1 #![allow(non_upper_case_globals)]
6 use crate::abi::{Abi, FnType, LlvmType, PassMode};
7 use crate::context::CodegenCx;
8 use crate::type_::Type;
9 use crate::type_of::LayoutLlvmExt;
10 use crate::builder::Builder;
11 use crate::value::Value;
12 use crate::va_arg::emit_va_arg;
13 use rustc_codegen_ssa::MemFlags;
14 use rustc_codegen_ssa::mir::place::PlaceRef;
15 use rustc_codegen_ssa::mir::operand::{OperandRef, OperandValue};
16 use rustc_codegen_ssa::glue;
17 use rustc_codegen_ssa::base::{to_immediate, wants_msvc_seh, compare_simd_types};
18 use rustc::ty::{self, Ty};
19 use rustc::ty::layout::{self, LayoutOf, HasTyCtxt, Primitive};
20 use rustc_codegen_ssa::common::{IntPredicate, TypeKind};
22 use syntax::ast::{self, FloatTy};
24 use rustc_codegen_ssa::traits::*;
26 use rustc::session::Session;
29 use std::cmp::Ordering;
30 use std::{iter, i128, u128};
32 fn get_simple_intrinsic(cx: &CodegenCx<'ll, '_>, name: &str) -> Option<&'ll Value> {
33 let llvm_name = match name {
34 "sqrtf32" => "llvm.sqrt.f32",
35 "sqrtf64" => "llvm.sqrt.f64",
36 "powif32" => "llvm.powi.f32",
37 "powif64" => "llvm.powi.f64",
38 "sinf32" => "llvm.sin.f32",
39 "sinf64" => "llvm.sin.f64",
40 "cosf32" => "llvm.cos.f32",
41 "cosf64" => "llvm.cos.f64",
42 "powf32" => "llvm.pow.f32",
43 "powf64" => "llvm.pow.f64",
44 "expf32" => "llvm.exp.f32",
45 "expf64" => "llvm.exp.f64",
46 "exp2f32" => "llvm.exp2.f32",
47 "exp2f64" => "llvm.exp2.f64",
48 "logf32" => "llvm.log.f32",
49 "logf64" => "llvm.log.f64",
50 "log10f32" => "llvm.log10.f32",
51 "log10f64" => "llvm.log10.f64",
52 "log2f32" => "llvm.log2.f32",
53 "log2f64" => "llvm.log2.f64",
54 "fmaf32" => "llvm.fma.f32",
55 "fmaf64" => "llvm.fma.f64",
56 "fabsf32" => "llvm.fabs.f32",
57 "fabsf64" => "llvm.fabs.f64",
58 "minnumf32" => "llvm.minnum.f32",
59 "minnumf64" => "llvm.minnum.f64",
60 "maxnumf32" => "llvm.maxnum.f32",
61 "maxnumf64" => "llvm.maxnum.f64",
62 "copysignf32" => "llvm.copysign.f32",
63 "copysignf64" => "llvm.copysign.f64",
64 "floorf32" => "llvm.floor.f32",
65 "floorf64" => "llvm.floor.f64",
66 "ceilf32" => "llvm.ceil.f32",
67 "ceilf64" => "llvm.ceil.f64",
68 "truncf32" => "llvm.trunc.f32",
69 "truncf64" => "llvm.trunc.f64",
70 "rintf32" => "llvm.rint.f32",
71 "rintf64" => "llvm.rint.f64",
72 "nearbyintf32" => "llvm.nearbyint.f32",
73 "nearbyintf64" => "llvm.nearbyint.f64",
74 "roundf32" => "llvm.round.f32",
75 "roundf64" => "llvm.round.f64",
76 "assume" => "llvm.assume",
77 "abort" => "llvm.trap",
80 Some(cx.get_intrinsic(&llvm_name))
83 impl IntrinsicCallMethods<'tcx> for Builder<'a, 'll, 'tcx> {
84 fn codegen_intrinsic_call(
87 fn_ty: &FnType<'tcx, Ty<'tcx>>,
88 args: &[OperandRef<'tcx, &'ll Value>],
94 let (def_id, substs) = match callee_ty.sty {
95 ty::FnDef(def_id, substs) => (def_id, substs),
96 _ => bug!("expected fn item type, found {}", callee_ty)
99 let sig = callee_ty.fn_sig(tcx);
100 let sig = tcx.normalize_erasing_late_bound_regions(ty::ParamEnv::reveal_all(), &sig);
101 let arg_tys = sig.inputs();
102 let ret_ty = sig.output();
103 let name = &*tcx.item_name(def_id).as_str();
105 let llret_ty = self.layout_of(ret_ty).llvm_type(self);
106 let result = PlaceRef::new_sized(llresult, fn_ty.ret.layout, fn_ty.ret.layout.align.abi);
108 let simple = get_simple_intrinsic(self, name);
109 let llval = match name {
110 _ if simple.is_some() => {
111 self.call(simple.unwrap(),
112 &args.iter().map(|arg| arg.immediate()).collect::<Vec<_>>(),
119 let expect = self.get_intrinsic(&("llvm.expect.i1"));
120 self.call(expect, &[args[0].immediate(), self.const_bool(true)], None)
123 let expect = self.get_intrinsic(&("llvm.expect.i1"));
124 self.call(expect, &[args[0].immediate(), self.const_bool(false)], None)
135 let llfn = self.get_intrinsic(&("llvm.debugtrap"));
136 self.call(llfn, &[], None)
139 let tp_ty = substs.type_at(0);
140 self.const_usize(self.size_of(tp_ty).bytes())
143 self.va_start(args[0].immediate())
146 self.va_end(args[0].immediate())
149 let intrinsic = self.cx().get_intrinsic(&("llvm.va_copy"));
150 self.call(intrinsic, &[args[0].immediate(), args[1].immediate()], None)
153 match fn_ty.ret.layout.abi {
154 layout::Abi::Scalar(ref scalar) => {
156 Primitive::Int(..) => {
157 if self.cx().size_of(ret_ty).bytes() < 4 {
158 // va_arg should not be called on a integer type
159 // less than 4 bytes in length. If it is, promote
160 // the integer to a `i32` and truncate the result
161 // back to the smaller type.
162 let promoted_result = emit_va_arg(self, args[0],
164 self.trunc(promoted_result, llret_ty)
166 emit_va_arg(self, args[0], ret_ty)
169 Primitive::Float(FloatTy::F64) |
170 Primitive::Pointer => {
171 emit_va_arg(self, args[0], ret_ty)
173 // `va_arg` should never be used with the return type f32.
174 Primitive::Float(FloatTy::F32) => {
175 bug!("the va_arg intrinsic does not work with `f32`")
180 bug!("the va_arg intrinsic does not work with non-scalar types")
185 let tp_ty = substs.type_at(0);
186 if let OperandValue::Pair(_, meta) = args[0].val {
187 let (llsize, _) = glue::size_and_align_of_dst(self, tp_ty, Some(meta));
190 self.const_usize(self.size_of(tp_ty).bytes())
194 let tp_ty = substs.type_at(0);
195 self.const_usize(self.align_of(tp_ty).bytes())
197 "min_align_of_val" => {
198 let tp_ty = substs.type_at(0);
199 if let OperandValue::Pair(_, meta) = args[0].val {
200 let (_, llalign) = glue::size_and_align_of_dst(self, tp_ty, Some(meta));
203 self.const_usize(self.align_of(tp_ty).bytes())
207 let tp_ty = substs.type_at(0);
208 self.const_usize(self.layout_of(tp_ty).align.pref.bytes())
211 let tp_ty = substs.type_at(0);
212 let ty_name = self.tcx.type_name(tp_ty);
213 OperandRef::from_const(self, ty_name).immediate_or_packed_pair(self)
216 self.const_u64(self.tcx.type_id_hash(substs.type_at(0)))
219 let ty = substs.type_at(0);
220 if !self.layout_of(ty).is_zst() {
221 // Just zero out the stack slot.
222 // If we store a zero constant, LLVM will drown in vreg allocation for large
223 // data structures, and the generated code will be awful. (A telltale sign of
224 // this is large quantities of `mov [byte ptr foo],0` in the generated code.)
236 // Effectively no-ops
241 let tp_ty = substs.type_at(0);
243 self.const_bool(self.type_needs_drop(tp_ty))
246 let ptr = args[0].immediate();
247 let offset = args[1].immediate();
248 self.inbounds_gep(ptr, &[offset])
251 let ptr = args[0].immediate();
252 let offset = args[1].immediate();
253 self.gep(ptr, &[offset])
256 "copy_nonoverlapping" => {
257 copy_intrinsic(self, false, false, substs.type_at(0),
258 args[1].immediate(), args[0].immediate(), args[2].immediate());
262 copy_intrinsic(self, true, false, substs.type_at(0),
263 args[1].immediate(), args[0].immediate(), args[2].immediate());
267 memset_intrinsic(self, false, substs.type_at(0),
268 args[0].immediate(), args[1].immediate(), args[2].immediate());
272 "volatile_copy_nonoverlapping_memory" => {
273 copy_intrinsic(self, false, true, substs.type_at(0),
274 args[0].immediate(), args[1].immediate(), args[2].immediate());
277 "volatile_copy_memory" => {
278 copy_intrinsic(self, true, true, substs.type_at(0),
279 args[0].immediate(), args[1].immediate(), args[2].immediate());
282 "volatile_set_memory" => {
283 memset_intrinsic(self, true, substs.type_at(0),
284 args[0].immediate(), args[1].immediate(), args[2].immediate());
287 "volatile_load" | "unaligned_volatile_load" => {
288 let tp_ty = substs.type_at(0);
289 let mut ptr = args[0].immediate();
290 if let PassMode::Cast(ty) = fn_ty.ret.mode {
291 ptr = self.pointercast(ptr, self.type_ptr_to(ty.llvm_type(self)));
293 let load = self.volatile_load(ptr);
294 let align = if name == "unaligned_volatile_load" {
297 self.align_of(tp_ty).bytes() as u32
300 llvm::LLVMSetAlignment(load, align);
302 to_immediate(self, load, self.layout_of(tp_ty))
304 "volatile_store" => {
305 let dst = args[0].deref(self.cx());
306 args[1].val.volatile_store(self, dst);
309 "unaligned_volatile_store" => {
310 let dst = args[0].deref(self.cx());
311 args[1].val.unaligned_volatile_store(self, dst);
314 "prefetch_read_data" | "prefetch_write_data" |
315 "prefetch_read_instruction" | "prefetch_write_instruction" => {
316 let expect = self.get_intrinsic(&("llvm.prefetch"));
317 let (rw, cache_type) = match name {
318 "prefetch_read_data" => (0, 1),
319 "prefetch_write_data" => (1, 1),
320 "prefetch_read_instruction" => (0, 0),
321 "prefetch_write_instruction" => (1, 0),
328 self.const_i32(cache_type)
331 "ctlz" | "ctlz_nonzero" | "cttz" | "cttz_nonzero" | "ctpop" | "bswap" |
332 "bitreverse" | "add_with_overflow" | "sub_with_overflow" |
333 "mul_with_overflow" | "overflowing_add" | "overflowing_sub" | "overflowing_mul" |
334 "unchecked_div" | "unchecked_rem" | "unchecked_shl" | "unchecked_shr" |
335 "unchecked_add" | "unchecked_sub" | "unchecked_mul" | "exact_div" |
336 "rotate_left" | "rotate_right" | "saturating_add" | "saturating_sub" => {
338 match int_type_width_signed(ty, self) {
339 Some((width, signed)) =>
342 let y = self.const_bool(false);
343 let llfn = self.get_intrinsic(
344 &format!("llvm.{}.i{}", name, width),
346 self.call(llfn, &[args[0].immediate(), y], None)
348 "ctlz_nonzero" | "cttz_nonzero" => {
349 let y = self.const_bool(true);
350 let llvm_name = &format!("llvm.{}.i{}", &name[..4], width);
351 let llfn = self.get_intrinsic(llvm_name);
352 self.call(llfn, &[args[0].immediate(), y], None)
354 "ctpop" => self.call(
355 self.get_intrinsic(&format!("llvm.ctpop.i{}", width)),
356 &[args[0].immediate()],
361 args[0].immediate() // byte swap a u8/i8 is just a no-op
365 &format!("llvm.bswap.i{}", width),
367 &[args[0].immediate()],
375 &format!("llvm.bitreverse.i{}", width),
377 &[args[0].immediate()],
381 "add_with_overflow" | "sub_with_overflow" | "mul_with_overflow" => {
382 let intrinsic = format!("llvm.{}{}.with.overflow.i{}",
383 if signed { 's' } else { 'u' },
385 let llfn = self.get_intrinsic(&intrinsic);
387 // Convert `i1` to a `bool`, and write it to the out parameter
388 let pair = self.call(llfn, &[
392 let val = self.extract_value(pair, 0);
393 let overflow = self.extract_value(pair, 1);
394 let overflow = self.zext(overflow, self.type_bool());
396 let dest = result.project_field(self, 0);
397 self.store(val, dest.llval, dest.align);
398 let dest = result.project_field(self, 1);
399 self.store(overflow, dest.llval, dest.align);
403 "overflowing_add" => self.add(args[0].immediate(), args[1].immediate()),
404 "overflowing_sub" => self.sub(args[0].immediate(), args[1].immediate()),
405 "overflowing_mul" => self.mul(args[0].immediate(), args[1].immediate()),
408 self.exactsdiv(args[0].immediate(), args[1].immediate())
410 self.exactudiv(args[0].immediate(), args[1].immediate())
414 self.sdiv(args[0].immediate(), args[1].immediate())
416 self.udiv(args[0].immediate(), args[1].immediate())
420 self.srem(args[0].immediate(), args[1].immediate())
422 self.urem(args[0].immediate(), args[1].immediate())
424 "unchecked_shl" => self.shl(args[0].immediate(), args[1].immediate()),
427 self.ashr(args[0].immediate(), args[1].immediate())
429 self.lshr(args[0].immediate(), args[1].immediate())
433 self.unchecked_sadd(args[0].immediate(), args[1].immediate())
435 self.unchecked_uadd(args[0].immediate(), args[1].immediate())
440 self.unchecked_ssub(args[0].immediate(), args[1].immediate())
442 self.unchecked_usub(args[0].immediate(), args[1].immediate())
447 self.unchecked_smul(args[0].immediate(), args[1].immediate())
449 self.unchecked_umul(args[0].immediate(), args[1].immediate())
452 "rotate_left" | "rotate_right" => {
453 let is_left = name == "rotate_left";
454 let val = args[0].immediate();
455 let raw_shift = args[1].immediate();
456 if llvm_util::get_major_version() >= 7 {
457 // rotate = funnel shift with first two args the same
458 let llvm_name = &format!("llvm.fsh{}.i{}",
459 if is_left { 'l' } else { 'r' }, width);
460 let llfn = self.get_intrinsic(llvm_name);
461 self.call(llfn, &[val, val, raw_shift], None)
463 // rotate_left: (X << (S % BW)) | (X >> ((BW - S) % BW))
464 // rotate_right: (X << ((BW - S) % BW)) | (X >> (S % BW))
465 let width = self.const_uint(
469 let shift = self.urem(raw_shift, width);
470 let width_minus_raw_shift = self.sub(width, raw_shift);
471 let inv_shift = self.urem(width_minus_raw_shift, width);
472 let shift1 = self.shl(
474 if is_left { shift } else { inv_shift },
476 let shift2 = self.lshr(
478 if !is_left { shift } else { inv_shift },
480 self.or(shift1, shift2)
483 "saturating_add" | "saturating_sub" => {
484 let is_add = name == "saturating_add";
485 let lhs = args[0].immediate();
486 let rhs = args[1].immediate();
487 if llvm_util::get_major_version() >= 8 {
488 let llvm_name = &format!("llvm.{}{}.sat.i{}",
489 if signed { 's' } else { 'u' },
490 if is_add { "add" } else { "sub" },
492 let llfn = self.get_intrinsic(llvm_name);
493 self.call(llfn, &[lhs, rhs], None)
495 let llvm_name = &format!("llvm.{}{}.with.overflow.i{}",
496 if signed { 's' } else { 'u' },
497 if is_add { "add" } else { "sub" },
499 let llfn = self.get_intrinsic(llvm_name);
500 let pair = self.call(llfn, &[lhs, rhs], None);
501 let val = self.extract_value(pair, 0);
502 let overflow = self.extract_value(pair, 1);
503 let llty = self.type_ix(width);
505 let limit = if signed {
506 let limit_lo = self.const_uint_big(
507 llty, (i128::MIN >> (128 - width)) as u128);
508 let limit_hi = self.const_uint_big(
509 llty, (i128::MAX >> (128 - width)) as u128);
511 IntPredicate::IntSLT, val, self.const_uint(llty, 0));
512 self.select(neg, limit_hi, limit_lo)
514 self.const_uint_big(llty, u128::MAX >> (128 - width))
516 self.const_uint(llty, 0)
518 self.select(overflow, limit, val)
524 span_invalid_monomorphization_error(
526 &format!("invalid monomorphization of `{}` intrinsic: \
527 expected basic integer type, found `{}`", name, ty));
533 "fadd_fast" | "fsub_fast" | "fmul_fast" | "fdiv_fast" | "frem_fast" => {
534 match float_type_width(arg_tys[0]) {
537 "fadd_fast" => self.fadd_fast(args[0].immediate(), args[1].immediate()),
538 "fsub_fast" => self.fsub_fast(args[0].immediate(), args[1].immediate()),
539 "fmul_fast" => self.fmul_fast(args[0].immediate(), args[1].immediate()),
540 "fdiv_fast" => self.fdiv_fast(args[0].immediate(), args[1].immediate()),
541 "frem_fast" => self.frem_fast(args[0].immediate(), args[1].immediate()),
545 span_invalid_monomorphization_error(
547 &format!("invalid monomorphization of `{}` intrinsic: \
548 expected basic float type, found `{}`", name, arg_tys[0]));
555 "discriminant_value" => {
556 args[0].deref(self.cx()).codegen_get_discr(self, ret_ty)
559 name if name.starts_with("simd_") => {
560 match generic_simd_intrinsic(self, name,
569 // This requires that atomic intrinsics follow a specific naming pattern:
570 // "atomic_<operation>[_<ordering>]", and no ordering means SeqCst
571 name if name.starts_with("atomic_") => {
572 use rustc_codegen_ssa::common::AtomicOrdering::*;
573 use rustc_codegen_ssa::common::
574 {SynchronizationScope, AtomicRmwBinOp};
576 let split: Vec<&str> = name.split('_').collect();
578 let is_cxchg = split[1] == "cxchg" || split[1] == "cxchgweak";
579 let (order, failorder) = match split.len() {
580 2 => (SequentiallyConsistent, SequentiallyConsistent),
581 3 => match split[2] {
582 "unordered" => (Unordered, Unordered),
583 "relaxed" => (Monotonic, Monotonic),
584 "acq" => (Acquire, Acquire),
585 "rel" => (Release, Monotonic),
586 "acqrel" => (AcquireRelease, Acquire),
587 "failrelaxed" if is_cxchg =>
588 (SequentiallyConsistent, Monotonic),
589 "failacq" if is_cxchg =>
590 (SequentiallyConsistent, Acquire),
591 _ => self.sess().fatal("unknown ordering in atomic intrinsic")
593 4 => match (split[2], split[3]) {
594 ("acq", "failrelaxed") if is_cxchg =>
595 (Acquire, Monotonic),
596 ("acqrel", "failrelaxed") if is_cxchg =>
597 (AcquireRelease, Monotonic),
598 _ => self.sess().fatal("unknown ordering in atomic intrinsic")
600 _ => self.sess().fatal("Atomic intrinsic not in correct format"),
603 let invalid_monomorphization = |ty| {
604 span_invalid_monomorphization_error(tcx.sess, span,
605 &format!("invalid monomorphization of `{}` intrinsic: \
606 expected basic integer type, found `{}`", name, ty));
610 "cxchg" | "cxchgweak" => {
611 let ty = substs.type_at(0);
612 if int_type_width_signed(ty, self).is_some() {
613 let weak = split[1] == "cxchgweak";
614 let pair = self.atomic_cmpxchg(
621 let val = self.extract_value(pair, 0);
622 let success = self.extract_value(pair, 1);
623 let success = self.zext(success, self.type_bool());
625 let dest = result.project_field(self, 0);
626 self.store(val, dest.llval, dest.align);
627 let dest = result.project_field(self, 1);
628 self.store(success, dest.llval, dest.align);
631 return invalid_monomorphization(ty);
636 let ty = substs.type_at(0);
637 if int_type_width_signed(ty, self).is_some() {
638 let size = self.size_of(ty);
639 self.atomic_load(args[0].immediate(), order, size)
641 return invalid_monomorphization(ty);
646 let ty = substs.type_at(0);
647 if int_type_width_signed(ty, self).is_some() {
648 let size = self.size_of(ty);
657 return invalid_monomorphization(ty);
662 self.atomic_fence(order, SynchronizationScope::CrossThread);
666 "singlethreadfence" => {
667 self.atomic_fence(order, SynchronizationScope::SingleThread);
671 // These are all AtomicRMW ops
673 let atom_op = match op {
674 "xchg" => AtomicRmwBinOp::AtomicXchg,
675 "xadd" => AtomicRmwBinOp::AtomicAdd,
676 "xsub" => AtomicRmwBinOp::AtomicSub,
677 "and" => AtomicRmwBinOp::AtomicAnd,
678 "nand" => AtomicRmwBinOp::AtomicNand,
679 "or" => AtomicRmwBinOp::AtomicOr,
680 "xor" => AtomicRmwBinOp::AtomicXor,
681 "max" => AtomicRmwBinOp::AtomicMax,
682 "min" => AtomicRmwBinOp::AtomicMin,
683 "umax" => AtomicRmwBinOp::AtomicUMax,
684 "umin" => AtomicRmwBinOp::AtomicUMin,
685 _ => self.sess().fatal("unknown atomic operation")
688 let ty = substs.type_at(0);
689 if int_type_width_signed(ty, self).is_some() {
697 return invalid_monomorphization(ty);
703 "nontemporal_store" => {
704 let dst = args[0].deref(self.cx());
705 args[1].val.nontemporal_store(self, dst);
709 _ => bug!("unknown intrinsic '{}'", name),
712 if !fn_ty.ret.is_ignore() {
713 if let PassMode::Cast(ty) = fn_ty.ret.mode {
714 let ptr_llty = self.type_ptr_to(ty.llvm_type(self));
715 let ptr = self.pointercast(result.llval, ptr_llty);
716 self.store(llval, ptr, result.align);
718 OperandRef::from_immediate_or_packed_pair(self, llval, result.layout)
719 .val.store(self, result);
724 fn abort(&mut self) {
725 let fnname = self.get_intrinsic(&("llvm.trap"));
726 self.call(fnname, &[], None);
729 fn assume(&mut self, val: Self::Value) {
730 let assume_intrinsic = self.get_intrinsic("llvm.assume");
731 self.call(assume_intrinsic, &[val], None);
734 fn expect(&mut self, cond: Self::Value, expected: bool) -> Self::Value {
735 let expect = self.get_intrinsic(&"llvm.expect.i1");
736 self.call(expect, &[cond, self.const_bool(expected)], None)
739 fn va_start(&mut self, va_list: &'ll Value) -> &'ll Value {
740 let intrinsic = self.cx().get_intrinsic("llvm.va_start");
741 self.call(intrinsic, &[va_list], None)
744 fn va_end(&mut self, va_list: &'ll Value) -> &'ll Value {
745 let intrinsic = self.cx().get_intrinsic("llvm.va_end");
746 self.call(intrinsic, &[va_list], None)
751 bx: &mut Builder<'a, 'll, 'tcx>,
759 let (size, align) = bx.size_and_align_of(ty);
760 let size = bx.mul(bx.const_usize(size.bytes()), count);
761 let flags = if volatile {
767 bx.memmove(dst, align, src, align, size, flags);
769 bx.memcpy(dst, align, src, align, size, flags);
774 bx: &mut Builder<'a, 'll, 'tcx>,
781 let (size, align) = bx.size_and_align_of(ty);
782 let size = bx.mul(bx.const_usize(size.bytes()), count);
783 let flags = if volatile {
788 bx.memset(dst, val, size, align, flags);
792 bx: &mut Builder<'a, 'll, 'tcx>,
795 local_ptr: &'ll Value,
798 if bx.sess().no_landing_pads() {
799 bx.call(func, &[data], None);
800 let ptr_align = bx.tcx().data_layout.pointer_align.abi;
801 bx.store(bx.const_null(bx.type_i8p()), dest, ptr_align);
802 } else if wants_msvc_seh(bx.sess()) {
803 codegen_msvc_try(bx, func, data, local_ptr, dest);
805 codegen_gnu_try(bx, func, data, local_ptr, dest);
809 // MSVC's definition of the `rust_try` function.
811 // This implementation uses the new exception handling instructions in LLVM
812 // which have support in LLVM for SEH on MSVC targets. Although these
813 // instructions are meant to work for all targets, as of the time of this
814 // writing, however, LLVM does not recommend the usage of these new instructions
815 // as the old ones are still more optimized.
817 bx: &mut Builder<'a, 'll, 'tcx>,
820 local_ptr: &'ll Value,
823 let llfn = get_rust_try_fn(bx, &mut |mut bx| {
824 bx.set_personality_fn(bx.eh_personality());
826 let mut normal = bx.build_sibling_block("normal");
827 let mut catchswitch = bx.build_sibling_block("catchswitch");
828 let mut catchpad = bx.build_sibling_block("catchpad");
829 let mut caught = bx.build_sibling_block("caught");
831 let func = llvm::get_param(bx.llfn(), 0);
832 let data = llvm::get_param(bx.llfn(), 1);
833 let local_ptr = llvm::get_param(bx.llfn(), 2);
835 // We're generating an IR snippet that looks like:
837 // declare i32 @rust_try(%func, %data, %ptr) {
838 // %slot = alloca i64*
839 // invoke %func(%data) to label %normal unwind label %catchswitch
845 // %cs = catchswitch within none [%catchpad] unwind to caller
848 // %tok = catchpad within %cs [%type_descriptor, 0, %slot]
849 // %ptr[0] = %slot[0]
850 // %ptr[1] = %slot[1]
851 // catchret from %tok to label %caught
857 // This structure follows the basic usage of throw/try/catch in LLVM.
858 // For example, compile this C++ snippet to see what LLVM generates:
860 // #include <stdint.h>
862 // int bar(void (*foo)(void), uint64_t *ret) {
866 // } catch(uint64_t a[2]) {
873 // More information can be found in libstd's seh.rs implementation.
874 let i64p = bx.type_ptr_to(bx.type_i64());
875 let ptr_align = bx.tcx().data_layout.pointer_align.abi;
876 let slot = bx.alloca(i64p, "slot", ptr_align);
877 bx.invoke(func, &[data], normal.llbb(), catchswitch.llbb(), None);
879 normal.ret(bx.const_i32(0));
881 let cs = catchswitch.catch_switch(None, None, 1);
882 catchswitch.add_handler(cs, catchpad.llbb());
884 let tydesc = match bx.tcx().lang_items().msvc_try_filter() {
885 Some(did) => bx.get_static(did),
886 None => bug!("msvc_try_filter not defined"),
888 let funclet = catchpad.catch_pad(cs, &[tydesc, bx.const_i32(0), slot]);
889 let addr = catchpad.load(slot, ptr_align);
891 let i64_align = bx.tcx().data_layout.i64_align.abi;
892 let arg1 = catchpad.load(addr, i64_align);
893 let val1 = bx.const_i32(1);
894 let gep1 = catchpad.inbounds_gep(addr, &[val1]);
895 let arg2 = catchpad.load(gep1, i64_align);
896 let local_ptr = catchpad.bitcast(local_ptr, i64p);
897 let gep2 = catchpad.inbounds_gep(local_ptr, &[val1]);
898 catchpad.store(arg1, local_ptr, i64_align);
899 catchpad.store(arg2, gep2, i64_align);
900 catchpad.catch_ret(&funclet, caught.llbb());
902 caught.ret(bx.const_i32(1));
905 // Note that no invoke is used here because by definition this function
906 // can't panic (that's what it's catching).
907 let ret = bx.call(llfn, &[func, data, local_ptr], None);
908 let i32_align = bx.tcx().data_layout.i32_align.abi;
909 bx.store(ret, dest, i32_align);
912 // Definition of the standard `try` function for Rust using the GNU-like model
913 // of exceptions (e.g., the normal semantics of LLVM's `landingpad` and `invoke`
916 // This codegen is a little surprising because we always call a shim
917 // function instead of inlining the call to `invoke` manually here. This is done
918 // because in LLVM we're only allowed to have one personality per function
919 // definition. The call to the `try` intrinsic is being inlined into the
920 // function calling it, and that function may already have other personality
921 // functions in play. By calling a shim we're guaranteed that our shim will have
922 // the right personality function.
924 bx: &mut Builder<'a, 'll, 'tcx>,
927 local_ptr: &'ll Value,
930 let llfn = get_rust_try_fn(bx, &mut |mut bx| {
931 // Codegens the shims described above:
934 // invoke %func(%args...) normal %normal unwind %catch
940 // (ptr, _) = landingpad
941 // store ptr, %local_ptr
944 // Note that the `local_ptr` data passed into the `try` intrinsic is
945 // expected to be `*mut *mut u8` for this to actually work, but that's
946 // managed by the standard library.
948 let mut then = bx.build_sibling_block("then");
949 let mut catch = bx.build_sibling_block("catch");
951 let func = llvm::get_param(bx.llfn(), 0);
952 let data = llvm::get_param(bx.llfn(), 1);
953 let local_ptr = llvm::get_param(bx.llfn(), 2);
954 bx.invoke(func, &[data], then.llbb(), catch.llbb(), None);
955 then.ret(bx.const_i32(0));
957 // Type indicator for the exception being thrown.
959 // The first value in this tuple is a pointer to the exception object
960 // being thrown. The second value is a "selector" indicating which of
961 // the landing pad clauses the exception's type had been matched to.
962 // rust_try ignores the selector.
963 let lpad_ty = bx.type_struct(&[bx.type_i8p(), bx.type_i32()], false);
964 let vals = catch.landing_pad(lpad_ty, bx.eh_personality(), 1);
965 catch.add_clause(vals, bx.const_null(bx.type_i8p()));
966 let ptr = catch.extract_value(vals, 0);
967 let ptr_align = bx.tcx().data_layout.pointer_align.abi;
968 let bitcast = catch.bitcast(local_ptr, bx.type_ptr_to(bx.type_i8p()));
969 catch.store(ptr, bitcast, ptr_align);
970 catch.ret(bx.const_i32(1));
973 // Note that no invoke is used here because by definition this function
974 // can't panic (that's what it's catching).
975 let ret = bx.call(llfn, &[func, data, local_ptr], None);
976 let i32_align = bx.tcx().data_layout.i32_align.abi;
977 bx.store(ret, dest, i32_align);
980 // Helper function to give a Block to a closure to codegen a shim function.
981 // This is currently primarily used for the `try` intrinsic functions above.
982 fn gen_fn<'ll, 'tcx>(
983 cx: &CodegenCx<'ll, 'tcx>,
985 inputs: Vec<Ty<'tcx>>,
987 codegen: &mut dyn FnMut(Builder<'_, 'll, 'tcx>),
989 let rust_fn_sig = ty::Binder::bind(cx.tcx.mk_fn_sig(
993 hir::Unsafety::Unsafe,
996 let llfn = cx.define_internal_fn(name, rust_fn_sig);
997 attributes::from_fn_attrs(cx, llfn, None, rust_fn_sig);
998 let bx = Builder::new_block(cx, llfn, "entry-block");
1003 // Helper function used to get a handle to the `__rust_try` function used to
1004 // catch exceptions.
1006 // This function is only generated once and is then cached.
1007 fn get_rust_try_fn<'ll, 'tcx>(
1008 cx: &CodegenCx<'ll, 'tcx>,
1009 codegen: &mut dyn FnMut(Builder<'_, 'll, 'tcx>),
1011 if let Some(llfn) = cx.rust_try_fn.get() {
1015 // Define the type up front for the signature of the rust_try function.
1017 let i8p = tcx.mk_mut_ptr(tcx.types.i8);
1018 let fn_ty = tcx.mk_fn_ptr(ty::Binder::bind(tcx.mk_fn_sig(
1022 hir::Unsafety::Unsafe,
1025 let output = tcx.types.i32;
1026 let rust_try = gen_fn(cx, "__rust_try", vec![fn_ty, i8p, i8p], output, codegen);
1027 cx.rust_try_fn.set(Some(rust_try));
1031 fn span_invalid_monomorphization_error(a: &Session, b: Span, c: &str) {
1032 span_err!(a, b, E0511, "{}", c);
1035 fn generic_simd_intrinsic(
1036 bx: &mut Builder<'a, 'll, 'tcx>,
1038 callee_ty: Ty<'tcx>,
1039 args: &[OperandRef<'tcx, &'ll Value>],
1041 llret_ty: &'ll Type,
1043 ) -> Result<&'ll Value, ()> {
1044 // macros for error handling:
1045 macro_rules! emit_error {
1049 ($msg: tt, $($fmt: tt)*) => {
1050 span_invalid_monomorphization_error(
1052 &format!(concat!("invalid monomorphization of `{}` intrinsic: ", $msg),
1057 macro_rules! return_error {
1060 emit_error!($($fmt)*);
1066 macro_rules! require {
1067 ($cond: expr, $($fmt: tt)*) => {
1069 return_error!($($fmt)*);
1074 macro_rules! require_simd {
1075 ($ty: expr, $position: expr) => {
1076 require!($ty.is_simd(), "expected SIMD {} type, found non-SIMD `{}`", $position, $ty)
1081 let sig = tcx.normalize_erasing_late_bound_regions(
1082 ty::ParamEnv::reveal_all(),
1083 &callee_ty.fn_sig(tcx),
1085 let arg_tys = sig.inputs();
1087 if name == "simd_select_bitmask" {
1088 let in_ty = arg_tys[0];
1089 let m_len = match in_ty.sty {
1090 // Note that this `.unwrap()` crashes for isize/usize, that's sort
1091 // of intentional as there's not currently a use case for that.
1092 ty::Int(i) => i.bit_width().unwrap(),
1093 ty::Uint(i) => i.bit_width().unwrap(),
1094 _ => return_error!("`{}` is not an integral type", in_ty),
1096 require_simd!(arg_tys[1], "argument");
1097 let v_len = arg_tys[1].simd_size(tcx);
1098 require!(m_len == v_len,
1099 "mismatched lengths: mask length `{}` != other vector length `{}`",
1102 let i1 = bx.type_i1();
1103 let i1xn = bx.type_vector(i1, m_len as u64);
1104 let m_i1s = bx.bitcast(args[0].immediate(), i1xn);
1105 return Ok(bx.select(m_i1s, args[1].immediate(), args[2].immediate()));
1108 // every intrinsic below takes a SIMD vector as its first argument
1109 require_simd!(arg_tys[0], "input");
1110 let in_ty = arg_tys[0];
1111 let in_elem = arg_tys[0].simd_type(tcx);
1112 let in_len = arg_tys[0].simd_size(tcx);
1114 let comparison = match name {
1115 "simd_eq" => Some(hir::BinOpKind::Eq),
1116 "simd_ne" => Some(hir::BinOpKind::Ne),
1117 "simd_lt" => Some(hir::BinOpKind::Lt),
1118 "simd_le" => Some(hir::BinOpKind::Le),
1119 "simd_gt" => Some(hir::BinOpKind::Gt),
1120 "simd_ge" => Some(hir::BinOpKind::Ge),
1124 if let Some(cmp_op) = comparison {
1125 require_simd!(ret_ty, "return");
1127 let out_len = ret_ty.simd_size(tcx);
1128 require!(in_len == out_len,
1129 "expected return type with length {} (same as input type `{}`), \
1130 found `{}` with length {}",
1133 require!(bx.type_kind(bx.element_type(llret_ty)) == TypeKind::Integer,
1134 "expected return type with integer elements, found `{}` with non-integer `{}`",
1136 ret_ty.simd_type(tcx));
1138 return Ok(compare_simd_types(bx,
1139 args[0].immediate(),
1140 args[1].immediate(),
1146 if name.starts_with("simd_shuffle") {
1147 let n: usize = name["simd_shuffle".len()..].parse().unwrap_or_else(|_|
1148 span_bug!(span, "bad `simd_shuffle` instruction only caught in codegen?"));
1150 require_simd!(ret_ty, "return");
1152 let out_len = ret_ty.simd_size(tcx);
1153 require!(out_len == n,
1154 "expected return type of length {}, found `{}` with length {}",
1155 n, ret_ty, out_len);
1156 require!(in_elem == ret_ty.simd_type(tcx),
1157 "expected return element type `{}` (element of input `{}`), \
1158 found `{}` with element type `{}`",
1160 ret_ty, ret_ty.simd_type(tcx));
1162 let total_len = in_len as u128 * 2;
1164 let vector = args[2].immediate();
1166 let indices: Option<Vec<_>> = (0..n)
1169 let val = bx.const_get_elt(vector, i as u64);
1170 match bx.const_to_opt_u128(val, true) {
1172 emit_error!("shuffle index #{} is not a constant", arg_idx);
1175 Some(idx) if idx >= total_len => {
1176 emit_error!("shuffle index #{} is out of bounds (limit {})",
1177 arg_idx, total_len);
1180 Some(idx) => Some(bx.const_i32(idx as i32)),
1184 let indices = match indices {
1186 None => return Ok(bx.const_null(llret_ty))
1189 return Ok(bx.shuffle_vector(args[0].immediate(),
1190 args[1].immediate(),
1191 bx.const_vector(&indices)))
1194 if name == "simd_insert" {
1195 require!(in_elem == arg_tys[2],
1196 "expected inserted type `{}` (element of input `{}`), found `{}`",
1197 in_elem, in_ty, arg_tys[2]);
1198 return Ok(bx.insert_element(args[0].immediate(),
1199 args[2].immediate(),
1200 args[1].immediate()))
1202 if name == "simd_extract" {
1203 require!(ret_ty == in_elem,
1204 "expected return type `{}` (element of input `{}`), found `{}`",
1205 in_elem, in_ty, ret_ty);
1206 return Ok(bx.extract_element(args[0].immediate(), args[1].immediate()))
1209 if name == "simd_select" {
1210 let m_elem_ty = in_elem;
1212 require_simd!(arg_tys[1], "argument");
1213 let v_len = arg_tys[1].simd_size(tcx);
1214 require!(m_len == v_len,
1215 "mismatched lengths: mask length `{}` != other vector length `{}`",
1218 match m_elem_ty.sty {
1220 _ => return_error!("mask element type is `{}`, expected `i_`", m_elem_ty)
1222 // truncate the mask to a vector of i1s
1223 let i1 = bx.type_i1();
1224 let i1xn = bx.type_vector(i1, m_len as u64);
1225 let m_i1s = bx.trunc(args[0].immediate(), i1xn);
1226 return Ok(bx.select(m_i1s, args[1].immediate(), args[2].immediate()));
1229 if name == "simd_bitmask" {
1230 // The `fn simd_bitmask(vector) -> unsigned integer` intrinsic takes a
1231 // vector mask and returns an unsigned integer containing the most
1232 // significant bit (MSB) of each lane.
1233 use rustc_target::abi::HasDataLayout;
1235 // If the vector has less than 8 lanes, an u8 is returned with zeroed
1237 let expected_int_bits = in_len.max(8);
1239 ty::Uint(i) if i.bit_width() == Some(expected_int_bits) => (),
1241 "bitmask `{}`, expected `u{}`",
1242 ret_ty, expected_int_bits
1246 // Integer vector <i{in_bitwidth} x in_len>:
1247 let (i_xn, in_elem_bitwidth) = match in_elem.sty {
1249 args[0].immediate(),
1250 i.bit_width().unwrap_or(bx.data_layout().pointer_size.bits() as _)
1253 args[0].immediate(),
1254 i.bit_width().unwrap_or(bx.data_layout().pointer_size.bits() as _)
1257 "vector argument `{}`'s element type `{}`, expected integer element type",
1262 // Shift the MSB to the right by "in_elem_bitwidth - 1" into the first bit position.
1263 let shift_indices = vec![
1264 bx.cx.const_int(bx.type_ix(in_elem_bitwidth as _), (in_elem_bitwidth - 1) as _); in_len
1266 let i_xn_msb = bx.lshr(i_xn, bx.const_vector(shift_indices.as_slice()));
1267 // Truncate vector to an <i1 x N>
1268 let i1xn = bx.trunc(i_xn_msb, bx.type_vector(bx.type_i1(), in_len as _));
1269 // Bitcast <i1 x N> to iN:
1270 let i_ = bx.bitcast(i1xn, bx.type_ix(in_len as _));
1271 // Zero-extend iN to the bitmask type:
1272 return Ok(bx.zext(i_, bx.type_ix(expected_int_bits as _)));
1275 fn simd_simple_float_intrinsic(
1277 in_elem: &::rustc::ty::TyS<'_>,
1278 in_ty: &::rustc::ty::TyS<'_>,
1280 bx: &mut Builder<'a, 'll, 'tcx>,
1282 args: &[OperandRef<'tcx, &'ll Value>],
1283 ) -> Result<&'ll Value, ()> {
1284 macro_rules! emit_error {
1288 ($msg: tt, $($fmt: tt)*) => {
1289 span_invalid_monomorphization_error(
1291 &format!(concat!("invalid monomorphization of `{}` intrinsic: ", $msg),
1295 macro_rules! return_error {
1298 emit_error!($($fmt)*);
1303 let ety = match in_elem.sty {
1304 ty::Float(f) if f.bit_width() == 32 => {
1305 if in_len < 2 || in_len > 16 {
1307 "unsupported floating-point vector `{}` with length `{}` \
1308 out-of-range [2, 16]",
1313 ty::Float(f) if f.bit_width() == 64 => {
1314 if in_len < 2 || in_len > 8 {
1315 return_error!("unsupported floating-point vector `{}` with length `{}` \
1316 out-of-range [2, 8]",
1322 return_error!("unsupported element type `{}` of floating-point vector `{}`",
1326 return_error!("`{}` is not a floating-point type", in_ty);
1330 let llvm_name = &format!("llvm.{0}.v{1}{2}", name, in_len, ety);
1331 let intrinsic = bx.get_intrinsic(&llvm_name);
1332 let c = bx.call(intrinsic,
1333 &args.iter().map(|arg| arg.immediate()).collect::<Vec<_>>(),
1335 unsafe { llvm::LLVMRustSetHasUnsafeAlgebra(c) };
1341 return simd_simple_float_intrinsic("sqrt", in_elem, in_ty, in_len, bx, span, args);
1344 return simd_simple_float_intrinsic("sin", in_elem, in_ty, in_len, bx, span, args);
1347 return simd_simple_float_intrinsic("cos", in_elem, in_ty, in_len, bx, span, args);
1350 return simd_simple_float_intrinsic("fabs", in_elem, in_ty, in_len, bx, span, args);
1353 return simd_simple_float_intrinsic("floor", in_elem, in_ty, in_len, bx, span, args);
1356 return simd_simple_float_intrinsic("ceil", in_elem, in_ty, in_len, bx, span, args);
1359 return simd_simple_float_intrinsic("exp", in_elem, in_ty, in_len, bx, span, args);
1362 return simd_simple_float_intrinsic("exp2", in_elem, in_ty, in_len, bx, span, args);
1365 return simd_simple_float_intrinsic("log10", in_elem, in_ty, in_len, bx, span, args);
1368 return simd_simple_float_intrinsic("log2", in_elem, in_ty, in_len, bx, span, args);
1371 return simd_simple_float_intrinsic("log", in_elem, in_ty, in_len, bx, span, args);
1374 return simd_simple_float_intrinsic("powi", in_elem, in_ty, in_len, bx, span, args);
1377 return simd_simple_float_intrinsic("pow", in_elem, in_ty, in_len, bx, span, args);
1380 return simd_simple_float_intrinsic("fma", in_elem, in_ty, in_len, bx, span, args);
1382 _ => { /* fallthrough */ }
1386 // https://github.com/llvm-mirror/llvm/blob/master/include/llvm/IR/Function.h#L182
1387 // https://github.com/llvm-mirror/llvm/blob/master/include/llvm/IR/Intrinsics.h#L81
1388 fn llvm_vector_str(elem_ty: Ty<'_>, vec_len: usize, no_pointers: usize) -> String {
1389 let p0s: String = "p0".repeat(no_pointers);
1391 ty::Int(v) => format!("v{}{}i{}", vec_len, p0s, v.bit_width().unwrap()),
1392 ty::Uint(v) => format!("v{}{}i{}", vec_len, p0s, v.bit_width().unwrap()),
1393 ty::Float(v) => format!("v{}{}f{}", vec_len, p0s, v.bit_width()),
1394 _ => unreachable!(),
1398 fn llvm_vector_ty(cx: &CodegenCx<'ll, '_>, elem_ty: Ty<'_>, vec_len: usize,
1399 mut no_pointers: usize) -> &'ll Type {
1400 // FIXME: use cx.layout_of(ty).llvm_type() ?
1401 let mut elem_ty = match elem_ty.sty {
1402 ty::Int(v) => cx.type_int_from_ty( v),
1403 ty::Uint(v) => cx.type_uint_from_ty( v),
1404 ty::Float(v) => cx.type_float_from_ty( v),
1405 _ => unreachable!(),
1407 while no_pointers > 0 {
1408 elem_ty = cx.type_ptr_to(elem_ty);
1411 cx.type_vector(elem_ty, vec_len as u64)
1415 if name == "simd_gather" {
1416 // simd_gather(values: <N x T>, pointers: <N x *_ T>,
1417 // mask: <N x i{M}>) -> <N x T>
1418 // * N: number of elements in the input vectors
1419 // * T: type of the element to load
1420 // * M: any integer width is supported, will be truncated to i1
1422 // All types must be simd vector types
1423 require_simd!(in_ty, "first");
1424 require_simd!(arg_tys[1], "second");
1425 require_simd!(arg_tys[2], "third");
1426 require_simd!(ret_ty, "return");
1428 // Of the same length:
1429 require!(in_len == arg_tys[1].simd_size(tcx),
1430 "expected {} argument with length {} (same as input type `{}`), \
1431 found `{}` with length {}", "second", in_len, in_ty, arg_tys[1],
1432 arg_tys[1].simd_size(tcx));
1433 require!(in_len == arg_tys[2].simd_size(tcx),
1434 "expected {} argument with length {} (same as input type `{}`), \
1435 found `{}` with length {}", "third", in_len, in_ty, arg_tys[2],
1436 arg_tys[2].simd_size(tcx));
1438 // The return type must match the first argument type
1439 require!(ret_ty == in_ty,
1440 "expected return type `{}`, found `{}`",
1443 // This counts how many pointers
1444 fn ptr_count(t: Ty<'_>) -> usize {
1446 ty::RawPtr(p) => 1 + ptr_count(p.ty),
1452 fn non_ptr(t: Ty<'_>) -> Ty<'_> {
1454 ty::RawPtr(p) => non_ptr(p.ty),
1459 // The second argument must be a simd vector with an element type that's a pointer
1460 // to the element type of the first argument
1461 let (pointer_count, underlying_ty) = match arg_tys[1].simd_type(tcx).sty {
1462 ty::RawPtr(p) if p.ty == in_elem => (ptr_count(arg_tys[1].simd_type(tcx)),
1463 non_ptr(arg_tys[1].simd_type(tcx))),
1465 require!(false, "expected element type `{}` of second argument `{}` \
1466 to be a pointer to the element type `{}` of the first \
1467 argument `{}`, found `{}` != `*_ {}`",
1468 arg_tys[1].simd_type(tcx), arg_tys[1], in_elem, in_ty,
1469 arg_tys[1].simd_type(tcx), in_elem);
1473 assert!(pointer_count > 0);
1474 assert_eq!(pointer_count - 1, ptr_count(arg_tys[0].simd_type(tcx)));
1475 assert_eq!(underlying_ty, non_ptr(arg_tys[0].simd_type(tcx)));
1477 // The element type of the third argument must be a signed integer type of any width:
1478 match arg_tys[2].simd_type(tcx).sty {
1481 require!(false, "expected element type `{}` of third argument `{}` \
1482 to be a signed integer type",
1483 arg_tys[2].simd_type(tcx), arg_tys[2]);
1487 // Alignment of T, must be a constant integer value:
1488 let alignment_ty = bx.type_i32();
1489 let alignment = bx.const_i32(bx.align_of(in_elem).bytes() as i32);
1491 // Truncate the mask vector to a vector of i1s:
1492 let (mask, mask_ty) = {
1493 let i1 = bx.type_i1();
1494 let i1xn = bx.type_vector(i1, in_len as u64);
1495 (bx.trunc(args[2].immediate(), i1xn), i1xn)
1498 // Type of the vector of pointers:
1499 let llvm_pointer_vec_ty = llvm_vector_ty(bx, underlying_ty, in_len, pointer_count);
1500 let llvm_pointer_vec_str = llvm_vector_str(underlying_ty, in_len, pointer_count);
1502 // Type of the vector of elements:
1503 let llvm_elem_vec_ty = llvm_vector_ty(bx, underlying_ty, in_len, pointer_count - 1);
1504 let llvm_elem_vec_str = llvm_vector_str(underlying_ty, in_len, pointer_count - 1);
1506 let llvm_intrinsic = format!("llvm.masked.gather.{}.{}",
1507 llvm_elem_vec_str, llvm_pointer_vec_str);
1508 let f = bx.declare_cfn(&llvm_intrinsic,
1510 llvm_pointer_vec_ty,
1513 llvm_elem_vec_ty], llvm_elem_vec_ty));
1514 llvm::SetUnnamedAddr(f, false);
1515 let v = bx.call(f, &[args[1].immediate(), alignment, mask, args[0].immediate()],
1520 if name == "simd_scatter" {
1521 // simd_scatter(values: <N x T>, pointers: <N x *mut T>,
1522 // mask: <N x i{M}>) -> ()
1523 // * N: number of elements in the input vectors
1524 // * T: type of the element to load
1525 // * M: any integer width is supported, will be truncated to i1
1527 // All types must be simd vector types
1528 require_simd!(in_ty, "first");
1529 require_simd!(arg_tys[1], "second");
1530 require_simd!(arg_tys[2], "third");
1532 // Of the same length:
1533 require!(in_len == arg_tys[1].simd_size(tcx),
1534 "expected {} argument with length {} (same as input type `{}`), \
1535 found `{}` with length {}", "second", in_len, in_ty, arg_tys[1],
1536 arg_tys[1].simd_size(tcx));
1537 require!(in_len == arg_tys[2].simd_size(tcx),
1538 "expected {} argument with length {} (same as input type `{}`), \
1539 found `{}` with length {}", "third", in_len, in_ty, arg_tys[2],
1540 arg_tys[2].simd_size(tcx));
1542 // This counts how many pointers
1543 fn ptr_count(t: Ty<'_>) -> usize {
1545 ty::RawPtr(p) => 1 + ptr_count(p.ty),
1551 fn non_ptr(t: Ty<'_>) -> Ty<'_> {
1553 ty::RawPtr(p) => non_ptr(p.ty),
1558 // The second argument must be a simd vector with an element type that's a pointer
1559 // to the element type of the first argument
1560 let (pointer_count, underlying_ty) = match arg_tys[1].simd_type(tcx).sty {
1561 ty::RawPtr(p) if p.ty == in_elem && p.mutbl == hir::MutMutable
1562 => (ptr_count(arg_tys[1].simd_type(tcx)),
1563 non_ptr(arg_tys[1].simd_type(tcx))),
1565 require!(false, "expected element type `{}` of second argument `{}` \
1566 to be a pointer to the element type `{}` of the first \
1567 argument `{}`, found `{}` != `*mut {}`",
1568 arg_tys[1].simd_type(tcx), arg_tys[1], in_elem, in_ty,
1569 arg_tys[1].simd_type(tcx), in_elem);
1573 assert!(pointer_count > 0);
1574 assert_eq!(pointer_count - 1, ptr_count(arg_tys[0].simd_type(tcx)));
1575 assert_eq!(underlying_ty, non_ptr(arg_tys[0].simd_type(tcx)));
1577 // The element type of the third argument must be a signed integer type of any width:
1578 match arg_tys[2].simd_type(tcx).sty {
1581 require!(false, "expected element type `{}` of third argument `{}` \
1582 to be a signed integer type",
1583 arg_tys[2].simd_type(tcx), arg_tys[2]);
1587 // Alignment of T, must be a constant integer value:
1588 let alignment_ty = bx.type_i32();
1589 let alignment = bx.const_i32(bx.align_of(in_elem).bytes() as i32);
1591 // Truncate the mask vector to a vector of i1s:
1592 let (mask, mask_ty) = {
1593 let i1 = bx.type_i1();
1594 let i1xn = bx.type_vector(i1, in_len as u64);
1595 (bx.trunc(args[2].immediate(), i1xn), i1xn)
1598 let ret_t = bx.type_void();
1600 // Type of the vector of pointers:
1601 let llvm_pointer_vec_ty = llvm_vector_ty(bx, underlying_ty, in_len, pointer_count);
1602 let llvm_pointer_vec_str = llvm_vector_str(underlying_ty, in_len, pointer_count);
1604 // Type of the vector of elements:
1605 let llvm_elem_vec_ty = llvm_vector_ty(bx, underlying_ty, in_len, pointer_count - 1);
1606 let llvm_elem_vec_str = llvm_vector_str(underlying_ty, in_len, pointer_count - 1);
1608 let llvm_intrinsic = format!("llvm.masked.scatter.{}.{}",
1609 llvm_elem_vec_str, llvm_pointer_vec_str);
1610 let f = bx.declare_cfn(&llvm_intrinsic,
1611 bx.type_func(&[llvm_elem_vec_ty,
1612 llvm_pointer_vec_ty,
1615 llvm::SetUnnamedAddr(f, false);
1616 let v = bx.call(f, &[args[0].immediate(), args[1].immediate(), alignment, mask],
1621 macro_rules! arith_red {
1622 ($name:tt : $integer_reduce:ident, $float_reduce:ident, $ordered:expr) => {
1624 require!(ret_ty == in_elem,
1625 "expected return type `{}` (element of input `{}`), found `{}`",
1626 in_elem, in_ty, ret_ty);
1627 return match in_elem.sty {
1628 ty::Int(_) | ty::Uint(_) => {
1629 let r = bx.$integer_reduce(args[0].immediate());
1631 // if overflow occurs, the result is the
1632 // mathematical result modulo 2^n:
1633 if name.contains("mul") {
1634 Ok(bx.mul(args[1].immediate(), r))
1636 Ok(bx.add(args[1].immediate(), r))
1639 Ok(bx.$integer_reduce(args[0].immediate()))
1643 let acc = if $ordered {
1644 // ordered arithmetic reductions take an accumulator
1647 // unordered arithmetic reductions use the identity accumulator
1648 let identity_acc = if $name.contains("mul") { 1.0 } else { 0.0 };
1649 match f.bit_width() {
1650 32 => bx.const_real(bx.type_f32(), identity_acc),
1651 64 => bx.const_real(bx.type_f64(), identity_acc),
1654 unsupported {} from `{}` with element `{}` of size `{}` to `{}`"#,
1655 $name, in_ty, in_elem, v, ret_ty
1660 Ok(bx.$float_reduce(acc, args[0].immediate()))
1664 "unsupported {} from `{}` with element `{}` to `{}`",
1665 $name, in_ty, in_elem, ret_ty
1673 arith_red!("simd_reduce_add_ordered": vector_reduce_add, vector_reduce_fadd, true);
1674 arith_red!("simd_reduce_mul_ordered": vector_reduce_mul, vector_reduce_fmul, true);
1675 arith_red!("simd_reduce_add_unordered": vector_reduce_add, vector_reduce_fadd_fast, false);
1676 arith_red!("simd_reduce_mul_unordered": vector_reduce_mul, vector_reduce_fmul_fast, false);
1678 macro_rules! minmax_red {
1679 ($name:tt: $int_red:ident, $float_red:ident) => {
1681 require!(ret_ty == in_elem,
1682 "expected return type `{}` (element of input `{}`), found `{}`",
1683 in_elem, in_ty, ret_ty);
1684 return match in_elem.sty {
1686 Ok(bx.$int_red(args[0].immediate(), true))
1689 Ok(bx.$int_red(args[0].immediate(), false))
1692 Ok(bx.$float_red(args[0].immediate()))
1695 return_error!("unsupported {} from `{}` with element `{}` to `{}`",
1696 $name, in_ty, in_elem, ret_ty)
1704 minmax_red!("simd_reduce_min": vector_reduce_min, vector_reduce_fmin);
1705 minmax_red!("simd_reduce_max": vector_reduce_max, vector_reduce_fmax);
1707 minmax_red!("simd_reduce_min_nanless": vector_reduce_min, vector_reduce_fmin_fast);
1708 minmax_red!("simd_reduce_max_nanless": vector_reduce_max, vector_reduce_fmax_fast);
1710 macro_rules! bitwise_red {
1711 ($name:tt : $red:ident, $boolean:expr) => {
1713 let input = if !$boolean {
1714 require!(ret_ty == in_elem,
1715 "expected return type `{}` (element of input `{}`), found `{}`",
1716 in_elem, in_ty, ret_ty);
1720 ty::Int(_) | ty::Uint(_) => {},
1722 return_error!("unsupported {} from `{}` with element `{}` to `{}`",
1723 $name, in_ty, in_elem, ret_ty)
1727 // boolean reductions operate on vectors of i1s:
1728 let i1 = bx.type_i1();
1729 let i1xn = bx.type_vector(i1, in_len as u64);
1730 bx.trunc(args[0].immediate(), i1xn)
1732 return match in_elem.sty {
1733 ty::Int(_) | ty::Uint(_) => {
1734 let r = bx.$red(input);
1739 bx.zext(r, bx.type_bool())
1744 return_error!("unsupported {} from `{}` with element `{}` to `{}`",
1745 $name, in_ty, in_elem, ret_ty)
1752 bitwise_red!("simd_reduce_and": vector_reduce_and, false);
1753 bitwise_red!("simd_reduce_or": vector_reduce_or, false);
1754 bitwise_red!("simd_reduce_xor": vector_reduce_xor, false);
1755 bitwise_red!("simd_reduce_all": vector_reduce_and, true);
1756 bitwise_red!("simd_reduce_any": vector_reduce_or, true);
1758 if name == "simd_cast" {
1759 require_simd!(ret_ty, "return");
1760 let out_len = ret_ty.simd_size(tcx);
1761 require!(in_len == out_len,
1762 "expected return type with length {} (same as input type `{}`), \
1763 found `{}` with length {}",
1766 // casting cares about nominal type, not just structural type
1767 let out_elem = ret_ty.simd_type(tcx);
1769 if in_elem == out_elem { return Ok(args[0].immediate()); }
1771 enum Style { Float, Int(/* is signed? */ bool), Unsupported }
1773 let (in_style, in_width) = match in_elem.sty {
1774 // vectors of pointer-sized integers should've been
1775 // disallowed before here, so this unwrap is safe.
1776 ty::Int(i) => (Style::Int(true), i.bit_width().unwrap()),
1777 ty::Uint(u) => (Style::Int(false), u.bit_width().unwrap()),
1778 ty::Float(f) => (Style::Float, f.bit_width()),
1779 _ => (Style::Unsupported, 0)
1781 let (out_style, out_width) = match out_elem.sty {
1782 ty::Int(i) => (Style::Int(true), i.bit_width().unwrap()),
1783 ty::Uint(u) => (Style::Int(false), u.bit_width().unwrap()),
1784 ty::Float(f) => (Style::Float, f.bit_width()),
1785 _ => (Style::Unsupported, 0)
1788 match (in_style, out_style) {
1789 (Style::Int(in_is_signed), Style::Int(_)) => {
1790 return Ok(match in_width.cmp(&out_width) {
1791 Ordering::Greater => bx.trunc(args[0].immediate(), llret_ty),
1792 Ordering::Equal => args[0].immediate(),
1793 Ordering::Less => if in_is_signed {
1794 bx.sext(args[0].immediate(), llret_ty)
1796 bx.zext(args[0].immediate(), llret_ty)
1800 (Style::Int(in_is_signed), Style::Float) => {
1801 return Ok(if in_is_signed {
1802 bx.sitofp(args[0].immediate(), llret_ty)
1804 bx.uitofp(args[0].immediate(), llret_ty)
1807 (Style::Float, Style::Int(out_is_signed)) => {
1808 return Ok(if out_is_signed {
1809 bx.fptosi(args[0].immediate(), llret_ty)
1811 bx.fptoui(args[0].immediate(), llret_ty)
1814 (Style::Float, Style::Float) => {
1815 return Ok(match in_width.cmp(&out_width) {
1816 Ordering::Greater => bx.fptrunc(args[0].immediate(), llret_ty),
1817 Ordering::Equal => args[0].immediate(),
1818 Ordering::Less => bx.fpext(args[0].immediate(), llret_ty)
1821 _ => {/* Unsupported. Fallthrough. */}
1824 "unsupported cast from `{}` with element `{}` to `{}` with element `{}`",
1828 macro_rules! arith {
1829 ($($name: ident: $($($p: ident),* => $call: ident),*;)*) => {
1830 $(if name == stringify!($name) {
1832 $($(ty::$p(_))|* => {
1833 return Ok(bx.$call(args[0].immediate(), args[1].immediate()))
1838 "unsupported operation on `{}` with element `{}`",
1845 simd_add: Uint, Int => add, Float => fadd;
1846 simd_sub: Uint, Int => sub, Float => fsub;
1847 simd_mul: Uint, Int => mul, Float => fmul;
1848 simd_div: Uint => udiv, Int => sdiv, Float => fdiv;
1849 simd_rem: Uint => urem, Int => srem, Float => frem;
1850 simd_shl: Uint, Int => shl;
1851 simd_shr: Uint => lshr, Int => ashr;
1852 simd_and: Uint, Int => and;
1853 simd_or: Uint, Int => or;
1854 simd_xor: Uint, Int => xor;
1855 simd_fmax: Float => maxnum;
1856 simd_fmin: Float => minnum;
1860 if name == "simd_saturating_add" || name == "simd_saturating_sub" {
1861 let lhs = args[0].immediate();
1862 let rhs = args[1].immediate();
1863 let is_add = name == "simd_saturating_add";
1864 let ptr_bits = bx.tcx().data_layout.pointer_size.bits() as _;
1865 let (signed, elem_width, elem_ty) = match in_elem.sty {
1869 i.bit_width().unwrap_or(ptr_bits),
1870 bx.cx.type_int_from_ty(i)
1875 i.bit_width().unwrap_or(ptr_bits),
1876 bx.cx.type_uint_from_ty(i)
1880 "expected element type `{}` of vector type `{}` \
1881 to be a signed or unsigned integer type",
1882 arg_tys[0].simd_type(tcx), arg_tys[0]
1886 let llvm_intrinsic = &format!(
1887 "llvm.{}{}.sat.v{}i{}",
1888 if signed { 's' } else { 'u' },
1889 if is_add { "add" } else { "sub" },
1892 let vec_ty = bx.cx.type_vector(elem_ty, in_len as u64);
1894 let f = bx.declare_cfn(
1896 bx.type_func(&[vec_ty, vec_ty], vec_ty)
1898 llvm::SetUnnamedAddr(f, false);
1899 let v = bx.call(f, &[lhs, rhs], None);
1903 span_bug!(span, "unknown SIMD intrinsic");
1906 // Returns the width of an int Ty, and if it's signed or not
1907 // Returns None if the type is not an integer
1908 // FIXME: there’s multiple of this functions, investigate using some of the already existing
1910 fn int_type_width_signed(ty: Ty<'_>, cx: &CodegenCx<'_, '_>) -> Option<(u64, bool)> {
1912 ty::Int(t) => Some((match t {
1913 ast::IntTy::Isize => cx.tcx.sess.target.isize_ty.bit_width().unwrap() as u64,
1914 ast::IntTy::I8 => 8,
1915 ast::IntTy::I16 => 16,
1916 ast::IntTy::I32 => 32,
1917 ast::IntTy::I64 => 64,
1918 ast::IntTy::I128 => 128,
1920 ty::Uint(t) => Some((match t {
1921 ast::UintTy::Usize => cx.tcx.sess.target.usize_ty.bit_width().unwrap() as u64,
1922 ast::UintTy::U8 => 8,
1923 ast::UintTy::U16 => 16,
1924 ast::UintTy::U32 => 32,
1925 ast::UintTy::U64 => 64,
1926 ast::UintTy::U128 => 128,
1932 // Returns the width of a float Ty
1933 // Returns None if the type is not a float
1934 fn float_type_width(ty: Ty<'_>) -> Option<u64> {
1936 ty::Float(t) => Some(t.bit_width() as u64),