1 use crate::abi::{Abi, FnAbi, LlvmType, PassMode};
2 use crate::builder::Builder;
3 use crate::context::CodegenCx;
6 use crate::type_::Type;
7 use crate::type_of::LayoutLlvmExt;
8 use crate::va_arg::emit_va_arg;
9 use crate::value::Value;
10 use rustc::ty::layout::{self, FnAbiExt, HasTyCtxt, LayoutOf, Primitive};
11 use rustc::ty::{self, Ty};
12 use rustc::{bug, span_bug};
13 use rustc_codegen_ssa::base::{compare_simd_types, to_immediate, wants_msvc_seh};
14 use rustc_codegen_ssa::common::{IntPredicate, TypeKind};
15 use rustc_codegen_ssa::glue;
16 use rustc_codegen_ssa::mir::operand::{OperandRef, OperandValue};
17 use rustc_codegen_ssa::mir::place::PlaceRef;
18 use rustc_codegen_ssa::MemFlags;
20 use rustc_target::abi::HasDataLayout;
23 use rustc_codegen_ssa::common::span_invalid_monomorphization_error;
24 use rustc_codegen_ssa::traits::*;
28 use std::cmp::Ordering;
29 use std::{i128, iter, u128};
31 fn get_simple_intrinsic(cx: &CodegenCx<'ll, '_>, name: &str) -> Option<&'ll Value> {
32 let llvm_name = match name {
33 "sqrtf32" => "llvm.sqrt.f32",
34 "sqrtf64" => "llvm.sqrt.f64",
35 "powif32" => "llvm.powi.f32",
36 "powif64" => "llvm.powi.f64",
37 "sinf32" => "llvm.sin.f32",
38 "sinf64" => "llvm.sin.f64",
39 "cosf32" => "llvm.cos.f32",
40 "cosf64" => "llvm.cos.f64",
41 "powf32" => "llvm.pow.f32",
42 "powf64" => "llvm.pow.f64",
43 "expf32" => "llvm.exp.f32",
44 "expf64" => "llvm.exp.f64",
45 "exp2f32" => "llvm.exp2.f32",
46 "exp2f64" => "llvm.exp2.f64",
47 "logf32" => "llvm.log.f32",
48 "logf64" => "llvm.log.f64",
49 "log10f32" => "llvm.log10.f32",
50 "log10f64" => "llvm.log10.f64",
51 "log2f32" => "llvm.log2.f32",
52 "log2f64" => "llvm.log2.f64",
53 "fmaf32" => "llvm.fma.f32",
54 "fmaf64" => "llvm.fma.f64",
55 "fabsf32" => "llvm.fabs.f32",
56 "fabsf64" => "llvm.fabs.f64",
57 "minnumf32" => "llvm.minnum.f32",
58 "minnumf64" => "llvm.minnum.f64",
59 "maxnumf32" => "llvm.maxnum.f32",
60 "maxnumf64" => "llvm.maxnum.f64",
61 "copysignf32" => "llvm.copysign.f32",
62 "copysignf64" => "llvm.copysign.f64",
63 "floorf32" => "llvm.floor.f32",
64 "floorf64" => "llvm.floor.f64",
65 "ceilf32" => "llvm.ceil.f32",
66 "ceilf64" => "llvm.ceil.f64",
67 "truncf32" => "llvm.trunc.f32",
68 "truncf64" => "llvm.trunc.f64",
69 "rintf32" => "llvm.rint.f32",
70 "rintf64" => "llvm.rint.f64",
71 "nearbyintf32" => "llvm.nearbyint.f32",
72 "nearbyintf64" => "llvm.nearbyint.f64",
73 "roundf32" => "llvm.round.f32",
74 "roundf64" => "llvm.round.f64",
75 "assume" => "llvm.assume",
76 "abort" => "llvm.trap",
79 Some(cx.get_intrinsic(&llvm_name))
82 impl IntrinsicCallMethods<'tcx> for Builder<'a, 'll, 'tcx> {
83 fn codegen_intrinsic_call(
85 instance: ty::Instance<'tcx>,
86 fn_abi: &FnAbi<'tcx, Ty<'tcx>>,
87 args: &[OperandRef<'tcx, &'ll Value>],
92 let callee_ty = instance.monomorphic_ty(tcx);
94 let (def_id, substs) = match callee_ty.kind {
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_abi.ret.layout);
108 let simple = get_simple_intrinsic(self, name);
109 let llval = match name {
110 _ if simple.is_some() => self.call(
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)
137 let llfn = self.get_intrinsic(&("llvm.debugtrap"));
138 self.call(llfn, &[], None)
140 "va_start" => self.va_start(args[0].immediate()),
141 "va_end" => self.va_end(args[0].immediate()),
143 let intrinsic = self.cx().get_intrinsic(&("llvm.va_copy"));
144 self.call(intrinsic, &[args[0].immediate(), args[1].immediate()], None)
147 match fn_abi.ret.layout.abi {
148 layout::Abi::Scalar(ref scalar) => {
150 Primitive::Int(..) => {
151 if self.cx().size_of(ret_ty).bytes() < 4 {
152 // `va_arg` should not be called on a integer type
153 // less than 4 bytes in length. If it is, promote
154 // the integer to a `i32` and truncate the result
155 // back to the smaller type.
156 let promoted_result = emit_va_arg(self, args[0], tcx.types.i32);
157 self.trunc(promoted_result, llret_ty)
159 emit_va_arg(self, args[0], ret_ty)
162 Primitive::F64 | Primitive::Pointer => {
163 emit_va_arg(self, args[0], ret_ty)
165 // `va_arg` should never be used with the return type f32.
166 Primitive::F32 => bug!("the va_arg intrinsic does not work with `f32`"),
169 _ => bug!("the va_arg intrinsic does not work with non-scalar types"),
173 let tp_ty = substs.type_at(0);
174 if let OperandValue::Pair(_, meta) = args[0].val {
175 let (llsize, _) = glue::size_and_align_of_dst(self, tp_ty, Some(meta));
178 self.const_usize(self.size_of(tp_ty).bytes())
181 "min_align_of_val" => {
182 let tp_ty = substs.type_at(0);
183 if let OperandValue::Pair(_, meta) = args[0].val {
184 let (_, llalign) = glue::size_and_align_of_dst(self, tp_ty, Some(meta));
187 self.const_usize(self.align_of(tp_ty).bytes())
190 "size_of" | "pref_align_of" | "min_align_of" | "needs_drop" | "type_id"
194 .const_eval_instance(ty::ParamEnv::reveal_all(), instance, None)
196 OperandRef::from_const(self, ty_name).immediate_or_packed_pair(self)
199 let ty = substs.type_at(0);
200 if !self.layout_of(ty).is_zst() {
201 // Just zero out the stack slot.
202 // If we store a zero constant, LLVM will drown in vreg allocation for large
203 // data structures, and the generated code will be awful. (A telltale sign of
204 // this is large quantities of `mov [byte ptr foo],0` in the generated code.)
216 // Effectively no-ops
217 "uninit" | "forget" => {
221 let ptr = args[0].immediate();
222 let offset = args[1].immediate();
223 self.inbounds_gep(ptr, &[offset])
226 let ptr = args[0].immediate();
227 let offset = args[1].immediate();
228 self.gep(ptr, &[offset])
231 "copy_nonoverlapping" => {
267 "volatile_copy_nonoverlapping_memory" => {
279 "volatile_copy_memory" => {
291 "volatile_set_memory" => {
302 "volatile_load" | "unaligned_volatile_load" => {
303 let tp_ty = substs.type_at(0);
304 let mut ptr = args[0].immediate();
305 if let PassMode::Cast(ty) = fn_abi.ret.mode {
306 ptr = self.pointercast(ptr, self.type_ptr_to(ty.llvm_type(self)));
308 let load = self.volatile_load(ptr);
309 let align = if name == "unaligned_volatile_load" {
312 self.align_of(tp_ty).bytes() as u32
315 llvm::LLVMSetAlignment(load, align);
317 to_immediate(self, load, self.layout_of(tp_ty))
319 "volatile_store" => {
320 let dst = args[0].deref(self.cx());
321 args[1].val.volatile_store(self, dst);
324 "unaligned_volatile_store" => {
325 let dst = args[0].deref(self.cx());
326 args[1].val.unaligned_volatile_store(self, dst);
330 | "prefetch_write_data"
331 | "prefetch_read_instruction"
332 | "prefetch_write_instruction" => {
333 let expect = self.get_intrinsic(&("llvm.prefetch"));
334 let (rw, cache_type) = match name {
335 "prefetch_read_data" => (0, 1),
336 "prefetch_write_data" => (1, 1),
337 "prefetch_read_instruction" => (0, 0),
338 "prefetch_write_instruction" => (1, 0),
347 self.const_i32(cache_type),
352 "ctlz" | "ctlz_nonzero" | "cttz" | "cttz_nonzero" | "ctpop" | "bswap"
353 | "bitreverse" | "add_with_overflow" | "sub_with_overflow" | "mul_with_overflow"
354 | "wrapping_add" | "wrapping_sub" | "wrapping_mul" | "unchecked_div"
355 | "unchecked_rem" | "unchecked_shl" | "unchecked_shr" | "unchecked_add"
356 | "unchecked_sub" | "unchecked_mul" | "exact_div" | "rotate_left" | "rotate_right"
357 | "saturating_add" | "saturating_sub" => {
359 match int_type_width_signed(ty, self) {
360 Some((width, signed)) => match name {
362 let y = self.const_bool(false);
363 let llfn = self.get_intrinsic(&format!("llvm.{}.i{}", name, width));
364 self.call(llfn, &[args[0].immediate(), y], None)
366 "ctlz_nonzero" | "cttz_nonzero" => {
367 let y = self.const_bool(true);
368 let llvm_name = &format!("llvm.{}.i{}", &name[..4], width);
369 let llfn = self.get_intrinsic(llvm_name);
370 self.call(llfn, &[args[0].immediate(), y], None)
372 "ctpop" => self.call(
373 self.get_intrinsic(&format!("llvm.ctpop.i{}", width)),
374 &[args[0].immediate()],
379 args[0].immediate() // byte swap a u8/i8 is just a no-op
382 self.get_intrinsic(&format!("llvm.bswap.i{}", width)),
383 &[args[0].immediate()],
388 "bitreverse" => self.call(
389 self.get_intrinsic(&format!("llvm.bitreverse.i{}", width)),
390 &[args[0].immediate()],
393 "add_with_overflow" | "sub_with_overflow" | "mul_with_overflow" => {
394 let intrinsic = format!(
395 "llvm.{}{}.with.overflow.i{}",
396 if signed { 's' } else { 'u' },
400 let llfn = self.get_intrinsic(&intrinsic);
402 // Convert `i1` to a `bool`, and write it to the out parameter
404 self.call(llfn, &[args[0].immediate(), args[1].immediate()], None);
405 let val = self.extract_value(pair, 0);
406 let overflow = self.extract_value(pair, 1);
407 let overflow = self.zext(overflow, self.type_bool());
409 let dest = result.project_field(self, 0);
410 self.store(val, dest.llval, dest.align);
411 let dest = result.project_field(self, 1);
412 self.store(overflow, dest.llval, dest.align);
416 "wrapping_add" => self.add(args[0].immediate(), args[1].immediate()),
417 "wrapping_sub" => self.sub(args[0].immediate(), args[1].immediate()),
418 "wrapping_mul" => self.mul(args[0].immediate(), args[1].immediate()),
421 self.exactsdiv(args[0].immediate(), args[1].immediate())
423 self.exactudiv(args[0].immediate(), args[1].immediate())
428 self.sdiv(args[0].immediate(), args[1].immediate())
430 self.udiv(args[0].immediate(), args[1].immediate())
435 self.srem(args[0].immediate(), args[1].immediate())
437 self.urem(args[0].immediate(), args[1].immediate())
440 "unchecked_shl" => self.shl(args[0].immediate(), args[1].immediate()),
443 self.ashr(args[0].immediate(), args[1].immediate())
445 self.lshr(args[0].immediate(), args[1].immediate())
450 self.unchecked_sadd(args[0].immediate(), args[1].immediate())
452 self.unchecked_uadd(args[0].immediate(), args[1].immediate())
457 self.unchecked_ssub(args[0].immediate(), args[1].immediate())
459 self.unchecked_usub(args[0].immediate(), args[1].immediate())
464 self.unchecked_smul(args[0].immediate(), args[1].immediate())
466 self.unchecked_umul(args[0].immediate(), args[1].immediate())
469 "rotate_left" | "rotate_right" => {
470 let is_left = name == "rotate_left";
471 let val = args[0].immediate();
472 let raw_shift = args[1].immediate();
473 // rotate = funnel shift with first two args the same
475 &format!("llvm.fsh{}.i{}", if is_left { 'l' } else { 'r' }, width);
476 let llfn = self.get_intrinsic(llvm_name);
477 self.call(llfn, &[val, val, raw_shift], None)
479 "saturating_add" | "saturating_sub" => {
480 let is_add = name == "saturating_add";
481 let lhs = args[0].immediate();
482 let rhs = args[1].immediate();
483 if llvm_util::get_major_version() >= 8 {
484 let llvm_name = &format!(
486 if signed { 's' } else { 'u' },
487 if is_add { "add" } else { "sub" },
490 let llfn = self.get_intrinsic(llvm_name);
491 self.call(llfn, &[lhs, rhs], None)
493 let llvm_name = &format!(
494 "llvm.{}{}.with.overflow.i{}",
495 if signed { 's' } else { 'u' },
496 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 {
507 .const_uint_big(llty, (i128::MIN >> (128 - width)) as u128);
509 .const_uint_big(llty, (i128::MAX >> (128 - width)) as u128);
511 IntPredicate::IntSLT,
513 self.const_uint(llty, 0),
515 self.select(neg, limit_hi, limit_lo)
517 self.const_uint_big(llty, u128::MAX >> (128 - width))
519 self.const_uint(llty, 0)
521 self.select(overflow, limit, val)
527 span_invalid_monomorphization_error(
531 "invalid monomorphization of `{}` intrinsic: \
532 expected basic integer type, found `{}`",
540 "fadd_fast" | "fsub_fast" | "fmul_fast" | "fdiv_fast" | "frem_fast" => {
541 match float_type_width(arg_tys[0]) {
542 Some(_width) => match name {
543 "fadd_fast" => self.fadd_fast(args[0].immediate(), args[1].immediate()),
544 "fsub_fast" => self.fsub_fast(args[0].immediate(), args[1].immediate()),
545 "fmul_fast" => self.fmul_fast(args[0].immediate(), args[1].immediate()),
546 "fdiv_fast" => self.fdiv_fast(args[0].immediate(), args[1].immediate()),
547 "frem_fast" => self.frem_fast(args[0].immediate(), args[1].immediate()),
551 span_invalid_monomorphization_error(
555 "invalid monomorphization of `{}` intrinsic: \
556 expected basic float type, found `{}`",
565 "float_to_int_approx_unchecked" => {
566 if float_type_width(arg_tys[0]).is_none() {
567 span_invalid_monomorphization_error(
571 "invalid monomorphization of `float_to_int_approx_unchecked` \
572 intrinsic: expected basic float type, \
579 match int_type_width_signed(ret_ty, self.cx) {
580 Some((width, signed)) => {
582 self.fptosi(args[0].immediate(), self.cx.type_ix(width))
584 self.fptoui(args[0].immediate(), self.cx.type_ix(width))
588 span_invalid_monomorphization_error(
592 "invalid monomorphization of `float_to_int_approx_unchecked` \
593 intrinsic: expected basic integer type, \
603 "discriminant_value" => args[0].deref(self.cx()).codegen_get_discr(self, ret_ty),
605 name if name.starts_with("simd_") => {
606 match generic_simd_intrinsic(self, name, callee_ty, args, ret_ty, llret_ty, span) {
611 // This requires that atomic intrinsics follow a specific naming pattern:
612 // "atomic_<operation>[_<ordering>]", and no ordering means SeqCst
613 name if name.starts_with("atomic_") => {
614 use rustc_codegen_ssa::common::AtomicOrdering::*;
615 use rustc_codegen_ssa::common::{AtomicRmwBinOp, SynchronizationScope};
617 let split: Vec<&str> = name.split('_').collect();
619 let is_cxchg = split[1] == "cxchg" || split[1] == "cxchgweak";
620 let (order, failorder) = match split.len() {
621 2 => (SequentiallyConsistent, SequentiallyConsistent),
622 3 => match split[2] {
623 "unordered" => (Unordered, Unordered),
624 "relaxed" => (Monotonic, Monotonic),
625 "acq" => (Acquire, Acquire),
626 "rel" => (Release, Monotonic),
627 "acqrel" => (AcquireRelease, Acquire),
628 "failrelaxed" if is_cxchg => (SequentiallyConsistent, Monotonic),
629 "failacq" if is_cxchg => (SequentiallyConsistent, Acquire),
630 _ => self.sess().fatal("unknown ordering in atomic intrinsic"),
632 4 => match (split[2], split[3]) {
633 ("acq", "failrelaxed") if is_cxchg => (Acquire, Monotonic),
634 ("acqrel", "failrelaxed") if is_cxchg => (AcquireRelease, Monotonic),
635 _ => self.sess().fatal("unknown ordering in atomic intrinsic"),
637 _ => self.sess().fatal("Atomic intrinsic not in correct format"),
640 let invalid_monomorphization = |ty| {
641 span_invalid_monomorphization_error(
645 "invalid monomorphization of `{}` intrinsic: \
646 expected basic integer type, found `{}`",
653 "cxchg" | "cxchgweak" => {
654 let ty = substs.type_at(0);
655 if int_type_width_signed(ty, self).is_some() {
656 let weak = split[1] == "cxchgweak";
657 let pair = self.atomic_cmpxchg(
665 let val = self.extract_value(pair, 0);
666 let success = self.extract_value(pair, 1);
667 let success = self.zext(success, self.type_bool());
669 let dest = result.project_field(self, 0);
670 self.store(val, dest.llval, dest.align);
671 let dest = result.project_field(self, 1);
672 self.store(success, dest.llval, dest.align);
675 return invalid_monomorphization(ty);
680 let ty = substs.type_at(0);
681 if int_type_width_signed(ty, self).is_some() {
682 let size = self.size_of(ty);
683 self.atomic_load(args[0].immediate(), order, size)
685 return invalid_monomorphization(ty);
690 let ty = substs.type_at(0);
691 if int_type_width_signed(ty, self).is_some() {
692 let size = self.size_of(ty);
701 return invalid_monomorphization(ty);
706 self.atomic_fence(order, SynchronizationScope::CrossThread);
710 "singlethreadfence" => {
711 self.atomic_fence(order, SynchronizationScope::SingleThread);
715 // These are all AtomicRMW ops
717 let atom_op = match op {
718 "xchg" => AtomicRmwBinOp::AtomicXchg,
719 "xadd" => AtomicRmwBinOp::AtomicAdd,
720 "xsub" => AtomicRmwBinOp::AtomicSub,
721 "and" => AtomicRmwBinOp::AtomicAnd,
722 "nand" => AtomicRmwBinOp::AtomicNand,
723 "or" => AtomicRmwBinOp::AtomicOr,
724 "xor" => AtomicRmwBinOp::AtomicXor,
725 "max" => AtomicRmwBinOp::AtomicMax,
726 "min" => AtomicRmwBinOp::AtomicMin,
727 "umax" => AtomicRmwBinOp::AtomicUMax,
728 "umin" => AtomicRmwBinOp::AtomicUMin,
729 _ => self.sess().fatal("unknown atomic operation"),
732 let ty = substs.type_at(0);
733 if int_type_width_signed(ty, self).is_some() {
741 return invalid_monomorphization(ty);
747 "nontemporal_store" => {
748 let dst = args[0].deref(self.cx());
749 args[1].val.nontemporal_store(self, dst);
753 "ptr_offset_from" => {
754 let ty = substs.type_at(0);
755 let pointee_size = self.size_of(ty);
757 // This is the same sequence that Clang emits for pointer subtraction.
758 // It can be neither `nsw` nor `nuw` because the input is treated as
759 // unsigned but then the output is treated as signed, so neither works.
760 let a = args[0].immediate();
761 let b = args[1].immediate();
762 let a = self.ptrtoint(a, self.type_isize());
763 let b = self.ptrtoint(b, self.type_isize());
764 let d = self.sub(a, b);
765 let pointee_size = self.const_usize(pointee_size.bytes());
766 // this is where the signed magic happens (notice the `s` in `exactsdiv`)
767 self.exactsdiv(d, pointee_size)
770 _ => bug!("unknown intrinsic '{}'", name),
773 if !fn_abi.ret.is_ignore() {
774 if let PassMode::Cast(ty) = fn_abi.ret.mode {
775 let ptr_llty = self.type_ptr_to(ty.llvm_type(self));
776 let ptr = self.pointercast(result.llval, ptr_llty);
777 self.store(llval, ptr, result.align);
779 OperandRef::from_immediate_or_packed_pair(self, llval, result.layout)
781 .store(self, result);
786 fn abort(&mut self) {
787 let fnname = self.get_intrinsic(&("llvm.trap"));
788 self.call(fnname, &[], None);
791 fn assume(&mut self, val: Self::Value) {
792 let assume_intrinsic = self.get_intrinsic("llvm.assume");
793 self.call(assume_intrinsic, &[val], None);
796 fn expect(&mut self, cond: Self::Value, expected: bool) -> Self::Value {
797 let expect = self.get_intrinsic(&"llvm.expect.i1");
798 self.call(expect, &[cond, self.const_bool(expected)], None)
801 fn sideeffect(&mut self) {
802 if self.tcx.sess.opts.debugging_opts.insert_sideeffect {
803 let fnname = self.get_intrinsic(&("llvm.sideeffect"));
804 self.call(fnname, &[], None);
808 fn va_start(&mut self, va_list: &'ll Value) -> &'ll Value {
809 let intrinsic = self.cx().get_intrinsic("llvm.va_start");
810 self.call(intrinsic, &[va_list], None)
813 fn va_end(&mut self, va_list: &'ll Value) -> &'ll Value {
814 let intrinsic = self.cx().get_intrinsic("llvm.va_end");
815 self.call(intrinsic, &[va_list], None)
820 bx: &mut Builder<'a, 'll, 'tcx>,
828 let (size, align) = bx.size_and_align_of(ty);
829 let size = bx.mul(bx.const_usize(size.bytes()), count);
830 let flags = if volatile { MemFlags::VOLATILE } else { MemFlags::empty() };
832 bx.memmove(dst, align, src, align, size, flags);
834 bx.memcpy(dst, align, src, align, size, flags);
839 bx: &mut Builder<'a, 'll, 'tcx>,
846 let (size, align) = bx.size_and_align_of(ty);
847 let size = bx.mul(bx.const_usize(size.bytes()), count);
848 let flags = if volatile { MemFlags::VOLATILE } else { MemFlags::empty() };
849 bx.memset(dst, val, size, align, flags);
853 bx: &mut Builder<'a, 'll, 'tcx>,
856 local_ptr: &'ll Value,
859 if bx.sess().no_landing_pads() {
860 bx.call(func, &[data], None);
861 let ptr_align = bx.tcx().data_layout.pointer_align.abi;
862 bx.store(bx.const_null(bx.type_i8p()), dest, ptr_align);
863 } else if wants_msvc_seh(bx.sess()) {
864 codegen_msvc_try(bx, func, data, local_ptr, dest);
866 codegen_gnu_try(bx, func, data, local_ptr, dest);
870 // MSVC's definition of the `rust_try` function.
872 // This implementation uses the new exception handling instructions in LLVM
873 // which have support in LLVM for SEH on MSVC targets. Although these
874 // instructions are meant to work for all targets, as of the time of this
875 // writing, however, LLVM does not recommend the usage of these new instructions
876 // as the old ones are still more optimized.
878 bx: &mut Builder<'a, 'll, 'tcx>,
881 local_ptr: &'ll Value,
884 let llfn = get_rust_try_fn(bx, &mut |mut bx| {
885 bx.set_personality_fn(bx.eh_personality());
888 let mut normal = bx.build_sibling_block("normal");
889 let mut catchswitch = bx.build_sibling_block("catchswitch");
890 let mut catchpad = bx.build_sibling_block("catchpad");
891 let mut caught = bx.build_sibling_block("caught");
893 let func = llvm::get_param(bx.llfn(), 0);
894 let data = llvm::get_param(bx.llfn(), 1);
895 let local_ptr = llvm::get_param(bx.llfn(), 2);
897 // We're generating an IR snippet that looks like:
899 // declare i32 @rust_try(%func, %data, %ptr) {
900 // %slot = alloca [2 x i64]
901 // invoke %func(%data) to label %normal unwind label %catchswitch
907 // %cs = catchswitch within none [%catchpad] unwind to caller
910 // %tok = catchpad within %cs [%type_descriptor, 0, %slot]
911 // %ptr[0] = %slot[0]
912 // %ptr[1] = %slot[1]
913 // catchret from %tok to label %caught
919 // This structure follows the basic usage of throw/try/catch in LLVM.
920 // For example, compile this C++ snippet to see what LLVM generates:
922 // #include <stdint.h>
924 // struct rust_panic {
928 // int bar(void (*foo)(void), uint64_t *ret) {
932 // } catch(rust_panic a) {
939 // More information can be found in libstd's seh.rs implementation.
940 let i64_2 = bx.type_array(bx.type_i64(), 2);
941 let i64_align = bx.tcx().data_layout.i64_align.abi;
942 let slot = bx.alloca(i64_2, i64_align);
943 bx.invoke(func, &[data], normal.llbb(), catchswitch.llbb(), None);
945 normal.ret(bx.const_i32(0));
947 let cs = catchswitch.catch_switch(None, None, 1);
948 catchswitch.add_handler(cs, catchpad.llbb());
950 let tydesc = match bx.tcx().lang_items().eh_catch_typeinfo() {
951 Some(did) => bx.get_static(did),
952 None => bug!("eh_catch_typeinfo not defined, but needed for SEH unwinding"),
954 let funclet = catchpad.catch_pad(cs, &[tydesc, bx.const_i32(0), slot]);
956 let payload = catchpad.load(slot, i64_align);
957 let local_ptr = catchpad.bitcast(local_ptr, bx.type_ptr_to(i64_2));
958 catchpad.store(payload, local_ptr, i64_align);
959 catchpad.catch_ret(&funclet, caught.llbb());
961 caught.ret(bx.const_i32(1));
964 // Note that no invoke is used here because by definition this function
965 // can't panic (that's what it's catching).
966 let ret = bx.call(llfn, &[func, data, local_ptr], None);
967 let i32_align = bx.tcx().data_layout.i32_align.abi;
968 bx.store(ret, dest, i32_align);
971 // Definition of the standard `try` function for Rust using the GNU-like model
972 // of exceptions (e.g., the normal semantics of LLVM's `landingpad` and `invoke`
975 // This codegen is a little surprising because we always call a shim
976 // function instead of inlining the call to `invoke` manually here. This is done
977 // because in LLVM we're only allowed to have one personality per function
978 // definition. The call to the `try` intrinsic is being inlined into the
979 // function calling it, and that function may already have other personality
980 // functions in play. By calling a shim we're guaranteed that our shim will have
981 // the right personality function.
983 bx: &mut Builder<'a, 'll, 'tcx>,
986 local_ptr: &'ll Value,
989 let llfn = get_rust_try_fn(bx, &mut |mut bx| {
990 // Codegens the shims described above:
993 // invoke %func(%args...) normal %normal unwind %catch
999 // (ptr, _) = landingpad
1000 // store ptr, %local_ptr
1003 // Note that the `local_ptr` data passed into the `try` intrinsic is
1004 // expected to be `*mut *mut u8` for this to actually work, but that's
1005 // managed by the standard library.
1009 let mut then = bx.build_sibling_block("then");
1010 let mut catch = bx.build_sibling_block("catch");
1012 let func = llvm::get_param(bx.llfn(), 0);
1013 let data = llvm::get_param(bx.llfn(), 1);
1014 let local_ptr = llvm::get_param(bx.llfn(), 2);
1015 bx.invoke(func, &[data], then.llbb(), catch.llbb(), None);
1016 then.ret(bx.const_i32(0));
1018 // Type indicator for the exception being thrown.
1020 // The first value in this tuple is a pointer to the exception object
1021 // being thrown. The second value is a "selector" indicating which of
1022 // the landing pad clauses the exception's type had been matched to.
1023 // rust_try ignores the selector.
1024 let lpad_ty = bx.type_struct(&[bx.type_i8p(), bx.type_i32()], false);
1025 let vals = catch.landing_pad(lpad_ty, bx.eh_personality(), 1);
1026 let tydesc = match bx.tcx().lang_items().eh_catch_typeinfo() {
1028 let tydesc = bx.get_static(tydesc);
1029 bx.bitcast(tydesc, bx.type_i8p())
1031 None => bx.const_null(bx.type_i8p()),
1033 catch.add_clause(vals, tydesc);
1034 let ptr = catch.extract_value(vals, 0);
1035 let ptr_align = bx.tcx().data_layout.pointer_align.abi;
1036 let bitcast = catch.bitcast(local_ptr, bx.type_ptr_to(bx.type_i8p()));
1037 catch.store(ptr, bitcast, ptr_align);
1038 catch.ret(bx.const_i32(1));
1041 // Note that no invoke is used here because by definition this function
1042 // can't panic (that's what it's catching).
1043 let ret = bx.call(llfn, &[func, data, local_ptr], None);
1044 let i32_align = bx.tcx().data_layout.i32_align.abi;
1045 bx.store(ret, dest, i32_align);
1048 // Helper function to give a Block to a closure to codegen a shim function.
1049 // This is currently primarily used for the `try` intrinsic functions above.
1050 fn gen_fn<'ll, 'tcx>(
1051 cx: &CodegenCx<'ll, 'tcx>,
1053 inputs: Vec<Ty<'tcx>>,
1055 codegen: &mut dyn FnMut(Builder<'_, 'll, 'tcx>),
1057 let rust_fn_sig = ty::Binder::bind(cx.tcx.mk_fn_sig(
1061 hir::Unsafety::Unsafe,
1064 let fn_abi = FnAbi::of_fn_ptr(cx, rust_fn_sig, &[]);
1065 let llfn = cx.declare_fn(name, &fn_abi);
1066 // FIXME(eddyb) find a nicer way to do this.
1067 unsafe { llvm::LLVMRustSetLinkage(llfn, llvm::Linkage::InternalLinkage) };
1068 let bx = Builder::new_block(cx, llfn, "entry-block");
1073 // Helper function used to get a handle to the `__rust_try` function used to
1074 // catch exceptions.
1076 // This function is only generated once and is then cached.
1077 fn get_rust_try_fn<'ll, 'tcx>(
1078 cx: &CodegenCx<'ll, 'tcx>,
1079 codegen: &mut dyn FnMut(Builder<'_, 'll, 'tcx>),
1081 if let Some(llfn) = cx.rust_try_fn.get() {
1085 // Define the type up front for the signature of the rust_try function.
1087 let i8p = tcx.mk_mut_ptr(tcx.types.i8);
1088 let fn_ty = tcx.mk_fn_ptr(ty::Binder::bind(tcx.mk_fn_sig(
1092 hir::Unsafety::Unsafe,
1095 let output = tcx.types.i32;
1096 let rust_try = gen_fn(cx, "__rust_try", vec![fn_ty, i8p, i8p], output, codegen);
1097 cx.rust_try_fn.set(Some(rust_try));
1101 fn generic_simd_intrinsic(
1102 bx: &mut Builder<'a, 'll, 'tcx>,
1104 callee_ty: Ty<'tcx>,
1105 args: &[OperandRef<'tcx, &'ll Value>],
1107 llret_ty: &'ll Type,
1109 ) -> Result<&'ll Value, ()> {
1110 // macros for error handling:
1111 macro_rules! emit_error {
1115 ($msg: tt, $($fmt: tt)*) => {
1116 span_invalid_monomorphization_error(
1118 &format!(concat!("invalid monomorphization of `{}` intrinsic: ", $msg),
1123 macro_rules! return_error {
1126 emit_error!($($fmt)*);
1132 macro_rules! require {
1133 ($cond: expr, $($fmt: tt)*) => {
1135 return_error!($($fmt)*);
1140 macro_rules! require_simd {
1141 ($ty: expr, $position: expr) => {
1142 require!($ty.is_simd(), "expected SIMD {} type, found non-SIMD `{}`", $position, $ty)
1148 .normalize_erasing_late_bound_regions(ty::ParamEnv::reveal_all(), &callee_ty.fn_sig(tcx));
1149 let arg_tys = sig.inputs();
1151 if name == "simd_select_bitmask" {
1152 let in_ty = arg_tys[0];
1153 let m_len = match in_ty.kind {
1154 // Note that this `.unwrap()` crashes for isize/usize, that's sort
1155 // of intentional as there's not currently a use case for that.
1156 ty::Int(i) => i.bit_width().unwrap() as u64,
1157 ty::Uint(i) => i.bit_width().unwrap() as u64,
1158 _ => return_error!("`{}` is not an integral type", in_ty),
1160 require_simd!(arg_tys[1], "argument");
1161 let v_len = arg_tys[1].simd_size(tcx);
1164 "mismatched lengths: mask length `{}` != other vector length `{}`",
1168 let i1 = bx.type_i1();
1169 let i1xn = bx.type_vector(i1, m_len);
1170 let m_i1s = bx.bitcast(args[0].immediate(), i1xn);
1171 return Ok(bx.select(m_i1s, args[1].immediate(), args[2].immediate()));
1174 // every intrinsic below takes a SIMD vector as its first argument
1175 require_simd!(arg_tys[0], "input");
1176 let in_ty = arg_tys[0];
1177 let in_elem = arg_tys[0].simd_type(tcx);
1178 let in_len = arg_tys[0].simd_size(tcx);
1180 let comparison = match name {
1181 "simd_eq" => Some(hir::BinOpKind::Eq),
1182 "simd_ne" => Some(hir::BinOpKind::Ne),
1183 "simd_lt" => Some(hir::BinOpKind::Lt),
1184 "simd_le" => Some(hir::BinOpKind::Le),
1185 "simd_gt" => Some(hir::BinOpKind::Gt),
1186 "simd_ge" => Some(hir::BinOpKind::Ge),
1190 if let Some(cmp_op) = comparison {
1191 require_simd!(ret_ty, "return");
1193 let out_len = ret_ty.simd_size(tcx);
1196 "expected return type with length {} (same as input type `{}`), \
1197 found `{}` with length {}",
1204 bx.type_kind(bx.element_type(llret_ty)) == TypeKind::Integer,
1205 "expected return type with integer elements, found `{}` with non-integer `{}`",
1207 ret_ty.simd_type(tcx)
1210 return Ok(compare_simd_types(
1212 args[0].immediate(),
1213 args[1].immediate(),
1220 if name.starts_with("simd_shuffle") {
1221 let n: u64 = name["simd_shuffle".len()..].parse().unwrap_or_else(|_| {
1222 span_bug!(span, "bad `simd_shuffle` instruction only caught in codegen?")
1225 require_simd!(ret_ty, "return");
1227 let out_len = ret_ty.simd_size(tcx);
1230 "expected return type of length {}, found `{}` with length {}",
1236 in_elem == ret_ty.simd_type(tcx),
1237 "expected return element type `{}` (element of input `{}`), \
1238 found `{}` with element type `{}`",
1242 ret_ty.simd_type(tcx)
1245 let total_len = u128::from(in_len) * 2;
1247 let vector = args[2].immediate();
1249 let indices: Option<Vec<_>> = (0..n)
1252 let val = bx.const_get_elt(vector, i as u64);
1253 match bx.const_to_opt_u128(val, true) {
1255 emit_error!("shuffle index #{} is not a constant", arg_idx);
1258 Some(idx) if idx >= total_len => {
1260 "shuffle index #{} is out of bounds (limit {})",
1266 Some(idx) => Some(bx.const_i32(idx as i32)),
1270 let indices = match indices {
1272 None => return Ok(bx.const_null(llret_ty)),
1275 return Ok(bx.shuffle_vector(
1276 args[0].immediate(),
1277 args[1].immediate(),
1278 bx.const_vector(&indices),
1282 if name == "simd_insert" {
1284 in_elem == arg_tys[2],
1285 "expected inserted type `{}` (element of input `{}`), found `{}`",
1290 return Ok(bx.insert_element(
1291 args[0].immediate(),
1292 args[2].immediate(),
1293 args[1].immediate(),
1296 if name == "simd_extract" {
1299 "expected return type `{}` (element of input `{}`), found `{}`",
1304 return Ok(bx.extract_element(args[0].immediate(), args[1].immediate()));
1307 if name == "simd_select" {
1308 let m_elem_ty = in_elem;
1310 require_simd!(arg_tys[1], "argument");
1311 let v_len = arg_tys[1].simd_size(tcx);
1314 "mismatched lengths: mask length `{}` != other vector length `{}`",
1318 match m_elem_ty.kind {
1320 _ => return_error!("mask element type is `{}`, expected `i_`", m_elem_ty),
1322 // truncate the mask to a vector of i1s
1323 let i1 = bx.type_i1();
1324 let i1xn = bx.type_vector(i1, m_len as u64);
1325 let m_i1s = bx.trunc(args[0].immediate(), i1xn);
1326 return Ok(bx.select(m_i1s, args[1].immediate(), args[2].immediate()));
1329 if name == "simd_bitmask" {
1330 // The `fn simd_bitmask(vector) -> unsigned integer` intrinsic takes a
1331 // vector mask and returns an unsigned integer containing the most
1332 // significant bit (MSB) of each lane.
1334 // If the vector has less than 8 lanes, an u8 is returned with zeroed
1336 let expected_int_bits = in_len.max(8);
1338 ty::Uint(i) if i.bit_width() == Some(expected_int_bits as usize) => (),
1339 _ => return_error!("bitmask `{}`, expected `u{}`", ret_ty, expected_int_bits),
1342 // Integer vector <i{in_bitwidth} x in_len>:
1343 let (i_xn, in_elem_bitwidth) = match in_elem.kind {
1345 args[0].immediate(),
1346 i.bit_width().unwrap_or(bx.data_layout().pointer_size.bits() as _),
1349 args[0].immediate(),
1350 i.bit_width().unwrap_or(bx.data_layout().pointer_size.bits() as _),
1353 "vector argument `{}`'s element type `{}`, expected integer element type",
1359 // Shift the MSB to the right by "in_elem_bitwidth - 1" into the first bit position.
1362 bx.cx.const_int(bx.type_ix(in_elem_bitwidth as _), (in_elem_bitwidth - 1) as _);
1365 let i_xn_msb = bx.lshr(i_xn, bx.const_vector(shift_indices.as_slice()));
1366 // Truncate vector to an <i1 x N>
1367 let i1xn = bx.trunc(i_xn_msb, bx.type_vector(bx.type_i1(), in_len as _));
1368 // Bitcast <i1 x N> to iN:
1369 let i_ = bx.bitcast(i1xn, bx.type_ix(in_len as _));
1370 // Zero-extend iN to the bitmask type:
1371 return Ok(bx.zext(i_, bx.type_ix(expected_int_bits as _)));
1374 fn simd_simple_float_intrinsic(
1376 in_elem: &::rustc::ty::TyS<'_>,
1377 in_ty: &::rustc::ty::TyS<'_>,
1379 bx: &mut Builder<'a, 'll, 'tcx>,
1381 args: &[OperandRef<'tcx, &'ll Value>],
1382 ) -> Result<&'ll Value, ()> {
1383 macro_rules! emit_error {
1387 ($msg: tt, $($fmt: tt)*) => {
1388 span_invalid_monomorphization_error(
1390 &format!(concat!("invalid monomorphization of `{}` intrinsic: ", $msg),
1394 macro_rules! return_error {
1397 emit_error!($($fmt)*);
1402 let ety = match in_elem.kind {
1403 ty::Float(f) if f.bit_width() == 32 => {
1404 if in_len < 2 || in_len > 16 {
1406 "unsupported floating-point vector `{}` with length `{}` \
1407 out-of-range [2, 16]",
1414 ty::Float(f) if f.bit_width() == 64 => {
1415 if in_len < 2 || in_len > 8 {
1417 "unsupported floating-point vector `{}` with length `{}` \
1418 out-of-range [2, 8]",
1427 "unsupported element type `{}` of floating-point vector `{}`",
1433 return_error!("`{}` is not a floating-point type", in_ty);
1437 let llvm_name = &format!("llvm.{0}.v{1}{2}", name, in_len, ety);
1438 let intrinsic = bx.get_intrinsic(&llvm_name);
1440 bx.call(intrinsic, &args.iter().map(|arg| arg.immediate()).collect::<Vec<_>>(), None);
1441 unsafe { llvm::LLVMRustSetHasUnsafeAlgebra(c) };
1447 return simd_simple_float_intrinsic("sqrt", in_elem, in_ty, in_len, bx, span, args);
1450 return simd_simple_float_intrinsic("sin", in_elem, in_ty, in_len, bx, span, args);
1453 return simd_simple_float_intrinsic("cos", in_elem, in_ty, in_len, bx, span, args);
1456 return simd_simple_float_intrinsic("fabs", in_elem, in_ty, in_len, bx, span, args);
1459 return simd_simple_float_intrinsic("floor", in_elem, in_ty, in_len, bx, span, args);
1462 return simd_simple_float_intrinsic("ceil", in_elem, in_ty, in_len, bx, span, args);
1465 return simd_simple_float_intrinsic("exp", in_elem, in_ty, in_len, bx, span, args);
1468 return simd_simple_float_intrinsic("exp2", in_elem, in_ty, in_len, bx, span, args);
1471 return simd_simple_float_intrinsic("log10", in_elem, in_ty, in_len, bx, span, args);
1474 return simd_simple_float_intrinsic("log2", in_elem, in_ty, in_len, bx, span, args);
1477 return simd_simple_float_intrinsic("log", in_elem, in_ty, in_len, bx, span, args);
1480 return simd_simple_float_intrinsic("powi", in_elem, in_ty, in_len, bx, span, args);
1483 return simd_simple_float_intrinsic("pow", in_elem, in_ty, in_len, bx, span, args);
1486 return simd_simple_float_intrinsic("fma", in_elem, in_ty, in_len, bx, span, args);
1488 _ => { /* fallthrough */ }
1492 // https://github.com/llvm-mirror/llvm/blob/master/include/llvm/IR/Function.h#L182
1493 // https://github.com/llvm-mirror/llvm/blob/master/include/llvm/IR/Intrinsics.h#L81
1494 fn llvm_vector_str(elem_ty: Ty<'_>, vec_len: u64, no_pointers: usize) -> String {
1495 let p0s: String = "p0".repeat(no_pointers);
1496 match elem_ty.kind {
1497 ty::Int(v) => format!("v{}{}i{}", vec_len, p0s, v.bit_width().unwrap()),
1498 ty::Uint(v) => format!("v{}{}i{}", vec_len, p0s, v.bit_width().unwrap()),
1499 ty::Float(v) => format!("v{}{}f{}", vec_len, p0s, v.bit_width()),
1500 _ => unreachable!(),
1505 cx: &CodegenCx<'ll, '_>,
1508 mut no_pointers: usize,
1510 // FIXME: use cx.layout_of(ty).llvm_type() ?
1511 let mut elem_ty = match elem_ty.kind {
1512 ty::Int(v) => cx.type_int_from_ty(v),
1513 ty::Uint(v) => cx.type_uint_from_ty(v),
1514 ty::Float(v) => cx.type_float_from_ty(v),
1515 _ => unreachable!(),
1517 while no_pointers > 0 {
1518 elem_ty = cx.type_ptr_to(elem_ty);
1521 cx.type_vector(elem_ty, vec_len)
1524 if name == "simd_gather" {
1525 // simd_gather(values: <N x T>, pointers: <N x *_ T>,
1526 // mask: <N x i{M}>) -> <N x T>
1527 // * N: number of elements in the input vectors
1528 // * T: type of the element to load
1529 // * M: any integer width is supported, will be truncated to i1
1531 // All types must be simd vector types
1532 require_simd!(in_ty, "first");
1533 require_simd!(arg_tys[1], "second");
1534 require_simd!(arg_tys[2], "third");
1535 require_simd!(ret_ty, "return");
1537 // Of the same length:
1539 in_len == arg_tys[1].simd_size(tcx),
1540 "expected {} argument with length {} (same as input type `{}`), \
1541 found `{}` with length {}",
1546 arg_tys[1].simd_size(tcx)
1549 in_len == arg_tys[2].simd_size(tcx),
1550 "expected {} argument with length {} (same as input type `{}`), \
1551 found `{}` with length {}",
1556 arg_tys[2].simd_size(tcx)
1559 // The return type must match the first argument type
1560 require!(ret_ty == in_ty, "expected return type `{}`, found `{}`", in_ty, ret_ty);
1562 // This counts how many pointers
1563 fn ptr_count(t: Ty<'_>) -> usize {
1565 ty::RawPtr(p) => 1 + ptr_count(p.ty),
1571 fn non_ptr(t: Ty<'_>) -> Ty<'_> {
1573 ty::RawPtr(p) => non_ptr(p.ty),
1578 // The second argument must be a simd vector with an element type that's a pointer
1579 // to the element type of the first argument
1580 let (pointer_count, underlying_ty) = match arg_tys[1].simd_type(tcx).kind {
1581 ty::RawPtr(p) if p.ty == in_elem => {
1582 (ptr_count(arg_tys[1].simd_type(tcx)), non_ptr(arg_tys[1].simd_type(tcx)))
1587 "expected element type `{}` of second argument `{}` \
1588 to be a pointer to the element type `{}` of the first \
1589 argument `{}`, found `{}` != `*_ {}`",
1590 arg_tys[1].simd_type(tcx),
1594 arg_tys[1].simd_type(tcx),
1600 assert!(pointer_count > 0);
1601 assert_eq!(pointer_count - 1, ptr_count(arg_tys[0].simd_type(tcx)));
1602 assert_eq!(underlying_ty, non_ptr(arg_tys[0].simd_type(tcx)));
1604 // The element type of the third argument must be a signed integer type of any width:
1605 match arg_tys[2].simd_type(tcx).kind {
1610 "expected element type `{}` of third argument `{}` \
1611 to be a signed integer type",
1612 arg_tys[2].simd_type(tcx),
1618 // Alignment of T, must be a constant integer value:
1619 let alignment_ty = bx.type_i32();
1620 let alignment = bx.const_i32(bx.align_of(in_elem).bytes() as i32);
1622 // Truncate the mask vector to a vector of i1s:
1623 let (mask, mask_ty) = {
1624 let i1 = bx.type_i1();
1625 let i1xn = bx.type_vector(i1, in_len);
1626 (bx.trunc(args[2].immediate(), i1xn), i1xn)
1629 // Type of the vector of pointers:
1630 let llvm_pointer_vec_ty = llvm_vector_ty(bx, underlying_ty, in_len, pointer_count);
1631 let llvm_pointer_vec_str = llvm_vector_str(underlying_ty, in_len, pointer_count);
1633 // Type of the vector of elements:
1634 let llvm_elem_vec_ty = llvm_vector_ty(bx, underlying_ty, in_len, pointer_count - 1);
1635 let llvm_elem_vec_str = llvm_vector_str(underlying_ty, in_len, pointer_count - 1);
1637 let llvm_intrinsic =
1638 format!("llvm.masked.gather.{}.{}", llvm_elem_vec_str, llvm_pointer_vec_str);
1639 let f = bx.declare_cfn(
1642 &[llvm_pointer_vec_ty, alignment_ty, mask_ty, llvm_elem_vec_ty],
1646 llvm::SetUnnamedAddr(f, false);
1647 let v = bx.call(f, &[args[1].immediate(), alignment, mask, args[0].immediate()], None);
1651 if name == "simd_scatter" {
1652 // simd_scatter(values: <N x T>, pointers: <N x *mut T>,
1653 // mask: <N x i{M}>) -> ()
1654 // * N: number of elements in the input vectors
1655 // * T: type of the element to load
1656 // * M: any integer width is supported, will be truncated to i1
1658 // All types must be simd vector types
1659 require_simd!(in_ty, "first");
1660 require_simd!(arg_tys[1], "second");
1661 require_simd!(arg_tys[2], "third");
1663 // Of the same length:
1665 in_len == arg_tys[1].simd_size(tcx),
1666 "expected {} argument with length {} (same as input type `{}`), \
1667 found `{}` with length {}",
1672 arg_tys[1].simd_size(tcx)
1675 in_len == arg_tys[2].simd_size(tcx),
1676 "expected {} argument with length {} (same as input type `{}`), \
1677 found `{}` with length {}",
1682 arg_tys[2].simd_size(tcx)
1685 // This counts how many pointers
1686 fn ptr_count(t: Ty<'_>) -> usize {
1688 ty::RawPtr(p) => 1 + ptr_count(p.ty),
1694 fn non_ptr(t: Ty<'_>) -> Ty<'_> {
1696 ty::RawPtr(p) => non_ptr(p.ty),
1701 // The second argument must be a simd vector with an element type that's a pointer
1702 // to the element type of the first argument
1703 let (pointer_count, underlying_ty) = match arg_tys[1].simd_type(tcx).kind {
1704 ty::RawPtr(p) if p.ty == in_elem && p.mutbl == hir::Mutability::Mut => {
1705 (ptr_count(arg_tys[1].simd_type(tcx)), non_ptr(arg_tys[1].simd_type(tcx)))
1710 "expected element type `{}` of second argument `{}` \
1711 to be a pointer to the element type `{}` of the first \
1712 argument `{}`, found `{}` != `*mut {}`",
1713 arg_tys[1].simd_type(tcx),
1717 arg_tys[1].simd_type(tcx),
1723 assert!(pointer_count > 0);
1724 assert_eq!(pointer_count - 1, ptr_count(arg_tys[0].simd_type(tcx)));
1725 assert_eq!(underlying_ty, non_ptr(arg_tys[0].simd_type(tcx)));
1727 // The element type of the third argument must be a signed integer type of any width:
1728 match arg_tys[2].simd_type(tcx).kind {
1733 "expected element type `{}` of third argument `{}` \
1734 to be a signed integer type",
1735 arg_tys[2].simd_type(tcx),
1741 // Alignment of T, must be a constant integer value:
1742 let alignment_ty = bx.type_i32();
1743 let alignment = bx.const_i32(bx.align_of(in_elem).bytes() as i32);
1745 // Truncate the mask vector to a vector of i1s:
1746 let (mask, mask_ty) = {
1747 let i1 = bx.type_i1();
1748 let i1xn = bx.type_vector(i1, in_len);
1749 (bx.trunc(args[2].immediate(), i1xn), i1xn)
1752 let ret_t = bx.type_void();
1754 // Type of the vector of pointers:
1755 let llvm_pointer_vec_ty = llvm_vector_ty(bx, underlying_ty, in_len, pointer_count);
1756 let llvm_pointer_vec_str = llvm_vector_str(underlying_ty, in_len, pointer_count);
1758 // Type of the vector of elements:
1759 let llvm_elem_vec_ty = llvm_vector_ty(bx, underlying_ty, in_len, pointer_count - 1);
1760 let llvm_elem_vec_str = llvm_vector_str(underlying_ty, in_len, pointer_count - 1);
1762 let llvm_intrinsic =
1763 format!("llvm.masked.scatter.{}.{}", llvm_elem_vec_str, llvm_pointer_vec_str);
1764 let f = bx.declare_cfn(
1766 bx.type_func(&[llvm_elem_vec_ty, llvm_pointer_vec_ty, alignment_ty, mask_ty], ret_t),
1768 llvm::SetUnnamedAddr(f, false);
1769 let v = bx.call(f, &[args[0].immediate(), args[1].immediate(), alignment, mask], None);
1773 macro_rules! arith_red {
1774 ($name:tt : $integer_reduce:ident, $float_reduce:ident, $ordered:expr) => {
1778 "expected return type `{}` (element of input `{}`), found `{}`",
1783 return match in_elem.kind {
1784 ty::Int(_) | ty::Uint(_) => {
1785 let r = bx.$integer_reduce(args[0].immediate());
1787 // if overflow occurs, the result is the
1788 // mathematical result modulo 2^n:
1789 if name.contains("mul") {
1790 Ok(bx.mul(args[1].immediate(), r))
1792 Ok(bx.add(args[1].immediate(), r))
1795 Ok(bx.$integer_reduce(args[0].immediate()))
1799 let acc = if $ordered {
1800 // ordered arithmetic reductions take an accumulator
1803 // unordered arithmetic reductions use the identity accumulator
1804 let identity_acc = if $name.contains("mul") { 1.0 } else { 0.0 };
1805 match f.bit_width() {
1806 32 => bx.const_real(bx.type_f32(), identity_acc),
1807 64 => bx.const_real(bx.type_f64(), identity_acc),
1810 unsupported {} from `{}` with element `{}` of size `{}` to `{}`"#,
1819 Ok(bx.$float_reduce(acc, args[0].immediate()))
1822 "unsupported {} from `{}` with element `{}` to `{}`",
1833 arith_red!("simd_reduce_add_ordered": vector_reduce_add, vector_reduce_fadd, true);
1834 arith_red!("simd_reduce_mul_ordered": vector_reduce_mul, vector_reduce_fmul, true);
1835 arith_red!("simd_reduce_add_unordered": vector_reduce_add, vector_reduce_fadd_fast, false);
1836 arith_red!("simd_reduce_mul_unordered": vector_reduce_mul, vector_reduce_fmul_fast, false);
1838 macro_rules! minmax_red {
1839 ($name:tt: $int_red:ident, $float_red:ident) => {
1843 "expected return type `{}` (element of input `{}`), found `{}`",
1848 return match in_elem.kind {
1849 ty::Int(_i) => Ok(bx.$int_red(args[0].immediate(), true)),
1850 ty::Uint(_u) => Ok(bx.$int_red(args[0].immediate(), false)),
1851 ty::Float(_f) => Ok(bx.$float_red(args[0].immediate())),
1853 "unsupported {} from `{}` with element `{}` to `{}`",
1864 minmax_red!("simd_reduce_min": vector_reduce_min, vector_reduce_fmin);
1865 minmax_red!("simd_reduce_max": vector_reduce_max, vector_reduce_fmax);
1867 minmax_red!("simd_reduce_min_nanless": vector_reduce_min, vector_reduce_fmin_fast);
1868 minmax_red!("simd_reduce_max_nanless": vector_reduce_max, vector_reduce_fmax_fast);
1870 macro_rules! bitwise_red {
1871 ($name:tt : $red:ident, $boolean:expr) => {
1873 let input = if !$boolean {
1876 "expected return type `{}` (element of input `{}`), found `{}`",
1883 match in_elem.kind {
1884 ty::Int(_) | ty::Uint(_) => {}
1886 "unsupported {} from `{}` with element `{}` to `{}`",
1894 // boolean reductions operate on vectors of i1s:
1895 let i1 = bx.type_i1();
1896 let i1xn = bx.type_vector(i1, in_len as u64);
1897 bx.trunc(args[0].immediate(), i1xn)
1899 return match in_elem.kind {
1900 ty::Int(_) | ty::Uint(_) => {
1901 let r = bx.$red(input);
1902 Ok(if !$boolean { r } else { bx.zext(r, bx.type_bool()) })
1905 "unsupported {} from `{}` with element `{}` to `{}`",
1916 bitwise_red!("simd_reduce_and": vector_reduce_and, false);
1917 bitwise_red!("simd_reduce_or": vector_reduce_or, false);
1918 bitwise_red!("simd_reduce_xor": vector_reduce_xor, false);
1919 bitwise_red!("simd_reduce_all": vector_reduce_and, true);
1920 bitwise_red!("simd_reduce_any": vector_reduce_or, true);
1922 if name == "simd_cast" {
1923 require_simd!(ret_ty, "return");
1924 let out_len = ret_ty.simd_size(tcx);
1927 "expected return type with length {} (same as input type `{}`), \
1928 found `{}` with length {}",
1934 // casting cares about nominal type, not just structural type
1935 let out_elem = ret_ty.simd_type(tcx);
1937 if in_elem == out_elem {
1938 return Ok(args[0].immediate());
1943 Int(/* is signed? */ bool),
1947 let (in_style, in_width) = match in_elem.kind {
1948 // vectors of pointer-sized integers should've been
1949 // disallowed before here, so this unwrap is safe.
1950 ty::Int(i) => (Style::Int(true), i.bit_width().unwrap()),
1951 ty::Uint(u) => (Style::Int(false), u.bit_width().unwrap()),
1952 ty::Float(f) => (Style::Float, f.bit_width()),
1953 _ => (Style::Unsupported, 0),
1955 let (out_style, out_width) = match out_elem.kind {
1956 ty::Int(i) => (Style::Int(true), i.bit_width().unwrap()),
1957 ty::Uint(u) => (Style::Int(false), u.bit_width().unwrap()),
1958 ty::Float(f) => (Style::Float, f.bit_width()),
1959 _ => (Style::Unsupported, 0),
1962 match (in_style, out_style) {
1963 (Style::Int(in_is_signed), Style::Int(_)) => {
1964 return Ok(match in_width.cmp(&out_width) {
1965 Ordering::Greater => bx.trunc(args[0].immediate(), llret_ty),
1966 Ordering::Equal => args[0].immediate(),
1969 bx.sext(args[0].immediate(), llret_ty)
1971 bx.zext(args[0].immediate(), llret_ty)
1976 (Style::Int(in_is_signed), Style::Float) => {
1977 return Ok(if in_is_signed {
1978 bx.sitofp(args[0].immediate(), llret_ty)
1980 bx.uitofp(args[0].immediate(), llret_ty)
1983 (Style::Float, Style::Int(out_is_signed)) => {
1984 return Ok(if out_is_signed {
1985 bx.fptosi(args[0].immediate(), llret_ty)
1987 bx.fptoui(args[0].immediate(), llret_ty)
1990 (Style::Float, Style::Float) => {
1991 return Ok(match in_width.cmp(&out_width) {
1992 Ordering::Greater => bx.fptrunc(args[0].immediate(), llret_ty),
1993 Ordering::Equal => args[0].immediate(),
1994 Ordering::Less => bx.fpext(args[0].immediate(), llret_ty),
1997 _ => { /* Unsupported. Fallthrough. */ }
2001 "unsupported cast from `{}` with element `{}` to `{}` with element `{}`",
2008 macro_rules! arith {
2009 ($($name: ident: $($($p: ident),* => $call: ident),*;)*) => {
2010 $(if name == stringify!($name) {
2011 match in_elem.kind {
2012 $($(ty::$p(_))|* => {
2013 return Ok(bx.$call(args[0].immediate(), args[1].immediate()))
2018 "unsupported operation on `{}` with element `{}`",
2025 simd_add: Uint, Int => add, Float => fadd;
2026 simd_sub: Uint, Int => sub, Float => fsub;
2027 simd_mul: Uint, Int => mul, Float => fmul;
2028 simd_div: Uint => udiv, Int => sdiv, Float => fdiv;
2029 simd_rem: Uint => urem, Int => srem, Float => frem;
2030 simd_shl: Uint, Int => shl;
2031 simd_shr: Uint => lshr, Int => ashr;
2032 simd_and: Uint, Int => and;
2033 simd_or: Uint, Int => or;
2034 simd_xor: Uint, Int => xor;
2035 simd_fmax: Float => maxnum;
2036 simd_fmin: Float => minnum;
2040 if name == "simd_saturating_add" || name == "simd_saturating_sub" {
2041 let lhs = args[0].immediate();
2042 let rhs = args[1].immediate();
2043 let is_add = name == "simd_saturating_add";
2044 let ptr_bits = bx.tcx().data_layout.pointer_size.bits() as _;
2045 let (signed, elem_width, elem_ty) = match in_elem.kind {
2046 ty::Int(i) => (true, i.bit_width().unwrap_or(ptr_bits), bx.cx.type_int_from_ty(i)),
2047 ty::Uint(i) => (false, i.bit_width().unwrap_or(ptr_bits), bx.cx.type_uint_from_ty(i)),
2050 "expected element type `{}` of vector type `{}` \
2051 to be a signed or unsigned integer type",
2052 arg_tys[0].simd_type(tcx),
2057 let llvm_intrinsic = &format!(
2058 "llvm.{}{}.sat.v{}i{}",
2059 if signed { 's' } else { 'u' },
2060 if is_add { "add" } else { "sub" },
2064 let vec_ty = bx.cx.type_vector(elem_ty, in_len as u64);
2066 let f = bx.declare_cfn(&llvm_intrinsic, bx.type_func(&[vec_ty, vec_ty], vec_ty));
2067 llvm::SetUnnamedAddr(f, false);
2068 let v = bx.call(f, &[lhs, rhs], None);
2072 span_bug!(span, "unknown SIMD intrinsic");
2075 // Returns the width of an int Ty, and if it's signed or not
2076 // Returns None if the type is not an integer
2077 // FIXME: there’s multiple of this functions, investigate using some of the already existing
2079 fn int_type_width_signed(ty: Ty<'_>, cx: &CodegenCx<'_, '_>) -> Option<(u64, bool)> {
2081 ty::Int(t) => Some((
2083 ast::IntTy::Isize => cx.tcx.sess.target.ptr_width as u64,
2084 ast::IntTy::I8 => 8,
2085 ast::IntTy::I16 => 16,
2086 ast::IntTy::I32 => 32,
2087 ast::IntTy::I64 => 64,
2088 ast::IntTy::I128 => 128,
2092 ty::Uint(t) => Some((
2094 ast::UintTy::Usize => cx.tcx.sess.target.ptr_width as u64,
2095 ast::UintTy::U8 => 8,
2096 ast::UintTy::U16 => 16,
2097 ast::UintTy::U32 => 32,
2098 ast::UintTy::U64 => 64,
2099 ast::UintTy::U128 => 128,
2107 // Returns the width of a float Ty
2108 // Returns None if the type is not a float
2109 fn float_type_width(ty: Ty<'_>) -> Option<u64> {
2111 ty::Float(t) => Some(t.bit_width() as u64),