1 // Copyright 2012-2014 The Rust Project Developers. See the COPYRIGHT
2 // file at the top-level directory of this distribution and at
3 // http://rust-lang.org/COPYRIGHT.
5 // Licensed under the Apache License, Version 2.0 <LICENSE-APACHE or
6 // http://www.apache.org/licenses/LICENSE-2.0> or the MIT license
7 // <LICENSE-MIT or http://opensource.org/licenses/MIT>, at your
8 // option. This file may not be copied, modified, or distributed
9 // except according to those terms.
11 #![allow(non_upper_case_globals)]
13 use intrinsics::{self, Intrinsic};
17 use abi::{Abi, FnType};
19 use mir::lvalue::LvalueRef;
27 use rustc::ty::{self, Ty};
30 use syntax::symbol::Symbol;
33 use rustc::session::Session;
36 use std::cmp::Ordering;
39 use mir::lvalue::Alignment;
41 fn get_simple_intrinsic(ccx: &CrateContext, name: &str) -> Option<ValueRef> {
42 let llvm_name = match name {
43 "sqrtf32" => "llvm.sqrt.f32",
44 "sqrtf64" => "llvm.sqrt.f64",
45 "powif32" => "llvm.powi.f32",
46 "powif64" => "llvm.powi.f64",
47 "sinf32" => "llvm.sin.f32",
48 "sinf64" => "llvm.sin.f64",
49 "cosf32" => "llvm.cos.f32",
50 "cosf64" => "llvm.cos.f64",
51 "powf32" => "llvm.pow.f32",
52 "powf64" => "llvm.pow.f64",
53 "expf32" => "llvm.exp.f32",
54 "expf64" => "llvm.exp.f64",
55 "exp2f32" => "llvm.exp2.f32",
56 "exp2f64" => "llvm.exp2.f64",
57 "logf32" => "llvm.log.f32",
58 "logf64" => "llvm.log.f64",
59 "log10f32" => "llvm.log10.f32",
60 "log10f64" => "llvm.log10.f64",
61 "log2f32" => "llvm.log2.f32",
62 "log2f64" => "llvm.log2.f64",
63 "fmaf32" => "llvm.fma.f32",
64 "fmaf64" => "llvm.fma.f64",
65 "fabsf32" => "llvm.fabs.f32",
66 "fabsf64" => "llvm.fabs.f64",
67 "copysignf32" => "llvm.copysign.f32",
68 "copysignf64" => "llvm.copysign.f64",
69 "floorf32" => "llvm.floor.f32",
70 "floorf64" => "llvm.floor.f64",
71 "ceilf32" => "llvm.ceil.f32",
72 "ceilf64" => "llvm.ceil.f64",
73 "truncf32" => "llvm.trunc.f32",
74 "truncf64" => "llvm.trunc.f64",
75 "rintf32" => "llvm.rint.f32",
76 "rintf64" => "llvm.rint.f64",
77 "nearbyintf32" => "llvm.nearbyint.f32",
78 "nearbyintf64" => "llvm.nearbyint.f64",
79 "roundf32" => "llvm.round.f32",
80 "roundf64" => "llvm.round.f64",
81 "assume" => "llvm.assume",
82 "abort" => "llvm.trap",
85 Some(ccx.get_intrinsic(&llvm_name))
88 /// Remember to add all intrinsics here, in librustc_typeck/check/mod.rs,
89 /// and in libcore/intrinsics.rs; if you need access to any llvm intrinsics,
90 /// add them to librustc_trans/trans/context.rs
91 pub fn trans_intrinsic_call<'a, 'tcx>(bcx: &Builder<'a, 'tcx>,
100 let (def_id, substs, sig) = match callee_ty.sty {
101 ty::TyFnDef(def_id, substs, sig) => (def_id, substs, sig),
102 _ => bug!("expected fn item type, found {}", callee_ty)
105 let sig = tcx.erase_late_bound_regions_and_normalize(&sig);
106 let arg_tys = sig.inputs();
107 let ret_ty = sig.output();
108 let name = &*tcx.item_name(def_id).as_str();
110 let llret_ty = type_of::type_of(ccx, ret_ty);
112 let simple = get_simple_intrinsic(ccx, name);
113 let llval = match name {
114 _ if simple.is_some() => {
115 bcx.call(simple.unwrap(), &llargs, None)
121 let expect = ccx.get_intrinsic(&("llvm.expect.i1"));
122 bcx.call(expect, &[llargs[0], C_bool(ccx, true)], None)
125 let expect = ccx.get_intrinsic(&("llvm.expect.i1"));
126 bcx.call(expect, &[llargs[0], C_bool(ccx, false)], None)
129 try_intrinsic(bcx, ccx, llargs[0], llargs[1], llargs[2], llresult);
133 let llfn = ccx.get_intrinsic(&("llvm.debugtrap"));
134 bcx.call(llfn, &[], None)
137 let tp_ty = substs.type_at(0);
138 let lltp_ty = type_of::type_of(ccx, tp_ty);
139 C_uint(ccx, machine::llsize_of_alloc(ccx, lltp_ty))
142 let tp_ty = substs.type_at(0);
143 if !bcx.ccx.shared().type_is_sized(tp_ty) {
145 glue::size_and_align_of_dst(bcx, tp_ty, llargs[1]);
148 let lltp_ty = type_of::type_of(ccx, tp_ty);
149 C_uint(ccx, machine::llsize_of_alloc(ccx, lltp_ty))
153 let tp_ty = substs.type_at(0);
154 C_uint(ccx, type_of::align_of(ccx, tp_ty))
156 "min_align_of_val" => {
157 let tp_ty = substs.type_at(0);
158 if !bcx.ccx.shared().type_is_sized(tp_ty) {
160 glue::size_and_align_of_dst(bcx, tp_ty, llargs[1]);
163 C_uint(ccx, type_of::align_of(ccx, tp_ty))
167 let tp_ty = substs.type_at(0);
168 let lltp_ty = type_of::type_of(ccx, tp_ty);
169 C_uint(ccx, machine::llalign_of_pref(ccx, lltp_ty))
172 let tp_ty = substs.type_at(0);
173 let ty_name = Symbol::intern(&tp_ty.to_string()).as_str();
174 C_str_slice(ccx, ty_name)
177 C_u64(ccx, ccx.tcx().type_id_hash(substs.type_at(0)))
180 let ty = substs.type_at(0);
181 if !type_is_zero_size(ccx, ty) {
182 // Just zero out the stack slot.
183 // If we store a zero constant, LLVM will drown in vreg allocation for large data
184 // structures, and the generated code will be awful. (A telltale sign of this is
185 // large quantities of `mov [byte ptr foo],0` in the generated code.)
186 memset_intrinsic(bcx, false, ty, llresult, C_u8(ccx, 0), C_uint(ccx, 1usize));
190 // Effectively no-ops
191 "uninit" | "forget" => {
195 let tp_ty = substs.type_at(0);
197 C_bool(ccx, bcx.ccx.shared().type_needs_drop(tp_ty))
201 let offset = llargs[1];
202 bcx.inbounds_gep(ptr, &[offset])
206 let offset = llargs[1];
207 bcx.gep(ptr, &[offset])
210 "copy_nonoverlapping" => {
211 copy_intrinsic(bcx, false, false, substs.type_at(0), llargs[1], llargs[0], llargs[2])
214 copy_intrinsic(bcx, true, false, substs.type_at(0), llargs[1], llargs[0], llargs[2])
217 memset_intrinsic(bcx, false, substs.type_at(0), llargs[0], llargs[1], llargs[2])
220 "volatile_copy_nonoverlapping_memory" => {
221 copy_intrinsic(bcx, false, true, substs.type_at(0), llargs[0], llargs[1], llargs[2])
223 "volatile_copy_memory" => {
224 copy_intrinsic(bcx, true, true, substs.type_at(0), llargs[0], llargs[1], llargs[2])
226 "volatile_set_memory" => {
227 memset_intrinsic(bcx, true, substs.type_at(0), llargs[0], llargs[1], llargs[2])
230 let tp_ty = substs.type_at(0);
231 let mut ptr = llargs[0];
232 if let Some(ty) = fn_ty.ret.cast {
233 ptr = bcx.pointercast(ptr, ty.ptr_to());
235 let load = bcx.volatile_load(ptr);
237 llvm::LLVMSetAlignment(load, type_of::align_of(ccx, tp_ty));
239 to_immediate(bcx, load, tp_ty)
241 "volatile_store" => {
242 let tp_ty = substs.type_at(0);
243 if type_is_fat_ptr(bcx.ccx, tp_ty) {
244 bcx.volatile_store(llargs[1], get_dataptr(bcx, llargs[0]));
245 bcx.volatile_store(llargs[2], get_meta(bcx, llargs[0]));
247 let val = if fn_ty.args[1].is_indirect() {
248 bcx.load(llargs[1], None)
250 from_immediate(bcx, llargs[1])
252 let ptr = bcx.pointercast(llargs[0], val_ty(val).ptr_to());
253 let store = bcx.volatile_store(val, ptr);
255 llvm::LLVMSetAlignment(store, type_of::align_of(ccx, tp_ty));
261 "ctlz" | "cttz" | "ctpop" | "bswap" |
262 "add_with_overflow" | "sub_with_overflow" | "mul_with_overflow" |
263 "overflowing_add" | "overflowing_sub" | "overflowing_mul" |
264 "unchecked_div" | "unchecked_rem" => {
265 let sty = &arg_tys[0].sty;
266 match int_type_width_signed(sty, ccx) {
267 Some((width, signed)) =>
270 let y = C_bool(bcx.ccx, false);
271 let llfn = ccx.get_intrinsic(&format!("llvm.{}.i{}", name, width));
272 bcx.call(llfn, &[llargs[0], y], None)
274 "ctpop" => bcx.call(ccx.get_intrinsic(&format!("llvm.ctpop.i{}", width)),
278 llargs[0] // byte swap a u8/i8 is just a no-op
280 bcx.call(ccx.get_intrinsic(&format!("llvm.bswap.i{}", width)),
284 "add_with_overflow" | "sub_with_overflow" | "mul_with_overflow" => {
285 let intrinsic = format!("llvm.{}{}.with.overflow.i{}",
286 if signed { 's' } else { 'u' },
288 let llfn = bcx.ccx.get_intrinsic(&intrinsic);
290 // Convert `i1` to a `bool`, and write it to the out parameter
291 let val = bcx.call(llfn, &[llargs[0], llargs[1]], None);
292 let result = bcx.extract_value(val, 0);
293 let overflow = bcx.zext(bcx.extract_value(val, 1), Type::bool(ccx));
294 bcx.store(result, bcx.struct_gep(llresult, 0), None);
295 bcx.store(overflow, bcx.struct_gep(llresult, 1), None);
299 "overflowing_add" => bcx.add(llargs[0], llargs[1]),
300 "overflowing_sub" => bcx.sub(llargs[0], llargs[1]),
301 "overflowing_mul" => bcx.mul(llargs[0], llargs[1]),
304 bcx.sdiv(llargs[0], llargs[1])
306 bcx.udiv(llargs[0], llargs[1])
310 bcx.srem(llargs[0], llargs[1])
312 bcx.urem(llargs[0], llargs[1])
317 span_invalid_monomorphization_error(
319 &format!("invalid monomorphization of `{}` intrinsic: \
320 expected basic integer type, found `{}`", name, sty));
326 "fadd_fast" | "fsub_fast" | "fmul_fast" | "fdiv_fast" | "frem_fast" => {
327 let sty = &arg_tys[0].sty;
328 match float_type_width(sty) {
331 "fadd_fast" => bcx.fadd_fast(llargs[0], llargs[1]),
332 "fsub_fast" => bcx.fsub_fast(llargs[0], llargs[1]),
333 "fmul_fast" => bcx.fmul_fast(llargs[0], llargs[1]),
334 "fdiv_fast" => bcx.fdiv_fast(llargs[0], llargs[1]),
335 "frem_fast" => bcx.frem_fast(llargs[0], llargs[1]),
339 span_invalid_monomorphization_error(
341 &format!("invalid monomorphization of `{}` intrinsic: \
342 expected basic float type, found `{}`", name, sty));
349 "discriminant_value" => {
350 let val_ty = substs.type_at(0);
352 ty::TyAdt(adt, ..) if adt.is_enum() => {
353 adt::trans_get_discr(bcx, val_ty, llargs[0], Alignment::AbiAligned,
354 Some(llret_ty), true)
356 _ => C_null(llret_ty)
359 name if name.starts_with("simd_") => {
360 generic_simd_intrinsic(bcx, name,
366 // This requires that atomic intrinsics follow a specific naming pattern:
367 // "atomic_<operation>[_<ordering>]", and no ordering means SeqCst
368 name if name.starts_with("atomic_") => {
369 use llvm::AtomicOrdering::*;
371 let split: Vec<&str> = name.split('_').collect();
373 let is_cxchg = split[1] == "cxchg" || split[1] == "cxchgweak";
374 let (order, failorder) = match split.len() {
375 2 => (SequentiallyConsistent, SequentiallyConsistent),
376 3 => match split[2] {
377 "unordered" => (Unordered, Unordered),
378 "relaxed" => (Monotonic, Monotonic),
379 "acq" => (Acquire, Acquire),
380 "rel" => (Release, Monotonic),
381 "acqrel" => (AcquireRelease, Acquire),
382 "failrelaxed" if is_cxchg =>
383 (SequentiallyConsistent, Monotonic),
384 "failacq" if is_cxchg =>
385 (SequentiallyConsistent, Acquire),
386 _ => ccx.sess().fatal("unknown ordering in atomic intrinsic")
388 4 => match (split[2], split[3]) {
389 ("acq", "failrelaxed") if is_cxchg =>
390 (Acquire, Monotonic),
391 ("acqrel", "failrelaxed") if is_cxchg =>
392 (AcquireRelease, Monotonic),
393 _ => ccx.sess().fatal("unknown ordering in atomic intrinsic")
395 _ => ccx.sess().fatal("Atomic intrinsic not in correct format"),
398 let invalid_monomorphization = |sty| {
399 span_invalid_monomorphization_error(tcx.sess, span,
400 &format!("invalid monomorphization of `{}` intrinsic: \
401 expected basic integer type, found `{}`", name, sty));
405 "cxchg" | "cxchgweak" => {
406 let sty = &substs.type_at(0).sty;
407 if int_type_width_signed(sty, ccx).is_some() {
408 let weak = if split[1] == "cxchgweak" { llvm::True } else { llvm::False };
409 let val = bcx.atomic_cmpxchg(llargs[0], llargs[1], llargs[2], order,
411 let result = bcx.extract_value(val, 0);
412 let success = bcx.zext(bcx.extract_value(val, 1), Type::bool(bcx.ccx));
413 bcx.store(result, bcx.struct_gep(llresult, 0), None);
414 bcx.store(success, bcx.struct_gep(llresult, 1), None);
416 invalid_monomorphization(sty);
422 let sty = &substs.type_at(0).sty;
423 if int_type_width_signed(sty, ccx).is_some() {
424 bcx.atomic_load(llargs[0], order)
426 invalid_monomorphization(sty);
432 let sty = &substs.type_at(0).sty;
433 if int_type_width_signed(sty, ccx).is_some() {
434 bcx.atomic_store(llargs[1], llargs[0], order);
436 invalid_monomorphization(sty);
442 bcx.atomic_fence(order, llvm::SynchronizationScope::CrossThread);
446 "singlethreadfence" => {
447 bcx.atomic_fence(order, llvm::SynchronizationScope::SingleThread);
451 // These are all AtomicRMW ops
453 let atom_op = match op {
454 "xchg" => llvm::AtomicXchg,
455 "xadd" => llvm::AtomicAdd,
456 "xsub" => llvm::AtomicSub,
457 "and" => llvm::AtomicAnd,
458 "nand" => llvm::AtomicNand,
459 "or" => llvm::AtomicOr,
460 "xor" => llvm::AtomicXor,
461 "max" => llvm::AtomicMax,
462 "min" => llvm::AtomicMin,
463 "umax" => llvm::AtomicUMax,
464 "umin" => llvm::AtomicUMin,
465 _ => ccx.sess().fatal("unknown atomic operation")
468 let sty = &substs.type_at(0).sty;
469 if int_type_width_signed(sty, ccx).is_some() {
470 bcx.atomic_rmw(atom_op, llargs[0], llargs[1], order)
472 invalid_monomorphization(sty);
480 let intr = match Intrinsic::find(&name) {
482 None => bug!("unknown intrinsic '{}'", name),
484 fn one<T>(x: Vec<T>) -> T {
485 assert_eq!(x.len(), 1);
486 x.into_iter().next().unwrap()
488 fn ty_to_type(ccx: &CrateContext, t: &intrinsics::Type,
489 any_changes_needed: &mut bool) -> Vec<Type> {
490 use intrinsics::Type::*;
492 Void => vec![Type::void(ccx)],
493 Integer(_signed, width, llvm_width) => {
494 *any_changes_needed |= width != llvm_width;
495 vec![Type::ix(ccx, llvm_width as u64)]
499 32 => vec![Type::f32(ccx)],
500 64 => vec![Type::f64(ccx)],
504 Pointer(ref t, ref llvm_elem, _const) => {
505 *any_changes_needed |= llvm_elem.is_some();
507 let t = llvm_elem.as_ref().unwrap_or(t);
508 let elem = one(ty_to_type(ccx, t, any_changes_needed));
511 Vector(ref t, ref llvm_elem, length) => {
512 *any_changes_needed |= llvm_elem.is_some();
514 let t = llvm_elem.as_ref().unwrap_or(t);
515 let elem = one(ty_to_type(ccx, t, any_changes_needed));
516 vec![Type::vector(&elem, length as u64)]
518 Aggregate(false, ref contents) => {
519 let elems = contents.iter()
520 .map(|t| one(ty_to_type(ccx, t, any_changes_needed)))
521 .collect::<Vec<_>>();
522 vec![Type::struct_(ccx, &elems, false)]
524 Aggregate(true, ref contents) => {
525 *any_changes_needed = true;
527 .flat_map(|t| ty_to_type(ccx, t, any_changes_needed))
533 // This allows an argument list like `foo, (bar, baz),
534 // qux` to be converted into `foo, bar, baz, qux`, integer
535 // arguments to be truncated as needed and pointers to be
537 fn modify_as_needed<'a, 'tcx>(bcx: &Builder<'a, 'tcx>,
538 t: &intrinsics::Type,
544 intrinsics::Type::Aggregate(true, ref contents) => {
545 // We found a tuple that needs squishing! So
546 // run over the tuple and load each field.
548 // This assumes the type is "simple", i.e. no
549 // destructors, and the contents are SIMD
551 assert!(!bcx.ccx.shared().type_needs_drop(arg_type));
552 let arg = LvalueRef::new_sized_ty(llarg, arg_type, Alignment::AbiAligned);
553 (0..contents.len()).map(|i| {
554 let (ptr, align) = arg.trans_field_ptr(bcx, i);
555 bcx.load(ptr, align.to_align())
558 intrinsics::Type::Pointer(_, Some(ref llvm_elem), _) => {
559 let llvm_elem = one(ty_to_type(bcx.ccx, llvm_elem, &mut false));
560 vec![bcx.pointercast(llarg, llvm_elem.ptr_to())]
562 intrinsics::Type::Vector(_, Some(ref llvm_elem), length) => {
563 let llvm_elem = one(ty_to_type(bcx.ccx, llvm_elem, &mut false));
564 vec![bcx.bitcast(llarg, Type::vector(&llvm_elem, length as u64))]
566 intrinsics::Type::Integer(_, width, llvm_width) if width != llvm_width => {
567 // the LLVM intrinsic uses a smaller integer
568 // size than the C intrinsic's signature, so
569 // we have to trim it down here.
570 vec![bcx.trunc(llarg, Type::ix(bcx.ccx, llvm_width as u64))]
577 let mut any_changes_needed = false;
578 let inputs = intr.inputs.iter()
579 .flat_map(|t| ty_to_type(ccx, t, &mut any_changes_needed))
580 .collect::<Vec<_>>();
582 let mut out_changes = false;
583 let outputs = one(ty_to_type(ccx, &intr.output, &mut out_changes));
584 // outputting a flattened aggregate is nonsense
585 assert!(!out_changes);
587 let llargs = if !any_changes_needed {
588 // no aggregates to flatten, so no change needed
591 // there are some aggregates that need to be flattened
592 // in the LLVM call, so we need to run over the types
593 // again to find them and extract the arguments
597 .flat_map(|((t, llarg), ty)| modify_as_needed(bcx, t, ty, *llarg))
600 assert_eq!(inputs.len(), llargs.len());
602 let val = match intr.definition {
603 intrinsics::IntrinsicDef::Named(name) => {
604 let f = declare::declare_cfn(ccx,
606 Type::func(&inputs, &outputs));
607 bcx.call(f, &llargs, None)
612 intrinsics::Type::Aggregate(flatten, ref elems) => {
613 // the output is a tuple so we need to munge it properly
616 for i in 0..elems.len() {
617 let val = bcx.extract_value(val, i);
618 bcx.store(val, bcx.struct_gep(llresult, i), None);
627 if val_ty(llval) != Type::void(ccx) && machine::llsize_of_alloc(ccx, val_ty(llval)) != 0 {
628 if let Some(ty) = fn_ty.ret.cast {
629 let ptr = bcx.pointercast(llresult, ty.ptr_to());
630 bcx.store(llval, ptr, Some(type_of::align_of(ccx, ret_ty)));
632 store_ty(bcx, llval, llresult, Alignment::AbiAligned, ret_ty);
637 fn copy_intrinsic<'a, 'tcx>(bcx: &Builder<'a, 'tcx>,
646 let lltp_ty = type_of::type_of(ccx, tp_ty);
647 let align = C_i32(ccx, type_of::align_of(ccx, tp_ty) as i32);
648 let size = machine::llsize_of(ccx, lltp_ty);
649 let int_size = machine::llbitsize_of_real(ccx, ccx.int_type());
651 let operation = if allow_overlap {
657 let name = format!("llvm.{}.p0i8.p0i8.i{}", operation, int_size);
659 let dst_ptr = bcx.pointercast(dst, Type::i8p(ccx));
660 let src_ptr = bcx.pointercast(src, Type::i8p(ccx));
661 let llfn = ccx.get_intrinsic(&name);
666 bcx.mul(size, count),
668 C_bool(ccx, volatile)],
672 fn memset_intrinsic<'a, 'tcx>(
673 bcx: &Builder<'a, 'tcx>,
681 let align = C_i32(ccx, type_of::align_of(ccx, ty) as i32);
682 let lltp_ty = type_of::type_of(ccx, ty);
683 let size = machine::llsize_of(ccx, lltp_ty);
684 let dst = bcx.pointercast(dst, Type::i8p(ccx));
685 call_memset(bcx, dst, val, bcx.mul(size, count), align, volatile)
688 fn try_intrinsic<'a, 'tcx>(
689 bcx: &Builder<'a, 'tcx>,
696 if bcx.sess().no_landing_pads() {
697 bcx.call(func, &[data], None);
698 bcx.store(C_null(Type::i8p(&bcx.ccx)), dest, None);
699 } else if wants_msvc_seh(bcx.sess()) {
700 trans_msvc_try(bcx, ccx, func, data, local_ptr, dest);
702 trans_gnu_try(bcx, ccx, func, data, local_ptr, dest);
706 // MSVC's definition of the `rust_try` function.
708 // This implementation uses the new exception handling instructions in LLVM
709 // which have support in LLVM for SEH on MSVC targets. Although these
710 // instructions are meant to work for all targets, as of the time of this
711 // writing, however, LLVM does not recommend the usage of these new instructions
712 // as the old ones are still more optimized.
713 fn trans_msvc_try<'a, 'tcx>(bcx: &Builder<'a, 'tcx>,
719 let llfn = get_rust_try_fn(ccx, &mut |bcx| {
722 bcx.set_personality_fn(bcx.ccx.eh_personality());
724 let normal = bcx.build_sibling_block("normal");
725 let catchswitch = bcx.build_sibling_block("catchswitch");
726 let catchpad = bcx.build_sibling_block("catchpad");
727 let caught = bcx.build_sibling_block("caught");
729 let func = llvm::get_param(bcx.llfn(), 0);
730 let data = llvm::get_param(bcx.llfn(), 1);
731 let local_ptr = llvm::get_param(bcx.llfn(), 2);
733 // We're generating an IR snippet that looks like:
735 // declare i32 @rust_try(%func, %data, %ptr) {
736 // %slot = alloca i64*
737 // invoke %func(%data) to label %normal unwind label %catchswitch
743 // %cs = catchswitch within none [%catchpad] unwind to caller
746 // %tok = catchpad within %cs [%type_descriptor, 0, %slot]
747 // %ptr[0] = %slot[0]
748 // %ptr[1] = %slot[1]
749 // catchret from %tok to label %caught
755 // This structure follows the basic usage of throw/try/catch in LLVM.
756 // For example, compile this C++ snippet to see what LLVM generates:
758 // #include <stdint.h>
760 // int bar(void (*foo)(void), uint64_t *ret) {
764 // } catch(uint64_t a[2]) {
771 // More information can be found in libstd's seh.rs implementation.
772 let i64p = Type::i64(ccx).ptr_to();
773 let slot = bcx.alloca(i64p, "slot");
774 bcx.invoke(func, &[data], normal.llbb(), catchswitch.llbb(),
777 normal.ret(C_i32(ccx, 0));
779 let cs = catchswitch.catch_switch(None, None, 1);
780 catchswitch.add_handler(cs, catchpad.llbb());
783 let tydesc = match tcx.lang_items.msvc_try_filter() {
784 Some(did) => ::consts::get_static(ccx, did),
785 None => bug!("msvc_try_filter not defined"),
787 let tok = catchpad.catch_pad(cs, &[tydesc, C_i32(ccx, 0), slot]);
788 let addr = catchpad.load(slot, None);
789 let arg1 = catchpad.load(addr, None);
790 let val1 = C_i32(ccx, 1);
791 let arg2 = catchpad.load(catchpad.inbounds_gep(addr, &[val1]), None);
792 let local_ptr = catchpad.bitcast(local_ptr, i64p);
793 catchpad.store(arg1, local_ptr, None);
794 catchpad.store(arg2, catchpad.inbounds_gep(local_ptr, &[val1]), None);
795 catchpad.catch_ret(tok, caught.llbb());
797 caught.ret(C_i32(ccx, 1));
800 // Note that no invoke is used here because by definition this function
801 // can't panic (that's what it's catching).
802 let ret = bcx.call(llfn, &[func, data, local_ptr], None);
803 bcx.store(ret, dest, None);
806 // Definition of the standard "try" function for Rust using the GNU-like model
807 // of exceptions (e.g. the normal semantics of LLVM's landingpad and invoke
810 // This translation is a little surprising because we always call a shim
811 // function instead of inlining the call to `invoke` manually here. This is done
812 // because in LLVM we're only allowed to have one personality per function
813 // definition. The call to the `try` intrinsic is being inlined into the
814 // function calling it, and that function may already have other personality
815 // functions in play. By calling a shim we're guaranteed that our shim will have
816 // the right personality function.
817 fn trans_gnu_try<'a, 'tcx>(bcx: &Builder<'a, 'tcx>,
823 let llfn = get_rust_try_fn(ccx, &mut |bcx| {
826 // Translates the shims described above:
829 // invoke %func(%args...) normal %normal unwind %catch
835 // (ptr, _) = landingpad
836 // store ptr, %local_ptr
839 // Note that the `local_ptr` data passed into the `try` intrinsic is
840 // expected to be `*mut *mut u8` for this to actually work, but that's
841 // managed by the standard library.
843 let then = bcx.build_sibling_block("then");
844 let catch = bcx.build_sibling_block("catch");
846 let func = llvm::get_param(bcx.llfn(), 0);
847 let data = llvm::get_param(bcx.llfn(), 1);
848 let local_ptr = llvm::get_param(bcx.llfn(), 2);
849 bcx.invoke(func, &[data], then.llbb(), catch.llbb(), None);
850 then.ret(C_i32(ccx, 0));
852 // Type indicator for the exception being thrown.
854 // The first value in this tuple is a pointer to the exception object
855 // being thrown. The second value is a "selector" indicating which of
856 // the landing pad clauses the exception's type had been matched to.
857 // rust_try ignores the selector.
858 let lpad_ty = Type::struct_(ccx, &[Type::i8p(ccx), Type::i32(ccx)],
860 let vals = catch.landing_pad(lpad_ty, bcx.ccx.eh_personality(), 1, catch.llfn());
861 catch.add_clause(vals, C_null(Type::i8p(ccx)));
862 let ptr = catch.extract_value(vals, 0);
863 catch.store(ptr, catch.bitcast(local_ptr, Type::i8p(ccx).ptr_to()), None);
864 catch.ret(C_i32(ccx, 1));
867 // Note that no invoke is used here because by definition this function
868 // can't panic (that's what it's catching).
869 let ret = bcx.call(llfn, &[func, data, local_ptr], None);
870 bcx.store(ret, dest, None);
873 // Helper function to give a Block to a closure to translate a shim function.
874 // This is currently primarily used for the `try` intrinsic functions above.
875 fn gen_fn<'a, 'tcx>(ccx: &CrateContext<'a, 'tcx>,
877 inputs: Vec<Ty<'tcx>>,
879 trans: &mut for<'b> FnMut(Builder<'b, 'tcx>))
881 let rust_fn_ty = ccx.tcx().mk_fn_ptr(ty::Binder(ccx.tcx().mk_fn_sig(
885 hir::Unsafety::Unsafe,
888 let llfn = declare::define_internal_fn(ccx, name, rust_fn_ty);
889 let bcx = Builder::new_block(ccx, llfn, "entry-block");
894 // Helper function used to get a handle to the `__rust_try` function used to
897 // This function is only generated once and is then cached.
898 fn get_rust_try_fn<'a, 'tcx>(ccx: &CrateContext<'a, 'tcx>,
899 trans: &mut for<'b> FnMut(Builder<'b, 'tcx>))
901 if let Some(llfn) = ccx.rust_try_fn().get() {
905 // Define the type up front for the signature of the rust_try function.
907 let i8p = tcx.mk_mut_ptr(tcx.types.i8);
908 let fn_ty = tcx.mk_fn_ptr(ty::Binder(tcx.mk_fn_sig(
912 hir::Unsafety::Unsafe,
915 let output = tcx.types.i32;
916 let rust_try = gen_fn(ccx, "__rust_try", vec![fn_ty, i8p, i8p], output, trans);
917 ccx.rust_try_fn().set(Some(rust_try));
921 fn span_invalid_monomorphization_error(a: &Session, b: Span, c: &str) {
922 span_err!(a, b, E0511, "{}", c);
925 fn generic_simd_intrinsic<'a, 'tcx>(
926 bcx: &Builder<'a, 'tcx>,
934 // macros for error handling:
935 macro_rules! emit_error {
939 ($msg: tt, $($fmt: tt)*) => {
940 span_invalid_monomorphization_error(
942 &format!(concat!("invalid monomorphization of `{}` intrinsic: ",
947 macro_rules! require {
948 ($cond: expr, $($fmt: tt)*) => {
950 emit_error!($($fmt)*);
951 return C_nil(bcx.ccx)
955 macro_rules! require_simd {
956 ($ty: expr, $position: expr) => {
957 require!($ty.is_simd(), "expected SIMD {} type, found non-SIMD `{}`", $position, $ty)
964 let sig = tcx.erase_late_bound_regions_and_normalize(&callee_ty.fn_sig());
965 let arg_tys = sig.inputs();
967 // every intrinsic takes a SIMD vector as its first argument
968 require_simd!(arg_tys[0], "input");
969 let in_ty = arg_tys[0];
970 let in_elem = arg_tys[0].simd_type(tcx);
971 let in_len = arg_tys[0].simd_size(tcx);
973 let comparison = match name {
974 "simd_eq" => Some(hir::BiEq),
975 "simd_ne" => Some(hir::BiNe),
976 "simd_lt" => Some(hir::BiLt),
977 "simd_le" => Some(hir::BiLe),
978 "simd_gt" => Some(hir::BiGt),
979 "simd_ge" => Some(hir::BiGe),
983 if let Some(cmp_op) = comparison {
984 require_simd!(ret_ty, "return");
986 let out_len = ret_ty.simd_size(tcx);
987 require!(in_len == out_len,
988 "expected return type with length {} (same as input type `{}`), \
989 found `{}` with length {}",
992 require!(llret_ty.element_type().kind() == llvm::Integer,
993 "expected return type with integer elements, found `{}` with non-integer `{}`",
995 ret_ty.simd_type(tcx));
997 return compare_simd_types(bcx,
1005 if name.starts_with("simd_shuffle") {
1006 let n: usize = match name["simd_shuffle".len()..].parse() {
1008 Err(_) => span_bug!(span,
1009 "bad `simd_shuffle` instruction only caught in trans?")
1012 require_simd!(ret_ty, "return");
1014 let out_len = ret_ty.simd_size(tcx);
1015 require!(out_len == n,
1016 "expected return type of length {}, found `{}` with length {}",
1017 n, ret_ty, out_len);
1018 require!(in_elem == ret_ty.simd_type(tcx),
1019 "expected return element type `{}` (element of input `{}`), \
1020 found `{}` with element type `{}`",
1022 ret_ty, ret_ty.simd_type(tcx));
1024 let total_len = in_len as u128 * 2;
1026 let vector = llargs[2];
1028 let indices: Option<Vec<_>> = (0..n)
1031 let val = const_get_elt(vector, &[i as libc::c_uint]);
1032 match const_to_opt_u128(val, true) {
1034 emit_error!("shuffle index #{} is not a constant", arg_idx);
1037 Some(idx) if idx >= total_len => {
1038 emit_error!("shuffle index #{} is out of bounds (limit {})",
1039 arg_idx, total_len);
1042 Some(idx) => Some(C_i32(bcx.ccx, idx as i32)),
1046 let indices = match indices {
1048 None => return C_null(llret_ty)
1051 return bcx.shuffle_vector(llargs[0], llargs[1], C_vector(&indices))
1054 if name == "simd_insert" {
1055 require!(in_elem == arg_tys[2],
1056 "expected inserted type `{}` (element of input `{}`), found `{}`",
1057 in_elem, in_ty, arg_tys[2]);
1058 return bcx.insert_element(llargs[0], llargs[2], llargs[1])
1060 if name == "simd_extract" {
1061 require!(ret_ty == in_elem,
1062 "expected return type `{}` (element of input `{}`), found `{}`",
1063 in_elem, in_ty, ret_ty);
1064 return bcx.extract_element(llargs[0], llargs[1])
1067 if name == "simd_cast" {
1068 require_simd!(ret_ty, "return");
1069 let out_len = ret_ty.simd_size(tcx);
1070 require!(in_len == out_len,
1071 "expected return type with length {} (same as input type `{}`), \
1072 found `{}` with length {}",
1075 // casting cares about nominal type, not just structural type
1076 let out_elem = ret_ty.simd_type(tcx);
1078 if in_elem == out_elem { return llargs[0]; }
1080 enum Style { Float, Int(/* is signed? */ bool), Unsupported }
1082 let (in_style, in_width) = match in_elem.sty {
1083 // vectors of pointer-sized integers should've been
1084 // disallowed before here, so this unwrap is safe.
1085 ty::TyInt(i) => (Style::Int(true), i.bit_width().unwrap()),
1086 ty::TyUint(u) => (Style::Int(false), u.bit_width().unwrap()),
1087 ty::TyFloat(f) => (Style::Float, f.bit_width()),
1088 _ => (Style::Unsupported, 0)
1090 let (out_style, out_width) = match out_elem.sty {
1091 ty::TyInt(i) => (Style::Int(true), i.bit_width().unwrap()),
1092 ty::TyUint(u) => (Style::Int(false), u.bit_width().unwrap()),
1093 ty::TyFloat(f) => (Style::Float, f.bit_width()),
1094 _ => (Style::Unsupported, 0)
1097 match (in_style, out_style) {
1098 (Style::Int(in_is_signed), Style::Int(_)) => {
1099 return match in_width.cmp(&out_width) {
1100 Ordering::Greater => bcx.trunc(llargs[0], llret_ty),
1101 Ordering::Equal => llargs[0],
1102 Ordering::Less => if in_is_signed {
1103 bcx.sext(llargs[0], llret_ty)
1105 bcx.zext(llargs[0], llret_ty)
1109 (Style::Int(in_is_signed), Style::Float) => {
1110 return if in_is_signed {
1111 bcx.sitofp(llargs[0], llret_ty)
1113 bcx.uitofp(llargs[0], llret_ty)
1116 (Style::Float, Style::Int(out_is_signed)) => {
1117 return if out_is_signed {
1118 bcx.fptosi(llargs[0], llret_ty)
1120 bcx.fptoui(llargs[0], llret_ty)
1123 (Style::Float, Style::Float) => {
1124 return match in_width.cmp(&out_width) {
1125 Ordering::Greater => bcx.fptrunc(llargs[0], llret_ty),
1126 Ordering::Equal => llargs[0],
1127 Ordering::Less => bcx.fpext(llargs[0], llret_ty)
1130 _ => {/* Unsupported. Fallthrough. */}
1133 "unsupported cast from `{}` with element `{}` to `{}` with element `{}`",
1137 macro_rules! arith {
1138 ($($name: ident: $($($p: ident),* => $call: ident),*;)*) => {
1140 if name == stringify!($name) {
1144 return bcx.$call(llargs[0], llargs[1])
1150 "unsupported operation on `{}` with element `{}`",
1157 simd_add: TyUint, TyInt => add, TyFloat => fadd;
1158 simd_sub: TyUint, TyInt => sub, TyFloat => fsub;
1159 simd_mul: TyUint, TyInt => mul, TyFloat => fmul;
1160 simd_div: TyFloat => fdiv;
1161 simd_shl: TyUint, TyInt => shl;
1162 simd_shr: TyUint => lshr, TyInt => ashr;
1163 simd_and: TyUint, TyInt => and;
1164 simd_or: TyUint, TyInt => or;
1165 simd_xor: TyUint, TyInt => xor;
1167 span_bug!(span, "unknown SIMD intrinsic");
1170 // Returns the width of an int TypeVariant, and if it's signed or not
1171 // Returns None if the type is not an integer
1172 // FIXME: there’s multiple of this functions, investigate using some of the already existing
1174 fn int_type_width_signed<'tcx>(sty: &ty::TypeVariants<'tcx>, ccx: &CrateContext)
1175 -> Option<(u64, bool)> {
1176 use rustc::ty::{TyInt, TyUint};
1178 TyInt(t) => Some((match t {
1180 match &ccx.tcx().sess.target.target.target_pointer_width[..] {
1184 tws => bug!("Unsupported target word size for isize: {}", tws),
1187 ast::IntTy::I8 => 8,
1188 ast::IntTy::I16 => 16,
1189 ast::IntTy::I32 => 32,
1190 ast::IntTy::I64 => 64,
1191 ast::IntTy::I128 => 128,
1193 TyUint(t) => Some((match t {
1194 ast::UintTy::Us => {
1195 match &ccx.tcx().sess.target.target.target_pointer_width[..] {
1199 tws => bug!("Unsupported target word size for usize: {}", tws),
1202 ast::UintTy::U8 => 8,
1203 ast::UintTy::U16 => 16,
1204 ast::UintTy::U32 => 32,
1205 ast::UintTy::U64 => 64,
1206 ast::UintTy::U128 => 128,
1212 // Returns the width of a float TypeVariant
1213 // Returns None if the type is not a float
1214 fn float_type_width<'tcx>(sty: &ty::TypeVariants<'tcx>)
1216 use rustc::ty::TyFloat;
1218 TyFloat(t) => Some(match t {
1219 ast::FloatTy::F32 => 32,
1220 ast::FloatTy::F64 => 64,