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