;
 537                     }
 538                 }
 539             }
 540             "fadd_fast" | "fsub_fast" | "fmul_fast" | "fdiv_fast" | "frem_fast" => {
 541                 match float_type_width(arg_tys[0]) {
 542                     Some(_width) => match name {
 543                         "fadd_fast" => self.fadd_fast(args[0].immediate(), args[1].immediate()),
 544                         "fsub_fast" => self.fsub_fast(args[0].immediate(), args[1].immediate()),
 545                         "fmul_fast" => self.fmul_fast(args[0].immediate(), args[1].immediate()),
 546                         "fdiv_fast" => self.fdiv_fast(args[0].immediate(), args[1].immediate()),
 547                         "frem_fast" => self.frem_fast(args[0].immediate(), args[1].immediate()),
 548                         _ => bug!(),
 549                     },
 550                     None => {
 551                         span_invalid_monomorphization_error(
 552                             tcx.sess,
 553                             span,
 554                             &format!(
 555                                 "invalid monomorphization of `{}` intrinsic: \
 556                                       expected basic float type, found `{}`",
 557                                 name, arg_tys[0]
 558                             ),
 559                         );
 560                         return;
 561                     }
 562                 }
 563             }
 564
 565             "float_to_int_approx_unchecked" => {
 566                 if float_type_width(arg_tys[0]).is_none() {
 567                     span_invalid_monomorphization_error(
 568                         tcx.sess,
 569                         span,
 570                         &format!(
 571                             "invalid monomorphization of `float_to_int_approx_unchecked` \
 572                                   intrinsic: expected basic float type, \
 573                                   found `{}`",
 574                             arg_tys[0]
 575                         ),
 576                     );
 577                     return;
 578                 }
 579                 match int_type_width_signed(ret_ty, self.cx) {
 580                     Some((width, signed)) => {
 581                         if signed {
 582                             self.fptosi(args[0].immediate(), self.cx.type_ix(width))
 583                         } else {
 584                             self.fptoui(args[0].immediate(), self.cx.type_ix(width))
 585                         }
 586                     }
 587                     None => {
 588                         span_invalid_monomorphization_error(
 589                             tcx.sess,
 590                             span,
 591                             &format!(
 592                                 "invalid monomorphization of `float_to_int_approx_unchecked` \
 593                                       intrinsic:  expected basic integer type, \
 594                                       found `{}`",
 595                                 ret_ty
 596                             ),
 597                         );
 598                         return;
 599                     }
 600                 }
 601             }
 602
 603             "discriminant_value" => args[0].deref(self.cx()).codegen_get_discr(self, ret_ty),
 604
 605             name if name.starts_with("simd_") => {
 606                 match generic_simd_intrinsic(self, name, callee_ty, args, ret_ty, llret_ty, span) {
 607                     Ok(llval) => llval,
 608                     Err(()) => return,
 609                 }
 610             }
 611             // This requires that atomic intrinsics follow a specific naming pattern:
 612             // "atomic_<operation>[_<ordering>]", and no ordering means SeqCst
 613             name if name.starts_with("atomic_") => {
 614                 use rustc_codegen_ssa::common::AtomicOrdering::*;
 615                 use rustc_codegen_ssa::common::{AtomicRmwBinOp, SynchronizationScope};
 616
 617                 let split: Vec<&str> = name.split('_').collect();
 618
 619                 let is_cxchg = split[1] == "cxchg" || split[1] == "cxchgweak";
 620                 let (order, failorder) = match split.len() {
 621                     2 => (SequentiallyConsistent, SequentiallyConsistent),
 622                     3 => match split[2] {
 623                         "unordered" => (Unordered, Unordered),
 624                         "relaxed" => (Monotonic, Monotonic),
 625                         "acq" => (Acquire, Acquire),
 626                         "rel" => (Release, Monotonic),
 627                         "acqrel" => (AcquireRelease, Acquire),
 628                         "failrelaxed" if is_cxchg => (SequentiallyConsistent, Monotonic),
 629                         "failacq" if is_cxchg => (SequentiallyConsistent, Acquire),
 630                         _ => self.sess().fatal("unknown ordering in atomic intrinsic"),
 631                     },
 632                     4 => match (split[2], split[3]) {
 633                         ("acq", "failrelaxed") if is_cxchg => (Acquire, Monotonic),
 634                         ("acqrel", "failrelaxed") if is_cxchg => (AcquireRelease, Monotonic),
 635                         _ => self.sess().fatal("unknown ordering in atomic intrinsic"),
 636                     },
 637                     _ => self.sess().fatal("Atomic intrinsic not in correct format"),
 638                 };
 639
 640                 let invalid_monomorphization = |ty| {
 641                     span_invalid_monomorphization_error(
 642                         tcx.sess,
 643                         span,
 644                         &format!(
 645                             "invalid monomorphization of `{}` intrinsic: \
 646                                   expected basic integer type, found `{}`",
 647                             name, ty
 648                         ),
 649                     );
 650                 };
 651
 652                 match split[1] {
 653                     "cxchg" | "cxchgweak" => {
 654                         let ty = substs.type_at(0);
 655                         if int_type_width_signed(ty, self).is_some() {
 656                             let weak = split[1] == "cxchgweak";
 657                             let pair = self.atomic_cmpxchg(
 658                                 args[0].immediate(),
 659                                 args[1].immediate(),
 660                                 args[2].immediate(),
 661                                 order,
 662                                 failorder,
 663                                 weak,
 664                             );
 665                             let val = self.extract_value(pair, 0);
 666                             let success = self.extract_value(pair, 1);
 667                             let success = self.zext(success, self.type_bool());
 668
 669                             let dest = result.project_field(self, 0);
 670                             self.store(val, dest.llval, dest.align);
 671                             let dest = result.project_field(self, 1);
 672                             self.store(success, dest.llval, dest.align);
 673                             return;
 674                         } else {
 675                             return invalid_monomorphization(ty);
 676                         }
 677                     }
 678
 679                     "load" => {
 680                         let ty = substs.type_at(0);
 681                         if int_type_width_signed(ty, self).is_some() {
 682                             let size = self.size_of(ty);
 683                             self.atomic_load(args[0].immediate(), order, size)
 684                         } else {
 685                             return invalid_monomorphization(ty);
 686                         }
 687                     }
 688
 689                     "store" => {
 690                         let ty = substs.type_at(0);
 691                         if int_type_width_signed(ty, self).is_some() {
 692                             let size = self.size_of(ty);
 693                             self.atomic_store(
 694                                 args[1].immediate(),
 695                                 args[0].immediate(),
 696                                 order,
 697                                 size,
 698                             );
 699                             return;
 700                         } else {
 701                             return invalid_monomorphization(ty);
 702                         }
 703                     }
 704
 705                     "fence" => {
 706                         self.atomic_fence(order, SynchronizationScope::CrossThread);
 707                         return;
 708                     }
 709
 710                     "singlethreadfence" => {
 711                         self.atomic_fence(order, SynchronizationScope::SingleThread);
 712                         return;
 713                     }
 714
 715                     // These are all AtomicRMW ops
 716                     op => {
 717                         let atom_op = match op {
 718                             "xchg" => AtomicRmwBinOp::AtomicXchg,
 719                             "xadd" => AtomicRmwBinOp::AtomicAdd,
 720                             "xsub" => AtomicRmwBinOp::AtomicSub,
 721                             "and" => AtomicRmwBinOp::AtomicAnd,
 722                             "nand" => AtomicRmwBinOp::AtomicNand,
 723                             "or" => AtomicRmwBinOp::AtomicOr,
 724                             "xor" => AtomicRmwBinOp::AtomicXor,
 725                             "max" => AtomicRmwBinOp::AtomicMax,
 726                             "min" => AtomicRmwBinOp::AtomicMin,
 727                             "umax" => AtomicRmwBinOp::AtomicUMax,
 728                             "umin" => AtomicRmwBinOp::AtomicUMin,
 729                             _ => self.sess().fatal("unknown atomic operation"),
 730                         };
 731
 732                         let ty = substs.type_at(0);
 733                         if int_type_width_signed(ty, self).is_some() {
 734                             self.atomic_rmw(
 735                                 atom_op,
 736                                 args[0].immediate(),
 737                                 args[1].immediate(),
 738                                 order,
 739                             )
 740                         } else {
 741                             return invalid_monomorphization(ty);
 742                         }
 743                     }
 744                 }
 745             }
 746
 747             "nontemporal_store" => {
 748                 let dst = args[0].deref(self.cx());
 749                 args[1].val.nontemporal_store(self, dst);
 750                 return;
 751             }
 752
 753             "ptr_offset_from" => {
 754                 let ty = substs.type_at(0);
 755                 let pointee_size = self.size_of(ty);
 756
 757                 // This is the same sequence that Clang emits for pointer subtraction.
 758                 // It can be neither `nsw` nor `nuw` because the input is treated as
 759                 // unsigned but then the output is treated as signed, so neither works.
 760                 let a = args[0].immediate();
 761                 let b = args[1].immediate();
 762                 let a = self.ptrtoint(a, self.type_isize());
 763                 let b = self.ptrtoint(b, self.type_isize());
 764                 let d = self.sub(a, b);
 765                 let pointee_size = self.const_usize(pointee_size.bytes());
 766                 // this is where the signed magic happens (notice the `s` in `exactsdiv`)
 767                 self.exactsdiv(d, pointee_size)
 768             }
 769
 770             _ => bug!("unknown intrinsic '{}'", name),
 771         };
 772
 773         if !fn_abi.ret.is_ignore() {
 774             if let PassMode::Cast(ty) = fn_abi.ret.mode {
 775                 let ptr_llty = self.type_ptr_to(ty.llvm_type(self));
 776                 let ptr = self.pointercast(result.llval, ptr_llty);
 777                 self.store(llval, ptr, result.align);
 778             } else {
 779                 OperandRef::from_immediate_or_packed_pair(self, llval, result.layout)
 780                     .val
 781                     .store(self, result);
 782             }
 783         }
 784     }
 785
 786     fn abort(&mut self) {
 787         let fnname = self.get_intrinsic(&("llvm.trap"));
 788         self.call(fnname, &[], None);
 789     }
 790
 791     fn assume(&mut self, val: Self::Value) {
 792         let assume_intrinsic = self.get_intrinsic("llvm.assume");
 793         self.call(assume_intrinsic, &[val], None);
 794     }
 795
 796     fn expect(&mut self, cond: Self::Value, expected: bool) -> Self::Value {
 797         let expect = self.get_intrinsic(&"llvm.expect.i1");
 798         self.call(expect, &[cond, self.const_bool(expected)], None)
 799     }
 800
 801     fn sideeffect(&mut self) {
 802         if self.tcx.sess.opts.debugging_opts.insert_sideeffect {
 803             let fnname = self.get_intrinsic(&("llvm.sideeffect"));
 804             self.call(fnname, &[], None);
 805         }
 806     }
 807
 808     fn va_start(&mut self, va_list: &'ll Value) -> &'ll Value {
 809         let intrinsic = self.cx().get_intrinsic("llvm.va_start");
 810         self.call(intrinsic, &[va_list], None)
 811     }
 812
 813     fn va_end(&mut self, va_list: &'ll Value) -> &'ll Value {
 814         let intrinsic = self.cx().get_intrinsic("llvm.va_end");
 815         self.call(intrinsic, &[va_list], None)
 816     }
 817 }
 818
 819 fn copy_intrinsic(
 820     bx: &mut Builder<'a, 'll, 'tcx>,
 821     allow_overlap: bool,
 822     volatile: bool,
 823     ty: Ty<'tcx>,
 824     dst: &'ll Value,
 825     src: &'ll Value,
 826     count: &'ll Value,
 827 ) {
 828     let (size, align) = bx.size_and_align_of(ty);
 829     let size = bx.mul(bx.const_usize(size.bytes()), count);
 830     let flags = if volatile { MemFlags::VOLATILE } else { MemFlags::empty() };
 831     if allow_overlap {
 832         bx.memmove(dst, align, src, align, size, flags);
 833     } else {
 834         bx.memcpy(dst, align, src, align, size, flags);
 835     }
 836 }
 837
 838 fn memset_intrinsic(
 839     bx: &mut Builder<'a, 'll, 'tcx>,
 840     volatile: bool,
 841     ty: Ty<'tcx>,
 842     dst: &'ll Value,
 843     val: &'ll Value,
 844     count: &'ll Value,
 845 ) {
 846     let (size, align) = bx.size_and_align_of(ty);
 847     let size = bx.mul(bx.const_usize(size.bytes()), count);
 848     let flags = if volatile { MemFlags::VOLATILE } else { MemFlags::empty() };
 849     bx.memset(dst, val, size, align, flags);
 850 }
 851
 852 fn try_intrinsic(
 853     bx: &mut Builder<'a, 'll, 'tcx>,
 854     func: &'ll Value,
 855     data: &'ll Value,
 856     local_ptr: &'ll Value,
 857     dest: &'ll Value,
 858 ) {
 859     if bx.sess().no_landing_pads() {
 860         bx.call(func, &[data], None);
 861         let ptr_align = bx.tcx().data_layout.pointer_align.abi;
 862         bx.store(bx.const_null(bx.type_i8p()), dest, ptr_align);
 863     } else if wants_msvc_seh(bx.sess()) {
 864         codegen_msvc_try(bx, func, data, local_ptr, dest);
 865     } else {
 866         codegen_gnu_try(bx, func, data, local_ptr, dest);
 867     }
 868 }
 869
 870 // MSVC's definition of the `rust_try` function.
 871 //
 872 // This implementation uses the new exception handling instructions in LLVM
 873 // which have support in LLVM for SEH on MSVC targets. Although these
 874 // instructions are meant to work for all targets, as of the time of this
 875 // writing, however, LLVM does not recommend the usage of these new instructions
 876 // as the old ones are still more optimized.
 877 fn codegen_msvc_try(
 878     bx: &mut Builder<'a, 'll, 'tcx>,
 879     func: &'ll Value,
 880     data: &'ll Value,
 881     local_ptr: &'ll Value,
 882     dest: &'ll Value,
 883 ) {
 884     let llfn = get_rust_try_fn(bx, &mut |mut bx| {
 885         bx.set_personality_fn(bx.eh_personality());
 886         bx.sideeffect();
 887
 888         let mut normal = bx.build_sibling_block("normal");
 889         let mut catchswitch = bx.build_sibling_block("catchswitch");
 890         let mut catchpad = bx.build_sibling_block("catchpad");
 891         let mut caught = bx.build_sibling_block("caught");
 892
 893         let func = llvm::get_param(bx.llfn(), 0);
 894         let data = llvm::get_param(bx.llfn(), 1);
 895         let local_ptr = llvm::get_param(bx.llfn(), 2);
 896
 897         // We're generating an IR snippet that looks like:
 898         //
 899         //   declare i32 @rust_try(%func, %data, %ptr) {
 900         //      %slot = alloca [2 x i64]
 901         //      invoke %func(%data) to label %normal unwind label %catchswitch
 902         //
 903         //   normal:
 904         //      ret i32 0
 905         //
 906         //   catchswitch:
 907         //      %cs = catchswitch within none [%catchpad] unwind to caller
 908         //
 909         //   catchpad:
 910         //      %tok = catchpad within %cs [%type_descriptor, 0, %slot]
 911         //      %ptr[0] = %slot[0]
 912         //      %ptr[1] = %slot[1]
 913         //      catchret from %tok to label %caught
 914         //
 915         //   caught:
 916         //      ret i32 1
 917         //   }
 918         //
 919         // This structure follows the basic usage of throw/try/catch in LLVM.
 920         // For example, compile this C++ snippet to see what LLVM generates:
 921         //
 922         //      #include <stdint.h>
 923         //
 924         //      struct rust_panic {
 925         //          uint64_t x[2];
 926         //      }
 927         //
 928         //      int bar(void (*foo)(void), uint64_t *ret) {
 929         //          try {
 930         //              foo();
 931         //              return 0;
 932         //          } catch(rust_panic a) {
 933         //              ret[0] = a.x[0];
 934         //              ret[1] = a.x[1];
 935         //              return 1;
 936         //          }
 937         //      }
 938         //
 939         // More information can be found in libstd's seh.rs implementation.
 940         let i64_2 = bx.type_array(bx.type_i64(), 2);
 941         let i64_align = bx.tcx().data_layout.i64_align.abi;
 942         let slot = bx.alloca(i64_2, i64_align);
 943         bx.invoke(func, &[data], normal.llbb(), catchswitch.llbb(), None);
 944
 945         normal.ret(bx.const_i32(0));
 946
 947         let cs = catchswitch.catch_switch(None, None, 1);
 948         catchswitch.add_handler(cs, catchpad.llbb());
 949
 950         let tydesc = match bx.tcx().lang_items().eh_catch_typeinfo() {
 951             Some(did) => bx.get_static(did),
 952             None => bug!("eh_catch_typeinfo not defined, but needed for SEH unwinding"),
 953         };
 954         let funclet = catchpad.catch_pad(cs, &[tydesc, bx.const_i32(0), slot]);
 955
 956         let payload = catchpad.load(slot, i64_align);
 957         let local_ptr = catchpad.bitcast(local_ptr, bx.type_ptr_to(i64_2));
 958         catchpad.store(payload, local_ptr, i64_align);
 959         catchpad.catch_ret(&funclet, caught.llbb());
 960
 961         caught.ret(bx.const_i32(1));
 962     });
 963
 964     // Note that no invoke is used here because by definition this function
 965     // can't panic (that's what it's catching).
 966     let ret = bx.call(llfn, &[func, data, local_ptr], None);
 967     let i32_align = bx.tcx().data_layout.i32_align.abi;
 968     bx.store(ret, dest, i32_align);
 969 }
 970
 971 // Definition of the standard `try` function for Rust using the GNU-like model
 972 // of exceptions (e.g., the normal semantics of LLVM's `landingpad` and `invoke`
 973 // instructions).
 974 //
 975 // This codegen is a little surprising because we always call a shim
 976 // function instead of inlining the call to `invoke` manually here. This is done
 977 // because in LLVM we're only allowed to have one personality per function
 978 // definition. The call to the `try` intrinsic is being inlined into the
 979 // function calling it, and that function may already have other personality
 980 // functions in play. By calling a shim we're guaranteed that our shim will have
 981 // the right personality function.
 982 fn codegen_gnu_try(
 983     bx: &mut Builder<'a, 'll, 'tcx>,
 984     func: &'ll Value,
 985     data: &'ll Value,
 986     local_ptr: &'ll Value,
 987     dest: &'ll Value,
 988 ) {
 989     let llfn = get_rust_try_fn(bx, &mut |mut bx| {
 990         // Codegens the shims described above:
 991         //
 992         //   bx:
 993         //      invoke %func(%args...) normal %normal unwind %catch
 994         //
 995         //   normal:
 996         //      ret 0
 997         //
 998         //   catch:
 999         //      (ptr, _) = landingpad
1000         //      store ptr, %local_ptr
1001         //      ret 1
1002         //
1003         // Note that the `local_ptr` data passed into the `try` intrinsic is
1004         // expected to be `*mut *mut u8` for this to actually work, but that's
1005         // managed by the standard library.
1006
1007         bx.sideeffect();
1008
1009         let mut then = bx.build_sibling_block("then");
1010         let mut catch = bx.build_sibling_block("catch");
1011
1012         let func = llvm::get_param(bx.llfn(), 0);
1013         let data = llvm::get_param(bx.llfn(), 1);
1014         let local_ptr = llvm::get_param(bx.llfn(), 2);
1015         bx.invoke(func, &[data], then.llbb(), catch.llbb(), None);
1016         then.ret(bx.const_i32(0));
1017
1018         // Type indicator for the exception being thrown.
1019         //
1020         // The first value in this tuple is a pointer to the exception object
1021         // being thrown.  The second value is a "selector" indicating which of
1022         // the landing pad clauses the exception's type had been matched to.
1023         // rust_try ignores the selector.
1024         let lpad_ty = bx.type_struct(&[bx.type_i8p(), bx.type_i32()], false);
1025         let vals = catch.landing_pad(lpad_ty, bx.eh_personality(), 1);
1026         let tydesc = match bx.tcx().lang_items().eh_catch_typeinfo() {
1027             Some(tydesc) => {
1028                 let tydesc = bx.get_static(tydesc);
1029                 bx.bitcast(tydesc, bx.type_i8p())
1030             }
1031             None => bx.const_null(bx.type_i8p()),
1032         };
1033         catch.add_clause(vals, tydesc);
1034         let ptr = catch.extract_value(vals, 0);
1035         let ptr_align = bx.tcx().data_layout.pointer_align.abi;
1036         let bitcast = catch.bitcast(local_ptr, bx.type_ptr_to(bx.type_i8p()));
1037         catch.store(ptr, bitcast, ptr_align);
1038         catch.ret(bx.const_i32(1));
1039     });
1040
1041     // Note that no invoke is used here because by definition this function
1042     // can't panic (that's what it's catching).
1043     let ret = bx.call(llfn, &[func, data, local_ptr], None);
1044     let i32_align = bx.tcx().data_layout.i32_align.abi;
1045     bx.store(ret, dest, i32_align);
1046 }
1047
1048 // Helper function to give a Block to a closure to codegen a shim function.
1049 // This is currently primarily used for the `try` intrinsic functions above.
1050 fn gen_fn<'ll, 'tcx>(
1051     cx: &CodegenCx<'ll, 'tcx>,
1052     name: &str,
1053     inputs: Vec<Ty<'tcx>>,
1054     output: Ty<'tcx>,
1055     codegen: &mut dyn FnMut(Builder<'_, 'll, 'tcx>),
1056 ) -> &'ll Value {
1057     let rust_fn_sig = ty::Binder::bind(cx.tcx.mk_fn_sig(
1058         inputs.into_iter(),
1059         output,
1060         false,
1061         hir::Unsafety::Unsafe,
1062         Abi::Rust,
1063     ));
1064     let fn_abi = FnAbi::of_fn_ptr(cx, rust_fn_sig, &[]);
1065     let llfn = cx.declare_fn(name, &fn_abi);
1066     // FIXME(eddyb) find a nicer way to do this.
1067     unsafe { llvm::LLVMRustSetLinkage(llfn, llvm::Linkage::InternalLinkage) };
1068     let bx = Builder::new_block(cx, llfn, "entry-block");
1069     codegen(bx);
1070     llfn
1071 }
1072
1073 // Helper function used to get a handle to the `__rust_try` function used to
1074 // catch exceptions.
1075 //
1076 // This function is only generated once and is then cached.
1077 fn get_rust_try_fn<'ll, 'tcx>(
1078     cx: &CodegenCx<'ll, 'tcx>,
1079     codegen: &mut dyn FnMut(Builder<'_, 'll, 'tcx>),
1080 ) -> &'ll Value {
1081     if let Some(llfn) = cx.rust_try_fn.get() {
1082         return llfn;
1083     }
1084
1085     // Define the type up front for the signature of the rust_try function.
1086     let tcx = cx.tcx;
1087     let i8p = tcx.mk_mut_ptr(tcx.types.i8);
1088     let fn_ty = tcx.mk_fn_ptr(ty::Binder::bind(tcx.mk_fn_sig(
1089         iter::once(i8p),
1090         tcx.mk_unit(),
1091         false,
1092         hir::Unsafety::Unsafe,
1093         Abi::Rust,
1094     )));
1095     let output = tcx.types.i32;
1096     let rust_try = gen_fn(cx, "__rust_try", vec![fn_ty, i8p, i8p], output, codegen);
1097     cx.rust_try_fn.set(Some(rust_try));
1098     rust_try
1099 }
1100
1101 fn generic_simd_intrinsic(
1102     bx: &mut Builder<'a, 'll, 'tcx>,
1103     name: &str,
1104     callee_ty: Ty<'tcx>,
1105     args: &[OperandRef<'tcx, &'ll Value>],
1106     ret_ty: Ty<'tcx>,
1107     llret_ty: &'ll Type,
1108     span: Span,
1109 ) -> Result<&'ll Value, ()> {
1110     // macros for error handling:
1111     macro_rules! emit_error {
1112         ($msg: tt) => {
1113             emit_error!($msg, )
1114         };
1115         ($msg: tt, $($fmt: tt)*) => {
1116             span_invalid_monomorphization_error(
1117                 bx.sess(), span,
1118                 &format!(concat!("invalid monomorphization of `{}` intrinsic: ", $msg),
1119                          name, $($fmt)*));
1120         }
1121     }
1122
1123     macro_rules! return_error {
1124         ($($fmt: tt)*) => {
1125             {
1126                 emit_error!($($fmt)*);
1127                 return Err(());
1128             }
1129         }
1130     }
1131
1132     macro_rules! require {
1133         ($cond: expr, $($fmt: tt)*) => {
1134             if !$cond {
1135                 return_error!($($fmt)*);
1136             }
1137         };
1138     }
1139
1140     macro_rules! require_simd {
1141         ($ty: expr, $position: expr) => {
1142             require!($ty.is_simd(), "expected SIMD {} type, found non-SIMD `{}`", $position, $ty)
1143         };
1144     }
1145
1146     let tcx = bx.tcx();
1147     let sig = tcx
1148         .normalize_erasing_late_bound_regions(ty::ParamEnv::reveal_all(), &callee_ty.fn_sig(tcx));
1149     let arg_tys = sig.inputs();
1150
1151     if name == "simd_select_bitmask" {
1152         let in_ty = arg_tys[0];
1153         let m_len = match in_ty.kind {
1154             // Note that this `.unwrap()` crashes for isize/usize, that's sort
1155             // of intentional as there's not currently a use case for that.
1156             ty::Int(i) => i.bit_width().unwrap() as u64,
1157             ty::Uint(i) => i.bit_width().unwrap() as u64,
1158             _ => return_error!("`{}` is not an integral type", in_ty),
1159         };
1160         require_simd!(arg_tys[1], "argument");
1161         let v_len = arg_tys[1].simd_size(tcx);
1162         require!(
1163             m_len == v_len,
1164             "mismatched lengths: mask length `{}` != other vector length `{}`",
1165             m_len,
1166             v_len
1167         );
1168         let i1 = bx.type_i1();
1169         let i1xn = bx.type_vector(i1, m_len);
1170         let m_i1s = bx.bitcast(args[0].immediate(), i1xn);
1171         return Ok(bx.select(m_i1s, args[1].immediate(), args[2].immediate()));
1172     }
1173
1174     // every intrinsic below takes a SIMD vector as its first argument
1175     require_simd!(arg_tys[0], "input");
1176     let in_ty = arg_tys[0];
1177     let in_elem = arg_tys[0].simd_type(tcx);
1178     let in_len = arg_tys[0].simd_size(tcx);
1179
1180     let comparison = match name {
1181         "simd_eq" => Some(hir::BinOpKind::Eq),
1182         "simd_ne" => Some(hir::BinOpKind::Ne),
1183         "simd_lt" => Some(hir::BinOpKind::Lt),
1184         "simd_le" => Some(hir::BinOpKind::Le),
1185         "simd_gt" => Some(hir::BinOpKind::Gt),
1186         "simd_ge" => Some(hir::BinOpKind::Ge),
1187         _ => None,
1188     };
1189
1190     if let Some(cmp_op) = comparison {
1191         require_simd!(ret_ty, "return");
1192
1193         let out_len = ret_ty.simd_size(tcx);
1194         require!(
1195             in_len == out_len,
1196             "expected return type with length {} (same as input type `{}`), \
1197                   found `{}` with length {}",
1198             in_len,
1199             in_ty,
1200             ret_ty,
1201             out_len
1202         );
1203         require!(
1204             bx.type_kind(bx.element_type(llret_ty)) == TypeKind::Integer,
1205             "expected return type with integer elements, found `{}` with non-integer `{}`",
1206             ret_ty,
1207             ret_ty.simd_type(tcx)
1208         );
1209
1210         return Ok(compare_simd_types(
1211             bx,
1212             args[0].immediate(),
1213             args[1].immediate(),
1214             in_elem,
1215             llret_ty,
1216             cmp_op,
1217         ));
1218     }
1219
1220     if name.starts_with("simd_shuffle") {
1221         let n: u64 = name["simd_shuffle".len()..].parse().unwrap_or_else(|_| {
1222             span_bug!(span, "bad `simd_shuffle` instruction only caught in codegen?")
1223         });
1224
1225         require_simd!(ret_ty, "return");
1226
1227         let out_len = ret_ty.simd_size(tcx);
1228         require!(
1229             out_len == n,
1230             "expected return type of length {}, found `{}` with length {}",
1231             n,
1232             ret_ty,
1233             out_len
1234         );
1235         require!(
1236             in_elem == ret_ty.simd_type(tcx),
1237             "expected return element type `{}` (element of input `{}`), \
1238                   found `{}` with element type `{}`",
1239             in_elem,
1240             in_ty,
1241             ret_ty,
1242             ret_ty.simd_type(tcx)
1243         );
1244
1245         let total_len = u128::from(in_len) * 2;
1246
1247         let vector = args[2].immediate();
1248
1249         let indices: Option<Vec<_>> = (0..n)
1250             .map(|i| {
1251                 let arg_idx = i;
1252                 let val = bx.const_get_elt(vector, i as u64);
1253                 match bx.const_to_opt_u128(val, true) {
1254                     None => {
1255                         emit_error!("shuffle index #{} is not a constant", arg_idx);
1256                         None
1257                     }
1258                     Some(idx) if idx >= total_len => {
1259                         emit_error!(
1260                             "shuffle index #{} is out of bounds (limit {})",
1261                             arg_idx,
1262                             total_len
1263                         );
1264                         None
1265                     }
1266                     Some(idx) => Some(bx.const_i32(idx as i32)),
1267                 }
1268             })
1269             .collect();
1270         let indices = match indices {
1271             Some(i) => i,
1272             None => return Ok(bx.const_null(llret_ty)),
1273         };
1274
1275         return Ok(bx.shuffle_vector(
1276             args[0].immediate(),
1277             args[1].immediate(),
1278             bx.const_vector(&indices),
1279         ));
1280     }
1281
1282     if name == "simd_insert" {
1283         require!(
1284             in_elem == arg_tys[2],
1285             "expected inserted type `{}` (element of input `{}`), found `{}`",
1286             in_elem,
1287             in_ty,
1288             arg_tys[2]
1289         );
1290         return Ok(bx.insert_element(
1291             args[0].immediate(),
1292             args[2].immediate(),
1293             args[1].immediate(),
1294         ));
1295     }
1296     if name == "simd_extract" {
1297         require!(
1298             ret_ty == in_elem,
1299             "expected return type `{}` (element of input `{}`), found `{}`",
1300             in_elem,
1301             in_ty,
1302             ret_ty
1303         );
1304         return Ok(bx.extract_element(args[0].immediate(), args[1].immediate()));
1305     }
1306
1307     if name == "simd_select" {
1308         let m_elem_ty = in_elem;
1309         let m_len = in_len;
1310         require_simd!(arg_tys[1], "argument");
1311         let v_len = arg_tys[1].simd_size(tcx);
1312         require!(
1313             m_len == v_len,
1314             "mismatched lengths: mask length `{}` != other vector length `{}`",
1315             m_len,
1316             v_len
1317         );
1318         match m_elem_ty.kind {
1319             ty::Int(_) => {}
1320             _ => return_error!("mask element type is `{}`, expected `i_`", m_elem_ty),
1321         }
1322         // truncate the mask to a vector of i1s
1323         let i1 = bx.type_i1();
1324         let i1xn = bx.type_vector(i1, m_len as u64);
1325         let m_i1s = bx.trunc(args[0].immediate(), i1xn);
1326         return Ok(bx.select(m_i1s, args[1].immediate(), args[2].immediate()));
1327     }
1328
1329     if name == "simd_bitmask" {
1330         // The `fn simd_bitmask(vector) -> unsigned integer` intrinsic takes a
1331         // vector mask and returns an unsigned integer containing the most
1332         // significant bit (MSB) of each lane.
1333
1334         // If the vector has less than 8 lanes, an u8 is returned with zeroed
1335         // trailing bits.
1336         let expected_int_bits = in_len.max(8);
1337         match ret_ty.kind {
1338             ty::Uint(i) if i.bit_width() == Some(expected_int_bits as usize) => (),
1339             _ => return_error!("bitmask `{}`, expected `u{}`", ret_ty, expected_int_bits),
1340         }
1341
1342         // Integer vector <i{in_bitwidth} x in_len>:
1343         let (i_xn, in_elem_bitwidth) = match in_elem.kind {
1344             ty::Int(i) => (
1345                 args[0].immediate(),
1346                 i.bit_width().unwrap_or(bx.data_layout().pointer_size.bits() as _),
1347             ),
1348             ty::Uint(i) => (
1349                 args[0].immediate(),
1350                 i.bit_width().unwrap_or(bx.data_layout().pointer_size.bits() as _),
1351             ),
1352             _ => return_error!(
1353                 "vector argument `{}`'s element type `{}`, expected integer element type",
1354                 in_ty,
1355                 in_elem
1356             ),
1357         };
1358
1359         // Shift the MSB to the right by "in_elem_bitwidth - 1" into the first bit position.
1360         let shift_indices =
1361             vec![
1362                 bx.cx.const_int(bx.type_ix(in_elem_bitwidth as _), (in_elem_bitwidth - 1) as _);
1363                 in_len as _
1364             ];
1365         let i_xn_msb = bx.lshr(i_xn, bx.const_vector(shift_indices.as_slice()));
1366         // Truncate vector to an <i1 x N>
1367         let i1xn = bx.trunc(i_xn_msb, bx.type_vector(bx.type_i1(), in_len as _));
1368         // Bitcast <i1 x N> to iN:
1369         let i_ = bx.bitcast(i1xn, bx.type_ix(in_len as _));
1370         // Zero-extend iN to the bitmask type:
1371         return Ok(bx.zext(i_, bx.type_ix(expected_int_bits as _)));
1372     }
1373
1374     fn simd_simple_float_intrinsic(
1375         name: &str,
1376         in_elem: &::rustc::ty::TyS<'_>,
1377         in_ty: &::rustc::ty::TyS<'_>,
1378         in_len: u64,
1379         bx: &mut Builder<'a, 'll, 'tcx>,
1380         span: Span,
1381         args: &[OperandRef<'tcx, &'ll Value>],
1382     ) -> Result<&'ll Value, ()> {
1383         macro_rules! emit_error {
1384             ($msg: tt) => {
1385                 emit_error!($msg, )
1386             };
1387             ($msg: tt, $($fmt: tt)*) => {
1388                 span_invalid_monomorphization_error(
1389                     bx.sess(), span,
1390                     &format!(concat!("invalid monomorphization of `{}` intrinsic: ", $msg),
1391                              name, $($fmt)*));
1392             }
1393         }
1394         macro_rules! return_error {
1395             ($($fmt: tt)*) => {
1396                 {
1397                     emit_error!($($fmt)*);
1398                     return Err(());
1399                 }
1400             }
1401         }
1402         let ety = match in_elem.kind {
1403             ty::Float(f) if f.bit_width() == 32 => {
1404                 if in_len < 2 || in_len > 16 {
1405                     return_error!(
1406                         "unsupported floating-point vector `{}` with length `{}` \
1407                          out-of-range [2, 16]",
1408                         in_ty,
1409                         in_len
1410                     );
1411                 }
1412                 "f32"
1413             }
1414             ty::Float(f) if f.bit_width() == 64 => {
1415                 if in_len < 2 || in_len > 8 {
1416                     return_error!(
1417                         "unsupported floating-point vector `{}` with length `{}` \
1418                                    out-of-range [2, 8]",
1419                         in_ty,
1420                         in_len
1421                     );
1422                 }
1423                 "f64"
1424             }
1425             ty::Float(f) => {
1426                 return_error!(
1427                     "unsupported element type `{}` of floating-point vector `{}`",
1428                     f.name_str(),
1429                     in_ty
1430                 );
1431             }
1432             _ => {
1433                 return_error!("`{}` is not a floating-point type", in_ty);
1434             }
1435         };
1436
1437         let llvm_name = &format!("llvm.{0}.v{1}{2}", name, in_len, ety);
1438         let intrinsic = bx.get_intrinsic(&llvm_name);
1439         let c =
1440             bx.call(intrinsic, &args.iter().map(|arg| arg.immediate()).collect::<Vec<_>>(), None);
1441         unsafe { llvm::LLVMRustSetHasUnsafeAlgebra(c) };
1442         Ok(c)
1443     }
1444
1445     match name {
1446         "simd_fsqrt" => {
1447             return simd_simple_float_intrinsic("sqrt", in_elem, in_ty, in_len, bx, span, args);
1448         }
1449         "simd_fsin" => {
1450             return simd_simple_float_intrinsic("sin", in_elem, in_ty, in_len, bx, span, args);
1451         }
1452         "simd_fcos" => {
1453             return simd_simple_float_intrinsic("cos", in_elem, in_ty, in_len, bx, span, args);
1454         }
1455         "simd_fabs" => {
1456             return simd_simple_float_intrinsic("fabs", in_elem, in_ty, in_len, bx, span, args);
1457         }
1458         "simd_floor" => {
1459             return simd_simple_float_intrinsic("floor", in_elem, in_ty, in_len, bx, span, args);
1460         }
1461         "simd_ceil" => {
1462             return simd_simple_float_intrinsic("ceil", in_elem, in_ty, in_len, bx, span, args);
1463         }
1464         "simd_fexp" => {
1465             return simd_simple_float_intrinsic("exp", in_elem, in_ty, in_len, bx, span, args);
1466         }
1467         "simd_fexp2" => {
1468             return simd_simple_float_intrinsic("exp2", in_elem, in_ty, in_len, bx, span, args);
1469         }
1470         "simd_flog10" => {
1471             return simd_simple_float_intrinsic("log10", in_elem, in_ty, in_len, bx, span, args);
1472         }
1473         "simd_flog2" => {
1474             return simd_simple_float_intrinsic("log2", in_elem, in_ty, in_len, bx, span, args);
1475         }
1476         "simd_flog" => {
1477             return simd_simple_float_intrinsic("log", in_elem, in_ty, in_len, bx, span, args);
1478         }
1479         "simd_fpowi" => {
1480             return simd_simple_float_intrinsic("powi", in_elem, in_ty, in_len, bx, span, args);
1481         }
1482         "simd_fpow" => {
1483             return simd_simple_float_intrinsic("pow", in_elem, in_ty, in_len, bx, span, args);
1484         }
1485         "simd_fma" => {
1486             return simd_simple_float_intrinsic("fma", in_elem, in_ty, in_len, bx, span, args);
1487         }
1488         _ => { /* fallthrough */ }
1489     }
1490
1491     // FIXME: use:
1492     //  https://github.com/llvm-mirror/llvm/blob/master/include/llvm/IR/Function.h#L182
1493     //  https://github.com/llvm-mirror/llvm/blob/master/include/llvm/IR/Intrinsics.h#L81
1494     fn llvm_vector_str(elem_ty: Ty<'_>, vec_len: u64, no_pointers: usize) -> String {
1495         let p0s: String = "p0".repeat(no_pointers);
1496         match elem_ty.kind {
1497             ty::Int(v) => format!("v{}{}i{}", vec_len, p0s, v.bit_width().unwrap()),
1498             ty::Uint(v) => format!("v{}{}i{}", vec_len, p0s, v.bit_width().unwrap()),
1499             ty::Float(v) => format!("v{}{}f{}", vec_len, p0s, v.bit_width()),
1500             _ => unreachable!(),
1501         }
1502     }
1503
1504     fn llvm_vector_ty(
1505         cx: &CodegenCx<'ll, '_>,
1506         elem_ty: Ty<'_>,
1507         vec_len: u64,
1508         mut no_pointers: usize,
1509     ) -> &'ll Type {
1510         // FIXME: use cx.layout_of(ty).llvm_type() ?
1511         let mut elem_ty = match elem_ty.kind {
1512             ty::Int(v) => cx.type_int_from_ty(v),
1513             ty::Uint(v) => cx.type_uint_from_ty(v),
1514             ty::Float(v) => cx.type_float_from_ty(v),
1515             _ => unreachable!(),
1516         };
1517         while no_pointers > 0 {
1518             elem_ty = cx.type_ptr_to(elem_ty);
1519             no_pointers -= 1;
1520         }
1521         cx.type_vector(elem_ty, vec_len)
1522     }
1523
1524     if name == "simd_gather" {
1525         // simd_gather(values: <N x T>, pointers: <N x *_ T>,
1526         //             mask: <N x i{M}>) -> <N x T>
1527         // * N: number of elements in the input vectors
1528         // * T: type of the element to load
1529         // * M: any integer width is supported, will be truncated to i1
1530
1531         // All types must be simd vector types
1532         require_simd!(in_ty, "first");
1533         require_simd!(arg_tys[1], "second");
1534         require_simd!(arg_tys[2], "third");
1535         require_simd!(ret_ty, "return");
1536
1537         // Of the same length:
1538         require!(
1539             in_len == arg_tys[1].simd_size(tcx),
1540             "expected {} argument with length {} (same as input type `{}`), \
1541                   found `{}` with length {}",
1542             "second",
1543             in_len,
1544             in_ty,
1545             arg_tys[1],
1546             arg_tys[1].simd_size(tcx)
1547         );
1548         require!(
1549             in_len == arg_tys[2].simd_size(tcx),
1550             "expected {} argument with length {} (same as input type `{}`), \
1551                   found `{}` with length {}",
1552             "third",
1553             in_len,
1554             in_ty,
1555             arg_tys[2],
1556             arg_tys[2].simd_size(tcx)
1557         );
1558
1559         // The return type must match the first argument type
1560         require!(ret_ty == in_ty, "expected return type `{}`, found `{}`", in_ty, ret_ty);
1561
1562         // This counts how many pointers
1563         fn ptr_count(t: Ty<'_>) -> usize {
1564             match t.kind {
1565                 ty::RawPtr(p) => 1 + ptr_count(p.ty),
1566                 _ => 0,
1567             }
1568         }
1569
1570         // Non-ptr type
1571         fn non_ptr(t: Ty<'_>) -> Ty<'_> {
1572             match t.kind {
1573                 ty::RawPtr(p) => non_ptr(p.ty),
1574                 _ => t,
1575             }
1576         }
1577
1578         // The second argument must be a simd vector with an element type that's a pointer
1579         // to the element type of the first argument
1580         let (pointer_count, underlying_ty) = match arg_tys[1].simd_type(tcx).kind {
1581             ty::RawPtr(p) if p.ty == in_elem => {
1582                 (ptr_count(arg_tys[1].simd_type(tcx)), non_ptr(arg_tys[1].simd_type(tcx)))
1583             }
1584             _ => {
1585                 require!(
1586                     false,
1587                     "expected element type `{}` of second argument `{}` \
1588                                  to be a pointer to the element type `{}` of the first \
1589                                  argument `{}`, found `{}` != `*_ {}`",
1590                     arg_tys[1].simd_type(tcx),
1591                     arg_tys[1],
1592                     in_elem,
1593                     in_ty,
1594                     arg_tys[1].simd_type(tcx),
1595                     in_elem
1596                 );
1597                 unreachable!();
1598             }
1599         };
1600         assert!(pointer_count > 0);
1601         assert_eq!(pointer_count - 1, ptr_count(arg_tys[0].simd_type(tcx)));
1602         assert_eq!(underlying_ty, non_ptr(arg_tys[0].simd_type(tcx)));
1603
1604         // The element type of the third argument must be a signed integer type of any width:
1605         match arg_tys[2].simd_type(tcx).kind {
1606             ty::Int(_) => (),
1607             _ => {
1608                 require!(
1609                     false,
1610                     "expected element type `{}` of third argument `{}` \
1611                                  to be a signed integer type",
1612                     arg_tys[2].simd_type(tcx),
1613                     arg_tys[2]
1614                 );
1615             }
1616         }
1617
1618         // Alignment of T, must be a constant integer value:
1619         let alignment_ty = bx.type_i32();
1620         let alignment = bx.const_i32(bx.align_of(in_elem).bytes() as i32);
1621
1622         // Truncate the mask vector to a vector of i1s:
1623         let (mask, mask_ty) = {
1624             let i1 = bx.type_i1();
1625             let i1xn = bx.type_vector(i1, in_len);
1626             (bx.trunc(args[2].immediate(), i1xn), i1xn)
1627         };
1628
1629         // Type of the vector of pointers:
1630         let llvm_pointer_vec_ty = llvm_vector_ty(bx, underlying_ty, in_len, pointer_count);
1631         let llvm_pointer_vec_str = llvm_vector_str(underlying_ty, in_len, pointer_count);
1632
1633         // Type of the vector of elements:
1634         let llvm_elem_vec_ty = llvm_vector_ty(bx, underlying_ty, in_len, pointer_count - 1);
1635         let llvm_elem_vec_str = llvm_vector_str(underlying_ty, in_len, pointer_count - 1);
1636
1637         let llvm_intrinsic =
1638             format!("llvm.masked.gather.{}.{}", llvm_elem_vec_str, llvm_pointer_vec_str);
1639         let f = bx.declare_cfn(
1640             &llvm_intrinsic,
1641             bx.type_func(
1642                 &[llvm_pointer_vec_ty, alignment_ty, mask_ty, llvm_elem_vec_ty],
1643                 llvm_elem_vec_ty,
1644             ),
1645         );
1646         llvm::SetUnnamedAddr(f, false);
1647         let v = bx.call(f, &[args[1].immediate(), alignment, mask, args[0].immediate()], None);
1648         return Ok(v);
1649     }
1650
1651     if name == "simd_scatter" {
1652         // simd_scatter(values: <N x T>, pointers: <N x *mut T>,
1653         //             mask: <N x i{M}>) -> ()
1654         // * N: number of elements in the input vectors
1655         // * T: type of the element to load
1656         // * M: any integer width is supported, will be truncated to i1
1657
1658         // All types must be simd vector types
1659         require_simd!(in_ty, "first");
1660         require_simd!(arg_tys[1], "second");
1661         require_simd!(arg_tys[2], "third");
1662
1663         // Of the same length:
1664         require!(
1665             in_len == arg_tys[1].simd_size(tcx),
1666             "expected {} argument with length {} (same as input type `{}`), \
1667                   found `{}` with length {}",
1668             "second",
1669             in_len,
1670             in_ty,
1671             arg_tys[1],
1672             arg_tys[1].simd_size(tcx)
1673         );
1674         require!(
1675             in_len == arg_tys[2].simd_size(tcx),
1676             "expected {} argument with length {} (same as input type `{}`), \
1677                   found `{}` with length {}",
1678             "third",
1679             in_len,
1680             in_ty,
1681             arg_tys[2],
1682             arg_tys[2].simd_size(tcx)
1683         );
1684
1685         // This counts how many pointers
1686         fn ptr_count(t: Ty<'_>) -> usize {
1687             match t.kind {
1688                 ty::RawPtr(p) => 1 + ptr_count(p.ty),
1689                 _ => 0,
1690             }
1691         }
1692
1693         // Non-ptr type
1694         fn non_ptr(t: Ty<'_>) -> Ty<'_> {
1695             match t.kind {
1696                 ty::RawPtr(p) => non_ptr(p.ty),
1697                 _ => t,
1698             }
1699         }
1700
1701         // The second argument must be a simd vector with an element type that's a pointer
1702         // to the element type of the first argument
1703         let (pointer_count, underlying_ty) = match arg_tys[1].simd_type(tcx).kind {
1704             ty::RawPtr(p) if p.ty == in_elem && p.mutbl == hir::Mutability::Mut => {
1705                 (ptr_count(arg_tys[1].simd_type(tcx)), non_ptr(arg_tys[1].simd_type(tcx)))
1706             }
1707             _ => {
1708                 require!(
1709                     false,
1710                     "expected element type `{}` of second argument `{}` \
1711                                  to be a pointer to the element type `{}` of the first \
1712                                  argument `{}`, found `{}` != `*mut {}`",
1713                     arg_tys[1].simd_type(tcx),
1714                     arg_tys[1],
1715                     in_elem,
1716                     in_ty,
1717                     arg_tys[1].simd_type(tcx),
1718                     in_elem
1719                 );
1720                 unreachable!();
1721             }
1722         };
1723         assert!(pointer_count > 0);
1724         assert_eq!(pointer_count - 1, ptr_count(arg_tys[0].simd_type(tcx)));
1725         assert_eq!(underlying_ty, non_ptr(arg_tys[0].simd_type(tcx)));
1726
1727         // The element type of the third argument must be a signed integer type of any width:
1728         match arg_tys[2].simd_type(tcx).kind {
1729             ty::Int(_) => (),
1730             _ => {
1731                 require!(
1732                     false,
1733                     "expected element type `{}` of third argument `{}` \
1734                                  to be a signed integer type",
1735                     arg_tys[2].simd_type(tcx),
1736                     arg_tys[2]
1737                 );
1738             }
1739         }
1740
1741         // Alignment of T, must be a constant integer value:
1742         let alignment_ty = bx.type_i32();
1743         let alignment = bx.const_i32(bx.align_of(in_elem).bytes() as i32);
1744
1745         // Truncate the mask vector to a vector of i1s:
1746         let (mask, mask_ty) = {
1747             let i1 = bx.type_i1();
1748             let i1xn = bx.type_vector(i1, in_len);
1749             (bx.trunc(args[2].immediate(), i1xn), i1xn)
1750         };
1751
1752         let ret_t = bx.type_void();
1753
1754         // Type of the vector of pointers:
1755         let llvm_pointer_vec_ty = llvm_vector_ty(bx, underlying_ty, in_len, pointer_count);
1756         let llvm_pointer_vec_str = llvm_vector_str(underlying_ty, in_len, pointer_count);
1757
1758         // Type of the vector of elements:
1759         let llvm_elem_vec_ty = llvm_vector_ty(bx, underlying_ty, in_len, pointer_count - 1);
1760         let llvm_elem_vec_str = llvm_vector_str(underlying_ty, in_len, pointer_count - 1);
1761
1762         let llvm_intrinsic =
1763             format!("llvm.masked.scatter.{}.{}", llvm_elem_vec_str, llvm_pointer_vec_str);
1764         let f = bx.declare_cfn(
1765             &llvm_intrinsic,
1766             bx.type_func(&[llvm_elem_vec_ty, llvm_pointer_vec_ty, alignment_ty, mask_ty], ret_t),
1767         );
1768         llvm::SetUnnamedAddr(f, false);
1769         let v = bx.call(f, &[args[0].immediate(), args[1].immediate(), alignment, mask], None);
1770         return Ok(v);
1771     }
1772
1773     macro_rules! arith_red {
1774         ($name:tt : $integer_reduce:ident, $float_reduce:ident, $ordered:expr) => {
1775             if name == $name {
1776                 require!(
1777                     ret_ty == in_elem,
1778                     "expected return type `{}` (element of input `{}`), found `{}`",
1779                     in_elem,
1780                     in_ty,
1781                     ret_ty
1782                 );
1783                 return match in_elem.kind {
1784                     ty::Int(_) | ty::Uint(_) => {
1785                         let r = bx.$integer_reduce(args[0].immediate());
1786                         if $ordered {
1787                             // if overflow occurs, the result is the
1788                             // mathematical result modulo 2^n:
1789                             if name.contains("mul") {
1790                                 Ok(bx.mul(args[1].immediate(), r))
1791                             } else {
1792                                 Ok(bx.add(args[1].immediate(), r))
1793                             }
1794                         } else {
1795                             Ok(bx.$integer_reduce(args[0].immediate()))
1796                         }
1797                     }
1798                     ty::Float(f) => {
1799                         let acc = if $ordered {
1800                             // ordered arithmetic reductions take an accumulator
1801                             args[1].immediate()
1802                         } else {
1803                             // unordered arithmetic reductions use the identity accumulator
1804                             let identity_acc = if $name.contains("mul") { 1.0 } else { 0.0 };
1805                             match f.bit_width() {
1806                                 32 => bx.const_real(bx.type_f32(), identity_acc),
1807                                 64 => bx.const_real(bx.type_f64(), identity_acc),
1808                                 v => return_error!(
1809                                     r#"
1810 unsupported {} from `{}` with element `{}` of size `{}` to `{}`"#,
1811                                     $name,
1812                                     in_ty,
1813                                     in_elem,
1814                                     v,
1815                                     ret_ty
1816                                 ),
1817                             }
1818                         };
1819                         Ok(bx.$float_reduce(acc, args[0].immediate()))
1820                     }
1821                     _ => return_error!(
1822                         "unsupported {} from `{}` with element `{}` to `{}`",
1823                         $name,
1824                         in_ty,
1825                         in_elem,
1826                         ret_ty
1827                     ),
1828                 };
1829             }
1830         };
1831     }
1832
1833     arith_red!("simd_reduce_add_ordered": vector_reduce_add, vector_reduce_fadd, true);
1834     arith_red!("simd_reduce_mul_ordered": vector_reduce_mul, vector_reduce_fmul, true);
1835     arith_red!("simd_reduce_add_unordered": vector_reduce_add, vector_reduce_fadd_fast, false);
1836     arith_red!("simd_reduce_mul_unordered": vector_reduce_mul, vector_reduce_fmul_fast, false);
1837
1838     macro_rules! minmax_red {
1839         ($name:tt: $int_red:ident, $float_red:ident) => {
1840             if name == $name {
1841                 require!(
1842                     ret_ty == in_elem,
1843                     "expected return type `{}` (element of input `{}`), found `{}`",
1844                     in_elem,
1845                     in_ty,
1846                     ret_ty
1847                 );
1848                 return match in_elem.kind {
1849                     ty::Int(_i) => Ok(bx.$int_red(args[0].immediate(), true)),
1850                     ty::Uint(_u) => Ok(bx.$int_red(args[0].immediate(), false)),
1851                     ty::Float(_f) => Ok(bx.$float_red(args[0].immediate())),
1852                     _ => return_error!(
1853                         "unsupported {} from `{}` with element `{}` to `{}`",
1854                         $name,
1855                         in_ty,
1856                         in_elem,
1857                         ret_ty
1858                     ),
1859                 };
1860             }
1861         };
1862     }
1863
1864     minmax_red!("simd_reduce_min": vector_reduce_min, vector_reduce_fmin);
1865     minmax_red!("simd_reduce_max": vector_reduce_max, vector_reduce_fmax);
1866
1867     minmax_red!("simd_reduce_min_nanless": vector_reduce_min, vector_reduce_fmin_fast);
1868     minmax_red!("simd_reduce_max_nanless": vector_reduce_max, vector_reduce_fmax_fast);
1869
1870     macro_rules! bitwise_red {
1871         ($name:tt : $red:ident, $boolean:expr) => {
1872             if name == $name {
1873                 let input = if !$boolean {
1874                     require!(
1875                         ret_ty == in_elem,
1876                         "expected return type `{}` (element of input `{}`), found `{}`",
1877                         in_elem,
1878                         in_ty,
1879                         ret_ty
1880                     );
1881                     args[0].immediate()
1882                 } else {
1883                     match in_elem.kind {
1884                         ty::Int(_) | ty::Uint(_) => {}
1885                         _ => return_error!(
1886                             "unsupported {} from `{}` with element `{}` to `{}`",
1887                             $name,
1888                             in_ty,
1889                             in_elem,
1890                             ret_ty
1891                         ),
1892                     }
1893
1894                     // boolean reductions operate on vectors of i1s:
1895                     let i1 = bx.type_i1();
1896                     let i1xn = bx.type_vector(i1, in_len as u64);
1897                     bx.trunc(args[0].immediate(), i1xn)
1898                 };
1899                 return match in_elem.kind {
1900                     ty::Int(_) | ty::Uint(_) => {
1901                         let r = bx.$red(input);
1902                         Ok(if !$boolean { r } else { bx.zext(r, bx.type_bool()) })
1903                     }
1904                     _ => return_error!(
1905                         "unsupported {} from `{}` with element `{}` to `{}`",
1906                         $name,
1907                         in_ty,
1908                         in_elem,
1909                         ret_ty
1910                     ),
1911                 };
1912             }
1913         };
1914     }
1915
1916     bitwise_red!("simd_reduce_and": vector_reduce_and, false);
1917     bitwise_red!("simd_reduce_or": vector_reduce_or, false);
1918     bitwise_red!("simd_reduce_xor": vector_reduce_xor, false);
1919     bitwise_red!("simd_reduce_all": vector_reduce_and, true);
1920     bitwise_red!("simd_reduce_any": vector_reduce_or, true);
1921
1922     if name == "simd_cast" {
1923         require_simd!(ret_ty, "return");
1924         let out_len = ret_ty.simd_size(tcx);
1925         require!(
1926             in_len == out_len,
1927             "expected return type with length {} (same as input type `{}`), \
1928                   found `{}` with length {}",
1929             in_len,
1930             in_ty,
1931             ret_ty,
1932             out_len
1933         );
1934         // casting cares about nominal type, not just structural type
1935         let out_elem = ret_ty.simd_type(tcx);
1936
1937         if in_elem == out_elem {
1938             return Ok(args[0].immediate());
1939         }
1940
1941         enum Style {
1942             Float,
1943             Int(/* is signed? */ bool),
1944             Unsupported,
1945         }
1946
1947         let (in_style, in_width) = match in_elem.kind {
1948             // vectors of pointer-sized integers should've been
1949             // disallowed before here, so this unwrap is safe.
1950             ty::Int(i) => (Style::Int(true), i.bit_width().unwrap()),
1951             ty::Uint(u) => (Style::Int(false), u.bit_width().unwrap()),
1952             ty::Float(f) => (Style::Float, f.bit_width()),
1953             _ => (Style::Unsupported, 0),
1954         };
1955         let (out_style, out_width) = match out_elem.kind {
1956             ty::Int(i) => (Style::Int(true), i.bit_width().unwrap()),
1957             ty::Uint(u) => (Style::Int(false), u.bit_width().unwrap()),
1958             ty::Float(f) => (Style::Float, f.bit_width()),
1959             _ => (Style::Unsupported, 0),
1960         };
1961
1962         match (in_style, out_style) {
1963             (Style::Int(in_is_signed), Style::Int(_)) => {
1964                 return Ok(match in_width.cmp(&out_width) {
1965                     Ordering::Greater => bx.trunc(args[0].immediate(), llret_ty),
1966                     Ordering::Equal => args[0].immediate(),
1967                     Ordering::Less => {
1968                         if in_is_signed {
1969                             bx.sext(args[0].immediate(), llret_ty)
1970                         } else {
1971                             bx.zext(args[0].immediate(), llret_ty)
1972                         }
1973                     }
1974                 });
1975             }
1976             (Style::Int(in_is_signed), Style::Float) => {
1977                 return Ok(if in_is_signed {
1978                     bx.sitofp(args[0].immediate(), llret_ty)
1979                 } else {
1980                     bx.uitofp(args[0].immediate(), llret_ty)
1981                 });
1982             }
1983             (Style::Float, Style::Int(out_is_signed)) => {
1984                 return Ok(if out_is_signed {
1985                     bx.fptosi(args[0].immediate(), llret_ty)
1986                 } else {
1987                     bx.fptoui(args[0].immediate(), llret_ty)
1988                 });
1989             }
1990             (Style::Float, Style::Float) => {
1991                 return Ok(match in_width.cmp(&out_width) {
1992                     Ordering::Greater => bx.fptrunc(args[0].immediate(), llret_ty),
1993                     Ordering::Equal => args[0].immediate(),
1994                     Ordering::Less => bx.fpext(args[0].immediate(), llret_ty),
1995                 });
1996             }
1997             _ => { /* Unsupported. Fallthrough. */ }
1998         }
1999         require!(
2000             false,
2001             "unsupported cast from `{}` with element `{}` to `{}` with element `{}`",
2002             in_ty,
2003             in_elem,
2004             ret_ty,
2005             out_elem
2006         );
2007     }
2008     macro_rules! arith {
2009         ($($name: ident: $($($p: ident),* => $call: ident),*;)*) => {
2010             $(if name == stringify!($name) {
2011                 match in_elem.kind {
2012                     $($(ty::$p(_))|* => {
2013                         return Ok(bx.$call(args[0].immediate(), args[1].immediate()))
2014                     })*
2015                     _ => {},
2016                 }
2017                 require!(false,
2018                          "unsupported operation on `{}` with element `{}`",
2019                          in_ty,
2020                          in_elem)
2021             })*
2022         }
2023     }
2024     arith! {
2025         simd_add: Uint, Int => add, Float => fadd;
2026         simd_sub: Uint, Int => sub, Float => fsub;
2027         simd_mul: Uint, Int => mul, Float => fmul;
2028         simd_div: Uint => udiv, Int => sdiv, Float => fdiv;
2029         simd_rem: Uint => urem, Int => srem, Float => frem;
2030         simd_shl: Uint, Int => shl;
2031         simd_shr: Uint => lshr, Int => ashr;
2032         simd_and: Uint, Int => and;
2033         simd_or: Uint, Int => or;
2034         simd_xor: Uint, Int => xor;
2035         simd_fmax: Float => maxnum;
2036         simd_fmin: Float => minnum;
2037
2038     }
2039
2040     if name == "simd_saturating_add" || name == "simd_saturating_sub" {
2041         let lhs = args[0].immediate();
2042         let rhs = args[1].immediate();
2043         let is_add = name == "simd_saturating_add";
2044         let ptr_bits = bx.tcx().data_layout.pointer_size.bits() as _;
2045         let (signed, elem_width, elem_ty) = match in_elem.kind {
2046             ty::Int(i) => (true, i.bit_width().unwrap_or(ptr_bits), bx.cx.type_int_from_ty(i)),
2047             ty::Uint(i) => (false, i.bit_width().unwrap_or(ptr_bits), bx.cx.type_uint_from_ty(i)),
2048             _ => {
2049                 return_error!(
2050                     "expected element type `{}` of vector type `{}` \
2051                      to be a signed or unsigned integer type",
2052                     arg_tys[0].simd_type(tcx),
2053                     arg_tys[0]
2054                 );
2055             }
2056         };
2057         let llvm_intrinsic = &format!(
2058             "llvm.{}{}.sat.v{}i{}",
2059             if signed { 's' } else { 'u' },
2060             if is_add { "add" } else { "sub" },
2061             in_len,
2062             elem_width
2063         );
2064         let vec_ty = bx.cx.type_vector(elem_ty, in_len as u64);
2065
2066         let f = bx.declare_cfn(&llvm_intrinsic, bx.type_func(&[vec_ty, vec_ty], vec_ty));
2067         llvm::SetUnnamedAddr(f, false);
2068         let v = bx.call(f, &[lhs, rhs], None);
2069         return Ok(v);
2070     }
2071
2072     span_bug!(span, "unknown SIMD intrinsic");
2073 }
2074
2075 // Returns the width of an int Ty, and if it's signed or not
2076 // Returns None if the type is not an integer
2077 // FIXME: there’s multiple of this functions, investigate using some of the already existing
2078 // stuffs.
2079 fn int_type_width_signed(ty: Ty<'_>, cx: &CodegenCx<'_, '_>) -> Option<(u64, bool)> {
2080     match ty.kind {
2081         ty::Int(t) => Some((
2082             match t {
2083                 ast::IntTy::Isize => cx.tcx.sess.target.ptr_width as u64,
2084                 ast::IntTy::I8 => 8,
2085                 ast::IntTy::I16 => 16,
2086                 ast::IntTy::I32 => 32,
2087                 ast::IntTy::I64 => 64,
2088                 ast::IntTy::I128 => 128,
2089             },
2090             true,
2091         )),
2092         ty::Uint(t) => Some((
2093             match t {
2094                 ast::UintTy::Usize => cx.tcx.sess.target.ptr_width as u64,
2095                 ast::UintTy::U8 => 8,
2096                 ast::UintTy::U16 => 16,
2097                 ast::UintTy::U32 => 32,
2098                 ast::UintTy::U64 => 64,
2099                 ast::UintTy::U128 => 128,
2100             },
2101             false,
2102         )),
2103         _ => None,
2104     }
2105 }
2106
2107 // Returns the width of a float Ty
2108 // Returns None if the type is not a float
2109 fn float_type_width(ty: Ty<'_>) -> Option<u64> {
2110     match ty.kind {
2111         ty::Float(t) => Some(t.bit_width() as u64),
2112         _ => None,
2113     }
2114 }