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