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         sym::ptr_mask => "llvm.ptrmask",
  75         _ => return None,
  76     };
  77     Some(cx.get_intrinsic(llvm_name))
  78 }
  79
  80 impl<'ll, 'tcx> IntrinsicCallMethods<'tcx> for Builder<'_, 'll, 'tcx> {
  81     fn codegen_intrinsic_call(
  82         &mut self,
  83         instance: ty::Instance<'tcx>,
  84         fn_abi: &FnAbi<'tcx, Ty<'tcx>>,
  85         args: &[OperandRef<'tcx, &'ll Value>],
  86         llresult: &'ll Value,
  87         span: Span,
  88     ) {
  89         let tcx = self.tcx;
  90         let callee_ty = instance.ty(tcx, ty::ParamEnv::reveal_all());
  91
  92         let ty::FnDef(def_id, substs) = *callee_ty.kind() else {
  93             bug!("expected fn item type, found {}", callee_ty);
  94         };
  95
  96         let sig = callee_ty.fn_sig(tcx);
  97         let sig = tcx.normalize_erasing_late_bound_regions(ty::ParamEnv::reveal_all(), sig);
  98         let arg_tys = sig.inputs();
  99         let ret_ty = sig.output();
 100         let name = tcx.item_name(def_id);
 101
 102         let llret_ty = self.layout_of(ret_ty).llvm_type(self);
 103         let result = PlaceRef::new_sized(llresult, fn_abi.ret.layout);
 104
 105         let simple = get_simple_intrinsic(self, name);
 106         let llval = match name {
 107             _ if simple.is_some() => {
 108                 let (simple_ty, simple_fn) = simple.unwrap();
 109                 self.call(
 110                     simple_ty,
 111                     None,
 112                     simple_fn,
 113                     &args.iter().map(|arg| arg.immediate()).collect::<Vec<_>>(),
 114                     None,
 115                 )
 116             }
 117             sym::likely => {
 118                 self.call_intrinsic("llvm.expect.i1", &[args[0].immediate(), self.const_bool(true)])
 119             }
 120             sym::unlikely => self
 121                 .call_intrinsic("llvm.expect.i1", &[args[0].immediate(), self.const_bool(false)]),
 122             kw::Try => {
 123                 try_intrinsic(
 124                     self,
 125                     args[0].immediate(),
 126                     args[1].immediate(),
 127                     args[2].immediate(),
 128                     llresult,
 129                 );
 130                 return;
 131             }
 132             sym::breakpoint => self.call_intrinsic("llvm.debugtrap", &[]),
 133             sym::va_copy => {
 134                 self.call_intrinsic("llvm.va_copy", &[args[0].immediate(), args[1].immediate()])
 135             }
 136             sym::va_arg => {
 137                 match fn_abi.ret.layout.abi {
 138                     abi::Abi::Scalar(scalar) => {
 139                         match scalar.primitive() {
 140                             Primitive::Int(..) => {
 141                                 if self.cx().size_of(ret_ty).bytes() < 4 {
 142                                     // `va_arg` should not be called on an integer type
 143                                     // less than 4 bytes in length. If it is, promote
 144                                     // the integer to an `i32` and truncate the result
 145                                     // back to the smaller type.
 146                                     let promoted_result = emit_va_arg(self, args[0], tcx.types.i32);
 147                                     self.trunc(promoted_result, llret_ty)
 148                                 } else {
 149                                     emit_va_arg(self, args[0], ret_ty)
 150                                 }
 151                             }
 152                             Primitive::F64 | Primitive::Pointer => {
 153                                 emit_va_arg(self, args[0], ret_ty)
 154                             }
 155                             // `va_arg` should never be used with the return type f32.
 156                             Primitive::F32 => bug!("the va_arg intrinsic does not work with `f32`"),
 157                         }
 158                     }
 159                     _ => bug!("the va_arg intrinsic does not work with non-scalar types"),
 160                 }
 161             }
 162
 163             sym::volatile_load | sym::unaligned_volatile_load => {
 164                 let tp_ty = substs.type_at(0);
 165                 let ptr = args[0].immediate();
 166                 let load = if let PassMode::Cast(ty, _) = &fn_abi.ret.mode {
 167                     let llty = ty.llvm_type(self);
 168                     let ptr = self.pointercast(ptr, self.type_ptr_to(llty));
 169                     self.volatile_load(llty, ptr)
 170                 } else {
 171                     self.volatile_load(self.layout_of(tp_ty).llvm_type(self), ptr)
 172                 };
 173                 let align = if name == sym::unaligned_volatile_load {
 174                     1
 175                 } else {
 176                     self.align_of(tp_ty).bytes() as u32
 177                 };
 178                 unsafe {
 179                     llvm::LLVMSetAlignment(load, align);
 180                 }
 181                 self.to_immediate(load, self.layout_of(tp_ty))
 182             }
 183             sym::volatile_store => {
 184                 let dst = args[0].deref(self.cx());
 185                 args[1].val.volatile_store(self, dst);
 186                 return;
 187             }
 188             sym::unaligned_volatile_store => {
 189                 let dst = args[0].deref(self.cx());
 190                 args[1].val.unaligned_volatile_store(self, dst);
 191                 return;
 192             }
 193             sym::prefetch_read_data
 194             | sym::prefetch_write_data
 195             | sym::prefetch_read_instruction
 196             | sym::prefetch_write_instruction => {
 197                 let (rw, cache_type) = match name {
 198                     sym::prefetch_read_data => (0, 1),
 199                     sym::prefetch_write_data => (1, 1),
 200                     sym::prefetch_read_instruction => (0, 0),
 201                     sym::prefetch_write_instruction => (1, 0),
 202                     _ => bug!(),
 203                 };
 204                 self.call_intrinsic(
 205                     "llvm.prefetch",
 206                     &[
 207                         args[0].immediate(),
 208                         self.const_i32(rw),
 209                         args[1].immediate(),
 210                         self.const_i32(cache_type),
 211                     ],
 212                 )
 213             }
 214             sym::ctlz
 215             | sym::ctlz_nonzero
 216             | sym::cttz
 217             | sym::cttz_nonzero
 218             | sym::ctpop
 219             | sym::bswap
 220             | sym::bitreverse
 221             | sym::rotate_left
 222             | sym::rotate_right
 223             | sym::saturating_add
 224             | sym::saturating_sub => {
 225                 let ty = arg_tys[0];
 226                 match int_type_width_signed(ty, self) {
 227                     Some((width, signed)) => match name {
 228                         sym::ctlz | sym::cttz => {
 229                             let y = self.const_bool(false);
 230                             self.call_intrinsic(
 231                                 &format!("llvm.{}.i{}", name, width),
 232                                 &[args[0].immediate(), y],
 233                             )
 234                         }
 235                         sym::ctlz_nonzero => {
 236                             let y = self.const_bool(true);
 237                             let llvm_name = &format!("llvm.ctlz.i{}", width);
 238                             self.call_intrinsic(llvm_name, &[args[0].immediate(), y])
 239                         }
 240                         sym::cttz_nonzero => {
 241                             let y = self.const_bool(true);
 242                             let llvm_name = &format!("llvm.cttz.i{}", width);
 243                             self.call_intrinsic(llvm_name, &[args[0].immediate(), y])
 244                         }
 245                         sym::ctpop => self.call_intrinsic(
 246                             &format!("llvm.ctpop.i{}", width),
 247                             &[args[0].immediate()],
 248                         ),
 249                         sym::bswap => {
 250                             if width == 8 {
 251                                 args[0].immediate() // byte swap a u8/i8 is just a no-op
 252                             } else {
 253                                 self.call_intrinsic(
 254                                     &format!("llvm.bswap.i{}", width),
 255                                     &[args[0].immediate()],
 256                                 )
 257                             }
 258                         }
 259                         sym::bitreverse => self.call_intrinsic(
 260                             &format!("llvm.bitreverse.i{}", width),
 261                             &[args[0].immediate()],
 262                         ),
 263                         sym::rotate_left | sym::rotate_right => {
 264                             let is_left = name == sym::rotate_left;
 265                             let val = args[0].immediate();
 266                             let raw_shift = args[1].immediate();
 267                             // rotate = funnel shift with first two args the same
 268                             let llvm_name =
 269                                 &format!("llvm.fsh{}.i{}", if is_left { 'l' } else { 'r' }, width);
 270                             self.call_intrinsic(llvm_name, &[val, val, raw_shift])
 271                         }
 272                         sym::saturating_add | sym::saturating_sub => {
 273                             let is_add = name == sym::saturating_add;
 274                             let lhs = args[0].immediate();
 275                             let rhs = args[1].immediate();
 276                             let llvm_name = &format!(
 277                                 "llvm.{}{}.sat.i{}",
 278                                 if signed { 's' } else { 'u' },
 279                                 if is_add { "add" } else { "sub" },
 280                                 width
 281                             );
 282                             self.call_intrinsic(llvm_name, &[lhs, rhs])
 283                         }
 284                         _ => bug!(),
 285                     },
 286                     None => {
 287                         span_invalid_monomorphization_error(
 288                             tcx.sess,
 289                             span,
 290                             &format!(
 291                                 "invalid monomorphization of `{}` intrinsic: \
 292                                       expected basic integer type, found `{}`",
 293                                 name, ty
 294                             ),
 295                         );
 296                         return;
 297                     }
 298                 }
 299             }
 300
 301             sym::raw_eq => {
 302                 use abi::Abi::*;
 303                 let tp_ty = substs.type_at(0);
 304                 let layout = self.layout_of(tp_ty).layout;
 305                 let use_integer_compare = match layout.abi() {
 306                     Scalar(_) | ScalarPair(_, _) => true,
 307                     Uninhabited | Vector { .. } => false,
 308                     Aggregate { .. } => {
 309                         // For rusty ABIs, small aggregates are actually passed
 310                         // as `RegKind::Integer` (see `FnAbi::adjust_for_abi`),
 311                         // so we re-use that same threshold here.
 312                         layout.size() <= self.data_layout().pointer_size * 2
 313                     }
 314                 };
 315
 316                 let a = args[0].immediate();
 317                 let b = args[1].immediate();
 318                 if layout.size().bytes() == 0 {
 319                     self.const_bool(true)
 320                 } else if use_integer_compare {
 321                     let integer_ty = self.type_ix(layout.size().bits());
 322                     let ptr_ty = self.type_ptr_to(integer_ty);
 323                     let a_ptr = self.bitcast(a, ptr_ty);
 324                     let a_val = self.load(integer_ty, a_ptr, layout.align().abi);
 325                     let b_ptr = self.bitcast(b, ptr_ty);
 326                     let b_val = self.load(integer_ty, b_ptr, layout.align().abi);
 327                     self.icmp(IntPredicate::IntEQ, a_val, b_val)
 328                 } else {
 329                     let i8p_ty = self.type_i8p();
 330                     let a_ptr = self.bitcast(a, i8p_ty);
 331                     let b_ptr = self.bitcast(b, i8p_ty);
 332                     let n = self.const_usize(layout.size().bytes());
 333                     let cmp = self.call_intrinsic("memcmp", &[a_ptr, b_ptr, n]);
 334                     match self.cx.sess().target.arch.as_ref() {
 335                         "avr" | "msp430" => self.icmp(IntPredicate::IntEQ, cmp, self.const_i16(0)),
 336                         _ => self.icmp(IntPredicate::IntEQ, cmp, self.const_i32(0)),
 337                     }
 338                 }
 339             }
 340
 341             sym::black_box => {
 342                 args[0].val.store(self, result);
 343                 let result_val_span = [result.llval];
 344                 // We need to "use" the argument in some way LLVM can't introspect, and on
 345                 // targets that support it we can typically leverage inline assembly to do
 346                 // this. LLVM's interpretation of inline assembly is that it's, well, a black
 347                 // box. This isn't the greatest implementation since it probably deoptimizes
 348                 // more than we want, but it's so far good enough.
 349                 //
 350                 // For zero-sized types, the location pointed to by the result may be
 351                 // uninitialized. Do not "use" the result in this case; instead just clobber
 352                 // the memory.
 353                 let (constraint, inputs): (&str, &[_]) = if result.layout.is_zst() {
 354                     ("~{memory}", &[])
 355                 } else {
 356                     ("r,~{memory}", &result_val_span)
 357                 };
 358                 crate::asm::inline_asm_call(
 359                     self,
 360                     "",
 361                     constraint,
 362                     inputs,
 363                     self.type_void(),
 364                     true,
 365                     false,
 366                     llvm::AsmDialect::Att,
 367                     &[span],
 368                     false,
 369                     None,
 370                 )
 371                 .unwrap_or_else(|| bug!("failed to generate inline asm call for `black_box`"));
 372
 373                 // We have copied the value to `result` already.
 374                 return;
 375             }
 376
 377             _ if name.as_str().starts_with("simd_") => {
 378                 match generic_simd_intrinsic(self, name, callee_ty, args, ret_ty, llret_ty, span) {
 379                     Ok(llval) => llval,
 380                     Err(()) => return,
 381                 }
 382             }
 383
 384             _ => bug!("unknown intrinsic '{}'", name),
 385         };
 386
 387         if !fn_abi.ret.is_ignore() {
 388             if let PassMode::Cast(ty, _) = &fn_abi.ret.mode {
 389                 let ptr_llty = self.type_ptr_to(ty.llvm_type(self));
 390                 let ptr = self.pointercast(result.llval, ptr_llty);
 391                 self.store(llval, ptr, result.align);
 392             } else {
 393                 OperandRef::from_immediate_or_packed_pair(self, llval, result.layout)
 394                     .val
 395                     .store(self, result);
 396             }
 397         }
 398     }
 399
 400     fn abort(&mut self) {
 401         self.call_intrinsic("llvm.trap", &[]);
 402     }
 403
 404     fn assume(&mut self, val: Self::Value) {
 405         self.call_intrinsic("llvm.assume", &[val]);
 406     }
 407
 408     fn expect(&mut self, cond: Self::Value, expected: bool) -> Self::Value {
 409         self.call_intrinsic("llvm.expect.i1", &[cond, self.const_bool(expected)])
 410     }
 411
 412     fn type_test(&mut self, pointer: Self::Value, typeid: Self::Value) -> Self::Value {
 413         // Test the called operand using llvm.type.test intrinsic. The LowerTypeTests link-time
 414         // optimization pass replaces calls to this intrinsic with code to test type membership.
 415         let i8p_ty = self.type_i8p();
 416         let bitcast = self.bitcast(pointer, i8p_ty);
 417         self.call_intrinsic("llvm.type.test", &[bitcast, typeid])
 418     }
 419
 420     fn type_checked_load(
 421         &mut self,
 422         llvtable: &'ll Value,
 423         vtable_byte_offset: u64,
 424         typeid: &'ll Value,
 425     ) -> Self::Value {
 426         let vtable_byte_offset = self.const_i32(vtable_byte_offset as i32);
 427         let type_checked_load =
 428             self.call_intrinsic("llvm.type.checked.load", &[llvtable, vtable_byte_offset, typeid]);
 429         self.extract_value(type_checked_load, 0)
 430     }
 431
 432     fn va_start(&mut self, va_list: &'ll Value) -> &'ll Value {
 433         self.call_intrinsic("llvm.va_start", &[va_list])
 434     }
 435
 436     fn va_end(&mut self, va_list: &'ll Value) -> &'ll Value {
 437         self.call_intrinsic("llvm.va_end", &[va_list])
 438     }
 439 }
 440
 441 fn try_intrinsic<'ll>(
 442     bx: &mut Builder<'_, 'll, '_>,
 443     try_func: &'ll Value,
 444     data: &'ll Value,
 445     catch_func: &'ll Value,
 446     dest: &'ll Value,
 447 ) {
 448     if bx.sess().panic_strategy() == PanicStrategy::Abort {
 449         let try_func_ty = bx.type_func(&[bx.type_i8p()], bx.type_void());
 450         bx.call(try_func_ty, None, try_func, &[data], None);
 451         // Return 0 unconditionally from the intrinsic call;
 452         // we can never unwind.
 453         let ret_align = bx.tcx().data_layout.i32_align.abi;
 454         bx.store(bx.const_i32(0), dest, ret_align);
 455     } else if wants_msvc_seh(bx.sess()) {
 456         codegen_msvc_try(bx, try_func, data, catch_func, dest);
 457     } else if bx.sess().target.os == "emscripten" {
 458         codegen_emcc_try(bx, try_func, data, catch_func, dest);
 459     } else {
 460         codegen_gnu_try(bx, try_func, data, catch_func, dest);
 461     }
 462 }
 463
 464 // MSVC's definition of the `rust_try` function.
 465 //
 466 // This implementation uses the new exception handling instructions in LLVM
 467 // which have support in LLVM for SEH on MSVC targets. Although these
 468 // instructions are meant to work for all targets, as of the time of this
 469 // writing, however, LLVM does not recommend the usage of these new instructions
 470 // as the old ones are still more optimized.
 471 fn codegen_msvc_try<'ll>(
 472     bx: &mut Builder<'_, 'll, '_>,
 473     try_func: &'ll Value,
 474     data: &'ll Value,
 475     catch_func: &'ll Value,
 476     dest: &'ll Value,
 477 ) {
 478     let (llty, llfn) = get_rust_try_fn(bx, &mut |mut bx| {
 479         bx.set_personality_fn(bx.eh_personality());
 480
 481         let normal = bx.append_sibling_block("normal");
 482         let catchswitch = bx.append_sibling_block("catchswitch");
 483         let catchpad_rust = bx.append_sibling_block("catchpad_rust");
 484         let catchpad_foreign = bx.append_sibling_block("catchpad_foreign");
 485         let caught = bx.append_sibling_block("caught");
 486
 487         let try_func = llvm::get_param(bx.llfn(), 0);
 488         let data = llvm::get_param(bx.llfn(), 1);
 489         let catch_func = llvm::get_param(bx.llfn(), 2);
 490
 491         // We're generating an IR snippet that looks like:
 492         //
 493         //   declare i32 @rust_try(%try_func, %data, %catch_func) {
 494         //      %slot = alloca i8*
 495         //      invoke %try_func(%data) to label %normal unwind label %catchswitch
 496         //
 497         //   normal:
 498         //      ret i32 0
 499         //
 500         //   catchswitch:
 501         //      %cs = catchswitch within none [%catchpad_rust, %catchpad_foreign] unwind to caller
 502         //
 503         //   catchpad_rust:
 504         //      %tok = catchpad within %cs [%type_descriptor, 8, %slot]
 505         //      %ptr = load %slot
 506         //      call %catch_func(%data, %ptr)
 507         //      catchret from %tok to label %caught
 508         //
 509         //   catchpad_foreign:
 510         //      %tok = catchpad within %cs [null, 64, null]
 511         //      call %catch_func(%data, null)
 512         //      catchret from %tok to label %caught
 513         //
 514         //   caught:
 515         //      ret i32 1
 516         //   }
 517         //
 518         // This structure follows the basic usage of throw/try/catch in LLVM.
 519         // For example, compile this C++ snippet to see what LLVM generates:
 520         //
 521         //      struct rust_panic {
 522         //          rust_panic(const rust_panic&);
 523         //          ~rust_panic();
 524         //
 525         //          void* x[2];
 526         //      };
 527         //
 528         //      int __rust_try(
 529         //          void (*try_func)(void*),
 530         //          void *data,
 531         //          void (*catch_func)(void*, void*) noexcept
 532         //      ) {
 533         //          try {
 534         //              try_func(data);
 535         //              return 0;
 536         //          } catch(rust_panic& a) {
 537         //              catch_func(data, &a);
 538         //              return 1;
 539         //          } catch(...) {
 540         //              catch_func(data, NULL);
 541         //              return 1;
 542         //          }
 543         //      }
 544         //
 545         // More information can be found in libstd's seh.rs implementation.
 546         let ptr_align = bx.tcx().data_layout.pointer_align.abi;
 547         let slot = bx.alloca(bx.type_i8p(), ptr_align);
 548         let try_func_ty = bx.type_func(&[bx.type_i8p()], bx.type_void());
 549         bx.invoke(try_func_ty, None, try_func, &[data], normal, catchswitch, None);
 550
 551         bx.switch_to_block(normal);
 552         bx.ret(bx.const_i32(0));
 553
 554         bx.switch_to_block(catchswitch);
 555         let cs = bx.catch_switch(None, None, &[catchpad_rust, catchpad_foreign]);
 556
 557         // We can't use the TypeDescriptor defined in libpanic_unwind because it
 558         // might be in another DLL and the SEH encoding only supports specifying
 559         // a TypeDescriptor from the current module.
 560         //
 561         // However this isn't an issue since the MSVC runtime uses string
 562         // comparison on the type name to match TypeDescriptors rather than
 563         // pointer equality.
 564         //
 565         // So instead we generate a new TypeDescriptor in each module that uses
 566         // `try` and let the linker merge duplicate definitions in the same
 567         // module.
 568         //
 569         // When modifying, make sure that the type_name string exactly matches
 570         // the one used in src/libpanic_unwind/seh.rs.
 571         let type_info_vtable = bx.declare_global("??_7type_info@@6B@", bx.type_i8p());
 572         let type_name = bx.const_bytes(b"rust_panic\0");
 573         let type_info =
 574             bx.const_struct(&[type_info_vtable, bx.const_null(bx.type_i8p()), type_name], false);
 575         let tydesc = bx.declare_global("__rust_panic_type_info", bx.val_ty(type_info));
 576         unsafe {
 577             llvm::LLVMRustSetLinkage(tydesc, llvm::Linkage::LinkOnceODRLinkage);
 578             llvm::SetUniqueComdat(bx.llmod, tydesc);
 579             llvm::LLVMSetInitializer(tydesc, type_info);
 580         }
 581
 582         // The flag value of 8 indicates that we are catching the exception by
 583         // reference instead of by value. We can't use catch by value because
 584         // that requires copying the exception object, which we don't support
 585         // since our exception object effectively contains a Box.
 586         //
 587         // Source: MicrosoftCXXABI::getAddrOfCXXCatchHandlerType in clang
 588         bx.switch_to_block(catchpad_rust);
 589         let flags = bx.const_i32(8);
 590         let funclet = bx.catch_pad(cs, &[tydesc, flags, slot]);
 591         let ptr = bx.load(bx.type_i8p(), slot, ptr_align);
 592         let catch_ty = bx.type_func(&[bx.type_i8p(), bx.type_i8p()], bx.type_void());
 593         bx.call(catch_ty, None, catch_func, &[data, ptr], Some(&funclet));
 594         bx.catch_ret(&funclet, caught);
 595
 596         // The flag value of 64 indicates a "catch-all".
 597         bx.switch_to_block(catchpad_foreign);
 598         let flags = bx.const_i32(64);
 599         let null = bx.const_null(bx.type_i8p());
 600         let funclet = bx.catch_pad(cs, &[null, flags, null]);
 601         bx.call(catch_ty, None, catch_func, &[data, null], Some(&funclet));
 602         bx.catch_ret(&funclet, caught);
 603
 604         bx.switch_to_block(caught);
 605         bx.ret(bx.const_i32(1));
 606     });
 607
 608     // Note that no invoke is used here because by definition this function
 609     // can't panic (that's what it's catching).
 610     let ret = bx.call(llty, None, llfn, &[try_func, data, catch_func], None);
 611     let i32_align = bx.tcx().data_layout.i32_align.abi;
 612     bx.store(ret, dest, i32_align);
 613 }
 614
 615 // Definition of the standard `try` function for Rust using the GNU-like model
 616 // of exceptions (e.g., the normal semantics of LLVM's `landingpad` and `invoke`
 617 // instructions).
 618 //
 619 // This codegen is a little surprising because we always call a shim
 620 // function instead of inlining the call to `invoke` manually here. This is done
 621 // because in LLVM we're only allowed to have one personality per function
 622 // definition. The call to the `try` intrinsic is being inlined into the
 623 // function calling it, and that function may already have other personality
 624 // functions in play. By calling a shim we're guaranteed that our shim will have
 625 // the right personality function.
 626 fn codegen_gnu_try<'ll>(
 627     bx: &mut Builder<'_, 'll, '_>,
 628     try_func: &'ll Value,
 629     data: &'ll Value,
 630     catch_func: &'ll Value,
 631     dest: &'ll Value,
 632 ) {
 633     let (llty, llfn) = get_rust_try_fn(bx, &mut |mut bx| {
 634         // Codegens the shims described above:
 635         //
 636         //   bx:
 637         //      invoke %try_func(%data) normal %normal unwind %catch
 638         //
 639         //   normal:
 640         //      ret 0
 641         //
 642         //   catch:
 643         //      (%ptr, _) = landingpad
 644         //      call %catch_func(%data, %ptr)
 645         //      ret 1
 646         let then = bx.append_sibling_block("then");
 647         let catch = bx.append_sibling_block("catch");
 648
 649         let try_func = llvm::get_param(bx.llfn(), 0);
 650         let data = llvm::get_param(bx.llfn(), 1);
 651         let catch_func = llvm::get_param(bx.llfn(), 2);
 652         let try_func_ty = bx.type_func(&[bx.type_i8p()], bx.type_void());
 653         bx.invoke(try_func_ty, None, try_func, &[data], then, catch, None);
 654
 655         bx.switch_to_block(then);
 656         bx.ret(bx.const_i32(0));
 657
 658         // Type indicator for the exception being thrown.
 659         //
 660         // The first value in this tuple is a pointer to the exception object
 661         // being thrown.  The second value is a "selector" indicating which of
 662         // the landing pad clauses the exception's type had been matched to.
 663         // rust_try ignores the selector.
 664         bx.switch_to_block(catch);
 665         let lpad_ty = bx.type_struct(&[bx.type_i8p(), bx.type_i32()], false);
 666         let vals = bx.landing_pad(lpad_ty, bx.eh_personality(), 1);
 667         let tydesc = bx.const_null(bx.type_i8p());
 668         bx.add_clause(vals, tydesc);
 669         let ptr = bx.extract_value(vals, 0);
 670         let catch_ty = bx.type_func(&[bx.type_i8p(), bx.type_i8p()], bx.type_void());
 671         bx.call(catch_ty, None, catch_func, &[data, ptr], None);
 672         bx.ret(bx.const_i32(1));
 673     });
 674
 675     // Note that no invoke is used here because by definition this function
 676     // can't panic (that's what it's catching).
 677     let ret = bx.call(llty, None, llfn, &[try_func, data, catch_func], None);
 678     let i32_align = bx.tcx().data_layout.i32_align.abi;
 679     bx.store(ret, dest, i32_align);
 680 }
 681
 682 // Variant of codegen_gnu_try used for emscripten where Rust panics are
 683 // implemented using C++ exceptions. Here we use exceptions of a specific type
 684 // (`struct rust_panic`) to represent Rust panics.
 685 fn codegen_emcc_try<'ll>(
 686     bx: &mut Builder<'_, 'll, '_>,
 687     try_func: &'ll Value,
 688     data: &'ll Value,
 689     catch_func: &'ll Value,
 690     dest: &'ll Value,
 691 ) {
 692     let (llty, llfn) = get_rust_try_fn(bx, &mut |mut bx| {
 693         // Codegens the shims described above:
 694         //
 695         //   bx:
 696         //      invoke %try_func(%data) normal %normal unwind %catch
 697         //
 698         //   normal:
 699         //      ret 0
 700         //
 701         //   catch:
 702         //      (%ptr, %selector) = landingpad
 703         //      %rust_typeid = @llvm.eh.typeid.for(@_ZTI10rust_panic)
 704         //      %is_rust_panic = %selector == %rust_typeid
 705         //      %catch_data = alloca { i8*, i8 }
 706         //      %catch_data[0] = %ptr
 707         //      %catch_data[1] = %is_rust_panic
 708         //      call %catch_func(%data, %catch_data)
 709         //      ret 1
 710         let then = bx.append_sibling_block("then");
 711         let catch = bx.append_sibling_block("catch");
 712
 713         let try_func = llvm::get_param(bx.llfn(), 0);
 714         let data = llvm::get_param(bx.llfn(), 1);
 715         let catch_func = llvm::get_param(bx.llfn(), 2);
 716         let try_func_ty = bx.type_func(&[bx.type_i8p()], bx.type_void());
 717         bx.invoke(try_func_ty, None, try_func, &[data], then, catch, None);
 718
 719         bx.switch_to_block(then);
 720         bx.ret(bx.const_i32(0));
 721
 722         // Type indicator for the exception being thrown.
 723         //
 724         // The first value in this tuple is a pointer to the exception object
 725         // being thrown.  The second value is a "selector" indicating which of
 726         // the landing pad clauses the exception's type had been matched to.
 727         bx.switch_to_block(catch);
 728         let tydesc = bx.eh_catch_typeinfo();
 729         let lpad_ty = bx.type_struct(&[bx.type_i8p(), bx.type_i32()], false);
 730         let vals = bx.landing_pad(lpad_ty, bx.eh_personality(), 2);
 731         bx.add_clause(vals, tydesc);
 732         bx.add_clause(vals, bx.const_null(bx.type_i8p()));
 733         let ptr = bx.extract_value(vals, 0);
 734         let selector = bx.extract_value(vals, 1);
 735
 736         // Check if the typeid we got is the one for a Rust panic.
 737         let rust_typeid = bx.call_intrinsic("llvm.eh.typeid.for", &[tydesc]);
 738         let is_rust_panic = bx.icmp(IntPredicate::IntEQ, selector, rust_typeid);
 739         let is_rust_panic = bx.zext(is_rust_panic, bx.type_bool());
 740
 741         // We need to pass two values to catch_func (ptr and is_rust_panic), so
 742         // create an alloca and pass a pointer to that.
 743         let ptr_align = bx.tcx().data_layout.pointer_align.abi;
 744         let i8_align = bx.tcx().data_layout.i8_align.abi;
 745         let catch_data_type = bx.type_struct(&[bx.type_i8p(), bx.type_bool()], false);
 746         let catch_data = bx.alloca(catch_data_type, ptr_align);
 747         let catch_data_0 =
 748             bx.inbounds_gep(catch_data_type, catch_data, &[bx.const_usize(0), bx.const_usize(0)]);
 749         bx.store(ptr, catch_data_0, ptr_align);
 750         let catch_data_1 =
 751             bx.inbounds_gep(catch_data_type, catch_data, &[bx.const_usize(0), bx.const_usize(1)]);
 752         bx.store(is_rust_panic, catch_data_1, i8_align);
 753         let catch_data = bx.bitcast(catch_data, bx.type_i8p());
 754
 755         let catch_ty = bx.type_func(&[bx.type_i8p(), bx.type_i8p()], bx.type_void());
 756         bx.call(catch_ty, None, catch_func, &[data, catch_data], None);
 757         bx.ret(bx.const_i32(1));
 758     });
 759
 760     // Note that no invoke is used here because by definition this function
 761     // can't panic (that's what it's catching).
 762     let ret = bx.call(llty, None, llfn, &[try_func, data, catch_func], None);
 763     let i32_align = bx.tcx().data_layout.i32_align.abi;
 764     bx.store(ret, dest, i32_align);
 765 }
 766
 767 // Helper function to give a Block to a closure to codegen a shim function.
 768 // This is currently primarily used for the `try` intrinsic functions above.
 769 fn gen_fn<'ll, 'tcx>(
 770     cx: &CodegenCx<'ll, 'tcx>,
 771     name: &str,
 772     rust_fn_sig: ty::PolyFnSig<'tcx>,
 773     codegen: &mut dyn FnMut(Builder<'_, 'll, 'tcx>),
 774 ) -> (&'ll Type, &'ll Value) {
 775     let fn_abi = cx.fn_abi_of_fn_ptr(rust_fn_sig, ty::List::empty());
 776     let llty = fn_abi.llvm_type(cx);
 777     let llfn = cx.declare_fn(name, fn_abi);
 778     cx.set_frame_pointer_type(llfn);
 779     cx.apply_target_cpu_attr(llfn);
 780     // FIXME(eddyb) find a nicer way to do this.
 781     unsafe { llvm::LLVMRustSetLinkage(llfn, llvm::Linkage::InternalLinkage) };
 782     let llbb = Builder::append_block(cx, llfn, "entry-block");
 783     let bx = Builder::build(cx, llbb);
 784     codegen(bx);
 785     (llty, llfn)
 786 }
 787
 788 // Helper function used to get a handle to the `__rust_try` function used to
 789 // catch exceptions.
 790 //
 791 // This function is only generated once and is then cached.
 792 fn get_rust_try_fn<'ll, 'tcx>(
 793     cx: &CodegenCx<'ll, 'tcx>,
 794     codegen: &mut dyn FnMut(Builder<'_, 'll, 'tcx>),
 795 ) -> (&'ll Type, &'ll Value) {
 796     if let Some(llfn) = cx.rust_try_fn.get() {
 797         return llfn;
 798     }
 799
 800     // Define the type up front for the signature of the rust_try function.
 801     let tcx = cx.tcx;
 802     let i8p = tcx.mk_mut_ptr(tcx.types.i8);
 803     // `unsafe fn(*mut i8) -> ()`
 804     let try_fn_ty = tcx.mk_fn_ptr(ty::Binder::dummy(tcx.mk_fn_sig(
 805         iter::once(i8p),
 806         tcx.mk_unit(),
 807         false,
 808         hir::Unsafety::Unsafe,
 809         Abi::Rust,
 810     )));
 811     // `unsafe fn(*mut i8, *mut i8) -> ()`
 812     let catch_fn_ty = tcx.mk_fn_ptr(ty::Binder::dummy(tcx.mk_fn_sig(
 813         [i8p, i8p].iter().cloned(),
 814         tcx.mk_unit(),
 815         false,
 816         hir::Unsafety::Unsafe,
 817         Abi::Rust,
 818     )));
 819     // `unsafe fn(unsafe fn(*mut i8) -> (), *mut i8, unsafe fn(*mut i8, *mut i8) -> ()) -> i32`
 820     let rust_fn_sig = ty::Binder::dummy(cx.tcx.mk_fn_sig(
 821         [try_fn_ty, i8p, catch_fn_ty].into_iter(),
 822         tcx.types.i32,
 823         false,
 824         hir::Unsafety::Unsafe,
 825         Abi::Rust,
 826     ));
 827     let rust_try = gen_fn(cx, "__rust_try", rust_fn_sig, codegen);
 828     cx.rust_try_fn.set(Some(rust_try));
 829     rust_try
 830 }
 831
 832 fn generic_simd_intrinsic<'ll, 'tcx>(
 833     bx: &mut Builder<'_, 'll, 'tcx>,
 834     name: Symbol,
 835     callee_ty: Ty<'tcx>,
 836     args: &[OperandRef<'tcx, &'ll Value>],
 837     ret_ty: Ty<'tcx>,
 838     llret_ty: &'ll Type,
 839     span: Span,
 840 ) -> Result<&'ll Value, ()> {
 841     // macros for error handling:
 842     #[allow(unused_macro_rules)]
 843     macro_rules! emit_error {
 844         ($msg: tt) => {
 845             emit_error!($msg, )
 846         };
 847         ($msg: tt, $($fmt: tt)*) => {
 848             span_invalid_monomorphization_error(
 849                 bx.sess(), span,
 850                 &format!(concat!("invalid monomorphization of `{}` intrinsic: ", $msg),
 851                          name, $($fmt)*));
 852         }
 853     }
 854
 855     macro_rules! return_error {
 856         ($($fmt: tt)*) => {
 857             {
 858                 emit_error!($($fmt)*);
 859                 return Err(());
 860             }
 861         }
 862     }
 863
 864     macro_rules! require {
 865         ($cond: expr, $($fmt: tt)*) => {
 866             if !$cond {
 867                 return_error!($($fmt)*);
 868             }
 869         };
 870     }
 871
 872     macro_rules! require_simd {
 873         ($ty: expr, $position: expr) => {
 874             require!($ty.is_simd(), "expected SIMD {} type, found non-SIMD `{}`", $position, $ty)
 875         };
 876     }
 877
 878     let tcx = bx.tcx();
 879     let sig =
 880         tcx.normalize_erasing_late_bound_regions(ty::ParamEnv::reveal_all(), callee_ty.fn_sig(tcx));
 881     let arg_tys = sig.inputs();
 882
 883     if name == sym::simd_select_bitmask {
 884         require_simd!(arg_tys[1], "argument");
 885         let (len, _) = arg_tys[1].simd_size_and_type(bx.tcx());
 886
 887         let expected_int_bits = (len.max(8) - 1).next_power_of_two();
 888         let expected_bytes = len / 8 + ((len % 8 > 0) as u64);
 889
 890         let mask_ty = arg_tys[0];
 891         let mask = match mask_ty.kind() {
 892             ty::Int(i) if i.bit_width() == Some(expected_int_bits) => args[0].immediate(),
 893             ty::Uint(i) if i.bit_width() == Some(expected_int_bits) => args[0].immediate(),
 894             ty::Array(elem, len)
 895                 if matches!(elem.kind(), ty::Uint(ty::UintTy::U8))
 896                     && len.try_eval_usize(bx.tcx, ty::ParamEnv::reveal_all())
 897                         == Some(expected_bytes) =>
 898             {
 899                 let place = PlaceRef::alloca(bx, args[0].layout);
 900                 args[0].val.store(bx, place);
 901                 let int_ty = bx.type_ix(expected_bytes * 8);
 902                 let ptr = bx.pointercast(place.llval, bx.cx.type_ptr_to(int_ty));
 903                 bx.load(int_ty, ptr, Align::ONE)
 904             }
 905             _ => return_error!(
 906                 "invalid bitmask `{}`, expected `u{}` or `[u8; {}]`",
 907                 mask_ty,
 908                 expected_int_bits,
 909                 expected_bytes
 910             ),
 911         };
 912
 913         let i1 = bx.type_i1();
 914         let im = bx.type_ix(len);
 915         let i1xn = bx.type_vector(i1, len);
 916         let m_im = bx.trunc(mask, im);
 917         let m_i1s = bx.bitcast(m_im, i1xn);
 918         return Ok(bx.select(m_i1s, args[1].immediate(), args[2].immediate()));
 919     }
 920
 921     // every intrinsic below takes a SIMD vector as its first argument
 922     require_simd!(arg_tys[0], "input");
 923     let in_ty = arg_tys[0];
 924
 925     let comparison = match name {
 926         sym::simd_eq => Some(hir::BinOpKind::Eq),
 927         sym::simd_ne => Some(hir::BinOpKind::Ne),
 928         sym::simd_lt => Some(hir::BinOpKind::Lt),
 929         sym::simd_le => Some(hir::BinOpKind::Le),
 930         sym::simd_gt => Some(hir::BinOpKind::Gt),
 931         sym::simd_ge => Some(hir::BinOpKind::Ge),
 932         _ => None,
 933     };
 934
 935     let (in_len, in_elem) = arg_tys[0].simd_size_and_type(bx.tcx());
 936     if let Some(cmp_op) = comparison {
 937         require_simd!(ret_ty, "return");
 938
 939         let (out_len, out_ty) = ret_ty.simd_size_and_type(bx.tcx());
 940         require!(
 941             in_len == out_len,
 942             "expected return type with length {} (same as input type `{}`), \
 943              found `{}` with length {}",
 944             in_len,
 945             in_ty,
 946             ret_ty,
 947             out_len
 948         );
 949         require!(
 950             bx.type_kind(bx.element_type(llret_ty)) == TypeKind::Integer,
 951             "expected return type with integer elements, found `{}` with non-integer `{}`",
 952             ret_ty,
 953             out_ty
 954         );
 955
 956         return Ok(compare_simd_types(
 957             bx,
 958             args[0].immediate(),
 959             args[1].immediate(),
 960             in_elem,
 961             llret_ty,
 962             cmp_op,
 963         ));
 964     }
 965
 966     if let Some(stripped) = name.as_str().strip_prefix("simd_shuffle") {
 967         // If this intrinsic is the older "simd_shuffleN" form, simply parse the integer.
 968         // If there is no suffix, use the index array length.
 969         let n: u64 = if stripped.is_empty() {
 970             // Make sure this is actually an array, since typeck only checks the length-suffixed
 971             // version of this intrinsic.
 972             match args[2].layout.ty.kind() {
 973                 ty::Array(ty, len) if matches!(ty.kind(), ty::Uint(ty::UintTy::U32)) => {
 974                     len.try_eval_usize(bx.cx.tcx, ty::ParamEnv::reveal_all()).unwrap_or_else(|| {
 975                         span_bug!(span, "could not evaluate shuffle index array length")
 976                     })
 977                 }
 978                 _ => return_error!(
 979                     "simd_shuffle index must be an array of `u32`, got `{}`",
 980                     args[2].layout.ty
 981                 ),
 982             }
 983         } else {
 984             stripped.parse().unwrap_or_else(|_| {
 985                 span_bug!(span, "bad `simd_shuffle` instruction only caught in codegen?")
 986             })
 987         };
 988
 989         require_simd!(ret_ty, "return");
 990         let (out_len, out_ty) = ret_ty.simd_size_and_type(bx.tcx());
 991         require!(
 992             out_len == n,
 993             "expected return type of length {}, found `{}` with length {}",
 994             n,
 995             ret_ty,
 996             out_len
 997         );
 998         require!(
 999             in_elem == out_ty,
1000             "expected return element type `{}` (element of input `{}`), \
1001              found `{}` with element type `{}`",
1002             in_elem,
1003             in_ty,
1004             ret_ty,
1005             out_ty
1006         );
1007
1008         let total_len = u128::from(in_len) * 2;
1009
1010         let vector = args[2].immediate();
1011
1012         let indices: Option<Vec<_>> = (0..n)
1013             .map(|i| {
1014                 let arg_idx = i;
1015                 let val = bx.const_get_elt(vector, i as u64);
1016                 match bx.const_to_opt_u128(val, true) {
1017                     None => {
1018                         emit_error!("shuffle index #{} is not a constant", arg_idx);
1019                         None
1020                     }
1021                     Some(idx) if idx >= total_len => {
1022                         emit_error!(
1023                             "shuffle index #{} is out of bounds (limit {})",
1024                             arg_idx,
1025                             total_len
1026                         );
1027                         None
1028                     }
1029                     Some(idx) => Some(bx.const_i32(idx as i32)),
1030                 }
1031             })
1032             .collect();
1033         let Some(indices) = indices else {
1034             return Ok(bx.const_null(llret_ty));
1035         };
1036
1037         return Ok(bx.shuffle_vector(
1038             args[0].immediate(),
1039             args[1].immediate(),
1040             bx.const_vector(&indices),
1041         ));
1042     }
1043
1044     if name == sym::simd_insert {
1045         require!(
1046             in_elem == arg_tys[2],
1047             "expected inserted type `{}` (element of input `{}`), found `{}`",
1048             in_elem,
1049             in_ty,
1050             arg_tys[2]
1051         );
1052         return Ok(bx.insert_element(
1053             args[0].immediate(),
1054             args[2].immediate(),
1055             args[1].immediate(),
1056         ));
1057     }
1058     if name == sym::simd_extract {
1059         require!(
1060             ret_ty == in_elem,
1061             "expected return type `{}` (element of input `{}`), found `{}`",
1062             in_elem,
1063             in_ty,
1064             ret_ty
1065         );
1066         return Ok(bx.extract_element(args[0].immediate(), args[1].immediate()));
1067     }
1068
1069     if name == sym::simd_select {
1070         let m_elem_ty = in_elem;
1071         let m_len = in_len;
1072         require_simd!(arg_tys[1], "argument");
1073         let (v_len, _) = arg_tys[1].simd_size_and_type(bx.tcx());
1074         require!(
1075             m_len == v_len,
1076             "mismatched lengths: mask length `{}` != other vector length `{}`",
1077             m_len,
1078             v_len
1079         );
1080         match m_elem_ty.kind() {
1081             ty::Int(_) => {}
1082             _ => return_error!("mask element type is `{}`, expected `i_`", m_elem_ty),
1083         }
1084         // truncate the mask to a vector of i1s
1085         let i1 = bx.type_i1();
1086         let i1xn = bx.type_vector(i1, m_len as u64);
1087         let m_i1s = bx.trunc(args[0].immediate(), i1xn);
1088         return Ok(bx.select(m_i1s, args[1].immediate(), args[2].immediate()));
1089     }
1090
1091     if name == sym::simd_bitmask {
1092         // The `fn simd_bitmask(vector) -> unsigned integer` intrinsic takes a
1093         // vector mask and returns the most significant bit (MSB) of each lane in the form
1094         // of either:
1095         // * an unsigned integer
1096         // * an array of `u8`
1097         // If the vector has less than 8 lanes, a u8 is returned with zeroed trailing bits.
1098         //
1099         // The bit order of the result depends on the byte endianness, LSB-first for little
1100         // endian and MSB-first for big endian.
1101         let expected_int_bits = in_len.max(8);
1102         let expected_bytes = expected_int_bits / 8 + ((expected_int_bits % 8 > 0) as u64);
1103
1104         // Integer vector <i{in_bitwidth} x in_len>:
1105         let (i_xn, in_elem_bitwidth) = match in_elem.kind() {
1106             ty::Int(i) => (
1107                 args[0].immediate(),
1108                 i.bit_width().unwrap_or_else(|| bx.data_layout().pointer_size.bits()),
1109             ),
1110             ty::Uint(i) => (
1111                 args[0].immediate(),
1112                 i.bit_width().unwrap_or_else(|| bx.data_layout().pointer_size.bits()),
1113             ),
1114             _ => return_error!(
1115                 "vector argument `{}`'s element type `{}`, expected integer element type",
1116                 in_ty,
1117                 in_elem
1118             ),
1119         };
1120
1121         // Shift the MSB to the right by "in_elem_bitwidth - 1" into the first bit position.
1122         let shift_indices =
1123             vec![
1124                 bx.cx.const_int(bx.type_ix(in_elem_bitwidth), (in_elem_bitwidth - 1) as _);
1125                 in_len as _
1126             ];
1127         let i_xn_msb = bx.lshr(i_xn, bx.const_vector(shift_indices.as_slice()));
1128         // Truncate vector to an <i1 x N>
1129         let i1xn = bx.trunc(i_xn_msb, bx.type_vector(bx.type_i1(), in_len));
1130         // Bitcast <i1 x N> to iN:
1131         let i_ = bx.bitcast(i1xn, bx.type_ix(in_len));
1132
1133         match ret_ty.kind() {
1134             ty::Uint(i) if i.bit_width() == Some(expected_int_bits) => {
1135                 // Zero-extend iN to the bitmask type:
1136                 return Ok(bx.zext(i_, bx.type_ix(expected_int_bits)));
1137             }
1138             ty::Array(elem, len)
1139                 if matches!(elem.kind(), ty::Uint(ty::UintTy::U8))
1140                     && len.try_eval_usize(bx.tcx, ty::ParamEnv::reveal_all())
1141                         == Some(expected_bytes) =>
1142             {
1143                 // Zero-extend iN to the array length:
1144                 let ze = bx.zext(i_, bx.type_ix(expected_bytes * 8));
1145
1146                 // Convert the integer to a byte array
1147                 let ptr = bx.alloca(bx.type_ix(expected_bytes * 8), Align::ONE);
1148                 bx.store(ze, ptr, Align::ONE);
1149                 let array_ty = bx.type_array(bx.type_i8(), expected_bytes);
1150                 let ptr = bx.pointercast(ptr, bx.cx.type_ptr_to(array_ty));
1151                 return Ok(bx.load(array_ty, ptr, Align::ONE));
1152             }
1153             _ => return_error!(
1154                 "cannot return `{}`, expected `u{}` or `[u8; {}]`",
1155                 ret_ty,
1156                 expected_int_bits,
1157                 expected_bytes
1158             ),
1159         }
1160     }
1161
1162     fn simd_simple_float_intrinsic<'ll, 'tcx>(
1163         name: Symbol,
1164         in_elem: Ty<'_>,
1165         in_ty: Ty<'_>,
1166         in_len: u64,
1167         bx: &mut Builder<'_, 'll, 'tcx>,
1168         span: Span,
1169         args: &[OperandRef<'tcx, &'ll Value>],
1170     ) -> Result<&'ll Value, ()> {
1171         #[allow(unused_macro_rules)]
1172         macro_rules! emit_error {
1173             ($msg: tt) => {
1174                 emit_error!($msg, )
1175             };
1176             ($msg: tt, $($fmt: tt)*) => {
1177                 span_invalid_monomorphization_error(
1178                     bx.sess(), span,
1179                     &format!(concat!("invalid monomorphization of `{}` intrinsic: ", $msg),
1180                              name, $($fmt)*));
1181             }
1182         }
1183         macro_rules! return_error {
1184             ($($fmt: tt)*) => {
1185                 {
1186                     emit_error!($($fmt)*);
1187                     return Err(());
1188                 }
1189             }
1190         }
1191
1192         let (elem_ty_str, elem_ty) = if let ty::Float(f) = in_elem.kind() {
1193             let elem_ty = bx.cx.type_float_from_ty(*f);
1194             match f.bit_width() {
1195                 32 => ("f32", elem_ty),
1196                 64 => ("f64", elem_ty),
1197                 _ => {
1198                     return_error!(
1199                         "unsupported element type `{}` of floating-point vector `{}`",
1200                         f.name_str(),
1201                         in_ty
1202                     );
1203                 }
1204             }
1205         } else {
1206             return_error!("`{}` is not a floating-point type", in_ty);
1207         };
1208
1209         let vec_ty = bx.type_vector(elem_ty, in_len);
1210
1211         let (intr_name, fn_ty) = match name {
1212             sym::simd_ceil => ("ceil", bx.type_func(&[vec_ty], vec_ty)),
1213             sym::simd_fabs => ("fabs", bx.type_func(&[vec_ty], vec_ty)),
1214             sym::simd_fcos => ("cos", bx.type_func(&[vec_ty], vec_ty)),
1215             sym::simd_fexp2 => ("exp2", bx.type_func(&[vec_ty], vec_ty)),
1216             sym::simd_fexp => ("exp", bx.type_func(&[vec_ty], vec_ty)),
1217             sym::simd_flog10 => ("log10", bx.type_func(&[vec_ty], vec_ty)),
1218             sym::simd_flog2 => ("log2", bx.type_func(&[vec_ty], vec_ty)),
1219             sym::simd_flog => ("log", bx.type_func(&[vec_ty], vec_ty)),
1220             sym::simd_floor => ("floor", bx.type_func(&[vec_ty], vec_ty)),
1221             sym::simd_fma => ("fma", bx.type_func(&[vec_ty, vec_ty, vec_ty], vec_ty)),
1222             sym::simd_fpowi => ("powi", bx.type_func(&[vec_ty, bx.type_i32()], vec_ty)),
1223             sym::simd_fpow => ("pow", bx.type_func(&[vec_ty, vec_ty], vec_ty)),
1224             sym::simd_fsin => ("sin", bx.type_func(&[vec_ty], vec_ty)),
1225             sym::simd_fsqrt => ("sqrt", bx.type_func(&[vec_ty], vec_ty)),
1226             sym::simd_round => ("round", bx.type_func(&[vec_ty], vec_ty)),
1227             sym::simd_trunc => ("trunc", bx.type_func(&[vec_ty], vec_ty)),
1228             _ => return_error!("unrecognized intrinsic `{}`", name),
1229         };
1230         let llvm_name = &format!("llvm.{0}.v{1}{2}", intr_name, in_len, elem_ty_str);
1231         let f = bx.declare_cfn(llvm_name, llvm::UnnamedAddr::No, fn_ty);
1232         let c = bx.call(
1233             fn_ty,
1234             None,
1235             f,
1236             &args.iter().map(|arg| arg.immediate()).collect::<Vec<_>>(),
1237             None,
1238         );
1239         Ok(c)
1240     }
1241
1242     if std::matches!(
1243         name,
1244         sym::simd_ceil
1245             | sym::simd_fabs
1246             | sym::simd_fcos
1247             | sym::simd_fexp2
1248             | sym::simd_fexp
1249             | sym::simd_flog10
1250             | sym::simd_flog2
1251             | sym::simd_flog
1252             | sym::simd_floor
1253             | sym::simd_fma
1254             | sym::simd_fpow
1255             | sym::simd_fpowi
1256             | sym::simd_fsin
1257             | sym::simd_fsqrt
1258             | sym::simd_round
1259             | sym::simd_trunc
1260     ) {
1261         return simd_simple_float_intrinsic(name, in_elem, in_ty, in_len, bx, span, args);
1262     }
1263
1264     // FIXME: use:
1265     //  https://github.com/llvm-mirror/llvm/blob/master/include/llvm/IR/Function.h#L182
1266     //  https://github.com/llvm-mirror/llvm/blob/master/include/llvm/IR/Intrinsics.h#L81
1267     fn llvm_vector_str(
1268         elem_ty: Ty<'_>,
1269         vec_len: u64,
1270         no_pointers: usize,
1271         bx: &Builder<'_, '_, '_>,
1272     ) -> String {
1273         let p0s: String = "p0".repeat(no_pointers);
1274         match *elem_ty.kind() {
1275             ty::Int(v) => format!(
1276                 "v{}{}i{}",
1277                 vec_len,
1278                 p0s,
1279                 // Normalize to prevent crash if v: IntTy::Isize
1280                 v.normalize(bx.target_spec().pointer_width).bit_width().unwrap()
1281             ),
1282             ty::Uint(v) => format!(
1283                 "v{}{}i{}",
1284                 vec_len,
1285                 p0s,
1286                 // Normalize to prevent crash if v: UIntTy::Usize
1287                 v.normalize(bx.target_spec().pointer_width).bit_width().unwrap()
1288             ),
1289             ty::Float(v) => format!("v{}{}f{}", vec_len, p0s, v.bit_width()),
1290             _ => unreachable!(),
1291         }
1292     }
1293
1294     fn llvm_vector_ty<'ll>(
1295         cx: &CodegenCx<'ll, '_>,
1296         elem_ty: Ty<'_>,
1297         vec_len: u64,
1298         mut no_pointers: usize,
1299     ) -> &'ll Type {
1300         // FIXME: use cx.layout_of(ty).llvm_type() ?
1301         let mut elem_ty = match *elem_ty.kind() {
1302             ty::Int(v) => cx.type_int_from_ty(v),
1303             ty::Uint(v) => cx.type_uint_from_ty(v),
1304             ty::Float(v) => cx.type_float_from_ty(v),
1305             _ => unreachable!(),
1306         };
1307         while no_pointers > 0 {
1308             elem_ty = cx.type_ptr_to(elem_ty);
1309             no_pointers -= 1;
1310         }
1311         cx.type_vector(elem_ty, vec_len)
1312     }
1313
1314     if name == sym::simd_gather {
1315         // simd_gather(values: <N x T>, pointers: <N x *_ T>,
1316         //             mask: <N x i{M}>) -> <N x T>
1317         // * N: number of elements in the input vectors
1318         // * T: type of the element to load
1319         // * M: any integer width is supported, will be truncated to i1
1320
1321         // All types must be simd vector types
1322         require_simd!(in_ty, "first");
1323         require_simd!(arg_tys[1], "second");
1324         require_simd!(arg_tys[2], "third");
1325         require_simd!(ret_ty, "return");
1326
1327         // Of the same length:
1328         let (out_len, _) = arg_tys[1].simd_size_and_type(bx.tcx());
1329         let (out_len2, _) = arg_tys[2].simd_size_and_type(bx.tcx());
1330         require!(
1331             in_len == out_len,
1332             "expected {} argument with length {} (same as input type `{}`), \
1333              found `{}` with length {}",
1334             "second",
1335             in_len,
1336             in_ty,
1337             arg_tys[1],
1338             out_len
1339         );
1340         require!(
1341             in_len == out_len2,
1342             "expected {} argument with length {} (same as input type `{}`), \
1343              found `{}` with length {}",
1344             "third",
1345             in_len,
1346             in_ty,
1347             arg_tys[2],
1348             out_len2
1349         );
1350
1351         // The return type must match the first argument type
1352         require!(ret_ty == in_ty, "expected return type `{}`, found `{}`", in_ty, ret_ty);
1353
1354         // This counts how many pointers
1355         fn ptr_count(t: Ty<'_>) -> usize {
1356             match t.kind() {
1357                 ty::RawPtr(p) => 1 + ptr_count(p.ty),
1358                 _ => 0,
1359             }
1360         }
1361
1362         // Non-ptr type
1363         fn non_ptr(t: Ty<'_>) -> Ty<'_> {
1364             match t.kind() {
1365                 ty::RawPtr(p) => non_ptr(p.ty),
1366                 _ => t,
1367             }
1368         }
1369
1370         // The second argument must be a simd vector with an element type that's a pointer
1371         // to the element type of the first argument
1372         let (_, element_ty0) = arg_tys[0].simd_size_and_type(bx.tcx());
1373         let (_, element_ty1) = arg_tys[1].simd_size_and_type(bx.tcx());
1374         let (pointer_count, underlying_ty) = match element_ty1.kind() {
1375             ty::RawPtr(p) if p.ty == in_elem => (ptr_count(element_ty1), non_ptr(element_ty1)),
1376             _ => {
1377                 require!(
1378                     false,
1379                     "expected element type `{}` of second argument `{}` \
1380                         to be a pointer to the element type `{}` of the first \
1381                         argument `{}`, found `{}` != `*_ {}`",
1382                     element_ty1,
1383                     arg_tys[1],
1384                     in_elem,
1385                     in_ty,
1386                     element_ty1,
1387                     in_elem
1388                 );
1389                 unreachable!();
1390             }
1391         };
1392         assert!(pointer_count > 0);
1393         assert_eq!(pointer_count - 1, ptr_count(element_ty0));
1394         assert_eq!(underlying_ty, non_ptr(element_ty0));
1395
1396         // The element type of the third argument must be a signed integer type of any width:
1397         let (_, element_ty2) = arg_tys[2].simd_size_and_type(bx.tcx());
1398         match element_ty2.kind() {
1399             ty::Int(_) => (),
1400             _ => {
1401                 require!(
1402                     false,
1403                     "expected element type `{}` of third argument `{}` \
1404                                  to be a signed integer type",
1405                     element_ty2,
1406                     arg_tys[2]
1407                 );
1408             }
1409         }
1410
1411         // Alignment of T, must be a constant integer value:
1412         let alignment_ty = bx.type_i32();
1413         let alignment = bx.const_i32(bx.align_of(in_elem).bytes() as i32);
1414
1415         // Truncate the mask vector to a vector of i1s:
1416         let (mask, mask_ty) = {
1417             let i1 = bx.type_i1();
1418             let i1xn = bx.type_vector(i1, in_len);
1419             (bx.trunc(args[2].immediate(), i1xn), i1xn)
1420         };
1421
1422         // Type of the vector of pointers:
1423         let llvm_pointer_vec_ty = llvm_vector_ty(bx, underlying_ty, in_len, pointer_count);
1424         let llvm_pointer_vec_str = llvm_vector_str(underlying_ty, in_len, pointer_count, bx);
1425
1426         // Type of the vector of elements:
1427         let llvm_elem_vec_ty = llvm_vector_ty(bx, underlying_ty, in_len, pointer_count - 1);
1428         let llvm_elem_vec_str = llvm_vector_str(underlying_ty, in_len, pointer_count - 1, bx);
1429
1430         let llvm_intrinsic =
1431             format!("llvm.masked.gather.{}.{}", llvm_elem_vec_str, llvm_pointer_vec_str);
1432         let fn_ty = bx.type_func(
1433             &[llvm_pointer_vec_ty, alignment_ty, mask_ty, llvm_elem_vec_ty],
1434             llvm_elem_vec_ty,
1435         );
1436         let f = bx.declare_cfn(&llvm_intrinsic, llvm::UnnamedAddr::No, fn_ty);
1437         let v = bx.call(
1438             fn_ty,
1439             None,
1440             f,
1441             &[args[1].immediate(), alignment, mask, args[0].immediate()],
1442             None,
1443         );
1444         return Ok(v);
1445     }
1446
1447     if name == sym::simd_scatter {
1448         // simd_scatter(values: <N x T>, pointers: <N x *mut T>,
1449         //             mask: <N x i{M}>) -> ()
1450         // * N: number of elements in the input vectors
1451         // * T: type of the element to load
1452         // * M: any integer width is supported, will be truncated to i1
1453
1454         // All types must be simd vector types
1455         require_simd!(in_ty, "first");
1456         require_simd!(arg_tys[1], "second");
1457         require_simd!(arg_tys[2], "third");
1458
1459         // Of the same length:
1460         let (element_len1, _) = arg_tys[1].simd_size_and_type(bx.tcx());
1461         let (element_len2, _) = arg_tys[2].simd_size_and_type(bx.tcx());
1462         require!(
1463             in_len == element_len1,
1464             "expected {} argument with length {} (same as input type `{}`), \
1465             found `{}` with length {}",
1466             "second",
1467             in_len,
1468             in_ty,
1469             arg_tys[1],
1470             element_len1
1471         );
1472         require!(
1473             in_len == element_len2,
1474             "expected {} argument with length {} (same as input type `{}`), \
1475             found `{}` with length {}",
1476             "third",
1477             in_len,
1478             in_ty,
1479             arg_tys[2],
1480             element_len2
1481         );
1482
1483         // This counts how many pointers
1484         fn ptr_count(t: Ty<'_>) -> usize {
1485             match t.kind() {
1486                 ty::RawPtr(p) => 1 + ptr_count(p.ty),
1487                 _ => 0,
1488             }
1489         }
1490
1491         // Non-ptr type
1492         fn non_ptr(t: Ty<'_>) -> Ty<'_> {
1493             match t.kind() {
1494                 ty::RawPtr(p) => non_ptr(p.ty),
1495                 _ => t,
1496             }
1497         }
1498
1499         // The second argument must be a simd vector with an element type that's a pointer
1500         // to the element type of the first argument
1501         let (_, element_ty0) = arg_tys[0].simd_size_and_type(bx.tcx());
1502         let (_, element_ty1) = arg_tys[1].simd_size_and_type(bx.tcx());
1503         let (_, element_ty2) = arg_tys[2].simd_size_and_type(bx.tcx());
1504         let (pointer_count, underlying_ty) = match element_ty1.kind() {
1505             ty::RawPtr(p) if p.ty == in_elem && p.mutbl.is_mut() => {
1506                 (ptr_count(element_ty1), non_ptr(element_ty1))
1507             }
1508             _ => {
1509                 require!(
1510                     false,
1511                     "expected element type `{}` of second argument `{}` \
1512                         to be a pointer to the element type `{}` of the first \
1513                         argument `{}`, found `{}` != `*mut {}`",
1514                     element_ty1,
1515                     arg_tys[1],
1516                     in_elem,
1517                     in_ty,
1518                     element_ty1,
1519                     in_elem
1520                 );
1521                 unreachable!();
1522             }
1523         };
1524         assert!(pointer_count > 0);
1525         assert_eq!(pointer_count - 1, ptr_count(element_ty0));
1526         assert_eq!(underlying_ty, non_ptr(element_ty0));
1527
1528         // The element type of the third argument must be a signed integer type of any width:
1529         match element_ty2.kind() {
1530             ty::Int(_) => (),
1531             _ => {
1532                 require!(
1533                     false,
1534                     "expected element type `{}` of third argument `{}` \
1535                          be a signed integer type",
1536                     element_ty2,
1537                     arg_tys[2]
1538                 );
1539             }
1540         }
1541
1542         // Alignment of T, must be a constant integer value:
1543         let alignment_ty = bx.type_i32();
1544         let alignment = bx.const_i32(bx.align_of(in_elem).bytes() as i32);
1545
1546         // Truncate the mask vector to a vector of i1s:
1547         let (mask, mask_ty) = {
1548             let i1 = bx.type_i1();
1549             let i1xn = bx.type_vector(i1, in_len);
1550             (bx.trunc(args[2].immediate(), i1xn), i1xn)
1551         };
1552
1553         let ret_t = bx.type_void();
1554
1555         // Type of the vector of pointers:
1556         let llvm_pointer_vec_ty = llvm_vector_ty(bx, underlying_ty, in_len, pointer_count);
1557         let llvm_pointer_vec_str = llvm_vector_str(underlying_ty, in_len, pointer_count, bx);
1558
1559         // Type of the vector of elements:
1560         let llvm_elem_vec_ty = llvm_vector_ty(bx, underlying_ty, in_len, pointer_count - 1);
1561         let llvm_elem_vec_str = llvm_vector_str(underlying_ty, in_len, pointer_count - 1, bx);
1562
1563         let llvm_intrinsic =
1564             format!("llvm.masked.scatter.{}.{}", llvm_elem_vec_str, llvm_pointer_vec_str);
1565         let fn_ty =
1566             bx.type_func(&[llvm_elem_vec_ty, llvm_pointer_vec_ty, alignment_ty, mask_ty], ret_t);
1567         let f = bx.declare_cfn(&llvm_intrinsic, llvm::UnnamedAddr::No, fn_ty);
1568         let v = bx.call(
1569             fn_ty,
1570             None,
1571             f,
1572             &[args[0].immediate(), args[1].immediate(), alignment, mask],
1573             None,
1574         );
1575         return Ok(v);
1576     }
1577
1578     macro_rules! arith_red {
1579         ($name:ident : $integer_reduce:ident, $float_reduce:ident, $ordered:expr, $op:ident,
1580          $identity:expr) => {
1581             if name == sym::$name {
1582                 require!(
1583                     ret_ty == in_elem,
1584                     "expected return type `{}` (element of input `{}`), found `{}`",
1585                     in_elem,
1586                     in_ty,
1587                     ret_ty
1588                 );
1589                 return match in_elem.kind() {
1590                     ty::Int(_) | ty::Uint(_) => {
1591                         let r = bx.$integer_reduce(args[0].immediate());
1592                         if $ordered {
1593                             // if overflow occurs, the result is the
1594                             // mathematical result modulo 2^n:
1595                             Ok(bx.$op(args[1].immediate(), r))
1596                         } else {
1597                             Ok(bx.$integer_reduce(args[0].immediate()))
1598                         }
1599                     }
1600                     ty::Float(f) => {
1601                         let acc = if $ordered {
1602                             // ordered arithmetic reductions take an accumulator
1603                             args[1].immediate()
1604                         } else {
1605                             // unordered arithmetic reductions use the identity accumulator
1606                             match f.bit_width() {
1607                                 32 => bx.const_real(bx.type_f32(), $identity),
1608                                 64 => bx.const_real(bx.type_f64(), $identity),
1609                                 v => return_error!(
1610                                     r#"
1611 unsupported {} from `{}` with element `{}` of size `{}` to `{}`"#,
1612                                     sym::$name,
1613                                     in_ty,
1614                                     in_elem,
1615                                     v,
1616                                     ret_ty
1617                                 ),
1618                             }
1619                         };
1620                         Ok(bx.$float_reduce(acc, args[0].immediate()))
1621                     }
1622                     _ => return_error!(
1623                         "unsupported {} from `{}` with element `{}` to `{}`",
1624                         sym::$name,
1625                         in_ty,
1626                         in_elem,
1627                         ret_ty
1628                     ),
1629                 };
1630             }
1631         };
1632     }
1633
1634     arith_red!(simd_reduce_add_ordered: vector_reduce_add, vector_reduce_fadd, true, add, 0.0);
1635     arith_red!(simd_reduce_mul_ordered: vector_reduce_mul, vector_reduce_fmul, true, mul, 1.0);
1636     arith_red!(
1637         simd_reduce_add_unordered: vector_reduce_add,
1638         vector_reduce_fadd_fast,
1639         false,
1640         add,
1641         0.0
1642     );
1643     arith_red!(
1644         simd_reduce_mul_unordered: vector_reduce_mul,
1645         vector_reduce_fmul_fast,
1646         false,
1647         mul,
1648         1.0
1649     );
1650
1651     macro_rules! minmax_red {
1652         ($name:ident: $int_red:ident, $float_red:ident) => {
1653             if name == sym::$name {
1654                 require!(
1655                     ret_ty == in_elem,
1656                     "expected return type `{}` (element of input `{}`), found `{}`",
1657                     in_elem,
1658                     in_ty,
1659                     ret_ty
1660                 );
1661                 return match in_elem.kind() {
1662                     ty::Int(_i) => Ok(bx.$int_red(args[0].immediate(), true)),
1663                     ty::Uint(_u) => Ok(bx.$int_red(args[0].immediate(), false)),
1664                     ty::Float(_f) => Ok(bx.$float_red(args[0].immediate())),
1665                     _ => return_error!(
1666                         "unsupported {} from `{}` with element `{}` to `{}`",
1667                         sym::$name,
1668                         in_ty,
1669                         in_elem,
1670                         ret_ty
1671                     ),
1672                 };
1673             }
1674         };
1675     }
1676
1677     minmax_red!(simd_reduce_min: vector_reduce_min, vector_reduce_fmin);
1678     minmax_red!(simd_reduce_max: vector_reduce_max, vector_reduce_fmax);
1679
1680     minmax_red!(simd_reduce_min_nanless: vector_reduce_min, vector_reduce_fmin_fast);
1681     minmax_red!(simd_reduce_max_nanless: vector_reduce_max, vector_reduce_fmax_fast);
1682
1683     macro_rules! bitwise_red {
1684         ($name:ident : $red:ident, $boolean:expr) => {
1685             if name == sym::$name {
1686                 let input = if !$boolean {
1687                     require!(
1688                         ret_ty == in_elem,
1689                         "expected return type `{}` (element of input `{}`), found `{}`",
1690                         in_elem,
1691                         in_ty,
1692                         ret_ty
1693                     );
1694                     args[0].immediate()
1695                 } else {
1696                     match in_elem.kind() {
1697                         ty::Int(_) | ty::Uint(_) => {}
1698                         _ => return_error!(
1699                             "unsupported {} from `{}` with element `{}` to `{}`",
1700                             sym::$name,
1701                             in_ty,
1702                             in_elem,
1703                             ret_ty
1704                         ),
1705                     }
1706
1707                     // boolean reductions operate on vectors of i1s:
1708                     let i1 = bx.type_i1();
1709                     let i1xn = bx.type_vector(i1, in_len as u64);
1710                     bx.trunc(args[0].immediate(), i1xn)
1711                 };
1712                 return match in_elem.kind() {
1713                     ty::Int(_) | ty::Uint(_) => {
1714                         let r = bx.$red(input);
1715                         Ok(if !$boolean { r } else { bx.zext(r, bx.type_bool()) })
1716                     }
1717                     _ => return_error!(
1718                         "unsupported {} from `{}` with element `{}` to `{}`",
1719                         sym::$name,
1720                         in_ty,
1721                         in_elem,
1722                         ret_ty
1723                     ),
1724                 };
1725             }
1726         };
1727     }
1728
1729     bitwise_red!(simd_reduce_and: vector_reduce_and, false);
1730     bitwise_red!(simd_reduce_or: vector_reduce_or, false);
1731     bitwise_red!(simd_reduce_xor: vector_reduce_xor, false);
1732     bitwise_red!(simd_reduce_all: vector_reduce_and, true);
1733     bitwise_red!(simd_reduce_any: vector_reduce_or, true);
1734
1735     if name == sym::simd_cast_ptr {
1736         require_simd!(ret_ty, "return");
1737         let (out_len, out_elem) = ret_ty.simd_size_and_type(bx.tcx());
1738         require!(
1739             in_len == out_len,
1740             "expected return type with length {} (same as input type `{}`), \
1741                   found `{}` with length {}",
1742             in_len,
1743             in_ty,
1744             ret_ty,
1745             out_len
1746         );
1747
1748         match in_elem.kind() {
1749             ty::RawPtr(p) => {
1750                 let (metadata, check_sized) = p.ty.ptr_metadata_ty(bx.tcx, |ty| {
1751                     bx.tcx.normalize_erasing_regions(ty::ParamEnv::reveal_all(), ty)
1752                 });
1753                 assert!(!check_sized); // we are in codegen, so we shouldn't see these types
1754                 require!(metadata.is_unit(), "cannot cast fat pointer `{}`", in_elem)
1755             }
1756             _ => return_error!("expected pointer, got `{}`", in_elem),
1757         }
1758         match out_elem.kind() {
1759             ty::RawPtr(p) => {
1760                 let (metadata, check_sized) = p.ty.ptr_metadata_ty(bx.tcx, |ty| {
1761                     bx.tcx.normalize_erasing_regions(ty::ParamEnv::reveal_all(), ty)
1762                 });
1763                 assert!(!check_sized); // we are in codegen, so we shouldn't see these types
1764                 require!(metadata.is_unit(), "cannot cast to fat pointer `{}`", out_elem)
1765             }
1766             _ => return_error!("expected pointer, got `{}`", out_elem),
1767         }
1768
1769         if in_elem == out_elem {
1770             return Ok(args[0].immediate());
1771         } else {
1772             return Ok(bx.pointercast(args[0].immediate(), llret_ty));
1773         }
1774     }
1775
1776     if name == sym::simd_expose_addr {
1777         require_simd!(ret_ty, "return");
1778         let (out_len, out_elem) = ret_ty.simd_size_and_type(bx.tcx());
1779         require!(
1780             in_len == out_len,
1781             "expected return type with length {} (same as input type `{}`), \
1782                   found `{}` with length {}",
1783             in_len,
1784             in_ty,
1785             ret_ty,
1786             out_len
1787         );
1788
1789         match in_elem.kind() {
1790             ty::RawPtr(_) => {}
1791             _ => return_error!("expected pointer, got `{}`", in_elem),
1792         }
1793         match out_elem.kind() {
1794             ty::Uint(ty::UintTy::Usize) => {}
1795             _ => return_error!("expected `usize`, got `{}`", out_elem),
1796         }
1797
1798         return Ok(bx.ptrtoint(args[0].immediate(), llret_ty));
1799     }
1800
1801     if name == sym::simd_from_exposed_addr {
1802         require_simd!(ret_ty, "return");
1803         let (out_len, out_elem) = ret_ty.simd_size_and_type(bx.tcx());
1804         require!(
1805             in_len == out_len,
1806             "expected return type with length {} (same as input type `{}`), \
1807                   found `{}` with length {}",
1808             in_len,
1809             in_ty,
1810             ret_ty,
1811             out_len
1812         );
1813
1814         match in_elem.kind() {
1815             ty::Uint(ty::UintTy::Usize) => {}
1816             _ => return_error!("expected `usize`, got `{}`", in_elem),
1817         }
1818         match out_elem.kind() {
1819             ty::RawPtr(_) => {}
1820             _ => return_error!("expected pointer, got `{}`", out_elem),
1821         }
1822
1823         return Ok(bx.inttoptr(args[0].immediate(), llret_ty));
1824     }
1825
1826     if name == sym::simd_cast || name == sym::simd_as {
1827         require_simd!(ret_ty, "return");
1828         let (out_len, out_elem) = ret_ty.simd_size_and_type(bx.tcx());
1829         require!(
1830             in_len == out_len,
1831             "expected return type with length {} (same as input type `{}`), \
1832                   found `{}` with length {}",
1833             in_len,
1834             in_ty,
1835             ret_ty,
1836             out_len
1837         );
1838         // casting cares about nominal type, not just structural type
1839         if in_elem == out_elem {
1840             return Ok(args[0].immediate());
1841         }
1842
1843         enum Style {
1844             Float,
1845             Int(/* is signed? */ bool),
1846             Unsupported,
1847         }
1848
1849         let (in_style, in_width) = match in_elem.kind() {
1850             // vectors of pointer-sized integers should've been
1851             // disallowed before here, so this unwrap is safe.
1852             ty::Int(i) => (
1853                 Style::Int(true),
1854                 i.normalize(bx.tcx().sess.target.pointer_width).bit_width().unwrap(),
1855             ),
1856             ty::Uint(u) => (
1857                 Style::Int(false),
1858                 u.normalize(bx.tcx().sess.target.pointer_width).bit_width().unwrap(),
1859             ),
1860             ty::Float(f) => (Style::Float, f.bit_width()),
1861             _ => (Style::Unsupported, 0),
1862         };
1863         let (out_style, out_width) = match out_elem.kind() {
1864             ty::Int(i) => (
1865                 Style::Int(true),
1866                 i.normalize(bx.tcx().sess.target.pointer_width).bit_width().unwrap(),
1867             ),
1868             ty::Uint(u) => (
1869                 Style::Int(false),
1870                 u.normalize(bx.tcx().sess.target.pointer_width).bit_width().unwrap(),
1871             ),
1872             ty::Float(f) => (Style::Float, f.bit_width()),
1873             _ => (Style::Unsupported, 0),
1874         };
1875
1876         match (in_style, out_style) {
1877             (Style::Int(in_is_signed), Style::Int(_)) => {
1878                 return Ok(match in_width.cmp(&out_width) {
1879                     Ordering::Greater => bx.trunc(args[0].immediate(), llret_ty),
1880                     Ordering::Equal => args[0].immediate(),
1881                     Ordering::Less => {
1882                         if in_is_signed {
1883                             bx.sext(args[0].immediate(), llret_ty)
1884                         } else {
1885                             bx.zext(args[0].immediate(), llret_ty)
1886                         }
1887                     }
1888                 });
1889             }
1890             (Style::Int(in_is_signed), Style::Float) => {
1891                 return Ok(if in_is_signed {
1892                     bx.sitofp(args[0].immediate(), llret_ty)
1893                 } else {
1894                     bx.uitofp(args[0].immediate(), llret_ty)
1895                 });
1896             }
1897             (Style::Float, Style::Int(out_is_signed)) => {
1898                 return Ok(match (out_is_signed, name == sym::simd_as) {
1899                     (false, false) => bx.fptoui(args[0].immediate(), llret_ty),
1900                     (true, false) => bx.fptosi(args[0].immediate(), llret_ty),
1901                     (_, true) => bx.cast_float_to_int(out_is_signed, args[0].immediate(), llret_ty),
1902                 });
1903             }
1904             (Style::Float, Style::Float) => {
1905                 return Ok(match in_width.cmp(&out_width) {
1906                     Ordering::Greater => bx.fptrunc(args[0].immediate(), llret_ty),
1907                     Ordering::Equal => args[0].immediate(),
1908                     Ordering::Less => bx.fpext(args[0].immediate(), llret_ty),
1909                 });
1910             }
1911             _ => { /* Unsupported. Fallthrough. */ }
1912         }
1913         require!(
1914             false,
1915             "unsupported cast from `{}` with element `{}` to `{}` with element `{}`",
1916             in_ty,
1917             in_elem,
1918             ret_ty,
1919             out_elem
1920         );
1921     }
1922     macro_rules! arith_binary {
1923         ($($name: ident: $($($p: ident),* => $call: ident),*;)*) => {
1924             $(if name == sym::$name {
1925                 match in_elem.kind() {
1926                     $($(ty::$p(_))|* => {
1927                         return Ok(bx.$call(args[0].immediate(), args[1].immediate()))
1928                     })*
1929                     _ => {},
1930                 }
1931                 require!(false,
1932                          "unsupported operation on `{}` with element `{}`",
1933                          in_ty,
1934                          in_elem)
1935             })*
1936         }
1937     }
1938     arith_binary! {
1939         simd_add: Uint, Int => add, Float => fadd;
1940         simd_sub: Uint, Int => sub, Float => fsub;
1941         simd_mul: Uint, Int => mul, Float => fmul;
1942         simd_div: Uint => udiv, Int => sdiv, Float => fdiv;
1943         simd_rem: Uint => urem, Int => srem, Float => frem;
1944         simd_shl: Uint, Int => shl;
1945         simd_shr: Uint => lshr, Int => ashr;
1946         simd_and: Uint, Int => and;
1947         simd_or: Uint, Int => or;
1948         simd_xor: Uint, Int => xor;
1949         simd_fmax: Float => maxnum;
1950         simd_fmin: Float => minnum;
1951
1952     }
1953     macro_rules! arith_unary {
1954         ($($name: ident: $($($p: ident),* => $call: ident),*;)*) => {
1955             $(if name == sym::$name {
1956                 match in_elem.kind() {
1957                     $($(ty::$p(_))|* => {
1958                         return Ok(bx.$call(args[0].immediate()))
1959                     })*
1960                     _ => {},
1961                 }
1962                 require!(false,
1963                          "unsupported operation on `{}` with element `{}`",
1964                          in_ty,
1965                          in_elem)
1966             })*
1967         }
1968     }
1969     arith_unary! {
1970         simd_neg: Int => neg, Float => fneg;
1971     }
1972
1973     if name == sym::simd_arith_offset {
1974         // This also checks that the first operand is a ptr type.
1975         let pointee = in_elem.builtin_deref(true).unwrap_or_else(|| {
1976             span_bug!(span, "must be called with a vector of pointer types as first argument")
1977         });
1978         let layout = bx.layout_of(pointee.ty);
1979         let ptrs = args[0].immediate();
1980         // The second argument must be a ptr-sized integer.
1981         // (We don't care about the signedness, this is wrapping anyway.)
1982         let (_offsets_len, offsets_elem) = arg_tys[1].simd_size_and_type(bx.tcx());
1983         if !matches!(offsets_elem.kind(), ty::Int(ty::IntTy::Isize) | ty::Uint(ty::UintTy::Usize)) {
1984             span_bug!(
1985                 span,
1986                 "must be called with a vector of pointer-sized integers as second argument"
1987             );
1988         }
1989         let offsets = args[1].immediate();
1990
1991         return Ok(bx.gep(bx.backend_type(layout), ptrs, &[offsets]));
1992     }
1993
1994     if name == sym::simd_saturating_add || name == sym::simd_saturating_sub {
1995         let lhs = args[0].immediate();
1996         let rhs = args[1].immediate();
1997         let is_add = name == sym::simd_saturating_add;
1998         let ptr_bits = bx.tcx().data_layout.pointer_size.bits() as _;
1999         let (signed, elem_width, elem_ty) = match *in_elem.kind() {
2000             ty::Int(i) => (true, i.bit_width().unwrap_or(ptr_bits), bx.cx.type_int_from_ty(i)),
2001             ty::Uint(i) => (false, i.bit_width().unwrap_or(ptr_bits), bx.cx.type_uint_from_ty(i)),
2002             _ => {
2003                 return_error!(
2004                     "expected element type `{}` of vector type `{}` \
2005                      to be a signed or unsigned integer type",
2006                     arg_tys[0].simd_size_and_type(bx.tcx()).1,
2007                     arg_tys[0]
2008                 );
2009             }
2010         };
2011         let llvm_intrinsic = &format!(
2012             "llvm.{}{}.sat.v{}i{}",
2013             if signed { 's' } else { 'u' },
2014             if is_add { "add" } else { "sub" },
2015             in_len,
2016             elem_width
2017         );
2018         let vec_ty = bx.cx.type_vector(elem_ty, in_len as u64);
2019
2020         let fn_ty = bx.type_func(&[vec_ty, vec_ty], vec_ty);
2021         let f = bx.declare_cfn(llvm_intrinsic, llvm::UnnamedAddr::No, fn_ty);
2022         let v = bx.call(fn_ty, None, f, &[lhs, rhs], None);
2023         return Ok(v);
2024     }
2025
2026     span_bug!(span, "unknown SIMD intrinsic");
2027 }
2028
2029 // Returns the width of an int Ty, and if it's signed or not
2030 // Returns None if the type is not an integer
2031 // FIXME: there’s multiple of this functions, investigate using some of the already existing
2032 // stuffs.
2033 fn int_type_width_signed(ty: Ty<'_>, cx: &CodegenCx<'_, '_>) -> Option<(u64, bool)> {
2034     match ty.kind() {
2035         ty::Int(t) => {
2036             Some((t.bit_width().unwrap_or(u64::from(cx.tcx.sess.target.pointer_width)), true))
2037         }
2038         ty::Uint(t) => {
2039             Some((t.bit_width().unwrap_or(u64::from(cx.tcx.sess.target.pointer_width)), false))
2040         }
2041         _ => None,
2042     }
2043 }