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_codegen_ssa::base::{compare_simd_types, wants_msvc_seh};
  11 use rustc_codegen_ssa::common::span_invalid_monomorphization_error;
  12 use rustc_codegen_ssa::common::{IntPredicate, TypeKind};
  13 use rustc_codegen_ssa::mir::operand::OperandRef;
  14 use rustc_codegen_ssa::mir::place::PlaceRef;
  15 use rustc_codegen_ssa::traits::*;
  16 use rustc_hir as hir;
  17 use rustc_middle::ty::layout::{FnAbiOf, HasTyCtxt, LayoutOf};
  18 use rustc_middle::ty::{self, Ty};
  19 use rustc_middle::{bug, span_bug};
  20 use rustc_span::{sym, symbol::kw, Span, Symbol};
  21 use rustc_target::abi::{self, Align, HasDataLayout, Primitive};
  22 use rustc_target::spec::{HasTargetSpec, PanicStrategy};
  23
  24 use std::cmp::Ordering;
  25 use std::iter;
  26
  27 fn get_simple_intrinsic<'ll>(
  28     cx: &CodegenCx<'ll, '_>,
  29     name: Symbol,
  30 ) -> Option<(&'ll Type, &'ll Value)> {
  31     let llvm_name = match name {
  32         sym::sqrtf32 => "llvm.sqrt.f32",
  33         sym::sqrtf64 => "llvm.sqrt.f64",
  34         sym::powif32 => "llvm.powi.f32",
  35         sym::powif64 => "llvm.powi.f64",
  36         sym::sinf32 => "llvm.sin.f32",
  37         sym::sinf64 => "llvm.sin.f64",
  38         sym::cosf32 => "llvm.cos.f32",
  39         sym::cosf64 => "llvm.cos.f64",
  40         sym::powf32 => "llvm.pow.f32",
  41         sym::powf64 => "llvm.pow.f64",
  42         sym::expf32 => "llvm.exp.f32",
  43         sym::expf64 => "llvm.exp.f64",
  44         sym::exp2f32 => "llvm.exp2.f32",
  45         sym::exp2f64 => "llvm.exp2.f64",
  46         sym::logf32 => "llvm.log.f32",
  47         sym::logf64 => "llvm.log.f64",
  48         sym::log10f32 => "llvm.log10.f32",
  49         sym::log10f64 => "llvm.log10.f64",
  50         sym::log2f32 => "llvm.log2.f32",
  51         sym::log2f64 => "llvm.log2.f64",
  52         sym::fmaf32 => "llvm.fma.f32",
  53         sym::fmaf64 => "llvm.fma.f64",
  54         sym::fabsf32 => "llvm.fabs.f32",
  55         sym::fabsf64 => "llvm.fabs.f64",
  56         sym::minnumf32 => "llvm.minnum.f32",
  57         sym::minnumf64 => "llvm.minnum.f64",
  58         sym::maxnumf32 => "llvm.maxnum.f32",
  59         sym::maxnumf64 => "llvm.maxnum.f64",
  60         sym::copysignf32 => "llvm.copysign.f32",
  61         sym::copysignf64 => "llvm.copysign.f64",
  62         sym::floorf32 => "llvm.floor.f32",
  63         sym::floorf64 => "llvm.floor.f64",
  64         sym::ceilf32 => "llvm.ceil.f32",
  65         sym::ceilf64 => "llvm.ceil.f64",
  66         sym::truncf32 => "llvm.trunc.f32",
  67         sym::truncf64 => "llvm.trunc.f64",
  68         sym::rintf32 => "llvm.rint.f32",
  69         sym::rintf64 => "llvm.rint.f64",
  70         sym::nearbyintf32 => "llvm.nearbyint.f32",
  71         sym::nearbyintf64 => "llvm.nearbyint.f64",
  72         sym::roundf32 => "llvm.round.f32",
  73         sym::roundf64 => "llvm.round.f64",
  74         _ => return None,
  75     };
  76     Some(cx.get_intrinsic(llvm_name))
  77 }
  78
  79 impl<'ll, 'tcx> IntrinsicCallMethods<'tcx> for Builder<'_, 'll, 'tcx> {
  80     fn codegen_intrinsic_call(
  81         &mut self,
  82         instance: ty::Instance<'tcx>,
  83         fn_abi: &FnAbi<'tcx, Ty<'tcx>>,
  84         args: &[OperandRef<'tcx, &'ll Value>],
  85         llresult: &'ll Value,
  86         span: Span,
  87     ) {
  88         let tcx = self.tcx;
  89         let callee_ty = instance.ty(tcx, ty::ParamEnv::reveal_all());
  90
  91         let ty::FnDef(def_id, substs) = *callee_ty.kind() else {
  92             bug!("expected fn item type, found {}", callee_ty);
  93         };
  94
  95         let sig = callee_ty.fn_sig(tcx);
  96         let sig = tcx.normalize_erasing_late_bound_regions(ty::ParamEnv::reveal_all(), sig);
  97         let arg_tys = sig.inputs();
  98         let ret_ty = sig.output();
  99         let name = tcx.item_name(def_id);
 100
 101         let llret_ty = self.layout_of(ret_ty).llvm_type(self);
 102         let result = PlaceRef::new_sized(llresult, fn_abi.ret.layout);
 103
 104         let simple = get_simple_intrinsic(self, name);
 105         let llval = match name {
 106             _ if simple.is_some() => {
 107                 let (simple_ty, simple_fn) = simple.unwrap();
 108                 self.call(
 109                     simple_ty,
 110                     simple_fn,
 111                     &args.iter().map(|arg| arg.immediate()).collect::<Vec<_>>(),
 112                     None,
 113                 )
 114             }
 115             sym::likely => {
 116                 self.call_intrinsic("llvm.expect.i1", &[args[0].immediate(), self.const_bool(true)])
 117             }
 118             sym::unlikely => self
 119                 .call_intrinsic("llvm.expect.i1", &[args[0].immediate(), self.const_bool(false)]),
 120             kw::Try => {
 121                 try_intrinsic(
 122                     self,
 123                     args[0].immediate(),
 124                     args[1].immediate(),
 125                     args[2].immediate(),
 126                     llresult,
 127                 );
 128                 return;
 129             }
 130             sym::breakpoint => self.call_intrinsic("llvm.debugtrap", &[]),
 131             sym::va_copy => {
 132                 self.call_intrinsic("llvm.va_copy", &[args[0].immediate(), args[1].immediate()])
 133             }
 134             sym::va_arg => {
 135                 match fn_abi.ret.layout.abi {
 136                     abi::Abi::Scalar(scalar) => {
 137                         match scalar.primitive() {
 138                             Primitive::Int(..) => {
 139                                 if self.cx().size_of(ret_ty).bytes() < 4 {
 140                                     // `va_arg` should not be called on an integer type
 141                                     // less than 4 bytes in length. If it is, promote
 142                                     // the integer to an `i32` and truncate the result
 143                                     // back to the smaller type.
 144                                     let promoted_result = emit_va_arg(self, args[0], tcx.types.i32);
 145                                     self.trunc(promoted_result, llret_ty)
 146                                 } else {
 147                                     emit_va_arg(self, args[0], ret_ty)
 148                                 }
 149                             }
 150                             Primitive::F64 | Primitive::Pointer => {
 151                                 emit_va_arg(self, args[0], ret_ty)
 152                             }
 153                             // `va_arg` should never be used with the return type f32.
 154                             Primitive::F32 => bug!("the va_arg intrinsic does not work with `f32`"),
 155                         }
 156                     }
 157                     _ => bug!("the va_arg intrinsic does not work with non-scalar types"),
 158                 }
 159             }
 160
 161             sym::volatile_load | sym::unaligned_volatile_load => {
 162                 let tp_ty = substs.type_at(0);
 163                 let ptr = args[0].immediate();
 164                 let load = if let PassMode::Cast(ty) = fn_abi.ret.mode {
 165                     let llty = ty.llvm_type(self);
 166                     let ptr = self.pointercast(ptr, self.type_ptr_to(llty));
 167                     self.volatile_load(llty, ptr)
 168                 } else {
 169                     self.volatile_load(self.layout_of(tp_ty).llvm_type(self), ptr)
 170                 };
 171                 let align = if name == sym::unaligned_volatile_load {
 172                     1
 173                 } else {
 174                     self.align_of(tp_ty).bytes() as u32
 175                 };
 176                 unsafe {
 177                     llvm::LLVMSetAlignment(load, align);
 178                 }
 179                 self.to_immediate(load, self.layout_of(tp_ty))
 180             }
 181             sym::volatile_store => {
 182                 let dst = args[0].deref(self.cx());
 183                 args[1].val.volatile_store(self, dst);
 184                 return;
 185             }
 186             sym::unaligned_volatile_store => {
 187                 let dst = args[0].deref(self.cx());
 188                 args[1].val.unaligned_volatile_store(self, dst);
 189                 return;
 190             }
 191             sym::prefetch_read_data
 192             | sym::prefetch_write_data
 193             | sym::prefetch_read_instruction
 194             | sym::prefetch_write_instruction => {
 195                 let (rw, cache_type) = match name {
 196                     sym::prefetch_read_data => (0, 1),
 197                     sym::prefetch_write_data => (1, 1),
 198                     sym::prefetch_read_instruction => (0, 0),
 199                     sym::prefetch_write_instruction => (1, 0),
 200                     _ => bug!(),
 201                 };
 202                 self.call_intrinsic(
 203                     "llvm.prefetch",
 204                     &[
 205                         args[0].immediate(),
 206                         self.const_i32(rw),
 207                         args[1].immediate(),
 208                         self.const_i32(cache_type),
 209                     ],
 210                 )
 211             }
 212             sym::ctlz
 213             | sym::ctlz_nonzero
 214             | sym::cttz
 215             | sym::cttz_nonzero
 216             | sym::ctpop
 217             | sym::bswap
 218             | sym::bitreverse
 219             | sym::rotate_left
 220             | sym::rotate_right
 221             | sym::saturating_add
 222             | sym::saturating_sub => {
 223                 let ty = arg_tys[0];
 224                 match int_type_width_signed(ty, self) {
 225                     Some((width, signed)) => match name {
 226                         sym::ctlz | sym::cttz => {
 227                             let y = self.const_bool(false);
 228                             self.call_intrinsic(
 229                                 &format!("llvm.{}.i{}", name, width),
 230                                 &[args[0].immediate(), y],
 231                             )
 232                         }
 233                         sym::ctlz_nonzero => {
 234                             let y = self.const_bool(true);
 235                             let llvm_name = &format!("llvm.ctlz.i{}", width);
 236                             self.call_intrinsic(llvm_name, &[args[0].immediate(), y])
 237                         }
 238                         sym::cttz_nonzero => {
 239                             let y = self.const_bool(true);
 240                             let llvm_name = &format!("llvm.cttz.i{}", width);
 241                             self.call_intrinsic(llvm_name, &[args[0].immediate(), y])
 242                         }
 243                         sym::ctpop => self.call_intrinsic(
 244                             &format!("llvm.ctpop.i{}", width),
 245                             &[args[0].immediate()],
 246                         ),
 247                         sym::bswap => {
 248                             if width == 8 {
 249                                 args[0].immediate() // byte swap a u8/i8 is just a no-op
 250                             } else {
 251                                 self.call_intrinsic(
 252                                     &format!("llvm.bswap.i{}", width),
 253                                     &[args[0].immediate()],
 254                                 )
 255                             }
 256                         }
 257                         sym::bitreverse => self.call_intrinsic(
 258                             &format!("llvm.bitreverse.i{}", width),
 259                             &[args[0].immediate()],
 260                         ),
 261                         sym::rotate_left | sym::rotate_right => {
 262                             let is_left = name == sym::rotate_left;
 263                             let val = args[0].immediate();
 264                             let raw_shift = args[1].immediate();
 265                             // rotate = funnel shift with first two args the same
 266                             let llvm_name =
 267                                 &format!("llvm.fsh{}.i{}", if is_left { 'l' } else { 'r' }, width);
 268                             self.call_intrinsic(llvm_name, &[val, val, raw_shift])
 269                         }
 270                         sym::saturating_add | sym::saturating_sub => {
 271                             let is_add = name == sym::saturating_add;
 272                             let lhs = args[0].immediate();
 273                             let rhs = args[1].immediate();
 274                             let llvm_name = &format!(
 275                                 "llvm.{}{}.sat.i{}",
 276                                 if signed { 's' } else { 'u' },
 277                                 if is_add { "add" } else { "sub" },
 278                                 width
 279                             );
 280                             self.call_intrinsic(llvm_name, &[lhs, rhs])
 281                         }
 282                         _ => bug!(),
 283                     },
 284                     None => {
 285                         span_invalid_monomorphization_error(
 286                             tcx.sess,
 287                             span,
 288                             &format!(
 289                                 "invalid monomorphization of `{}` intrinsic: \
 290                                       expected basic integer type, found `{}`",
 291                                 name, ty
 292                             ),
 293                         );
 294                         return;
 295                     }
 296                 }
 297             }
 298
 299             sym::raw_eq => {
 300                 use abi::Abi::*;
 301                 let tp_ty = substs.type_at(0);
 302                 let layout = self.layout_of(tp_ty).layout;
 303                 let use_integer_compare = match layout.abi() {
 304                     Scalar(_) | ScalarPair(_, _) => true,
 305                     Uninhabited | Vector { .. } => false,
 306                     Aggregate { .. } => {
 307                         // For rusty ABIs, small aggregates are actually passed
 308                         // as `RegKind::Integer` (see `FnAbi::adjust_for_abi`),
 309                         // so we re-use that same threshold here.
 310                         layout.size() <= self.data_layout().pointer_size * 2
 311                     }
 312                 };
 313
 314                 let a = args[0].immediate();
 315                 let b = args[1].immediate();
 316                 if layout.size().bytes() == 0 {
 317                     self.const_bool(true)
 318                 } else if use_integer_compare {
 319                     let integer_ty = self.type_ix(layout.size().bits());
 320                     let ptr_ty = self.type_ptr_to(integer_ty);
 321                     let a_ptr = self.bitcast(a, ptr_ty);
 322                     let a_val = self.load(integer_ty, a_ptr, layout.align().abi);
 323                     let b_ptr = self.bitcast(b, ptr_ty);
 324                     let b_val = self.load(integer_ty, b_ptr, layout.align().abi);
 325                     self.icmp(IntPredicate::IntEQ, a_val, b_val)
 326                 } else {
 327                     let i8p_ty = self.type_i8p();
 328                     let a_ptr = self.bitcast(a, i8p_ty);
 329                     let b_ptr = self.bitcast(b, i8p_ty);
 330                     let n = self.const_usize(layout.size().bytes());
 331                     let cmp = self.call_intrinsic("memcmp", &[a_ptr, b_ptr, n]);
 332                     match self.cx.sess().target.arch.as_ref() {
 333                         "avr" | "msp430" => self.icmp(IntPredicate::IntEQ, cmp, self.const_i16(0)),
 334                         _ => self.icmp(IntPredicate::IntEQ, cmp, self.const_i32(0)),
 335                     }
 336                 }
 337             }
 338
 339             sym::black_box => {
 340                 args[0].val.store(self, result);
 341
 342                 // We need to "use" the argument in some way LLVM can't introspect, and on
 343                 // targets that support it we can typically leverage inline assembly to do
 344                 // this. LLVM's interpretation of inline assembly is that it's, well, a black
 345                 // box. This isn't the greatest implementation since it probably deoptimizes
 346                 // more than we want, but it's so far good enough.
 347                 crate::asm::inline_asm_call(
 348                     self,
 349                     "",
 350                     "r,~{memory}",
 351                     &[result.llval],
 352                     self.type_void(),
 353                     true,
 354                     false,
 355                     llvm::AsmDialect::Att,
 356                     &[span],
 357                     false,
 358                     None,
 359                 )
 360                 .unwrap_or_else(|| bug!("failed to generate inline asm call for `black_box`"));
 361
 362                 // We have copied the value to `result` already.
 363                 return;
 364             }
 365
 366             _ if name.as_str().starts_with("simd_") => {
 367                 match generic_simd_intrinsic(self, name, callee_ty, args, ret_ty, llret_ty, span) {
 368                     Ok(llval) => llval,
 369                     Err(()) => return,
 370                 }
 371             }
 372
 373             _ => bug!("unknown intrinsic '{}'", name),
 374         };
 375
 376         if !fn_abi.ret.is_ignore() {
 377             if let PassMode::Cast(ty) = fn_abi.ret.mode {
 378                 let ptr_llty = self.type_ptr_to(ty.llvm_type(self));
 379                 let ptr = self.pointercast(result.llval, ptr_llty);
 380                 self.store(llval, ptr, result.align);
 381             } else {
 382                 OperandRef::from_immediate_or_packed_pair(self, llval, result.layout)
 383                     .val
 384                     .store(self, result);
 385             }
 386         }
 387     }
 388
 389     fn abort(&mut self) {
 390         self.call_intrinsic("llvm.trap", &[]);
 391     }
 392
 393     fn assume(&mut self, val: Self::Value) {
 394         self.call_intrinsic("llvm.assume", &[val]);
 395     }
 396
 397     fn expect(&mut self, cond: Self::Value, expected: bool) -> Self::Value {
 398         self.call_intrinsic("llvm.expect.i1", &[cond, self.const_bool(expected)])
 399     }
 400
 401     fn type_test(&mut self, pointer: Self::Value, typeid: Self::Value) -> Self::Value {
 402         // Test the called operand using llvm.type.test intrinsic. The LowerTypeTests link-time
 403         // optimization pass replaces calls to this intrinsic with code to test type membership.
 404         let i8p_ty = self.type_i8p();
 405         let bitcast = self.bitcast(pointer, i8p_ty);
 406         self.call_intrinsic("llvm.type.test", &[bitcast, typeid])
 407     }
 408
 409     fn va_start(&mut self, va_list: &'ll Value) -> &'ll Value {
 410         self.call_intrinsic("llvm.va_start", &[va_list])
 411     }
 412
 413     fn va_end(&mut self, va_list: &'ll Value) -> &'ll Value {
 414         self.call_intrinsic("llvm.va_end", &[va_list])
 415     }
 416 }
 417
 418 fn try_intrinsic<'ll>(
 419     bx: &mut Builder<'_, 'll, '_>,
 420     try_func: &'ll Value,
 421     data: &'ll Value,
 422     catch_func: &'ll Value,
 423     dest: &'ll Value,
 424 ) {
 425     if bx.sess().panic_strategy() == PanicStrategy::Abort {
 426         let try_func_ty = bx.type_func(&[bx.type_i8p()], bx.type_void());
 427         bx.call(try_func_ty, try_func, &[data], None);
 428         // Return 0 unconditionally from the intrinsic call;
 429         // we can never unwind.
 430         let ret_align = bx.tcx().data_layout.i32_align.abi;
 431         bx.store(bx.const_i32(0), dest, ret_align);
 432     } else if wants_msvc_seh(bx.sess()) {
 433         codegen_msvc_try(bx, try_func, data, catch_func, dest);
 434     } else if bx.sess().target.is_like_emscripten {
 435         codegen_emcc_try(bx, try_func, data, catch_func, dest);
 436     } else {
 437         codegen_gnu_try(bx, try_func, data, catch_func, dest);
 438     }
 439 }
 440
 441 // MSVC's definition of the `rust_try` function.
 442 //
 443 // This implementation uses the new exception handling instructions in LLVM
 444 // which have support in LLVM for SEH on MSVC targets. Although these
 445 // instructions are meant to work for all targets, as of the time of this
 446 // writing, however, LLVM does not recommend the usage of these new instructions
 447 // as the old ones are still more optimized.
 448 fn codegen_msvc_try<'ll>(
 449     bx: &mut Builder<'_, 'll, '_>,
 450     try_func: &'ll Value,
 451     data: &'ll Value,
 452     catch_func: &'ll Value,
 453     dest: &'ll Value,
 454 ) {
 455     let (llty, llfn) = get_rust_try_fn(bx, &mut |mut bx| {
 456         bx.set_personality_fn(bx.eh_personality());
 457
 458         let normal = bx.append_sibling_block("normal");
 459         let catchswitch = bx.append_sibling_block("catchswitch");
 460         let catchpad_rust = bx.append_sibling_block("catchpad_rust");
 461         let catchpad_foreign = bx.append_sibling_block("catchpad_foreign");
 462         let caught = bx.append_sibling_block("caught");
 463
 464         let try_func = llvm::get_param(bx.llfn(), 0);
 465         let data = llvm::get_param(bx.llfn(), 1);
 466         let catch_func = llvm::get_param(bx.llfn(), 2);
 467
 468         // We're generating an IR snippet that looks like:
 469         //
 470         //   declare i32 @rust_try(%try_func, %data, %catch_func) {
 471         //      %slot = alloca i8*
 472         //      invoke %try_func(%data) to label %normal unwind label %catchswitch
 473         //
 474         //   normal:
 475         //      ret i32 0
 476         //
 477         //   catchswitch:
 478         //      %cs = catchswitch within none [%catchpad_rust, %catchpad_foreign] unwind to caller
 479         //
 480         //   catchpad_rust:
 481         //      %tok = catchpad within %cs [%type_descriptor, 8, %slot]
 482         //      %ptr = load %slot
 483         //      call %catch_func(%data, %ptr)
 484         //      catchret from %tok to label %caught
 485         //
 486         //   catchpad_foreign:
 487         //      %tok = catchpad within %cs [null, 64, null]
 488         //      call %catch_func(%data, null)
 489         //      catchret from %tok to label %caught
 490         //
 491         //   caught:
 492         //      ret i32 1
 493         //   }
 494         //
 495         // This structure follows the basic usage of throw/try/catch in LLVM.
 496         // For example, compile this C++ snippet to see what LLVM generates:
 497         //
 498         //      struct rust_panic {
 499         //          rust_panic(const rust_panic&);
 500         //          ~rust_panic();
 501         //
 502         //          void* x[2];
 503         //      };
 504         //
 505         //      int __rust_try(
 506         //          void (*try_func)(void*),
 507         //          void *data,
 508         //          void (*catch_func)(void*, void*) noexcept
 509         //      ) {
 510         //          try {
 511         //              try_func(data);
 512         //              return 0;
 513         //          } catch(rust_panic& a) {
 514         //              catch_func(data, &a);
 515         //              return 1;
 516         //          } catch(...) {
 517         //              catch_func(data, NULL);
 518         //              return 1;
 519         //          }
 520         //      }
 521         //
 522         // More information can be found in libstd's seh.rs implementation.
 523         let ptr_align = bx.tcx().data_layout.pointer_align.abi;
 524         let slot = bx.alloca(bx.type_i8p(), ptr_align);
 525         let try_func_ty = bx.type_func(&[bx.type_i8p()], bx.type_void());
 526         bx.invoke(try_func_ty, try_func, &[data], normal, catchswitch, None);
 527
 528         bx.switch_to_block(normal);
 529         bx.ret(bx.const_i32(0));
 530
 531         bx.switch_to_block(catchswitch);
 532         let cs = bx.catch_switch(None, None, &[catchpad_rust, catchpad_foreign]);
 533
 534         // We can't use the TypeDescriptor defined in libpanic_unwind because it
 535         // might be in another DLL and the SEH encoding only supports specifying
 536         // a TypeDescriptor from the current module.
 537         //
 538         // However this isn't an issue since the MSVC runtime uses string
 539         // comparison on the type name to match TypeDescriptors rather than
 540         // pointer equality.
 541         //
 542         // So instead we generate a new TypeDescriptor in each module that uses
 543         // `try` and let the linker merge duplicate definitions in the same
 544         // module.
 545         //
 546         // When modifying, make sure that the type_name string exactly matches
 547         // the one used in src/libpanic_unwind/seh.rs.
 548         let type_info_vtable = bx.declare_global("??_7type_info@@6B@", bx.type_i8p());
 549         let type_name = bx.const_bytes(b"rust_panic\0");
 550         let type_info =
 551             bx.const_struct(&[type_info_vtable, bx.const_null(bx.type_i8p()), type_name], false);
 552         let tydesc = bx.declare_global("__rust_panic_type_info", bx.val_ty(type_info));
 553         unsafe {
 554             llvm::LLVMRustSetLinkage(tydesc, llvm::Linkage::LinkOnceODRLinkage);
 555             llvm::SetUniqueComdat(bx.llmod, tydesc);
 556             llvm::LLVMSetInitializer(tydesc, type_info);
 557         }
 558
 559         // The flag value of 8 indicates that we are catching the exception by
 560         // reference instead of by value. We can't use catch by value because
 561         // that requires copying the exception object, which we don't support
 562         // since our exception object effectively contains a Box.
 563         //
 564         // Source: MicrosoftCXXABI::getAddrOfCXXCatchHandlerType in clang
 565         bx.switch_to_block(catchpad_rust);
 566         let flags = bx.const_i32(8);
 567         let funclet = bx.catch_pad(cs, &[tydesc, flags, slot]);
 568         let ptr = bx.load(bx.type_i8p(), slot, ptr_align);
 569         let catch_ty = bx.type_func(&[bx.type_i8p(), bx.type_i8p()], bx.type_void());
 570         bx.call(catch_ty, catch_func, &[data, ptr], Some(&funclet));
 571         bx.catch_ret(&funclet, caught);
 572
 573         // The flag value of 64 indicates a "catch-all".
 574         bx.switch_to_block(catchpad_foreign);
 575         let flags = bx.const_i32(64);
 576         let null = bx.const_null(bx.type_i8p());
 577         let funclet = bx.catch_pad(cs, &[null, flags, null]);
 578         bx.call(catch_ty, catch_func, &[data, null], Some(&funclet));
 579         bx.catch_ret(&funclet, caught);
 580
 581         bx.switch_to_block(caught);
 582         bx.ret(bx.const_i32(1));
 583     });
 584
 585     // Note that no invoke is used here because by definition this function
 586     // can't panic (that's what it's catching).
 587     let ret = bx.call(llty, llfn, &[try_func, data, catch_func], None);
 588     let i32_align = bx.tcx().data_layout.i32_align.abi;
 589     bx.store(ret, dest, i32_align);
 590 }
 591
 592 // Definition of the standard `try` function for Rust using the GNU-like model
 593 // of exceptions (e.g., the normal semantics of LLVM's `landingpad` and `invoke`
 594 // instructions).
 595 //
 596 // This codegen is a little surprising because we always call a shim
 597 // function instead of inlining the call to `invoke` manually here. This is done
 598 // because in LLVM we're only allowed to have one personality per function
 599 // definition. The call to the `try` intrinsic is being inlined into the
 600 // function calling it, and that function may already have other personality
 601 // functions in play. By calling a shim we're guaranteed that our shim will have
 602 // the right personality function.
 603 fn codegen_gnu_try<'ll>(
 604     bx: &mut Builder<'_, 'll, '_>,
 605     try_func: &'ll Value,
 606     data: &'ll Value,
 607     catch_func: &'ll Value,
 608     dest: &'ll Value,
 609 ) {
 610     let (llty, llfn) = get_rust_try_fn(bx, &mut |mut bx| {
 611         // Codegens the shims described above:
 612         //
 613         //   bx:
 614         //      invoke %try_func(%data) normal %normal unwind %catch
 615         //
 616         //   normal:
 617         //      ret 0
 618         //
 619         //   catch:
 620         //      (%ptr, _) = landingpad
 621         //      call %catch_func(%data, %ptr)
 622         //      ret 1
 623         let then = bx.append_sibling_block("then");
 624         let catch = bx.append_sibling_block("catch");
 625
 626         let try_func = llvm::get_param(bx.llfn(), 0);
 627         let data = llvm::get_param(bx.llfn(), 1);
 628         let catch_func = llvm::get_param(bx.llfn(), 2);
 629         let try_func_ty = bx.type_func(&[bx.type_i8p()], bx.type_void());
 630         bx.invoke(try_func_ty, try_func, &[data], then, catch, None);
 631
 632         bx.switch_to_block(then);
 633         bx.ret(bx.const_i32(0));
 634
 635         // Type indicator for the exception being thrown.
 636         //
 637         // The first value in this tuple is a pointer to the exception object
 638         // being thrown.  The second value is a "selector" indicating which of
 639         // the landing pad clauses the exception's type had been matched to.
 640         // rust_try ignores the selector.
 641         bx.switch_to_block(catch);
 642         let lpad_ty = bx.type_struct(&[bx.type_i8p(), bx.type_i32()], false);
 643         let vals = bx.landing_pad(lpad_ty, bx.eh_personality(), 1);
 644         let tydesc = bx.const_null(bx.type_i8p());
 645         bx.add_clause(vals, tydesc);
 646         let ptr = bx.extract_value(vals, 0);
 647         let catch_ty = bx.type_func(&[bx.type_i8p(), bx.type_i8p()], bx.type_void());
 648         bx.call(catch_ty, catch_func, &[data, ptr], None);
 649         bx.ret(bx.const_i32(1));
 650     });
 651
 652     // Note that no invoke is used here because by definition this function
 653     // can't panic (that's what it's catching).
 654     let ret = bx.call(llty, llfn, &[try_func, data, catch_func], None);
 655     let i32_align = bx.tcx().data_layout.i32_align.abi;
 656     bx.store(ret, dest, i32_align);
 657 }
 658
 659 // Variant of codegen_gnu_try used for emscripten where Rust panics are
 660 // implemented using C++ exceptions. Here we use exceptions of a specific type
 661 // (`struct rust_panic`) to represent Rust panics.
 662 fn codegen_emcc_try<'ll>(
 663     bx: &mut Builder<'_, 'll, '_>,
 664     try_func: &'ll Value,
 665     data: &'ll Value,
 666     catch_func: &'ll Value,
 667     dest: &'ll Value,
 668 ) {
 669     let (llty, llfn) = get_rust_try_fn(bx, &mut |mut bx| {
 670         // Codegens the shims described above:
 671         //
 672         //   bx:
 673         //      invoke %try_func(%data) normal %normal unwind %catch
 674         //
 675         //   normal:
 676         //      ret 0
 677         //
 678         //   catch:
 679         //      (%ptr, %selector) = landingpad
 680         //      %rust_typeid = @llvm.eh.typeid.for(@_ZTI10rust_panic)
 681         //      %is_rust_panic = %selector == %rust_typeid
 682         //      %catch_data = alloca { i8*, i8 }
 683         //      %catch_data[0] = %ptr
 684         //      %catch_data[1] = %is_rust_panic
 685         //      call %catch_func(%data, %catch_data)
 686         //      ret 1
 687         let then = bx.append_sibling_block("then");
 688         let catch = bx.append_sibling_block("catch");
 689
 690         let try_func = llvm::get_param(bx.llfn(), 0);
 691         let data = llvm::get_param(bx.llfn(), 1);
 692         let catch_func = llvm::get_param(bx.llfn(), 2);
 693         let try_func_ty = bx.type_func(&[bx.type_i8p()], bx.type_void());
 694         bx.invoke(try_func_ty, try_func, &[data], then, catch, None);
 695
 696         bx.switch_to_block(then);
 697         bx.ret(bx.const_i32(0));
 698
 699         // Type indicator for the exception being thrown.
 700         //
 701         // The first value in this tuple is a pointer to the exception object
 702         // being thrown.  The second value is a "selector" indicating which of
 703         // the landing pad clauses the exception's type had been matched to.
 704         bx.switch_to_block(catch);
 705         let tydesc = bx.eh_catch_typeinfo();
 706         let lpad_ty = bx.type_struct(&[bx.type_i8p(), bx.type_i32()], false);
 707         let vals = bx.landing_pad(lpad_ty, bx.eh_personality(), 2);
 708         bx.add_clause(vals, tydesc);
 709         bx.add_clause(vals, bx.const_null(bx.type_i8p()));
 710         let ptr = bx.extract_value(vals, 0);
 711         let selector = bx.extract_value(vals, 1);
 712
 713         // Check if the typeid we got is the one for a Rust panic.
 714         let rust_typeid = bx.call_intrinsic("llvm.eh.typeid.for", &[tydesc]);
 715         let is_rust_panic = bx.icmp(IntPredicate::IntEQ, selector, rust_typeid);
 716         let is_rust_panic = bx.zext(is_rust_panic, bx.type_bool());
 717
 718         // We need to pass two values to catch_func (ptr and is_rust_panic), so
 719         // create an alloca and pass a pointer to that.
 720         let ptr_align = bx.tcx().data_layout.pointer_align.abi;
 721         let i8_align = bx.tcx().data_layout.i8_align.abi;
 722         let catch_data_type = bx.type_struct(&[bx.type_i8p(), bx.type_bool()], false);
 723         let catch_data = bx.alloca(catch_data_type, ptr_align);
 724         let catch_data_0 =
 725             bx.inbounds_gep(catch_data_type, catch_data, &[bx.const_usize(0), bx.const_usize(0)]);
 726         bx.store(ptr, catch_data_0, ptr_align);
 727         let catch_data_1 =
 728             bx.inbounds_gep(catch_data_type, catch_data, &[bx.const_usize(0), bx.const_usize(1)]);
 729         bx.store(is_rust_panic, catch_data_1, i8_align);
 730         let catch_data = bx.bitcast(catch_data, bx.type_i8p());
 731
 732         let catch_ty = bx.type_func(&[bx.type_i8p(), bx.type_i8p()], bx.type_void());
 733         bx.call(catch_ty, catch_func, &[data, catch_data], None);
 734         bx.ret(bx.const_i32(1));
 735     });
 736
 737     // Note that no invoke is used here because by definition this function
 738     // can't panic (that's what it's catching).
 739     let ret = bx.call(llty, llfn, &[try_func, data, catch_func], None);
 740     let i32_align = bx.tcx().data_layout.i32_align.abi;
 741     bx.store(ret, dest, i32_align);
 742 }
 743
 744 // Helper function to give a Block to a closure to codegen a shim function.
 745 // This is currently primarily used for the `try` intrinsic functions above.
 746 fn gen_fn<'ll, 'tcx>(
 747     cx: &CodegenCx<'ll, 'tcx>,
 748     name: &str,
 749     rust_fn_sig: ty::PolyFnSig<'tcx>,
 750     codegen: &mut dyn FnMut(Builder<'_, 'll, 'tcx>),
 751 ) -> (&'ll Type, &'ll Value) {
 752     let fn_abi = cx.fn_abi_of_fn_ptr(rust_fn_sig, ty::List::empty());
 753     let llty = fn_abi.llvm_type(cx);
 754     let llfn = cx.declare_fn(name, fn_abi);
 755     cx.set_frame_pointer_type(llfn);
 756     cx.apply_target_cpu_attr(llfn);
 757     // FIXME(eddyb) find a nicer way to do this.
 758     unsafe { llvm::LLVMRustSetLinkage(llfn, llvm::Linkage::InternalLinkage) };
 759     let llbb = Builder::append_block(cx, llfn, "entry-block");
 760     let bx = Builder::build(cx, llbb);
 761     codegen(bx);
 762     (llty, llfn)
 763 }
 764
 765 // Helper function used to get a handle to the `__rust_try` function used to
 766 // catch exceptions.
 767 //
 768 // This function is only generated once and is then cached.
 769 fn get_rust_try_fn<'ll, 'tcx>(
 770     cx: &CodegenCx<'ll, 'tcx>,
 771     codegen: &mut dyn FnMut(Builder<'_, 'll, 'tcx>),
 772 ) -> (&'ll Type, &'ll Value) {
 773     if let Some(llfn) = cx.rust_try_fn.get() {
 774         return llfn;
 775     }
 776
 777     // Define the type up front for the signature of the rust_try function.
 778     let tcx = cx.tcx;
 779     let i8p = tcx.mk_mut_ptr(tcx.types.i8);
 780     // `unsafe fn(*mut i8) -> ()`
 781     let try_fn_ty = tcx.mk_fn_ptr(ty::Binder::dummy(tcx.mk_fn_sig(
 782         iter::once(i8p),
 783         tcx.mk_unit(),
 784         false,
 785         hir::Unsafety::Unsafe,
 786         Abi::Rust,
 787     )));
 788     // `unsafe fn(*mut i8, *mut i8) -> ()`
 789     let catch_fn_ty = tcx.mk_fn_ptr(ty::Binder::dummy(tcx.mk_fn_sig(
 790         [i8p, i8p].iter().cloned(),
 791         tcx.mk_unit(),
 792         false,
 793         hir::Unsafety::Unsafe,
 794         Abi::Rust,
 795     )));
 796     // `unsafe fn(unsafe fn(*mut i8) -> (), *mut i8, unsafe fn(*mut i8, *mut i8) -> ()) -> i32`
 797     let rust_fn_sig = ty::Binder::dummy(cx.tcx.mk_fn_sig(
 798         [try_fn_ty, i8p, catch_fn_ty].into_iter(),
 799         tcx.types.i32,
 800         false,
 801         hir::Unsafety::Unsafe,
 802         Abi::Rust,
 803     ));
 804     let rust_try = gen_fn(cx, "__rust_try", rust_fn_sig, codegen);
 805     cx.rust_try_fn.set(Some(rust_try));
 806     rust_try
 807 }
 808
 809 fn generic_simd_intrinsic<'ll, 'tcx>(
 810     bx: &mut Builder<'_, 'll, 'tcx>,
 811     name: Symbol,
 812     callee_ty: Ty<'tcx>,
 813     args: &[OperandRef<'tcx, &'ll Value>],
 814     ret_ty: Ty<'tcx>,
 815     llret_ty: &'ll Type,
 816     span: Span,
 817 ) -> Result<&'ll Value, ()> {
 818     // macros for error handling:
 819     macro_rules! emit_error {
 820         ($msg: tt) => {
 821             emit_error!($msg, )
 822         };
 823         ($msg: tt, $($fmt: tt)*) => {
 824             span_invalid_monomorphization_error(
 825                 bx.sess(), span,
 826                 &format!(concat!("invalid monomorphization of `{}` intrinsic: ", $msg),
 827                          name, $($fmt)*));
 828         }
 829     }
 830
 831     macro_rules! return_error {
 832         ($($fmt: tt)*) => {
 833             {
 834                 emit_error!($($fmt)*);
 835                 return Err(());
 836             }
 837         }
 838     }
 839
 840     macro_rules! require {
 841         ($cond: expr, $($fmt: tt)*) => {
 842             if !$cond {
 843                 return_error!($($fmt)*);
 844             }
 845         };
 846     }
 847
 848     macro_rules! require_simd {
 849         ($ty: expr, $position: expr) => {
 850             require!($ty.is_simd(), "expected SIMD {} type, found non-SIMD `{}`", $position, $ty)
 851         };
 852     }
 853
 854     let tcx = bx.tcx();
 855     let sig =
 856         tcx.normalize_erasing_late_bound_regions(ty::ParamEnv::reveal_all(), callee_ty.fn_sig(tcx));
 857     let arg_tys = sig.inputs();
 858
 859     if name == sym::simd_select_bitmask {
 860         require_simd!(arg_tys[1], "argument");
 861         let (len, _) = arg_tys[1].simd_size_and_type(bx.tcx());
 862
 863         let expected_int_bits = (len.max(8) - 1).next_power_of_two();
 864         let expected_bytes = len / 8 + ((len % 8 > 0) as u64);
 865
 866         let mask_ty = arg_tys[0];
 867         let mask = match mask_ty.kind() {
 868             ty::Int(i) if i.bit_width() == Some(expected_int_bits) => args[0].immediate(),
 869             ty::Uint(i) if i.bit_width() == Some(expected_int_bits) => args[0].immediate(),
 870             ty::Array(elem, len)
 871                 if matches!(elem.kind(), ty::Uint(ty::UintTy::U8))
 872                     && len.try_eval_usize(bx.tcx, ty::ParamEnv::reveal_all())
 873                         == Some(expected_bytes) =>
 874             {
 875                 let place = PlaceRef::alloca(bx, args[0].layout);
 876                 args[0].val.store(bx, place);
 877                 let int_ty = bx.type_ix(expected_bytes * 8);
 878                 let ptr = bx.pointercast(place.llval, bx.cx.type_ptr_to(int_ty));
 879                 bx.load(int_ty, ptr, Align::ONE)
 880             }
 881             _ => return_error!(
 882                 "invalid bitmask `{}`, expected `u{}` or `[u8; {}]`",
 883                 mask_ty,
 884                 expected_int_bits,
 885                 expected_bytes
 886             ),
 887         };
 888
 889         let i1 = bx.type_i1();
 890         let im = bx.type_ix(len);
 891         let i1xn = bx.type_vector(i1, len);
 892         let m_im = bx.trunc(mask, im);
 893         let m_i1s = bx.bitcast(m_im, i1xn);
 894         return Ok(bx.select(m_i1s, args[1].immediate(), args[2].immediate()));
 895     }
 896
 897     // every intrinsic below takes a SIMD vector as its first argument
 898     require_simd!(arg_tys[0], "input");
 899     let in_ty = arg_tys[0];
 900
 901     let comparison = match name {
 902         sym::simd_eq => Some(hir::BinOpKind::Eq),
 903         sym::simd_ne => Some(hir::BinOpKind::Ne),
 904         sym::simd_lt => Some(hir::BinOpKind::Lt),
 905         sym::simd_le => Some(hir::BinOpKind::Le),
 906         sym::simd_gt => Some(hir::BinOpKind::Gt),
 907         sym::simd_ge => Some(hir::BinOpKind::Ge),
 908         _ => None,
 909     };
 910
 911     let (in_len, in_elem) = arg_tys[0].simd_size_and_type(bx.tcx());
 912     if let Some(cmp_op) = comparison {
 913         require_simd!(ret_ty, "return");
 914
 915         let (out_len, out_ty) = ret_ty.simd_size_and_type(bx.tcx());
 916         require!(
 917             in_len == out_len,
 918             "expected return type with length {} (same as input type `{}`), \
 919              found `{}` with length {}",
 920             in_len,
 921             in_ty,
 922             ret_ty,
 923             out_len
 924         );
 925         require!(
 926             bx.type_kind(bx.element_type(llret_ty)) == TypeKind::Integer,
 927             "expected return type with integer elements, found `{}` with non-integer `{}`",
 928             ret_ty,
 929             out_ty
 930         );
 931
 932         return Ok(compare_simd_types(
 933             bx,
 934             args[0].immediate(),
 935             args[1].immediate(),
 936             in_elem,
 937             llret_ty,
 938             cmp_op,
 939         ));
 940     }
 941
 942     if let Some(stripped) = name.as_str().strip_prefix("simd_shuffle") {
 943         // If this intrinsic is the older "simd_shuffleN" form, simply parse the integer.
 944         // If there is no suffix, use the index array length.
 945         let n: u64 = if stripped.is_empty() {
 946             // Make sure this is actually an array, since typeck only checks the length-suffixed
 947             // version of this intrinsic.
 948             match args[2].layout.ty.kind() {
 949                 ty::Array(ty, len) if matches!(ty.kind(), ty::Uint(ty::UintTy::U32)) => {
 950                     len.try_eval_usize(bx.cx.tcx, ty::ParamEnv::reveal_all()).unwrap_or_else(|| {
 951                         span_bug!(span, "could not evaluate shuffle index array length")
 952                     })
 953                 }
 954                 _ => return_error!(
 955                     "simd_shuffle index must be an array of `u32`, got `{}`",
 956                     args[2].layout.ty
 957                 ),
 958             }
 959         } else {
 960             stripped.parse().unwrap_or_else(|_| {
 961                 span_bug!(span, "bad `simd_shuffle` instruction only caught in codegen?")
 962             })
 963         };
 964
 965         require_simd!(ret_ty, "return");
 966         let (out_len, out_ty) = ret_ty.simd_size_and_type(bx.tcx());
 967         require!(
 968             out_len == n,
 969             "expected return type of length {}, found `{}` with length {}",
 970             n,
 971             ret_ty,
 972             out_len
 973         );
 974         require!(
 975             in_elem == out_ty,
 976             "expected return element type `{}` (element of input `{}`), \
 977              found `{}` with element type `{}`",
 978             in_elem,
 979             in_ty,
 980             ret_ty,
 981             out_ty
 982         );
 983
 984         let total_len = u128::from(in_len) * 2;
 985
 986         let vector = args[2].immediate();
 987
 988         let indices: Option<Vec<_>> = (0..n)
 989             .map(|i| {
 990                 let arg_idx = i;
 991                 let val = bx.const_get_elt(vector, i as u64);
 992                 match bx.const_to_opt_u128(val, true) {
 993                     None => {
 994                         emit_error!("shuffle index #{} is not a constant", arg_idx);
 995                         None
 996                     }
 997                     Some(idx) if idx >= total_len => {
 998                         emit_error!(
 999                             "shuffle index #{} is out of bounds (limit {})",
1000                             arg_idx,
1001                             total_len
1002                         );
1003                         None
1004                     }
1005                     Some(idx) => Some(bx.const_i32(idx as i32)),
1006                 }
1007             })
1008             .collect();
1009         let Some(indices) = indices else {
1010             return Ok(bx.const_null(llret_ty));
1011         };
1012
1013         return Ok(bx.shuffle_vector(
1014             args[0].immediate(),
1015             args[1].immediate(),
1016             bx.const_vector(&indices),
1017         ));
1018     }
1019
1020     if name == sym::simd_insert {
1021         require!(
1022             in_elem == arg_tys[2],
1023             "expected inserted type `{}` (element of input `{}`), found `{}`",
1024             in_elem,
1025             in_ty,
1026             arg_tys[2]
1027         );
1028         return Ok(bx.insert_element(
1029             args[0].immediate(),
1030             args[2].immediate(),
1031             args[1].immediate(),
1032         ));
1033     }
1034     if name == sym::simd_extract {
1035         require!(
1036             ret_ty == in_elem,
1037             "expected return type `{}` (element of input `{}`), found `{}`",
1038             in_elem,
1039             in_ty,
1040             ret_ty
1041         );
1042         return Ok(bx.extract_element(args[0].immediate(), args[1].immediate()));
1043     }
1044
1045     if name == sym::simd_select {
1046         let m_elem_ty = in_elem;
1047         let m_len = in_len;
1048         require_simd!(arg_tys[1], "argument");
1049         let (v_len, _) = arg_tys[1].simd_size_and_type(bx.tcx());
1050         require!(
1051             m_len == v_len,
1052             "mismatched lengths: mask length `{}` != other vector length `{}`",
1053             m_len,
1054             v_len
1055         );
1056         match m_elem_ty.kind() {
1057             ty::Int(_) => {}
1058             _ => return_error!("mask element type is `{}`, expected `i_`", m_elem_ty),
1059         }
1060         // truncate the mask to a vector of i1s
1061         let i1 = bx.type_i1();
1062         let i1xn = bx.type_vector(i1, m_len as u64);
1063         let m_i1s = bx.trunc(args[0].immediate(), i1xn);
1064         return Ok(bx.select(m_i1s, args[1].immediate(), args[2].immediate()));
1065     }
1066
1067     if name == sym::simd_bitmask {
1068         // The `fn simd_bitmask(vector) -> unsigned integer` intrinsic takes a
1069         // vector mask and returns the most significant bit (MSB) of each lane in the form
1070         // of either:
1071         // * an unsigned integer
1072         // * an array of `u8`
1073         // If the vector has less than 8 lanes, a u8 is returned with zeroed trailing bits.
1074         //
1075         // The bit order of the result depends on the byte endianness, LSB-first for little
1076         // endian and MSB-first for big endian.
1077         let expected_int_bits = in_len.max(8);
1078         let expected_bytes = expected_int_bits / 8 + ((expected_int_bits % 8 > 0) as u64);
1079
1080         // Integer vector <i{in_bitwidth} x in_len>:
1081         let (i_xn, in_elem_bitwidth) = match in_elem.kind() {
1082             ty::Int(i) => (
1083                 args[0].immediate(),
1084                 i.bit_width().unwrap_or_else(|| bx.data_layout().pointer_size.bits()),
1085             ),
1086             ty::Uint(i) => (
1087                 args[0].immediate(),
1088                 i.bit_width().unwrap_or_else(|| bx.data_layout().pointer_size.bits()),
1089             ),
1090             _ => return_error!(
1091                 "vector argument `{}`'s element type `{}`, expected integer element type",
1092                 in_ty,
1093                 in_elem
1094             ),
1095         };
1096
1097         // Shift the MSB to the right by "in_elem_bitwidth - 1" into the first bit position.
1098         let shift_indices =
1099             vec![
1100                 bx.cx.const_int(bx.type_ix(in_elem_bitwidth), (in_elem_bitwidth - 1) as _);
1101                 in_len as _
1102             ];
1103         let i_xn_msb = bx.lshr(i_xn, bx.const_vector(shift_indices.as_slice()));
1104         // Truncate vector to an <i1 x N>
1105         let i1xn = bx.trunc(i_xn_msb, bx.type_vector(bx.type_i1(), in_len));
1106         // Bitcast <i1 x N> to iN:
1107         let i_ = bx.bitcast(i1xn, bx.type_ix(in_len));
1108
1109         match ret_ty.kind() {
1110             ty::Uint(i) if i.bit_width() == Some(expected_int_bits) => {
1111                 // Zero-extend iN to the bitmask type:
1112                 return Ok(bx.zext(i_, bx.type_ix(expected_int_bits)));
1113             }
1114             ty::Array(elem, len)
1115                 if matches!(elem.kind(), ty::Uint(ty::UintTy::U8))
1116                     && len.try_eval_usize(bx.tcx, ty::ParamEnv::reveal_all())
1117                         == Some(expected_bytes) =>
1118             {
1119                 // Zero-extend iN to the array length:
1120                 let ze = bx.zext(i_, bx.type_ix(expected_bytes * 8));
1121
1122                 // Convert the integer to a byte array
1123                 let ptr = bx.alloca(bx.type_ix(expected_bytes * 8), Align::ONE);
1124                 bx.store(ze, ptr, Align::ONE);
1125                 let array_ty = bx.type_array(bx.type_i8(), expected_bytes);
1126                 let ptr = bx.pointercast(ptr, bx.cx.type_ptr_to(array_ty));
1127                 return Ok(bx.load(array_ty, ptr, Align::ONE));
1128             }
1129             _ => return_error!(
1130                 "cannot return `{}`, expected `u{}` or `[u8; {}]`",
1131                 ret_ty,
1132                 expected_int_bits,
1133                 expected_bytes
1134             ),
1135         }
1136     }
1137
1138     fn simd_simple_float_intrinsic<'ll, 'tcx>(
1139         name: Symbol,
1140         in_elem: Ty<'_>,
1141         in_ty: Ty<'_>,
1142         in_len: u64,
1143         bx: &mut Builder<'_, 'll, 'tcx>,
1144         span: Span,
1145         args: &[OperandRef<'tcx, &'ll Value>],
1146     ) -> Result<&'ll Value, ()> {
1147         macro_rules! emit_error {
1148             ($msg: tt) => {
1149                 emit_error!($msg, )
1150             };
1151             ($msg: tt, $($fmt: tt)*) => {
1152                 span_invalid_monomorphization_error(
1153                     bx.sess(), span,
1154                     &format!(concat!("invalid monomorphization of `{}` intrinsic: ", $msg),
1155                              name, $($fmt)*));
1156             }
1157         }
1158         macro_rules! return_error {
1159             ($($fmt: tt)*) => {
1160                 {
1161                     emit_error!($($fmt)*);
1162                     return Err(());
1163                 }
1164             }
1165         }
1166
1167         let (elem_ty_str, elem_ty) = if let ty::Float(f) = in_elem.kind() {
1168             let elem_ty = bx.cx.type_float_from_ty(*f);
1169             match f.bit_width() {
1170                 32 => ("f32", elem_ty),
1171                 64 => ("f64", elem_ty),
1172                 _ => {
1173                     return_error!(
1174                         "unsupported element type `{}` of floating-point vector `{}`",
1175                         f.name_str(),
1176                         in_ty
1177                     );
1178                 }
1179             }
1180         } else {
1181             return_error!("`{}` is not a floating-point type", in_ty);
1182         };
1183
1184         let vec_ty = bx.type_vector(elem_ty, in_len);
1185
1186         let (intr_name, fn_ty) = match name {
1187             sym::simd_ceil => ("ceil", bx.type_func(&[vec_ty], vec_ty)),
1188             sym::simd_fabs => ("fabs", bx.type_func(&[vec_ty], vec_ty)),
1189             sym::simd_fcos => ("cos", bx.type_func(&[vec_ty], vec_ty)),
1190             sym::simd_fexp2 => ("exp2", bx.type_func(&[vec_ty], vec_ty)),
1191             sym::simd_fexp => ("exp", bx.type_func(&[vec_ty], vec_ty)),
1192             sym::simd_flog10 => ("log10", bx.type_func(&[vec_ty], vec_ty)),
1193             sym::simd_flog2 => ("log2", bx.type_func(&[vec_ty], vec_ty)),
1194             sym::simd_flog => ("log", bx.type_func(&[vec_ty], vec_ty)),
1195             sym::simd_floor => ("floor", bx.type_func(&[vec_ty], vec_ty)),
1196             sym::simd_fma => ("fma", bx.type_func(&[vec_ty, vec_ty, vec_ty], vec_ty)),
1197             sym::simd_fpowi => ("powi", bx.type_func(&[vec_ty, bx.type_i32()], vec_ty)),
1198             sym::simd_fpow => ("pow", bx.type_func(&[vec_ty, vec_ty], vec_ty)),
1199             sym::simd_fsin => ("sin", bx.type_func(&[vec_ty], vec_ty)),
1200             sym::simd_fsqrt => ("sqrt", bx.type_func(&[vec_ty], vec_ty)),
1201             sym::simd_round => ("round", bx.type_func(&[vec_ty], vec_ty)),
1202             sym::simd_trunc => ("trunc", bx.type_func(&[vec_ty], vec_ty)),
1203             _ => return_error!("unrecognized intrinsic `{}`", name),
1204         };
1205         let llvm_name = &format!("llvm.{0}.v{1}{2}", intr_name, in_len, elem_ty_str);
1206         let f = bx.declare_cfn(llvm_name, llvm::UnnamedAddr::No, fn_ty);
1207         let c =
1208             bx.call(fn_ty, f, &args.iter().map(|arg| arg.immediate()).collect::<Vec<_>>(), None);
1209         Ok(c)
1210     }
1211
1212     if std::matches!(
1213         name,
1214         sym::simd_ceil
1215             | sym::simd_fabs
1216             | sym::simd_fcos
1217             | sym::simd_fexp2
1218             | sym::simd_fexp
1219             | sym::simd_flog10
1220             | sym::simd_flog2
1221             | sym::simd_flog
1222             | sym::simd_floor
1223             | sym::simd_fma
1224             | sym::simd_fpow
1225             | sym::simd_fpowi
1226             | sym::simd_fsin
1227             | sym::simd_fsqrt
1228             | sym::simd_round
1229             | sym::simd_trunc
1230     ) {
1231         return simd_simple_float_intrinsic(name, in_elem, in_ty, in_len, bx, span, args);
1232     }
1233
1234     // FIXME: use:
1235     //  https://github.com/llvm-mirror/llvm/blob/master/include/llvm/IR/Function.h#L182
1236     //  https://github.com/llvm-mirror/llvm/blob/master/include/llvm/IR/Intrinsics.h#L81
1237     fn llvm_vector_str(
1238         elem_ty: Ty<'_>,
1239         vec_len: u64,
1240         no_pointers: usize,
1241         bx: &Builder<'_, '_, '_>,
1242     ) -> String {
1243         let p0s: String = "p0".repeat(no_pointers);
1244         match *elem_ty.kind() {
1245             ty::Int(v) => format!(
1246                 "v{}{}i{}",
1247                 vec_len,
1248                 p0s,
1249                 // Normalize to prevent crash if v: IntTy::Isize
1250                 v.normalize(bx.target_spec().pointer_width).bit_width().unwrap()
1251             ),
1252             ty::Uint(v) => format!(
1253                 "v{}{}i{}",
1254                 vec_len,
1255                 p0s,
1256                 // Normalize to prevent crash if v: UIntTy::Usize
1257                 v.normalize(bx.target_spec().pointer_width).bit_width().unwrap()
1258             ),
1259             ty::Float(v) => format!("v{}{}f{}", vec_len, p0s, v.bit_width()),
1260             _ => unreachable!(),
1261         }
1262     }
1263
1264     fn llvm_vector_ty<'ll>(
1265         cx: &CodegenCx<'ll, '_>,
1266         elem_ty: Ty<'_>,
1267         vec_len: u64,
1268         mut no_pointers: usize,
1269     ) -> &'ll Type {
1270         // FIXME: use cx.layout_of(ty).llvm_type() ?
1271         let mut elem_ty = match *elem_ty.kind() {
1272             ty::Int(v) => cx.type_int_from_ty(v),
1273             ty::Uint(v) => cx.type_uint_from_ty(v),
1274             ty::Float(v) => cx.type_float_from_ty(v),
1275             _ => unreachable!(),
1276         };
1277         while no_pointers > 0 {
1278             elem_ty = cx.type_ptr_to(elem_ty);
1279             no_pointers -= 1;
1280         }
1281         cx.type_vector(elem_ty, vec_len)
1282     }
1283
1284     if name == sym::simd_gather {
1285         // simd_gather(values: <N x T>, pointers: <N x *_ T>,
1286         //             mask: <N x i{M}>) -> <N x T>
1287         // * N: number of elements in the input vectors
1288         // * T: type of the element to load
1289         // * M: any integer width is supported, will be truncated to i1
1290
1291         // All types must be simd vector types
1292         require_simd!(in_ty, "first");
1293         require_simd!(arg_tys[1], "second");
1294         require_simd!(arg_tys[2], "third");
1295         require_simd!(ret_ty, "return");
1296
1297         // Of the same length:
1298         let (out_len, _) = arg_tys[1].simd_size_and_type(bx.tcx());
1299         let (out_len2, _) = arg_tys[2].simd_size_and_type(bx.tcx());
1300         require!(
1301             in_len == out_len,
1302             "expected {} argument with length {} (same as input type `{}`), \
1303              found `{}` with length {}",
1304             "second",
1305             in_len,
1306             in_ty,
1307             arg_tys[1],
1308             out_len
1309         );
1310         require!(
1311             in_len == out_len2,
1312             "expected {} argument with length {} (same as input type `{}`), \
1313              found `{}` with length {}",
1314             "third",
1315             in_len,
1316             in_ty,
1317             arg_tys[2],
1318             out_len2
1319         );
1320
1321         // The return type must match the first argument type
1322         require!(ret_ty == in_ty, "expected return type `{}`, found `{}`", in_ty, ret_ty);
1323
1324         // This counts how many pointers
1325         fn ptr_count(t: Ty<'_>) -> usize {
1326             match t.kind() {
1327                 ty::RawPtr(p) => 1 + ptr_count(p.ty),
1328                 _ => 0,
1329             }
1330         }
1331
1332         // Non-ptr type
1333         fn non_ptr(t: Ty<'_>) -> Ty<'_> {
1334             match t.kind() {
1335                 ty::RawPtr(p) => non_ptr(p.ty),
1336                 _ => t,
1337             }
1338         }
1339
1340         // The second argument must be a simd vector with an element type that's a pointer
1341         // to the element type of the first argument
1342         let (_, element_ty0) = arg_tys[0].simd_size_and_type(bx.tcx());
1343         let (_, element_ty1) = arg_tys[1].simd_size_and_type(bx.tcx());
1344         let (pointer_count, underlying_ty) = match element_ty1.kind() {
1345             ty::RawPtr(p) if p.ty == in_elem => (ptr_count(element_ty1), non_ptr(element_ty1)),
1346             _ => {
1347                 require!(
1348                     false,
1349                     "expected element type `{}` of second argument `{}` \
1350                         to be a pointer to the element type `{}` of the first \
1351                         argument `{}`, found `{}` != `*_ {}`",
1352                     element_ty1,
1353                     arg_tys[1],
1354                     in_elem,
1355                     in_ty,
1356                     element_ty1,
1357                     in_elem
1358                 );
1359                 unreachable!();
1360             }
1361         };
1362         assert!(pointer_count > 0);
1363         assert_eq!(pointer_count - 1, ptr_count(element_ty0));
1364         assert_eq!(underlying_ty, non_ptr(element_ty0));
1365
1366         // The element type of the third argument must be a signed integer type of any width:
1367         let (_, element_ty2) = arg_tys[2].simd_size_and_type(bx.tcx());
1368         match element_ty2.kind() {
1369             ty::Int(_) => (),
1370             _ => {
1371                 require!(
1372                     false,
1373                     "expected element type `{}` of third argument `{}` \
1374                                  to be a signed integer type",
1375                     element_ty2,
1376                     arg_tys[2]
1377                 );
1378             }
1379         }
1380
1381         // Alignment of T, must be a constant integer value:
1382         let alignment_ty = bx.type_i32();
1383         let alignment = bx.const_i32(bx.align_of(in_elem).bytes() as i32);
1384
1385         // Truncate the mask vector to a vector of i1s:
1386         let (mask, mask_ty) = {
1387             let i1 = bx.type_i1();
1388             let i1xn = bx.type_vector(i1, in_len);
1389             (bx.trunc(args[2].immediate(), i1xn), i1xn)
1390         };
1391
1392         // Type of the vector of pointers:
1393         let llvm_pointer_vec_ty = llvm_vector_ty(bx, underlying_ty, in_len, pointer_count);
1394         let llvm_pointer_vec_str = llvm_vector_str(underlying_ty, in_len, pointer_count, bx);
1395
1396         // Type of the vector of elements:
1397         let llvm_elem_vec_ty = llvm_vector_ty(bx, underlying_ty, in_len, pointer_count - 1);
1398         let llvm_elem_vec_str = llvm_vector_str(underlying_ty, in_len, pointer_count - 1, bx);
1399
1400         let llvm_intrinsic =
1401             format!("llvm.masked.gather.{}.{}", llvm_elem_vec_str, llvm_pointer_vec_str);
1402         let fn_ty = bx.type_func(
1403             &[llvm_pointer_vec_ty, alignment_ty, mask_ty, llvm_elem_vec_ty],
1404             llvm_elem_vec_ty,
1405         );
1406         let f = bx.declare_cfn(&llvm_intrinsic, llvm::UnnamedAddr::No, fn_ty);
1407         let v =
1408             bx.call(fn_ty, f, &[args[1].immediate(), alignment, mask, args[0].immediate()], None);
1409         return Ok(v);
1410     }
1411
1412     if name == sym::simd_scatter {
1413         // simd_scatter(values: <N x T>, pointers: <N x *mut T>,
1414         //             mask: <N x i{M}>) -> ()
1415         // * N: number of elements in the input vectors
1416         // * T: type of the element to load
1417         // * M: any integer width is supported, will be truncated to i1
1418
1419         // All types must be simd vector types
1420         require_simd!(in_ty, "first");
1421         require_simd!(arg_tys[1], "second");
1422         require_simd!(arg_tys[2], "third");
1423
1424         // Of the same length:
1425         let (element_len1, _) = arg_tys[1].simd_size_and_type(bx.tcx());
1426         let (element_len2, _) = arg_tys[2].simd_size_and_type(bx.tcx());
1427         require!(
1428             in_len == element_len1,
1429             "expected {} argument with length {} (same as input type `{}`), \
1430             found `{}` with length {}",
1431             "second",
1432             in_len,
1433             in_ty,
1434             arg_tys[1],
1435             element_len1
1436         );
1437         require!(
1438             in_len == element_len2,
1439             "expected {} argument with length {} (same as input type `{}`), \
1440             found `{}` with length {}",
1441             "third",
1442             in_len,
1443             in_ty,
1444             arg_tys[2],
1445             element_len2
1446         );
1447
1448         // This counts how many pointers
1449         fn ptr_count(t: Ty<'_>) -> usize {
1450             match t.kind() {
1451                 ty::RawPtr(p) => 1 + ptr_count(p.ty),
1452                 _ => 0,
1453             }
1454         }
1455
1456         // Non-ptr type
1457         fn non_ptr(t: Ty<'_>) -> Ty<'_> {
1458             match t.kind() {
1459                 ty::RawPtr(p) => non_ptr(p.ty),
1460                 _ => t,
1461             }
1462         }
1463
1464         // The second argument must be a simd vector with an element type that's a pointer
1465         // to the element type of the first argument
1466         let (_, element_ty0) = arg_tys[0].simd_size_and_type(bx.tcx());
1467         let (_, element_ty1) = arg_tys[1].simd_size_and_type(bx.tcx());
1468         let (_, element_ty2) = arg_tys[2].simd_size_and_type(bx.tcx());
1469         let (pointer_count, underlying_ty) = match element_ty1.kind() {
1470             ty::RawPtr(p) if p.ty == in_elem && p.mutbl == hir::Mutability::Mut => {
1471                 (ptr_count(element_ty1), non_ptr(element_ty1))
1472             }
1473             _ => {
1474                 require!(
1475                     false,
1476                     "expected element type `{}` of second argument `{}` \
1477                         to be a pointer to the element type `{}` of the first \
1478                         argument `{}`, found `{}` != `*mut {}`",
1479                     element_ty1,
1480                     arg_tys[1],
1481                     in_elem,
1482                     in_ty,
1483                     element_ty1,
1484                     in_elem
1485                 );
1486                 unreachable!();
1487             }
1488         };
1489         assert!(pointer_count > 0);
1490         assert_eq!(pointer_count - 1, ptr_count(element_ty0));
1491         assert_eq!(underlying_ty, non_ptr(element_ty0));
1492
1493         // The element type of the third argument must be a signed integer type of any width:
1494         match element_ty2.kind() {
1495             ty::Int(_) => (),
1496             _ => {
1497                 require!(
1498                     false,
1499                     "expected element type `{}` of third argument `{}` \
1500                          be a signed integer type",
1501                     element_ty2,
1502                     arg_tys[2]
1503                 );
1504             }
1505         }
1506
1507         // Alignment of T, must be a constant integer value:
1508         let alignment_ty = bx.type_i32();
1509         let alignment = bx.const_i32(bx.align_of(in_elem).bytes() as i32);
1510
1511         // Truncate the mask vector to a vector of i1s:
1512         let (mask, mask_ty) = {
1513             let i1 = bx.type_i1();
1514             let i1xn = bx.type_vector(i1, in_len);
1515             (bx.trunc(args[2].immediate(), i1xn), i1xn)
1516         };
1517
1518         let ret_t = bx.type_void();
1519
1520         // Type of the vector of pointers:
1521         let llvm_pointer_vec_ty = llvm_vector_ty(bx, underlying_ty, in_len, pointer_count);
1522         let llvm_pointer_vec_str = llvm_vector_str(underlying_ty, in_len, pointer_count, bx);
1523
1524         // Type of the vector of elements:
1525         let llvm_elem_vec_ty = llvm_vector_ty(bx, underlying_ty, in_len, pointer_count - 1);
1526         let llvm_elem_vec_str = llvm_vector_str(underlying_ty, in_len, pointer_count - 1, bx);
1527
1528         let llvm_intrinsic =
1529             format!("llvm.masked.scatter.{}.{}", llvm_elem_vec_str, llvm_pointer_vec_str);
1530         let fn_ty =
1531             bx.type_func(&[llvm_elem_vec_ty, llvm_pointer_vec_ty, alignment_ty, mask_ty], ret_t);
1532         let f = bx.declare_cfn(&llvm_intrinsic, llvm::UnnamedAddr::No, fn_ty);
1533         let v =
1534             bx.call(fn_ty, f, &[args[0].immediate(), args[1].immediate(), alignment, mask], None);
1535         return Ok(v);
1536     }
1537
1538     macro_rules! arith_red {
1539         ($name:ident : $integer_reduce:ident, $float_reduce:ident, $ordered:expr, $op:ident,
1540          $identity:expr) => {
1541             if name == sym::$name {
1542                 require!(
1543                     ret_ty == in_elem,
1544                     "expected return type `{}` (element of input `{}`), found `{}`",
1545                     in_elem,
1546                     in_ty,
1547                     ret_ty
1548                 );
1549                 return match in_elem.kind() {
1550                     ty::Int(_) | ty::Uint(_) => {
1551                         let r = bx.$integer_reduce(args[0].immediate());
1552                         if $ordered {
1553                             // if overflow occurs, the result is the
1554                             // mathematical result modulo 2^n:
1555                             Ok(bx.$op(args[1].immediate(), r))
1556                         } else {
1557                             Ok(bx.$integer_reduce(args[0].immediate()))
1558                         }
1559                     }
1560                     ty::Float(f) => {
1561                         let acc = if $ordered {
1562                             // ordered arithmetic reductions take an accumulator
1563                             args[1].immediate()
1564                         } else {
1565                             // unordered arithmetic reductions use the identity accumulator
1566                             match f.bit_width() {
1567                                 32 => bx.const_real(bx.type_f32(), $identity),
1568                                 64 => bx.const_real(bx.type_f64(), $identity),
1569                                 v => return_error!(
1570                                     r#"
1571 unsupported {} from `{}` with element `{}` of size `{}` to `{}`"#,
1572                                     sym::$name,
1573                                     in_ty,
1574                                     in_elem,
1575                                     v,
1576                                     ret_ty
1577                                 ),
1578                             }
1579                         };
1580                         Ok(bx.$float_reduce(acc, args[0].immediate()))
1581                     }
1582                     _ => return_error!(
1583                         "unsupported {} from `{}` with element `{}` to `{}`",
1584                         sym::$name,
1585                         in_ty,
1586                         in_elem,
1587                         ret_ty
1588                     ),
1589                 };
1590             }
1591         };
1592     }
1593
1594     arith_red!(simd_reduce_add_ordered: vector_reduce_add, vector_reduce_fadd, true, add, 0.0);
1595     arith_red!(simd_reduce_mul_ordered: vector_reduce_mul, vector_reduce_fmul, true, mul, 1.0);
1596     arith_red!(
1597         simd_reduce_add_unordered: vector_reduce_add,
1598         vector_reduce_fadd_fast,
1599         false,
1600         add,
1601         0.0
1602     );
1603     arith_red!(
1604         simd_reduce_mul_unordered: vector_reduce_mul,
1605         vector_reduce_fmul_fast,
1606         false,
1607         mul,
1608         1.0
1609     );
1610
1611     macro_rules! minmax_red {
1612         ($name:ident: $int_red:ident, $float_red:ident) => {
1613             if name == sym::$name {
1614                 require!(
1615                     ret_ty == in_elem,
1616                     "expected return type `{}` (element of input `{}`), found `{}`",
1617                     in_elem,
1618                     in_ty,
1619                     ret_ty
1620                 );
1621                 return match in_elem.kind() {
1622                     ty::Int(_i) => Ok(bx.$int_red(args[0].immediate(), true)),
1623                     ty::Uint(_u) => Ok(bx.$int_red(args[0].immediate(), false)),
1624                     ty::Float(_f) => Ok(bx.$float_red(args[0].immediate())),
1625                     _ => return_error!(
1626                         "unsupported {} from `{}` with element `{}` to `{}`",
1627                         sym::$name,
1628                         in_ty,
1629                         in_elem,
1630                         ret_ty
1631                     ),
1632                 };
1633             }
1634         };
1635     }
1636
1637     minmax_red!(simd_reduce_min: vector_reduce_min, vector_reduce_fmin);
1638     minmax_red!(simd_reduce_max: vector_reduce_max, vector_reduce_fmax);
1639
1640     minmax_red!(simd_reduce_min_nanless: vector_reduce_min, vector_reduce_fmin_fast);
1641     minmax_red!(simd_reduce_max_nanless: vector_reduce_max, vector_reduce_fmax_fast);
1642
1643     macro_rules! bitwise_red {
1644         ($name:ident : $red:ident, $boolean:expr) => {
1645             if name == sym::$name {
1646                 let input = if !$boolean {
1647                     require!(
1648                         ret_ty == in_elem,
1649                         "expected return type `{}` (element of input `{}`), found `{}`",
1650                         in_elem,
1651                         in_ty,
1652                         ret_ty
1653                     );
1654                     args[0].immediate()
1655                 } else {
1656                     match in_elem.kind() {
1657                         ty::Int(_) | ty::Uint(_) => {}
1658                         _ => return_error!(
1659                             "unsupported {} from `{}` with element `{}` to `{}`",
1660                             sym::$name,
1661                             in_ty,
1662                             in_elem,
1663                             ret_ty
1664                         ),
1665                     }
1666
1667                     // boolean reductions operate on vectors of i1s:
1668                     let i1 = bx.type_i1();
1669                     let i1xn = bx.type_vector(i1, in_len as u64);
1670                     bx.trunc(args[0].immediate(), i1xn)
1671                 };
1672                 return match in_elem.kind() {
1673                     ty::Int(_) | ty::Uint(_) => {
1674                         let r = bx.$red(input);
1675                         Ok(if !$boolean { r } else { bx.zext(r, bx.type_bool()) })
1676                     }
1677                     _ => return_error!(
1678                         "unsupported {} from `{}` with element `{}` to `{}`",
1679                         sym::$name,
1680                         in_ty,
1681                         in_elem,
1682                         ret_ty
1683                     ),
1684                 };
1685             }
1686         };
1687     }
1688
1689     bitwise_red!(simd_reduce_and: vector_reduce_and, false);
1690     bitwise_red!(simd_reduce_or: vector_reduce_or, false);
1691     bitwise_red!(simd_reduce_xor: vector_reduce_xor, false);
1692     bitwise_red!(simd_reduce_all: vector_reduce_and, true);
1693     bitwise_red!(simd_reduce_any: vector_reduce_or, true);
1694
1695     if name == sym::simd_cast || name == sym::simd_as {
1696         require_simd!(ret_ty, "return");
1697         let (out_len, out_elem) = ret_ty.simd_size_and_type(bx.tcx());
1698         require!(
1699             in_len == out_len,
1700             "expected return type with length {} (same as input type `{}`), \
1701                   found `{}` with length {}",
1702             in_len,
1703             in_ty,
1704             ret_ty,
1705             out_len
1706         );
1707         // casting cares about nominal type, not just structural type
1708         if in_elem == out_elem {
1709             return Ok(args[0].immediate());
1710         }
1711
1712         enum Style {
1713             Float,
1714             Int(/* is signed? */ bool),
1715             Unsupported,
1716         }
1717
1718         let (in_style, in_width) = match in_elem.kind() {
1719             // vectors of pointer-sized integers should've been
1720             // disallowed before here, so this unwrap is safe.
1721             ty::Int(i) => (
1722                 Style::Int(true),
1723                 i.normalize(bx.tcx().sess.target.pointer_width).bit_width().unwrap(),
1724             ),
1725             ty::Uint(u) => (
1726                 Style::Int(false),
1727                 u.normalize(bx.tcx().sess.target.pointer_width).bit_width().unwrap(),
1728             ),
1729             ty::Float(f) => (Style::Float, f.bit_width()),
1730             _ => (Style::Unsupported, 0),
1731         };
1732         let (out_style, out_width) = match out_elem.kind() {
1733             ty::Int(i) => (
1734                 Style::Int(true),
1735                 i.normalize(bx.tcx().sess.target.pointer_width).bit_width().unwrap(),
1736             ),
1737             ty::Uint(u) => (
1738                 Style::Int(false),
1739                 u.normalize(bx.tcx().sess.target.pointer_width).bit_width().unwrap(),
1740             ),
1741             ty::Float(f) => (Style::Float, f.bit_width()),
1742             _ => (Style::Unsupported, 0),
1743         };
1744
1745         match (in_style, out_style) {
1746             (Style::Int(in_is_signed), Style::Int(_)) => {
1747                 return Ok(match in_width.cmp(&out_width) {
1748                     Ordering::Greater => bx.trunc(args[0].immediate(), llret_ty),
1749                     Ordering::Equal => args[0].immediate(),
1750                     Ordering::Less => {
1751                         if in_is_signed {
1752                             bx.sext(args[0].immediate(), llret_ty)
1753                         } else {
1754                             bx.zext(args[0].immediate(), llret_ty)
1755                         }
1756                     }
1757                 });
1758             }
1759             (Style::Int(in_is_signed), Style::Float) => {
1760                 return Ok(if in_is_signed {
1761                     bx.sitofp(args[0].immediate(), llret_ty)
1762                 } else {
1763                     bx.uitofp(args[0].immediate(), llret_ty)
1764                 });
1765             }
1766             (Style::Float, Style::Int(out_is_signed)) => {
1767                 return Ok(match (out_is_signed, name == sym::simd_as) {
1768                     (false, false) => bx.fptoui(args[0].immediate(), llret_ty),
1769                     (true, false) => bx.fptosi(args[0].immediate(), llret_ty),
1770                     (_, true) => bx.cast_float_to_int(out_is_signed, args[0].immediate(), llret_ty),
1771                 });
1772             }
1773             (Style::Float, Style::Float) => {
1774                 return Ok(match in_width.cmp(&out_width) {
1775                     Ordering::Greater => bx.fptrunc(args[0].immediate(), llret_ty),
1776                     Ordering::Equal => args[0].immediate(),
1777                     Ordering::Less => bx.fpext(args[0].immediate(), llret_ty),
1778                 });
1779             }
1780             _ => { /* Unsupported. Fallthrough. */ }
1781         }
1782         require!(
1783             false,
1784             "unsupported cast from `{}` with element `{}` to `{}` with element `{}`",
1785             in_ty,
1786             in_elem,
1787             ret_ty,
1788             out_elem
1789         );
1790     }
1791     macro_rules! arith_binary {
1792         ($($name: ident: $($($p: ident),* => $call: ident),*;)*) => {
1793             $(if name == sym::$name {
1794                 match in_elem.kind() {
1795                     $($(ty::$p(_))|* => {
1796                         return Ok(bx.$call(args[0].immediate(), args[1].immediate()))
1797                     })*
1798                     _ => {},
1799                 }
1800                 require!(false,
1801                          "unsupported operation on `{}` with element `{}`",
1802                          in_ty,
1803                          in_elem)
1804             })*
1805         }
1806     }
1807     arith_binary! {
1808         simd_add: Uint, Int => add, Float => fadd;
1809         simd_sub: Uint, Int => sub, Float => fsub;
1810         simd_mul: Uint, Int => mul, Float => fmul;
1811         simd_div: Uint => udiv, Int => sdiv, Float => fdiv;
1812         simd_rem: Uint => urem, Int => srem, Float => frem;
1813         simd_shl: Uint, Int => shl;
1814         simd_shr: Uint => lshr, Int => ashr;
1815         simd_and: Uint, Int => and;
1816         simd_or: Uint, Int => or;
1817         simd_xor: Uint, Int => xor;
1818         simd_fmax: Float => maxnum;
1819         simd_fmin: Float => minnum;
1820
1821     }
1822     macro_rules! arith_unary {
1823         ($($name: ident: $($($p: ident),* => $call: ident),*;)*) => {
1824             $(if name == sym::$name {
1825                 match in_elem.kind() {
1826                     $($(ty::$p(_))|* => {
1827                         return Ok(bx.$call(args[0].immediate()))
1828                     })*
1829                     _ => {},
1830                 }
1831                 require!(false,
1832                          "unsupported operation on `{}` with element `{}`",
1833                          in_ty,
1834                          in_elem)
1835             })*
1836         }
1837     }
1838     arith_unary! {
1839         simd_neg: Int => neg, Float => fneg;
1840     }
1841
1842     if name == sym::simd_saturating_add || name == sym::simd_saturating_sub {
1843         let lhs = args[0].immediate();
1844         let rhs = args[1].immediate();
1845         let is_add = name == sym::simd_saturating_add;
1846         let ptr_bits = bx.tcx().data_layout.pointer_size.bits() as _;
1847         let (signed, elem_width, elem_ty) = match *in_elem.kind() {
1848             ty::Int(i) => (true, i.bit_width().unwrap_or(ptr_bits), bx.cx.type_int_from_ty(i)),
1849             ty::Uint(i) => (false, i.bit_width().unwrap_or(ptr_bits), bx.cx.type_uint_from_ty(i)),
1850             _ => {
1851                 return_error!(
1852                     "expected element type `{}` of vector type `{}` \
1853                      to be a signed or unsigned integer type",
1854                     arg_tys[0].simd_size_and_type(bx.tcx()).1,
1855                     arg_tys[0]
1856                 );
1857             }
1858         };
1859         let llvm_intrinsic = &format!(
1860             "llvm.{}{}.sat.v{}i{}",
1861             if signed { 's' } else { 'u' },
1862             if is_add { "add" } else { "sub" },
1863             in_len,
1864             elem_width
1865         );
1866         let vec_ty = bx.cx.type_vector(elem_ty, in_len as u64);
1867
1868         let fn_ty = bx.type_func(&[vec_ty, vec_ty], vec_ty);
1869         let f = bx.declare_cfn(llvm_intrinsic, llvm::UnnamedAddr::No, fn_ty);
1870         let v = bx.call(fn_ty, f, &[lhs, rhs], None);
1871         return Ok(v);
1872     }
1873
1874     span_bug!(span, "unknown SIMD intrinsic");
1875 }
1876
1877 // Returns the width of an int Ty, and if it's signed or not
1878 // Returns None if the type is not an integer
1879 // FIXME: there’s multiple of this functions, investigate using some of the already existing
1880 // stuffs.
1881 fn int_type_width_signed(ty: Ty<'_>, cx: &CodegenCx<'_, '_>) -> Option<(u64, bool)> {
1882     match ty.kind() {
1883         ty::Int(t) => {
1884             Some((t.bit_width().unwrap_or(u64::from(cx.tcx.sess.target.pointer_width)), true))
1885         }
1886         ty::Uint(t) => {
1887             Some((t.bit_width().unwrap_or(u64::from(cx.tcx.sess.target.pointer_width)), false))
1888         }
1889         _ => None,
1890     }
1891 }