compiler/rustc_codegen_llvm/src/intrinsic.rs

   1 use crate::abi::{Abi, FnAbi, FnAbiLlvmExt, LlvmType, PassMode};
   2 use crate::builder::Builder;
   3 use crate::context::CodegenCx;
   4 use crate::llvm;
   5 use crate::type_::Type;
   6 use crate::type_of::LayoutLlvmExt;
   7 use crate::va_arg::emit_va_arg;
   8 use crate::value::Value;
   9
  10 use rustc_ast as ast;
  11 use rustc_codegen_ssa::base::{compare_simd_types, wants_msvc_seh};
  12 use rustc_codegen_ssa::common::span_invalid_monomorphization_error;
  13 use rustc_codegen_ssa::common::{IntPredicate, TypeKind};
  14 use rustc_codegen_ssa::mir::operand::OperandRef;
  15 use rustc_codegen_ssa::mir::place::PlaceRef;
  16 use rustc_codegen_ssa::traits::*;
  17 use rustc_hir as hir;
  18 use rustc_middle::ty::layout::{FnAbiOf, HasTyCtxt, LayoutOf};
  19 use rustc_middle::ty::{self, Ty};
  20 use rustc_middle::{bug, span_bug};
  21 use rustc_span::{sym, symbol::kw, Span, Symbol};
  22 use rustc_target::abi::{self, Align, HasDataLayout, Primitive};
  23 use rustc_target::spec::{HasTargetSpec, PanicStrategy};
  24
  25 use std::cmp::Ordering;
  26 use std::iter;
  27
  28 fn get_simple_intrinsic(cx: &CodegenCx<'ll, '_>, name: Symbol) -> Option<(&'ll Type, &'ll Value)> {
  29     let llvm_name = match name {
  30         sym::sqrtf32 => "llvm.sqrt.f32",
  31         sym::sqrtf64 => "llvm.sqrt.f64",
  32         sym::powif32 => "llvm.powi.f32",
  33         sym::powif64 => "llvm.powi.f64",
  34         sym::sinf32 => "llvm.sin.f32",
  35         sym::sinf64 => "llvm.sin.f64",
  36         sym::cosf32 => "llvm.cos.f32",
  37         sym::cosf64 => "llvm.cos.f64",
  38         sym::powf32 => "llvm.pow.f32",
  39         sym::powf64 => "llvm.pow.f64",
  40         sym::expf32 => "llvm.exp.f32",
  41         sym::expf64 => "llvm.exp.f64",
  42         sym::exp2f32 => "llvm.exp2.f32",
  43         sym::exp2f64 => "llvm.exp2.f64",
  44         sym::logf32 => "llvm.log.f32",
  45         sym::logf64 => "llvm.log.f64",
  46         sym::log10f32 => "llvm.log10.f32",
  47         sym::log10f64 => "llvm.log10.f64",
  48         sym::log2f32 => "llvm.log2.f32",
  49         sym::log2f64 => "llvm.log2.f64",
  50         sym::fmaf32 => "llvm.fma.f32",
  51         sym::fmaf64 => "llvm.fma.f64",
  52         sym::fabsf32 => "llvm.fabs.f32",
  53         sym::fabsf64 => "llvm.fabs.f64",
  54         sym::minnumf32 => "llvm.minnum.f32",
  55         sym::minnumf64 => "llvm.minnum.f64",
  56         sym::maxnumf32 => "llvm.maxnum.f32",
  57         sym::maxnumf64 => "llvm.maxnum.f64",
  58         sym::copysignf32 => "llvm.copysign.f32",
  59         sym::copysignf64 => "llvm.copysign.f64",
  60         sym::floorf32 => "llvm.floor.f32",
  61         sym::floorf64 => "llvm.floor.f64",
  62         sym::ceilf32 => "llvm.ceil.f32",
  63         sym::ceilf64 => "llvm.ceil.f64",
  64         sym::truncf32 => "llvm.trunc.f32",
  65         sym::truncf64 => "llvm.trunc.f64",
  66         sym::rintf32 => "llvm.rint.f32",
  67         sym::rintf64 => "llvm.rint.f64",
  68         sym::nearbyintf32 => "llvm.nearbyint.f32",
  69         sym::nearbyintf64 => "llvm.nearbyint.f64",
  70         sym::roundf32 => "llvm.round.f32",
  71         sym::roundf64 => "llvm.round.f64",
  72         _ => return None,
  73     };
  74     Some(cx.get_intrinsic(llvm_name))
  75 }
  76
  77 impl IntrinsicCallMethods<'tcx> for Builder<'a, 'll, 'tcx> {
  78     fn codegen_intrinsic_call(
  79         &mut self,
  80         instance: ty::Instance<'tcx>,
  81         fn_abi: &FnAbi<'tcx, Ty<'tcx>>,
  82         args: &[OperandRef<'tcx, &'ll Value>],
  83         llresult: &'ll Value,
  84         span: Span,
  85     ) {
  86         let tcx = self.tcx;
  87         let callee_ty = instance.ty(tcx, ty::ParamEnv::reveal_all());
  88
  89         let (def_id, substs) = match *callee_ty.kind() {
  90             ty::FnDef(def_id, substs) => (def_id, substs),
  91             _ => bug!("expected fn item type, found {}", callee_ty),
  92         };
  93
  94         let sig = callee_ty.fn_sig(tcx);
  95         let sig = tcx.normalize_erasing_late_bound_regions(ty::ParamEnv::reveal_all(), sig);
  96         let arg_tys = sig.inputs();
  97         let ret_ty = sig.output();
  98         let name = tcx.item_name(def_id);
  99
 100         let llret_ty = self.layout_of(ret_ty).llvm_type(self);
 101         let result = PlaceRef::new_sized(llresult, fn_abi.ret.layout);
 102
 103         let simple = get_simple_intrinsic(self, name);
 104         let llval = match name {
 105             _ if simple.is_some() => {
 106                 let (simple_ty, simple_fn) = simple.unwrap();
 107                 self.call(
 108                     simple_ty,
 109                     simple_fn,
 110                     &args.iter().map(|arg| arg.immediate()).collect::<Vec<_>>(),
 111                     None,
 112                 )
 113             }
 114             sym::likely => {
 115                 self.call_intrinsic("llvm.expect.i1", &[args[0].immediate(), self.const_bool(true)])
 116             }
 117             sym::unlikely => self
 118                 .call_intrinsic("llvm.expect.i1", &[args[0].immediate(), self.const_bool(false)]),
 119             kw::Try => {
 120                 try_intrinsic(
 121                     self,
 122                     args[0].immediate(),
 123                     args[1].immediate(),
 124                     args[2].immediate(),
 125                     llresult,
 126                 );
 127                 return;
 128             }
 129             sym::breakpoint => self.call_intrinsic("llvm.debugtrap", &[]),
 130             sym::va_copy => {
 131                 self.call_intrinsic("llvm.va_copy", &[args[0].immediate(), args[1].immediate()])
 132             }
 133             sym::va_arg => {
 134                 match fn_abi.ret.layout.abi {
 135                     abi::Abi::Scalar(scalar) => {
 136                         match scalar.value {
 137                             Primitive::Int(..) => {
 138                                 if self.cx().size_of(ret_ty).bytes() < 4 {
 139                                     // `va_arg` should not be called on an integer type
 140                                     // less than 4 bytes in length. If it is, promote
 141                                     // the integer to an `i32` and truncate the result
 142                                     // back to the smaller type.
 143                                     let promoted_result = emit_va_arg(self, args[0], tcx.types.i32);
 144                                     self.trunc(promoted_result, llret_ty)
 145                                 } else {
 146                                     emit_va_arg(self, args[0], ret_ty)
 147                                 }
 148                             }
 149                             Primitive::F64 | Primitive::Pointer => {
 150                                 emit_va_arg(self, args[0], ret_ty)
 151                             }
 152                             // `va_arg` should never be used with the return type f32.
 153                             Primitive::F32 => bug!("the va_arg intrinsic does not work with `f32`"),
 154                         }
 155                     }
 156                     _ => bug!("the va_arg intrinsic does not work with non-scalar types"),
 157                 }
 158             }
 159
 160             sym::volatile_load | sym::unaligned_volatile_load => {
 161                 let tp_ty = substs.type_at(0);
 162                 let ptr = args[0].immediate();
 163                 let load = if let PassMode::Cast(ty) = fn_abi.ret.mode {
 164                     let llty = ty.llvm_type(self);
 165                     let ptr = self.pointercast(ptr, self.type_ptr_to(llty));
 166                     self.volatile_load(llty, ptr)
 167                 } else {
 168                     self.volatile_load(self.layout_of(tp_ty).llvm_type(self), ptr)
 169                 };
 170                 let align = if name == sym::unaligned_volatile_load {
 171                     1
 172                 } else {
 173                     self.align_of(tp_ty).bytes() as u32
 174                 };
 175                 unsafe {
 176                     llvm::LLVMSetAlignment(load, align);
 177                 }
 178                 self.to_immediate(load, self.layout_of(tp_ty))
 179             }
 180             sym::volatile_store => {
 181                 let dst = args[0].deref(self.cx());
 182                 args[1].val.volatile_store(self, dst);
 183                 return;
 184             }
 185             sym::unaligned_volatile_store => {
 186                 let dst = args[0].deref(self.cx());
 187                 args[1].val.unaligned_volatile_store(self, dst);
 188                 return;
 189             }
 190             sym::prefetch_read_data
 191             | sym::prefetch_write_data
 192             | sym::prefetch_read_instruction
 193             | sym::prefetch_write_instruction => {
 194                 let (rw, cache_type) = match name {
 195                     sym::prefetch_read_data => (0, 1),
 196                     sym::prefetch_write_data => (1, 1),
 197                     sym::prefetch_read_instruction => (0, 0),
 198                     sym::prefetch_write_instruction => (1, 0),
 199                     _ => bug!(),
 200                 };
 201                 self.call_intrinsic(
 202                     "llvm.prefetch",
 203                     &[
 204                         args[0].immediate(),
 205                         self.const_i32(rw),
 206                         args[1].immediate(),
 207                         self.const_i32(cache_type),
 208                     ],
 209                 )
 210             }
 211             sym::ctlz
 212             | sym::ctlz_nonzero
 213             | sym::cttz
 214             | sym::cttz_nonzero
 215             | sym::ctpop
 216             | sym::bswap
 217             | sym::bitreverse
 218             | sym::rotate_left
 219             | sym::rotate_right
 220             | sym::saturating_add
 221             | sym::saturating_sub => {
 222                 let ty = arg_tys[0];
 223                 match int_type_width_signed(ty, self) {
 224                     Some((width, signed)) => match name {
 225                         sym::ctlz | sym::cttz => {
 226                             let y = self.const_bool(false);
 227                             self.call_intrinsic(
 228                                 &format!("llvm.{}.i{}", name, width),
 229                                 &[args[0].immediate(), y],
 230                             )
 231                         }
 232                         sym::ctlz_nonzero => {
 233                             let y = self.const_bool(true);
 234                             let llvm_name = &format!("llvm.ctlz.i{}", width);
 235                             self.call_intrinsic(llvm_name, &[args[0].immediate(), y])
 236                         }
 237                         sym::cttz_nonzero => {
 238                             let y = self.const_bool(true);
 239                             let llvm_name = &format!("llvm.cttz.i{}", width);
 240                             self.call_intrinsic(llvm_name, &[args[0].immediate(), y])
 241                         }
 242                         sym::ctpop => self.call_intrinsic(
 243                             &format!("llvm.ctpop.i{}", width),
 244                             &[args[0].immediate()],
 245                         ),
 246                         sym::bswap => {
 247                             if width == 8 {
 248                                 args[0].immediate() // byte swap a u8/i8 is just a no-op
 249                             } else {
 250                                 self.call_intrinsic(
 251                                     &format!("llvm.bswap.i{}", width),
 252                                     &[args[0].immediate()],
 253                                 )
 254                             }
 255                         }
 256                         sym::bitreverse => self.call_intrinsic(
 257                             &format!("llvm.bitreverse.i{}", width),
 258                             &[args[0].immediate()],
 259                         ),
 260                         sym::rotate_left | sym::rotate_right => {
 261                             let is_left = name == sym::rotate_left;
 262                             let val = args[0].immediate();
 263                             let raw_shift = args[1].immediate();
 264                             // rotate = funnel shift with first two args the same
 265                             let llvm_name =
 266                                 &format!("llvm.fsh{}.i{}", if is_left { 'l' } else { 'r' }, width);
 267                             self.call_intrinsic(llvm_name, &[val, val, raw_shift])
 268                         }
 269                         sym::saturating_add | sym::saturating_sub => {
 270                             let is_add = name == sym::saturating_add;
 271                             let lhs = args[0].immediate();
 272                             let rhs = args[1].immediate();
 273                             let llvm_name = &format!(
 274                                 "llvm.{}{}.sat.i{}",
 275                                 if signed { 's' } else { 'u' },
 276                                 if is_add { "add" } else { "sub" },
 277                                 width
 278                             );
 279                             self.call_intrinsic(llvm_name, &[lhs, rhs])
 280                         }
 281                         _ => bug!(),
 282                     },
 283                     None => {
 284                         span_invalid_monomorphization_error(
 285                             tcx.sess,
 286                             span,
 287                             &format!(
 288                                 "invalid monomorphization of `{}` intrinsic: \
 289                                       expected basic integer type, found `{}`",
 290                                 name, ty
 291                             ),
 292                         );
 293                         return;
 294                     }
 295                 }
 296             }
 297
 298             sym::raw_eq => {
 299                 use abi::Abi::*;
 300                 let tp_ty = substs.type_at(0);
 301                 let layout = self.layout_of(tp_ty).layout;
 302                 let use_integer_compare = match layout.abi {
 303                     Scalar(_) | ScalarPair(_, _) => true,
 304                     Uninhabited | Vector { .. } => false,
 305                     Aggregate { .. } => {
 306                         // For rusty ABIs, small aggregates are actually passed
 307                         // as `RegKind::Integer` (see `FnAbi::adjust_for_abi`),
 308                         // so we re-use that same threshold here.
 309                         layout.size <= self.data_layout().pointer_size * 2
 310                     }
 311                 };
 312
 313                 let a = args[0].immediate();
 314                 let b = args[1].immediate();
 315                 if layout.size.bytes() == 0 {
 316                     self.const_bool(true)
 317                 } else if use_integer_compare {
 318                     let integer_ty = self.type_ix(layout.size.bits());
 319                     let ptr_ty = self.type_ptr_to(integer_ty);
 320                     let a_ptr = self.bitcast(a, ptr_ty);
 321                     let a_val = self.load(integer_ty, a_ptr, layout.align.abi);
 322                     let b_ptr = self.bitcast(b, ptr_ty);
 323                     let b_val = self.load(integer_ty, b_ptr, layout.align.abi);
 324                     self.icmp(IntPredicate::IntEQ, a_val, b_val)
 325                 } else {
 326                     let i8p_ty = self.type_i8p();
 327                     let a_ptr = self.bitcast(a, i8p_ty);
 328                     let b_ptr = self.bitcast(b, i8p_ty);
 329                     let n = self.const_usize(layout.size.bytes());
 330                     let cmp = self.call_intrinsic("memcmp", &[a_ptr, b_ptr, n]);
 331                     self.icmp(IntPredicate::IntEQ, cmp, self.const_i32(0))
 332                 }
 333             }
 334
 335             sym::black_box => {
 336                 args[0].val.store(self, result);
 337
 338                 // We need to "use" the argument in some way LLVM can't introspect, and on
 339                 // targets that support it we can typically leverage inline assembly to do
 340                 // this. LLVM's interpretation of inline assembly is that it's, well, a black
 341                 // box. This isn't the greatest implementation since it probably deoptimizes
 342                 // more than we want, but it's so far good enough.
 343                 crate::asm::inline_asm_call(
 344                     self,
 345                     "",
 346                     "r,~{memory}",
 347                     &[result.llval],
 348                     self.type_void(),
 349                     true,
 350                     false,
 351                     ast::LlvmAsmDialect::Att,
 352                     &[span],
 353                 )
 354                 .unwrap_or_else(|| bug!("failed to generate inline asm call for `black_box`"));
 355
 356                 // We have copied the value to `result` already.
 357                 return;
 358             }
 359
 360             _ if name.as_str().starts_with("simd_") => {
 361                 match generic_simd_intrinsic(self, name, callee_ty, args, ret_ty, llret_ty, span) {
 362                     Ok(llval) => llval,
 363                     Err(()) => return,
 364                 }
 365             }
 366
 367             _ => bug!("unknown intrinsic '{}'", name),
 368         };
 369
 370         if !fn_abi.ret.is_ignore() {
 371             if let PassMode::Cast(ty) = fn_abi.ret.mode {
 372                 let ptr_llty = self.type_ptr_to(ty.llvm_type(self));
 373                 let ptr = self.pointercast(result.llval, ptr_llty);
 374                 self.store(llval, ptr, result.align);
 375             } else {
 376                 OperandRef::from_immediate_or_packed_pair(self, llval, result.layout)
 377                     .val
 378                     .store(self, result);
 379             }
 380         }
 381     }
 382
 383     fn abort(&mut self) {
 384         self.call_intrinsic("llvm.trap", &[]);
 385     }
 386
 387     fn assume(&mut self, val: Self::Value) {
 388         self.call_intrinsic("llvm.assume", &[val]);
 389     }
 390
 391     fn expect(&mut self, cond: Self::Value, expected: bool) -> Self::Value {
 392         self.call_intrinsic("llvm.expect.i1", &[cond, self.const_bool(expected)])
 393     }
 394
 395     fn type_test(&mut self, pointer: Self::Value, typeid: Self::Value) -> Self::Value {
 396         // Test the called operand using llvm.type.test intrinsic. The LowerTypeTests link-time
 397         // optimization pass replaces calls to this intrinsic with code to test type membership.
 398         let i8p_ty = self.type_i8p();
 399         let bitcast = self.bitcast(pointer, i8p_ty);
 400         self.call_intrinsic("llvm.type.test", &[bitcast, typeid])
 401     }
 402
 403     fn va_start(&mut self, va_list: &'ll Value) -> &'ll Value {
 404         self.call_intrinsic("llvm.va_start", &[va_list])
 405     }
 406
 407     fn va_end(&mut self, va_list: &'ll Value) -> &'ll Value {
 408         self.call_intrinsic("llvm.va_end", &[va_list])
 409     }
 410 }
 411
 412 fn try_intrinsic(
 413     bx: &mut Builder<'a, 'll, 'tcx>,
 414     try_func: &'ll Value,
 415     data: &'ll Value,
 416     catch_func: &'ll Value,
 417     dest: &'ll Value,
 418 ) {
 419     if bx.sess().panic_strategy() == PanicStrategy::Abort {
 420         let try_func_ty = bx.type_func(&[bx.type_i8p()], bx.type_void());
 421         bx.call(try_func_ty, try_func, &[data], None);
 422         // Return 0 unconditionally from the intrinsic call;
 423         // we can never unwind.
 424         let ret_align = bx.tcx().data_layout.i32_align.abi;
 425         bx.store(bx.const_i32(0), dest, ret_align);
 426     } else if wants_msvc_seh(bx.sess()) {
 427         codegen_msvc_try(bx, try_func, data, catch_func, dest);
 428     } else if bx.sess().target.is_like_emscripten {
 429         codegen_emcc_try(bx, try_func, data, catch_func, dest);
 430     } else {
 431         codegen_gnu_try(bx, try_func, data, catch_func, dest);
 432     }
 433 }
 434
 435 // MSVC's definition of the `rust_try` function.
 436 //
 437 // This implementation uses the new exception handling instructions in LLVM
 438 // which have support in LLVM for SEH on MSVC targets. Although these
 439 // instructions are meant to work for all targets, as of the time of this
 440 // writing, however, LLVM does not recommend the usage of these new instructions
 441 // as the old ones are still more optimized.
 442 fn codegen_msvc_try(
 443     bx: &mut Builder<'a, 'll, 'tcx>,
 444     try_func: &'ll Value,
 445     data: &'ll Value,
 446     catch_func: &'ll Value,
 447     dest: &'ll Value,
 448 ) {
 449     let (llty, llfn) = get_rust_try_fn(bx, &mut |mut bx| {
 450         bx.set_personality_fn(bx.eh_personality());
 451
 452         let mut normal = bx.build_sibling_block("normal");
 453         let mut catchswitch = bx.build_sibling_block("catchswitch");
 454         let mut catchpad_rust = bx.build_sibling_block("catchpad_rust");
 455         let mut catchpad_foreign = bx.build_sibling_block("catchpad_foreign");
 456         let mut caught = bx.build_sibling_block("caught");
 457
 458         let try_func = llvm::get_param(bx.llfn(), 0);
 459         let data = llvm::get_param(bx.llfn(), 1);
 460         let catch_func = llvm::get_param(bx.llfn(), 2);
 461
 462         // We're generating an IR snippet that looks like:
 463         //
 464         //   declare i32 @rust_try(%try_func, %data, %catch_func) {
 465         //      %slot = alloca i8*
 466         //      invoke %try_func(%data) to label %normal unwind label %catchswitch
 467         //
 468         //   normal:
 469         //      ret i32 0
 470         //
 471         //   catchswitch:
 472         //      %cs = catchswitch within none [%catchpad_rust, %catchpad_foreign] unwind to caller
 473         //
 474         //   catchpad_rust:
 475         //      %tok = catchpad within %cs [%type_descriptor, 8, %slot]
 476         //      %ptr = load %slot
 477         //      call %catch_func(%data, %ptr)
 478         //      catchret from %tok to label %caught
 479         //
 480         //   catchpad_foreign:
 481         //      %tok = catchpad within %cs [null, 64, null]
 482         //      call %catch_func(%data, null)
 483         //      catchret from %tok to label %caught
 484         //
 485         //   caught:
 486         //      ret i32 1
 487         //   }
 488         //
 489         // This structure follows the basic usage of throw/try/catch in LLVM.
 490         // For example, compile this C++ snippet to see what LLVM generates:
 491         //
 492         //      struct rust_panic {
 493         //          rust_panic(const rust_panic&);
 494         //          ~rust_panic();
 495         //
 496         //          void* x[2];
 497         //      };
 498         //
 499         //      int __rust_try(
 500         //          void (*try_func)(void*),
 501         //          void *data,
 502         //          void (*catch_func)(void*, void*) noexcept
 503         //      ) {
 504         //          try {
 505         //              try_func(data);
 506         //              return 0;
 507         //          } catch(rust_panic& a) {
 508         //              catch_func(data, &a);
 509         //              return 1;
 510         //          } catch(...) {
 511         //              catch_func(data, NULL);
 512         //              return 1;
 513         //          }
 514         //      }
 515         //
 516         // More information can be found in libstd's seh.rs implementation.
 517         let ptr_align = bx.tcx().data_layout.pointer_align.abi;
 518         let slot = bx.alloca(bx.type_i8p(), ptr_align);
 519         let try_func_ty = bx.type_func(&[bx.type_i8p()], bx.type_void());
 520         bx.invoke(try_func_ty, try_func, &[data], normal.llbb(), catchswitch.llbb(), None);
 521
 522         normal.ret(bx.const_i32(0));
 523
 524         let cs = catchswitch.catch_switch(None, None, 2);
 525         catchswitch.add_handler(cs, catchpad_rust.llbb());
 526         catchswitch.add_handler(cs, catchpad_foreign.llbb());
 527
 528         // We can't use the TypeDescriptor defined in libpanic_unwind because it
 529         // might be in another DLL and the SEH encoding only supports specifying
 530         // a TypeDescriptor from the current module.
 531         //
 532         // However this isn't an issue since the MSVC runtime uses string
 533         // comparison on the type name to match TypeDescriptors rather than
 534         // pointer equality.
 535         //
 536         // So instead we generate a new TypeDescriptor in each module that uses
 537         // `try` and let the linker merge duplicate definitions in the same
 538         // module.
 539         //
 540         // When modifying, make sure that the type_name string exactly matches
 541         // the one used in src/libpanic_unwind/seh.rs.
 542         let type_info_vtable = bx.declare_global("??_7type_info@@6B@", bx.type_i8p());
 543         let type_name = bx.const_bytes(b"rust_panic\0");
 544         let type_info =
 545             bx.const_struct(&[type_info_vtable, bx.const_null(bx.type_i8p()), type_name], false);
 546         let tydesc = bx.declare_global("__rust_panic_type_info", bx.val_ty(type_info));
 547         unsafe {
 548             llvm::LLVMRustSetLinkage(tydesc, llvm::Linkage::LinkOnceODRLinkage);
 549             llvm::SetUniqueComdat(bx.llmod, tydesc);
 550             llvm::LLVMSetInitializer(tydesc, type_info);
 551         }
 552
 553         // The flag value of 8 indicates that we are catching the exception by
 554         // reference instead of by value. We can't use catch by value because
 555         // that requires copying the exception object, which we don't support
 556         // since our exception object effectively contains a Box.
 557         //
 558         // Source: MicrosoftCXXABI::getAddrOfCXXCatchHandlerType in clang
 559         let flags = bx.const_i32(8);
 560         let funclet = catchpad_rust.catch_pad(cs, &[tydesc, flags, slot]);
 561         let ptr = catchpad_rust.load(bx.type_i8p(), slot, ptr_align);
 562         let catch_ty = bx.type_func(&[bx.type_i8p(), bx.type_i8p()], bx.type_void());
 563         catchpad_rust.call(catch_ty, catch_func, &[data, ptr], Some(&funclet));
 564         catchpad_rust.catch_ret(&funclet, caught.llbb());
 565
 566         // The flag value of 64 indicates a "catch-all".
 567         let flags = bx.const_i32(64);
 568         let null = bx.const_null(bx.type_i8p());
 569         let funclet = catchpad_foreign.catch_pad(cs, &[null, flags, null]);
 570         catchpad_foreign.call(catch_ty, catch_func, &[data, null], Some(&funclet));
 571         catchpad_foreign.catch_ret(&funclet, caught.llbb());
 572
 573         caught.ret(bx.const_i32(1));
 574     });
 575
 576     // Note that no invoke is used here because by definition this function
 577     // can't panic (that's what it's catching).
 578     let ret = bx.call(llty, llfn, &[try_func, data, catch_func], None);
 579     let i32_align = bx.tcx().data_layout.i32_align.abi;
 580     bx.store(ret, dest, i32_align);
 581 }
 582
 583 // Definition of the standard `try` function for Rust using the GNU-like model
 584 // of exceptions (e.g., the normal semantics of LLVM's `landingpad` and `invoke`
 585 // instructions).
 586 //
 587 // This codegen is a little surprising because we always call a shim
 588 // function instead of inlining the call to `invoke` manually here. This is done
 589 // because in LLVM we're only allowed to have one personality per function
 590 // definition. The call to the `try` intrinsic is being inlined into the
 591 // function calling it, and that function may already have other personality
 592 // functions in play. By calling a shim we're guaranteed that our shim will have
 593 // the right personality function.
 594 fn codegen_gnu_try(
 595     bx: &mut Builder<'a, 'll, 'tcx>,
 596     try_func: &'ll Value,
 597     data: &'ll Value,
 598     catch_func: &'ll Value,
 599     dest: &'ll Value,
 600 ) {
 601     let (llty, llfn) = get_rust_try_fn(bx, &mut |mut bx| {
 602         // Codegens the shims described above:
 603         //
 604         //   bx:
 605         //      invoke %try_func(%data) normal %normal unwind %catch
 606         //
 607         //   normal:
 608         //      ret 0
 609         //
 610         //   catch:
 611         //      (%ptr, _) = landingpad
 612         //      call %catch_func(%data, %ptr)
 613         //      ret 1
 614         let mut then = bx.build_sibling_block("then");
 615         let mut catch = bx.build_sibling_block("catch");
 616
 617         let try_func = llvm::get_param(bx.llfn(), 0);
 618         let data = llvm::get_param(bx.llfn(), 1);
 619         let catch_func = llvm::get_param(bx.llfn(), 2);
 620         let try_func_ty = bx.type_func(&[bx.type_i8p()], bx.type_void());
 621         bx.invoke(try_func_ty, try_func, &[data], then.llbb(), catch.llbb(), None);
 622         then.ret(bx.const_i32(0));
 623
 624         // Type indicator for the exception being thrown.
 625         //
 626         // The first value in this tuple is a pointer to the exception object
 627         // being thrown.  The second value is a "selector" indicating which of
 628         // the landing pad clauses the exception's type had been matched to.
 629         // rust_try ignores the selector.
 630         let lpad_ty = bx.type_struct(&[bx.type_i8p(), bx.type_i32()], false);
 631         let vals = catch.landing_pad(lpad_ty, bx.eh_personality(), 1);
 632         let tydesc = bx.const_null(bx.type_i8p());
 633         catch.add_clause(vals, tydesc);
 634         let ptr = catch.extract_value(vals, 0);
 635         let catch_ty = bx.type_func(&[bx.type_i8p(), bx.type_i8p()], bx.type_void());
 636         catch.call(catch_ty, catch_func, &[data, ptr], None);
 637         catch.ret(bx.const_i32(1));
 638     });
 639
 640     // Note that no invoke is used here because by definition this function
 641     // can't panic (that's what it's catching).
 642     let ret = bx.call(llty, llfn, &[try_func, data, catch_func], None);
 643     let i32_align = bx.tcx().data_layout.i32_align.abi;
 644     bx.store(ret, dest, i32_align);
 645 }
 646
 647 // Variant of codegen_gnu_try used for emscripten where Rust panics are
 648 // implemented using C++ exceptions. Here we use exceptions of a specific type
 649 // (`struct rust_panic`) to represent Rust panics.
 650 fn codegen_emcc_try(
 651     bx: &mut Builder<'a, 'll, 'tcx>,
 652     try_func: &'ll Value,
 653     data: &'ll Value,
 654     catch_func: &'ll Value,
 655     dest: &'ll Value,
 656 ) {
 657     let (llty, llfn) = get_rust_try_fn(bx, &mut |mut bx| {
 658         // Codegens the shims described above:
 659         //
 660         //   bx:
 661         //      invoke %try_func(%data) normal %normal unwind %catch
 662         //
 663         //   normal:
 664         //      ret 0
 665         //
 666         //   catch:
 667         //      (%ptr, %selector) = landingpad
 668         //      %rust_typeid = @llvm.eh.typeid.for(@_ZTI10rust_panic)
 669         //      %is_rust_panic = %selector == %rust_typeid
 670         //      %catch_data = alloca { i8*, i8 }
 671         //      %catch_data[0] = %ptr
 672         //      %catch_data[1] = %is_rust_panic
 673         //      call %catch_func(%data, %catch_data)
 674         //      ret 1
 675         let mut then = bx.build_sibling_block("then");
 676         let mut catch = bx.build_sibling_block("catch");
 677
 678         let try_func = llvm::get_param(bx.llfn(), 0);
 679         let data = llvm::get_param(bx.llfn(), 1);
 680         let catch_func = llvm::get_param(bx.llfn(), 2);
 681         let try_func_ty = bx.type_func(&[bx.type_i8p()], bx.type_void());
 682         bx.invoke(try_func_ty, try_func, &[data], then.llbb(), catch.llbb(), None);
 683         then.ret(bx.const_i32(0));
 684
 685         // Type indicator for the exception being thrown.
 686         //
 687         // The first value in this tuple is a pointer to the exception object
 688         // being thrown.  The second value is a "selector" indicating which of
 689         // the landing pad clauses the exception's type had been matched to.
 690         let tydesc = bx.eh_catch_typeinfo();
 691         let lpad_ty = bx.type_struct(&[bx.type_i8p(), bx.type_i32()], false);
 692         let vals = catch.landing_pad(lpad_ty, bx.eh_personality(), 2);
 693         catch.add_clause(vals, tydesc);
 694         catch.add_clause(vals, bx.const_null(bx.type_i8p()));
 695         let ptr = catch.extract_value(vals, 0);
 696         let selector = catch.extract_value(vals, 1);
 697
 698         // Check if the typeid we got is the one for a Rust panic.
 699         let rust_typeid = catch.call_intrinsic("llvm.eh.typeid.for", &[tydesc]);
 700         let is_rust_panic = catch.icmp(IntPredicate::IntEQ, selector, rust_typeid);
 701         let is_rust_panic = catch.zext(is_rust_panic, bx.type_bool());
 702
 703         // We need to pass two values to catch_func (ptr and is_rust_panic), so
 704         // create an alloca and pass a pointer to that.
 705         let ptr_align = bx.tcx().data_layout.pointer_align.abi;
 706         let i8_align = bx.tcx().data_layout.i8_align.abi;
 707         let catch_data_type = bx.type_struct(&[bx.type_i8p(), bx.type_bool()], false);
 708         let catch_data = catch.alloca(catch_data_type, ptr_align);
 709         let catch_data_0 = catch.inbounds_gep(
 710             catch_data_type,
 711             catch_data,
 712             &[bx.const_usize(0), bx.const_usize(0)],
 713         );
 714         catch.store(ptr, catch_data_0, ptr_align);
 715         let catch_data_1 = catch.inbounds_gep(
 716             catch_data_type,
 717             catch_data,
 718             &[bx.const_usize(0), bx.const_usize(1)],
 719         );
 720         catch.store(is_rust_panic, catch_data_1, i8_align);
 721         let catch_data = catch.bitcast(catch_data, bx.type_i8p());
 722
 723         let catch_ty = bx.type_func(&[bx.type_i8p(), bx.type_i8p()], bx.type_void());
 724         catch.call(catch_ty, catch_func, &[data, catch_data], None);
 725         catch.ret(bx.const_i32(1));
 726     });
 727
 728     // Note that no invoke is used here because by definition this function
 729     // can't panic (that's what it's catching).
 730     let ret = bx.call(llty, llfn, &[try_func, data, catch_func], None);
 731     let i32_align = bx.tcx().data_layout.i32_align.abi;
 732     bx.store(ret, dest, i32_align);
 733 }
 734
 735 // Helper function to give a Block to a closure to codegen a shim function.
 736 // This is currently primarily used for the `try` intrinsic functions above.
 737 fn gen_fn<'ll, 'tcx>(
 738     cx: &CodegenCx<'ll, 'tcx>,
 739     name: &str,
 740     rust_fn_sig: ty::PolyFnSig<'tcx>,
 741     codegen: &mut dyn FnMut(Builder<'_, 'll, 'tcx>),
 742 ) -> (&'ll Type, &'ll Value) {
 743     let fn_abi = cx.fn_abi_of_fn_ptr(rust_fn_sig, ty::List::empty());
 744     let llty = fn_abi.llvm_type(cx);
 745     let llfn = cx.declare_fn(name, fn_abi);
 746     cx.set_frame_pointer_type(llfn);
 747     cx.apply_target_cpu_attr(llfn);
 748     // FIXME(eddyb) find a nicer way to do this.
 749     unsafe { llvm::LLVMRustSetLinkage(llfn, llvm::Linkage::InternalLinkage) };
 750     let llbb = Builder::append_block(cx, llfn, "entry-block");
 751     let bx = Builder::build(cx, llbb);
 752     codegen(bx);
 753     (llty, llfn)
 754 }
 755
 756 // Helper function used to get a handle to the `__rust_try` function used to
 757 // catch exceptions.
 758 //
 759 // This function is only generated once and is then cached.
 760 fn get_rust_try_fn<'ll, 'tcx>(
 761     cx: &CodegenCx<'ll, 'tcx>,
 762     codegen: &mut dyn FnMut(Builder<'_, 'll, 'tcx>),
 763 ) -> (&'ll Type, &'ll Value) {
 764     if let Some(llfn) = cx.rust_try_fn.get() {
 765         return llfn;
 766     }
 767
 768     // Define the type up front for the signature of the rust_try function.
 769     let tcx = cx.tcx;
 770     let i8p = tcx.mk_mut_ptr(tcx.types.i8);
 771     // `unsafe fn(*mut i8) -> ()`
 772     let try_fn_ty = tcx.mk_fn_ptr(ty::Binder::dummy(tcx.mk_fn_sig(
 773         iter::once(i8p),
 774         tcx.mk_unit(),
 775         false,
 776         hir::Unsafety::Unsafe,
 777         Abi::Rust,
 778     )));
 779     // `unsafe fn(*mut i8, *mut i8) -> ()`
 780     let catch_fn_ty = tcx.mk_fn_ptr(ty::Binder::dummy(tcx.mk_fn_sig(
 781         [i8p, i8p].iter().cloned(),
 782         tcx.mk_unit(),
 783         false,
 784         hir::Unsafety::Unsafe,
 785         Abi::Rust,
 786     )));
 787     // `unsafe fn(unsafe fn(*mut i8) -> (), *mut i8, unsafe fn(*mut i8, *mut i8) -> ()) -> i32`
 788     let rust_fn_sig = ty::Binder::dummy(cx.tcx.mk_fn_sig(
 789         vec![try_fn_ty, i8p, catch_fn_ty].into_iter(),
 790         tcx.types.i32,
 791         false,
 792         hir::Unsafety::Unsafe,
 793         Abi::Rust,
 794     ));
 795     let rust_try = gen_fn(cx, "__rust_try", rust_fn_sig, codegen);
 796     cx.rust_try_fn.set(Some(rust_try));
 797     rust_try
 798 }
 799
 800 fn generic_simd_intrinsic(
 801     bx: &mut Builder<'a, 'll, 'tcx>,
 802     name: Symbol,
 803     callee_ty: Ty<'tcx>,
 804     args: &[OperandRef<'tcx, &'ll Value>],
 805     ret_ty: Ty<'tcx>,
 806     llret_ty: &'ll Type,
 807     span: Span,
 808 ) -> Result<&'ll Value, ()> {
 809     // macros for error handling:
 810     macro_rules! emit_error {
 811         ($msg: tt) => {
 812             emit_error!($msg, )
 813         };
 814         ($msg: tt, $($fmt: tt)*) => {
 815             span_invalid_monomorphization_error(
 816                 bx.sess(), span,
 817                 &format!(concat!("invalid monomorphization of `{}` intrinsic: ", $msg),
 818                          name, $($fmt)*));
 819         }
 820     }
 821
 822     macro_rules! return_error {
 823         ($($fmt: tt)*) => {
 824             {
 825                 emit_error!($($fmt)*);
 826                 return Err(());
 827             }
 828         }
 829     }
 830
 831     macro_rules! require {
 832         ($cond: expr, $($fmt: tt)*) => {
 833             if !$cond {
 834                 return_error!($($fmt)*);
 835             }
 836         };
 837     }
 838
 839     macro_rules! require_simd {
 840         ($ty: expr, $position: expr) => {
 841             require!($ty.is_simd(), "expected SIMD {} type, found non-SIMD `{}`", $position, $ty)
 842         };
 843     }
 844
 845     let tcx = bx.tcx();
 846     let sig =
 847         tcx.normalize_erasing_late_bound_regions(ty::ParamEnv::reveal_all(), callee_ty.fn_sig(tcx));
 848     let arg_tys = sig.inputs();
 849
 850     if name == sym::simd_select_bitmask {
 851         require_simd!(arg_tys[1], "argument");
 852         let (len, _) = arg_tys[1].simd_size_and_type(bx.tcx());
 853
 854         let expected_int_bits = (len.max(8) - 1).next_power_of_two();
 855         let expected_bytes = len / 8 + ((len % 8 > 0) as u64);
 856
 857         let mask_ty = arg_tys[0];
 858         let mask = match mask_ty.kind() {
 859             ty::Int(i) if i.bit_width() == Some(expected_int_bits) => args[0].immediate(),
 860             ty::Uint(i) if i.bit_width() == Some(expected_int_bits) => args[0].immediate(),
 861             ty::Array(elem, len)
 862                 if matches!(elem.kind(), ty::Uint(ty::UintTy::U8))
 863                     && len.try_eval_usize(bx.tcx, ty::ParamEnv::reveal_all())
 864                         == Some(expected_bytes) =>
 865             {
 866                 let place = PlaceRef::alloca(bx, args[0].layout);
 867                 args[0].val.store(bx, place);
 868                 let int_ty = bx.type_ix(expected_bytes * 8);
 869                 let ptr = bx.pointercast(place.llval, bx.cx.type_ptr_to(int_ty));
 870                 bx.load(int_ty, ptr, Align::ONE)
 871             }
 872             _ => return_error!(
 873                 "invalid bitmask `{}`, expected `u{}` or `[u8; {}]`",
 874                 mask_ty,
 875                 expected_int_bits,
 876                 expected_bytes
 877             ),
 878         };
 879
 880         let i1 = bx.type_i1();
 881         let im = bx.type_ix(len);
 882         let i1xn = bx.type_vector(i1, len);
 883         let m_im = bx.trunc(mask, im);
 884         let m_i1s = bx.bitcast(m_im, i1xn);
 885         return Ok(bx.select(m_i1s, args[1].immediate(), args[2].immediate()));
 886     }
 887
 888     // every intrinsic below takes a SIMD vector as its first argument
 889     require_simd!(arg_tys[0], "input");
 890     let in_ty = arg_tys[0];
 891
 892     let comparison = match name {
 893         sym::simd_eq => Some(hir::BinOpKind::Eq),
 894         sym::simd_ne => Some(hir::BinOpKind::Ne),
 895         sym::simd_lt => Some(hir::BinOpKind::Lt),
 896         sym::simd_le => Some(hir::BinOpKind::Le),
 897         sym::simd_gt => Some(hir::BinOpKind::Gt),
 898         sym::simd_ge => Some(hir::BinOpKind::Ge),
 899         _ => None,
 900     };
 901
 902     let (in_len, in_elem) = arg_tys[0].simd_size_and_type(bx.tcx());
 903     if let Some(cmp_op) = comparison {
 904         require_simd!(ret_ty, "return");
 905
 906         let (out_len, out_ty) = ret_ty.simd_size_and_type(bx.tcx());
 907         require!(
 908             in_len == out_len,
 909             "expected return type with length {} (same as input type `{}`), \
 910              found `{}` with length {}",
 911             in_len,
 912             in_ty,
 913             ret_ty,
 914             out_len
 915         );
 916         require!(
 917             bx.type_kind(bx.element_type(llret_ty)) == TypeKind::Integer,
 918             "expected return type with integer elements, found `{}` with non-integer `{}`",
 919             ret_ty,
 920             out_ty
 921         );
 922
 923         return Ok(compare_simd_types(
 924             bx,
 925             args[0].immediate(),
 926             args[1].immediate(),
 927             in_elem,
 928             llret_ty,
 929             cmp_op,
 930         ));
 931     }
 932
 933     if let Some(stripped) = name.as_str().strip_prefix("simd_shuffle") {
 934         // If this intrinsic is the older "simd_shuffleN" form, simply parse the integer.
 935         // If there is no suffix, use the index array length.
 936         let n: u64 = if stripped.is_empty() {
 937             // Make sure this is actually an array, since typeck only checks the length-suffixed
 938             // version of this intrinsic.
 939             match args[2].layout.ty.kind() {
 940                 ty::Array(ty, len) if matches!(ty.kind(), ty::Uint(ty::UintTy::U32)) => {
 941                     len.try_eval_usize(bx.cx.tcx, ty::ParamEnv::reveal_all()).unwrap_or_else(|| {
 942                         span_bug!(span, "could not evaluate shuffle index array length")
 943                     })
 944                 }
 945                 _ => return_error!(
 946                     "simd_shuffle index must be an array of `u32`, got `{}`",
 947                     args[2].layout.ty
 948                 ),
 949             }
 950         } else {
 951             stripped.parse().unwrap_or_else(|_| {
 952                 span_bug!(span, "bad `simd_shuffle` instruction only caught in codegen?")
 953             })
 954         };
 955
 956         require_simd!(ret_ty, "return");
 957         let (out_len, out_ty) = ret_ty.simd_size_and_type(bx.tcx());
 958         require!(
 959             out_len == n,
 960             "expected return type of length {}, found `{}` with length {}",
 961             n,
 962             ret_ty,
 963             out_len
 964         );
 965         require!(
 966             in_elem == out_ty,
 967             "expected return element type `{}` (element of input `{}`), \
 968              found `{}` with element type `{}`",
 969             in_elem,
 970             in_ty,
 971             ret_ty,
 972             out_ty
 973         );
 974
 975         let total_len = u128::from(in_len) * 2;
 976
 977         let vector = args[2].immediate();
 978
 979         let indices: Option<Vec<_>> = (0..n)
 980             .map(|i| {
 981                 let arg_idx = i;
 982                 let val = bx.const_get_elt(vector, i as u64);
 983                 match bx.const_to_opt_u128(val, true) {
 984                     None => {
 985                         emit_error!("shuffle index #{} is not a constant", arg_idx);
 986                         None
 987                     }
 988                     Some(idx) if idx >= total_len => {
 989                         emit_error!(
 990                             "shuffle index #{} is out of bounds (limit {})",
 991                             arg_idx,
 992                             total_len
 993                         );
 994                         None
 995                     }
 996                     Some(idx) => Some(bx.const_i32(idx as i32)),
 997                 }
 998             })
 999             .collect();
1000         let indices = match indices {
1001             Some(i) => i,
1002             None => return Ok(bx.const_null(llret_ty)),
1003         };
1004
1005         return Ok(bx.shuffle_vector(
1006             args[0].immediate(),
1007             args[1].immediate(),
1008             bx.const_vector(&indices),
1009         ));
1010     }
1011
1012     if name == sym::simd_insert {
1013         require!(
1014             in_elem == arg_tys[2],
1015             "expected inserted type `{}` (element of input `{}`), found `{}`",
1016             in_elem,
1017             in_ty,
1018             arg_tys[2]
1019         );
1020         return Ok(bx.insert_element(
1021             args[0].immediate(),
1022             args[2].immediate(),
1023             args[1].immediate(),
1024         ));
1025     }
1026     if name == sym::simd_extract {
1027         require!(
1028             ret_ty == in_elem,
1029             "expected return type `{}` (element of input `{}`), found `{}`",
1030             in_elem,
1031             in_ty,
1032             ret_ty
1033         );
1034         return Ok(bx.extract_element(args[0].immediate(), args[1].immediate()));
1035     }
1036
1037     if name == sym::simd_select {
1038         let m_elem_ty = in_elem;
1039         let m_len = in_len;
1040         require_simd!(arg_tys[1], "argument");
1041         let (v_len, _) = arg_tys[1].simd_size_and_type(bx.tcx());
1042         require!(
1043             m_len == v_len,
1044             "mismatched lengths: mask length `{}` != other vector length `{}`",
1045             m_len,
1046             v_len
1047         );
1048         match m_elem_ty.kind() {
1049             ty::Int(_) => {}
1050             _ => return_error!("mask element type is `{}`, expected `i_`", m_elem_ty),
1051         }
1052         // truncate the mask to a vector of i1s
1053         let i1 = bx.type_i1();
1054         let i1xn = bx.type_vector(i1, m_len as u64);
1055         let m_i1s = bx.trunc(args[0].immediate(), i1xn);
1056         return Ok(bx.select(m_i1s, args[1].immediate(), args[2].immediate()));
1057     }
1058
1059     if name == sym::simd_bitmask {
1060         // The `fn simd_bitmask(vector) -> unsigned integer` intrinsic takes a
1061         // vector mask and returns the most significant bit (MSB) of each lane in the form
1062         // of either:
1063         // * an unsigned integer
1064         // * an array of `u8`
1065         // If the vector has less than 8 lanes, a u8 is returned with zeroed trailing bits.
1066         //
1067         // The bit order of the result depends on the byte endianness, LSB-first for little
1068         // endian and MSB-first for big endian.
1069         let expected_int_bits = in_len.max(8);
1070         let expected_bytes = expected_int_bits / 8 + ((expected_int_bits % 8 > 0) as u64);
1071
1072         // Integer vector <i{in_bitwidth} x in_len>:
1073         let (i_xn, in_elem_bitwidth) = match in_elem.kind() {
1074             ty::Int(i) => (
1075                 args[0].immediate(),
1076                 i.bit_width().unwrap_or_else(|| bx.data_layout().pointer_size.bits()),
1077             ),
1078             ty::Uint(i) => (
1079                 args[0].immediate(),
1080                 i.bit_width().unwrap_or_else(|| bx.data_layout().pointer_size.bits()),
1081             ),
1082             _ => return_error!(
1083                 "vector argument `{}`'s element type `{}`, expected integer element type",
1084                 in_ty,
1085                 in_elem
1086             ),
1087         };
1088
1089         // Shift the MSB to the right by "in_elem_bitwidth - 1" into the first bit position.
1090         let shift_indices =
1091             vec![
1092                 bx.cx.const_int(bx.type_ix(in_elem_bitwidth), (in_elem_bitwidth - 1) as _);
1093                 in_len as _
1094             ];
1095         let i_xn_msb = bx.lshr(i_xn, bx.const_vector(shift_indices.as_slice()));
1096         // Truncate vector to an <i1 x N>
1097         let i1xn = bx.trunc(i_xn_msb, bx.type_vector(bx.type_i1(), in_len));
1098         // Bitcast <i1 x N> to iN:
1099         let i_ = bx.bitcast(i1xn, bx.type_ix(in_len));
1100
1101         match ret_ty.kind() {
1102             ty::Uint(i) if i.bit_width() == Some(expected_int_bits) => {
1103                 // Zero-extend iN to the bitmask type:
1104                 return Ok(bx.zext(i_, bx.type_ix(expected_int_bits)));
1105             }
1106             ty::Array(elem, len)
1107                 if matches!(elem.kind(), ty::Uint(ty::UintTy::U8))
1108                     && len.try_eval_usize(bx.tcx, ty::ParamEnv::reveal_all())
1109                         == Some(expected_bytes) =>
1110             {
1111                 // Zero-extend iN to the array lengh:
1112                 let ze = bx.zext(i_, bx.type_ix(expected_bytes * 8));
1113
1114                 // Convert the integer to a byte array
1115                 let ptr = bx.alloca(bx.type_ix(expected_bytes * 8), Align::ONE);
1116                 bx.store(ze, ptr, Align::ONE);
1117                 let array_ty = bx.type_array(bx.type_i8(), expected_bytes);
1118                 let ptr = bx.pointercast(ptr, bx.cx.type_ptr_to(array_ty));
1119                 return Ok(bx.load(array_ty, ptr, Align::ONE));
1120             }
1121             _ => return_error!(
1122                 "cannot return `{}`, expected `u{}` or `[u8; {}]`",
1123                 ret_ty,
1124                 expected_int_bits,
1125                 expected_bytes
1126             ),
1127         }
1128     }
1129
1130     fn simd_simple_float_intrinsic(
1131         name: Symbol,
1132         in_elem: &::rustc_middle::ty::TyS<'_>,
1133         in_ty: &::rustc_middle::ty::TyS<'_>,
1134         in_len: u64,
1135         bx: &mut Builder<'a, 'll, 'tcx>,
1136         span: Span,
1137         args: &[OperandRef<'tcx, &'ll Value>],
1138     ) -> Result<&'ll Value, ()> {
1139         macro_rules! emit_error {
1140             ($msg: tt) => {
1141                 emit_error!($msg, )
1142             };
1143             ($msg: tt, $($fmt: tt)*) => {
1144                 span_invalid_monomorphization_error(
1145                     bx.sess(), span,
1146                     &format!(concat!("invalid monomorphization of `{}` intrinsic: ", $msg),
1147                              name, $($fmt)*));
1148             }
1149         }
1150         macro_rules! return_error {
1151             ($($fmt: tt)*) => {
1152                 {
1153                     emit_error!($($fmt)*);
1154                     return Err(());
1155                 }
1156             }
1157         }
1158
1159         let (elem_ty_str, elem_ty) = if let ty::Float(f) = in_elem.kind() {
1160             let elem_ty = bx.cx.type_float_from_ty(*f);
1161             match f.bit_width() {
1162                 32 => ("f32", elem_ty),
1163                 64 => ("f64", elem_ty),
1164                 _ => {
1165                     return_error!(
1166                         "unsupported element type `{}` of floating-point vector `{}`",
1167                         f.name_str(),
1168                         in_ty
1169                     );
1170                 }
1171             }
1172         } else {
1173             return_error!("`{}` is not a floating-point type", in_ty);
1174         };
1175
1176         let vec_ty = bx.type_vector(elem_ty, in_len);
1177
1178         let (intr_name, fn_ty) = match name {
1179             sym::simd_ceil => ("ceil", bx.type_func(&[vec_ty], vec_ty)),
1180             sym::simd_fabs => ("fabs", bx.type_func(&[vec_ty], vec_ty)),
1181             sym::simd_fcos => ("cos", bx.type_func(&[vec_ty], vec_ty)),
1182             sym::simd_fexp2 => ("exp2", bx.type_func(&[vec_ty], vec_ty)),
1183             sym::simd_fexp => ("exp", bx.type_func(&[vec_ty], vec_ty)),
1184             sym::simd_flog10 => ("log10", bx.type_func(&[vec_ty], vec_ty)),
1185             sym::simd_flog2 => ("log2", bx.type_func(&[vec_ty], vec_ty)),
1186             sym::simd_flog => ("log", bx.type_func(&[vec_ty], vec_ty)),
1187             sym::simd_floor => ("floor", bx.type_func(&[vec_ty], vec_ty)),
1188             sym::simd_fma => ("fma", bx.type_func(&[vec_ty, vec_ty, vec_ty], vec_ty)),
1189             sym::simd_fpowi => ("powi", bx.type_func(&[vec_ty, bx.type_i32()], vec_ty)),
1190             sym::simd_fpow => ("pow", bx.type_func(&[vec_ty, vec_ty], vec_ty)),
1191             sym::simd_fsin => ("sin", bx.type_func(&[vec_ty], vec_ty)),
1192             sym::simd_fsqrt => ("sqrt", bx.type_func(&[vec_ty], vec_ty)),
1193             sym::simd_round => ("round", bx.type_func(&[vec_ty], vec_ty)),
1194             sym::simd_trunc => ("trunc", bx.type_func(&[vec_ty], vec_ty)),
1195             _ => return_error!("unrecognized intrinsic `{}`", name),
1196         };
1197         let llvm_name = &format!("llvm.{0}.v{1}{2}", intr_name, in_len, elem_ty_str);
1198         let f = bx.declare_cfn(llvm_name, llvm::UnnamedAddr::No, fn_ty);
1199         let c =
1200             bx.call(fn_ty, f, &args.iter().map(|arg| arg.immediate()).collect::<Vec<_>>(), None);
1201         Ok(c)
1202     }
1203
1204     if std::matches!(
1205         name,
1206         sym::simd_ceil
1207             | sym::simd_fabs
1208             | sym::simd_fcos
1209             | sym::simd_fexp2
1210             | sym::simd_fexp
1211             | sym::simd_flog10
1212             | sym::simd_flog2
1213             | sym::simd_flog
1214             | sym::simd_floor
1215             | sym::simd_fma
1216             | sym::simd_fpow
1217             | sym::simd_fpowi
1218             | sym::simd_fsin
1219             | sym::simd_fsqrt
1220             | sym::simd_round
1221             | sym::simd_trunc
1222     ) {
1223         return simd_simple_float_intrinsic(name, in_elem, in_ty, in_len, bx, span, args);
1224     }
1225
1226     // FIXME: use:
1227     //  https://github.com/llvm-mirror/llvm/blob/master/include/llvm/IR/Function.h#L182
1228     //  https://github.com/llvm-mirror/llvm/blob/master/include/llvm/IR/Intrinsics.h#L81
1229     fn llvm_vector_str(
1230         elem_ty: Ty<'_>,
1231         vec_len: u64,
1232         no_pointers: usize,
1233         bx: &Builder<'a, 'll, 'tcx>,
1234     ) -> String {
1235         let p0s: String = "p0".repeat(no_pointers);
1236         match *elem_ty.kind() {
1237             ty::Int(v) => format!(
1238                 "v{}{}i{}",
1239                 vec_len,
1240                 p0s,
1241                 // Normalize to prevent crash if v: IntTy::Isize
1242                 v.normalize(bx.target_spec().pointer_width).bit_width().unwrap()
1243             ),
1244             ty::Uint(v) => format!(
1245                 "v{}{}i{}",
1246                 vec_len,
1247                 p0s,
1248                 // Normalize to prevent crash if v: UIntTy::Usize
1249                 v.normalize(bx.target_spec().pointer_width).bit_width().unwrap()
1250             ),
1251             ty::Float(v) => format!("v{}{}f{}", vec_len, p0s, v.bit_width()),
1252             _ => unreachable!(),
1253         }
1254     }
1255
1256     fn llvm_vector_ty(
1257         cx: &CodegenCx<'ll, '_>,
1258         elem_ty: Ty<'_>,
1259         vec_len: u64,
1260         mut no_pointers: usize,
1261     ) -> &'ll Type {
1262         // FIXME: use cx.layout_of(ty).llvm_type() ?
1263         let mut elem_ty = match *elem_ty.kind() {
1264             ty::Int(v) => cx.type_int_from_ty(v),
1265             ty::Uint(v) => cx.type_uint_from_ty(v),
1266             ty::Float(v) => cx.type_float_from_ty(v),
1267             _ => unreachable!(),
1268         };
1269         while no_pointers > 0 {
1270             elem_ty = cx.type_ptr_to(elem_ty);
1271             no_pointers -= 1;
1272         }
1273         cx.type_vector(elem_ty, vec_len)
1274     }
1275
1276     if name == sym::simd_gather {
1277         // simd_gather(values: <N x T>, pointers: <N x *_ T>,
1278         //             mask: <N x i{M}>) -> <N x T>
1279         // * N: number of elements in the input vectors
1280         // * T: type of the element to load
1281         // * M: any integer width is supported, will be truncated to i1
1282
1283         // All types must be simd vector types
1284         require_simd!(in_ty, "first");
1285         require_simd!(arg_tys[1], "second");
1286         require_simd!(arg_tys[2], "third");
1287         require_simd!(ret_ty, "return");
1288
1289         // Of the same length:
1290         let (out_len, _) = arg_tys[1].simd_size_and_type(bx.tcx());
1291         let (out_len2, _) = arg_tys[2].simd_size_and_type(bx.tcx());
1292         require!(
1293             in_len == out_len,
1294             "expected {} argument with length {} (same as input type `{}`), \
1295              found `{}` with length {}",
1296             "second",
1297             in_len,
1298             in_ty,
1299             arg_tys[1],
1300             out_len
1301         );
1302         require!(
1303             in_len == out_len2,
1304             "expected {} argument with length {} (same as input type `{}`), \
1305              found `{}` with length {}",
1306             "third",
1307             in_len,
1308             in_ty,
1309             arg_tys[2],
1310             out_len2
1311         );
1312
1313         // The return type must match the first argument type
1314         require!(ret_ty == in_ty, "expected return type `{}`, found `{}`", in_ty, ret_ty);
1315
1316         // This counts how many pointers
1317         fn ptr_count(t: Ty<'_>) -> usize {
1318             match t.kind() {
1319                 ty::RawPtr(p) => 1 + ptr_count(p.ty),
1320                 _ => 0,
1321             }
1322         }
1323
1324         // Non-ptr type
1325         fn non_ptr(t: Ty<'_>) -> Ty<'_> {
1326             match t.kind() {
1327                 ty::RawPtr(p) => non_ptr(p.ty),
1328                 _ => t,
1329             }
1330         }
1331
1332         // The second argument must be a simd vector with an element type that's a pointer
1333         // to the element type of the first argument
1334         let (_, element_ty0) = arg_tys[0].simd_size_and_type(bx.tcx());
1335         let (_, element_ty1) = arg_tys[1].simd_size_and_type(bx.tcx());
1336         let (pointer_count, underlying_ty) = match element_ty1.kind() {
1337             ty::RawPtr(p) if p.ty == in_elem => (ptr_count(element_ty1), non_ptr(element_ty1)),
1338             _ => {
1339                 require!(
1340                     false,
1341                     "expected element type `{}` of second argument `{}` \
1342                         to be a pointer to the element type `{}` of the first \
1343                         argument `{}`, found `{}` != `*_ {}`",
1344                     element_ty1,
1345                     arg_tys[1],
1346                     in_elem,
1347                     in_ty,
1348                     element_ty1,
1349                     in_elem
1350                 );
1351                 unreachable!();
1352             }
1353         };
1354         assert!(pointer_count > 0);
1355         assert_eq!(pointer_count - 1, ptr_count(element_ty0));
1356         assert_eq!(underlying_ty, non_ptr(element_ty0));
1357
1358         // The element type of the third argument must be a signed integer type of any width:
1359         let (_, element_ty2) = arg_tys[2].simd_size_and_type(bx.tcx());
1360         match element_ty2.kind() {
1361             ty::Int(_) => (),
1362             _ => {
1363                 require!(
1364                     false,
1365                     "expected element type `{}` of third argument `{}` \
1366                                  to be a signed integer type",
1367                     element_ty2,
1368                     arg_tys[2]
1369                 );
1370             }
1371         }
1372
1373         // Alignment of T, must be a constant integer value:
1374         let alignment_ty = bx.type_i32();
1375         let alignment = bx.const_i32(bx.align_of(in_elem).bytes() as i32);
1376
1377         // Truncate the mask vector to a vector of i1s:
1378         let (mask, mask_ty) = {
1379             let i1 = bx.type_i1();
1380             let i1xn = bx.type_vector(i1, in_len);
1381             (bx.trunc(args[2].immediate(), i1xn), i1xn)
1382         };
1383
1384         // Type of the vector of pointers:
1385         let llvm_pointer_vec_ty = llvm_vector_ty(bx, underlying_ty, in_len, pointer_count);
1386         let llvm_pointer_vec_str = llvm_vector_str(underlying_ty, in_len, pointer_count, bx);
1387
1388         // Type of the vector of elements:
1389         let llvm_elem_vec_ty = llvm_vector_ty(bx, underlying_ty, in_len, pointer_count - 1);
1390         let llvm_elem_vec_str = llvm_vector_str(underlying_ty, in_len, pointer_count - 1, bx);
1391
1392         let llvm_intrinsic =
1393             format!("llvm.masked.gather.{}.{}", llvm_elem_vec_str, llvm_pointer_vec_str);
1394         let fn_ty = bx.type_func(
1395             &[llvm_pointer_vec_ty, alignment_ty, mask_ty, llvm_elem_vec_ty],
1396             llvm_elem_vec_ty,
1397         );
1398         let f = bx.declare_cfn(&llvm_intrinsic, llvm::UnnamedAddr::No, fn_ty);
1399         let v =
1400             bx.call(fn_ty, f, &[args[1].immediate(), alignment, mask, args[0].immediate()], None);
1401         return Ok(v);
1402     }
1403
1404     if name == sym::simd_scatter {
1405         // simd_scatter(values: <N x T>, pointers: <N x *mut T>,
1406         //             mask: <N x i{M}>) -> ()
1407         // * N: number of elements in the input vectors
1408         // * T: type of the element to load
1409         // * M: any integer width is supported, will be truncated to i1
1410
1411         // All types must be simd vector types
1412         require_simd!(in_ty, "first");
1413         require_simd!(arg_tys[1], "second");
1414         require_simd!(arg_tys[2], "third");
1415
1416         // Of the same length:
1417         let (element_len1, _) = arg_tys[1].simd_size_and_type(bx.tcx());
1418         let (element_len2, _) = arg_tys[2].simd_size_and_type(bx.tcx());
1419         require!(
1420             in_len == element_len1,
1421             "expected {} argument with length {} (same as input type `{}`), \
1422             found `{}` with length {}",
1423             "second",
1424             in_len,
1425             in_ty,
1426             arg_tys[1],
1427             element_len1
1428         );
1429         require!(
1430             in_len == element_len2,
1431             "expected {} argument with length {} (same as input type `{}`), \
1432             found `{}` with length {}",
1433             "third",
1434             in_len,
1435             in_ty,
1436             arg_tys[2],
1437             element_len2
1438         );
1439
1440         // This counts how many pointers
1441         fn ptr_count(t: Ty<'_>) -> usize {
1442             match t.kind() {
1443                 ty::RawPtr(p) => 1 + ptr_count(p.ty),
1444                 _ => 0,
1445             }
1446         }
1447
1448         // Non-ptr type
1449         fn non_ptr(t: Ty<'_>) -> Ty<'_> {
1450             match t.kind() {
1451                 ty::RawPtr(p) => non_ptr(p.ty),
1452                 _ => t,
1453             }
1454         }
1455
1456         // The second argument must be a simd vector with an element type that's a pointer
1457         // to the element type of the first argument
1458         let (_, element_ty0) = arg_tys[0].simd_size_and_type(bx.tcx());
1459         let (_, element_ty1) = arg_tys[1].simd_size_and_type(bx.tcx());
1460         let (_, element_ty2) = arg_tys[2].simd_size_and_type(bx.tcx());
1461         let (pointer_count, underlying_ty) = match element_ty1.kind() {
1462             ty::RawPtr(p) if p.ty == in_elem && p.mutbl == hir::Mutability::Mut => {
1463                 (ptr_count(element_ty1), non_ptr(element_ty1))
1464             }
1465             _ => {
1466                 require!(
1467                     false,
1468                     "expected element type `{}` of second argument `{}` \
1469                         to be a pointer to the element type `{}` of the first \
1470                         argument `{}`, found `{}` != `*mut {}`",
1471                     element_ty1,
1472                     arg_tys[1],
1473                     in_elem,
1474                     in_ty,
1475                     element_ty1,
1476                     in_elem
1477                 );
1478                 unreachable!();
1479             }
1480         };
1481         assert!(pointer_count > 0);
1482         assert_eq!(pointer_count - 1, ptr_count(element_ty0));
1483         assert_eq!(underlying_ty, non_ptr(element_ty0));
1484
1485         // The element type of the third argument must be a signed integer type of any width:
1486         match element_ty2.kind() {
1487             ty::Int(_) => (),
1488             _ => {
1489                 require!(
1490                     false,
1491                     "expected element type `{}` of third argument `{}` \
1492                          be a signed integer type",
1493                     element_ty2,
1494                     arg_tys[2]
1495                 );
1496             }
1497         }
1498
1499         // Alignment of T, must be a constant integer value:
1500         let alignment_ty = bx.type_i32();
1501         let alignment = bx.const_i32(bx.align_of(in_elem).bytes() as i32);
1502
1503         // Truncate the mask vector to a vector of i1s:
1504         let (mask, mask_ty) = {
1505             let i1 = bx.type_i1();
1506             let i1xn = bx.type_vector(i1, in_len);
1507             (bx.trunc(args[2].immediate(), i1xn), i1xn)
1508         };
1509
1510         let ret_t = bx.type_void();
1511
1512         // Type of the vector of pointers:
1513         let llvm_pointer_vec_ty = llvm_vector_ty(bx, underlying_ty, in_len, pointer_count);
1514         let llvm_pointer_vec_str = llvm_vector_str(underlying_ty, in_len, pointer_count, bx);
1515
1516         // Type of the vector of elements:
1517         let llvm_elem_vec_ty = llvm_vector_ty(bx, underlying_ty, in_len, pointer_count - 1);
1518         let llvm_elem_vec_str = llvm_vector_str(underlying_ty, in_len, pointer_count - 1, bx);
1519
1520         let llvm_intrinsic =
1521             format!("llvm.masked.scatter.{}.{}", llvm_elem_vec_str, llvm_pointer_vec_str);
1522         let fn_ty =
1523             bx.type_func(&[llvm_elem_vec_ty, llvm_pointer_vec_ty, alignment_ty, mask_ty], ret_t);
1524         let f = bx.declare_cfn(&llvm_intrinsic, llvm::UnnamedAddr::No, fn_ty);
1525         let v =
1526             bx.call(fn_ty, f, &[args[0].immediate(), args[1].immediate(), alignment, mask], None);
1527         return Ok(v);
1528     }
1529
1530     macro_rules! arith_red {
1531         ($name:ident : $integer_reduce:ident, $float_reduce:ident, $ordered:expr, $op:ident,
1532          $identity:expr) => {
1533             if name == sym::$name {
1534                 require!(
1535                     ret_ty == in_elem,
1536                     "expected return type `{}` (element of input `{}`), found `{}`",
1537                     in_elem,
1538                     in_ty,
1539                     ret_ty
1540                 );
1541                 return match in_elem.kind() {
1542                     ty::Int(_) | ty::Uint(_) => {
1543                         let r = bx.$integer_reduce(args[0].immediate());
1544                         if $ordered {
1545                             // if overflow occurs, the result is the
1546                             // mathematical result modulo 2^n:
1547                             Ok(bx.$op(args[1].immediate(), r))
1548                         } else {
1549                             Ok(bx.$integer_reduce(args[0].immediate()))
1550                         }
1551                     }
1552                     ty::Float(f) => {
1553                         let acc = if $ordered {
1554                             // ordered arithmetic reductions take an accumulator
1555                             args[1].immediate()
1556                         } else {
1557                             // unordered arithmetic reductions use the identity accumulator
1558                             match f.bit_width() {
1559                                 32 => bx.const_real(bx.type_f32(), $identity),
1560                                 64 => bx.const_real(bx.type_f64(), $identity),
1561                                 v => return_error!(
1562                                     r#"
1563 unsupported {} from `{}` with element `{}` of size `{}` to `{}`"#,
1564                                     sym::$name,
1565                                     in_ty,
1566                                     in_elem,
1567                                     v,
1568                                     ret_ty
1569                                 ),
1570                             }
1571                         };
1572                         Ok(bx.$float_reduce(acc, args[0].immediate()))
1573                     }
1574                     _ => return_error!(
1575                         "unsupported {} from `{}` with element `{}` to `{}`",
1576                         sym::$name,
1577                         in_ty,
1578                         in_elem,
1579                         ret_ty
1580                     ),
1581                 };
1582             }
1583         };
1584     }
1585
1586     arith_red!(simd_reduce_add_ordered: vector_reduce_add, vector_reduce_fadd, true, add, 0.0);
1587     arith_red!(simd_reduce_mul_ordered: vector_reduce_mul, vector_reduce_fmul, true, mul, 1.0);
1588     arith_red!(
1589         simd_reduce_add_unordered: vector_reduce_add,
1590         vector_reduce_fadd_fast,
1591         false,
1592         add,
1593         0.0
1594     );
1595     arith_red!(
1596         simd_reduce_mul_unordered: vector_reduce_mul,
1597         vector_reduce_fmul_fast,
1598         false,
1599         mul,
1600         1.0
1601     );
1602
1603     macro_rules! minmax_red {
1604         ($name:ident: $int_red:ident, $float_red:ident) => {
1605             if name == sym::$name {
1606                 require!(
1607                     ret_ty == in_elem,
1608                     "expected return type `{}` (element of input `{}`), found `{}`",
1609                     in_elem,
1610                     in_ty,
1611                     ret_ty
1612                 );
1613                 return match in_elem.kind() {
1614                     ty::Int(_i) => Ok(bx.$int_red(args[0].immediate(), true)),
1615                     ty::Uint(_u) => Ok(bx.$int_red(args[0].immediate(), false)),
1616                     ty::Float(_f) => Ok(bx.$float_red(args[0].immediate())),
1617                     _ => return_error!(
1618                         "unsupported {} from `{}` with element `{}` to `{}`",
1619                         sym::$name,
1620                         in_ty,
1621                         in_elem,
1622                         ret_ty
1623                     ),
1624                 };
1625             }
1626         };
1627     }
1628
1629     minmax_red!(simd_reduce_min: vector_reduce_min, vector_reduce_fmin);
1630     minmax_red!(simd_reduce_max: vector_reduce_max, vector_reduce_fmax);
1631
1632     minmax_red!(simd_reduce_min_nanless: vector_reduce_min, vector_reduce_fmin_fast);
1633     minmax_red!(simd_reduce_max_nanless: vector_reduce_max, vector_reduce_fmax_fast);
1634
1635     macro_rules! bitwise_red {
1636         ($name:ident : $red:ident, $boolean:expr) => {
1637             if name == sym::$name {
1638                 let input = if !$boolean {
1639                     require!(
1640                         ret_ty == in_elem,
1641                         "expected return type `{}` (element of input `{}`), found `{}`",
1642                         in_elem,
1643                         in_ty,
1644                         ret_ty
1645                     );
1646                     args[0].immediate()
1647                 } else {
1648                     match in_elem.kind() {
1649                         ty::Int(_) | ty::Uint(_) => {}
1650                         _ => return_error!(
1651                             "unsupported {} from `{}` with element `{}` to `{}`",
1652                             sym::$name,
1653                             in_ty,
1654                             in_elem,
1655                             ret_ty
1656                         ),
1657                     }
1658
1659                     // boolean reductions operate on vectors of i1s:
1660                     let i1 = bx.type_i1();
1661                     let i1xn = bx.type_vector(i1, in_len as u64);
1662                     bx.trunc(args[0].immediate(), i1xn)
1663                 };
1664                 return match in_elem.kind() {
1665                     ty::Int(_) | ty::Uint(_) => {
1666                         let r = bx.$red(input);
1667                         Ok(if !$boolean { r } else { bx.zext(r, bx.type_bool()) })
1668                     }
1669                     _ => return_error!(
1670                         "unsupported {} from `{}` with element `{}` to `{}`",
1671                         sym::$name,
1672                         in_ty,
1673                         in_elem,
1674                         ret_ty
1675                     ),
1676                 };
1677             }
1678         };
1679     }
1680
1681     bitwise_red!(simd_reduce_and: vector_reduce_and, false);
1682     bitwise_red!(simd_reduce_or: vector_reduce_or, false);
1683     bitwise_red!(simd_reduce_xor: vector_reduce_xor, false);
1684     bitwise_red!(simd_reduce_all: vector_reduce_and, true);
1685     bitwise_red!(simd_reduce_any: vector_reduce_or, true);
1686
1687     if name == sym::simd_cast {
1688         require_simd!(ret_ty, "return");
1689         let (out_len, out_elem) = ret_ty.simd_size_and_type(bx.tcx());
1690         require!(
1691             in_len == out_len,
1692             "expected return type with length {} (same as input type `{}`), \
1693                   found `{}` with length {}",
1694             in_len,
1695             in_ty,
1696             ret_ty,
1697             out_len
1698         );
1699         // casting cares about nominal type, not just structural type
1700         if in_elem == out_elem {
1701             return Ok(args[0].immediate());
1702         }
1703
1704         enum Style {
1705             Float,
1706             Int(/* is signed? */ bool),
1707             Unsupported,
1708         }
1709
1710         let (in_style, in_width) = match in_elem.kind() {
1711             // vectors of pointer-sized integers should've been
1712             // disallowed before here, so this unwrap is safe.
1713             ty::Int(i) => (Style::Int(true), i.bit_width().unwrap()),
1714             ty::Uint(u) => (Style::Int(false), u.bit_width().unwrap()),
1715             ty::Float(f) => (Style::Float, f.bit_width()),
1716             _ => (Style::Unsupported, 0),
1717         };
1718         let (out_style, out_width) = match out_elem.kind() {
1719             ty::Int(i) => (Style::Int(true), i.bit_width().unwrap()),
1720             ty::Uint(u) => (Style::Int(false), u.bit_width().unwrap()),
1721             ty::Float(f) => (Style::Float, f.bit_width()),
1722             _ => (Style::Unsupported, 0),
1723         };
1724
1725         match (in_style, out_style) {
1726             (Style::Int(in_is_signed), Style::Int(_)) => {
1727                 return Ok(match in_width.cmp(&out_width) {
1728                     Ordering::Greater => bx.trunc(args[0].immediate(), llret_ty),
1729                     Ordering::Equal => args[0].immediate(),
1730                     Ordering::Less => {
1731                         if in_is_signed {
1732                             bx.sext(args[0].immediate(), llret_ty)
1733                         } else {
1734                             bx.zext(args[0].immediate(), llret_ty)
1735                         }
1736                     }
1737                 });
1738             }
1739             (Style::Int(in_is_signed), Style::Float) => {
1740                 return Ok(if in_is_signed {
1741                     bx.sitofp(args[0].immediate(), llret_ty)
1742                 } else {
1743                     bx.uitofp(args[0].immediate(), llret_ty)
1744                 });
1745             }
1746             (Style::Float, Style::Int(out_is_signed)) => {
1747                 return Ok(if out_is_signed {
1748                     bx.fptosi(args[0].immediate(), llret_ty)
1749                 } else {
1750                     bx.fptoui(args[0].immediate(), llret_ty)
1751                 });
1752             }
1753             (Style::Float, Style::Float) => {
1754                 return Ok(match in_width.cmp(&out_width) {
1755                     Ordering::Greater => bx.fptrunc(args[0].immediate(), llret_ty),
1756                     Ordering::Equal => args[0].immediate(),
1757                     Ordering::Less => bx.fpext(args[0].immediate(), llret_ty),
1758                 });
1759             }
1760             _ => { /* Unsupported. Fallthrough. */ }
1761         }
1762         require!(
1763             false,
1764             "unsupported cast from `{}` with element `{}` to `{}` with element `{}`",
1765             in_ty,
1766             in_elem,
1767             ret_ty,
1768             out_elem
1769         );
1770     }
1771     macro_rules! arith_binary {
1772         ($($name: ident: $($($p: ident),* => $call: ident),*;)*) => {
1773             $(if name == sym::$name {
1774                 match in_elem.kind() {
1775                     $($(ty::$p(_))|* => {
1776                         return Ok(bx.$call(args[0].immediate(), args[1].immediate()))
1777                     })*
1778                     _ => {},
1779                 }
1780                 require!(false,
1781                          "unsupported operation on `{}` with element `{}`",
1782                          in_ty,
1783                          in_elem)
1784             })*
1785         }
1786     }
1787     arith_binary! {
1788         simd_add: Uint, Int => add, Float => fadd;
1789         simd_sub: Uint, Int => sub, Float => fsub;
1790         simd_mul: Uint, Int => mul, Float => fmul;
1791         simd_div: Uint => udiv, Int => sdiv, Float => fdiv;
1792         simd_rem: Uint => urem, Int => srem, Float => frem;
1793         simd_shl: Uint, Int => shl;
1794         simd_shr: Uint => lshr, Int => ashr;
1795         simd_and: Uint, Int => and;
1796         simd_or: Uint, Int => or;
1797         simd_xor: Uint, Int => xor;
1798         simd_fmax: Float => maxnum;
1799         simd_fmin: Float => minnum;
1800
1801     }
1802     macro_rules! arith_unary {
1803         ($($name: ident: $($($p: ident),* => $call: ident),*;)*) => {
1804             $(if name == sym::$name {
1805                 match in_elem.kind() {
1806                     $($(ty::$p(_))|* => {
1807                         return Ok(bx.$call(args[0].immediate()))
1808                     })*
1809                     _ => {},
1810                 }
1811                 require!(false,
1812                          "unsupported operation on `{}` with element `{}`",
1813                          in_ty,
1814                          in_elem)
1815             })*
1816         }
1817     }
1818     arith_unary! {
1819         simd_neg: Int => neg, Float => fneg;
1820     }
1821
1822     if name == sym::simd_saturating_add || name == sym::simd_saturating_sub {
1823         let lhs = args[0].immediate();
1824         let rhs = args[1].immediate();
1825         let is_add = name == sym::simd_saturating_add;
1826         let ptr_bits = bx.tcx().data_layout.pointer_size.bits() as _;
1827         let (signed, elem_width, elem_ty) = match *in_elem.kind() {
1828             ty::Int(i) => (true, i.bit_width().unwrap_or(ptr_bits), bx.cx.type_int_from_ty(i)),
1829             ty::Uint(i) => (false, i.bit_width().unwrap_or(ptr_bits), bx.cx.type_uint_from_ty(i)),
1830             _ => {
1831                 return_error!(
1832                     "expected element type `{}` of vector type `{}` \
1833                      to be a signed or unsigned integer type",
1834                     arg_tys[0].simd_size_and_type(bx.tcx()).1,
1835                     arg_tys[0]
1836                 );
1837             }
1838         };
1839         let llvm_intrinsic = &format!(
1840             "llvm.{}{}.sat.v{}i{}",
1841             if signed { 's' } else { 'u' },
1842             if is_add { "add" } else { "sub" },
1843             in_len,
1844             elem_width
1845         );
1846         let vec_ty = bx.cx.type_vector(elem_ty, in_len as u64);
1847
1848         let fn_ty = bx.type_func(&[vec_ty, vec_ty], vec_ty);
1849         let f = bx.declare_cfn(llvm_intrinsic, llvm::UnnamedAddr::No, fn_ty);
1850         let v = bx.call(fn_ty, f, &[lhs, rhs], None);
1851         return Ok(v);
1852     }
1853
1854     span_bug!(span, "unknown SIMD intrinsic");
1855 }
1856
1857 // Returns the width of an int Ty, and if it's signed or not
1858 // Returns None if the type is not an integer
1859 // FIXME: there’s multiple of this functions, investigate using some of the already existing
1860 // stuffs.
1861 fn int_type_width_signed(ty: Ty<'_>, cx: &CodegenCx<'_, '_>) -> Option<(u64, bool)> {
1862     match ty.kind() {
1863         ty::Int(t) => {
1864             Some((t.bit_width().unwrap_or(u64::from(cx.tcx.sess.target.pointer_width)), true))
1865         }
1866         ty::Uint(t) => {
1867             Some((t.bit_width().unwrap_or(u64::from(cx.tcx.sess.target.pointer_width)), false))
1868         }
1869         _ => None,
1870     }
1871 }