1 //! Codegen of intrinsics. This includes `extern "rust-intrinsic"`, `extern "platform-intrinsic"`
2 //! and LLVM intrinsics that have symbol names starting with `llvm.`.
4 macro_rules! intrinsic_args {
5 ($fx:expr, $args:expr => ($($arg:tt),*); $intrinsic:expr) => {
6 #[allow(unused_parens)]
7 let ($($arg),*) = if let [$($arg),*] = $args {
8 ($(codegen_operand($fx, $arg)),*)
10 $crate::intrinsics::bug_on_incorrect_arg_count($intrinsic);
21 pub(crate) use cpuid::codegen_cpuid_call;
22 pub(crate) use llvm::codegen_llvm_intrinsic_call;
24 use rustc_middle::ty::print::with_no_trimmed_paths;
25 use rustc_middle::ty::subst::SubstsRef;
26 use rustc_span::symbol::{kw, sym, Symbol};
28 use crate::prelude::*;
29 use cranelift_codegen::ir::AtomicRmwOp;
31 fn bug_on_incorrect_arg_count(intrinsic: impl std::fmt::Display) -> ! {
32 bug!("wrong number of args for intrinsic {}", intrinsic);
35 fn report_atomic_type_validation_error<'tcx>(
36 fx: &mut FunctionCx<'_, '_, 'tcx>,
44 "`{}` intrinsic: expected basic integer or raw pointer type, found `{:?}`",
48 // Prevent verifier error
49 fx.bcx.ins().trap(TrapCode::UnreachableCodeReached);
52 pub(crate) fn clif_vector_type<'tcx>(tcx: TyCtxt<'tcx>, layout: TyAndLayout<'tcx>) -> Option<Type> {
53 let (element, count) = match layout.abi {
54 Abi::Vector { element, count } => (element, count),
58 match scalar_to_clif_type(tcx, element).by(u32::try_from(count).unwrap()) {
59 // Cranelift currently only implements icmp for 128bit vectors.
60 Some(vector_ty) if vector_ty.bits() == 128 => Some(vector_ty),
65 fn simd_for_each_lane<'tcx>(
66 fx: &mut FunctionCx<'_, '_, 'tcx>,
69 f: &dyn Fn(&mut FunctionCx<'_, '_, 'tcx>, Ty<'tcx>, Ty<'tcx>, Value) -> Value,
71 let layout = val.layout();
73 let (lane_count, lane_ty) = layout.ty.simd_size_and_type(fx.tcx);
74 let lane_layout = fx.layout_of(lane_ty);
75 let (ret_lane_count, ret_lane_ty) = ret.layout().ty.simd_size_and_type(fx.tcx);
76 let ret_lane_layout = fx.layout_of(ret_lane_ty);
77 assert_eq!(lane_count, ret_lane_count);
79 for lane_idx in 0..lane_count {
80 let lane = val.value_lane(fx, lane_idx).load_scalar(fx);
82 let res_lane = f(fx, lane_layout.ty, ret_lane_layout.ty, lane);
83 let res_lane = CValue::by_val(res_lane, ret_lane_layout);
85 ret.place_lane(fx, lane_idx).write_cvalue(fx, res_lane);
89 fn simd_pair_for_each_lane_typed<'tcx>(
90 fx: &mut FunctionCx<'_, '_, 'tcx>,
94 f: &dyn Fn(&mut FunctionCx<'_, '_, 'tcx>, CValue<'tcx>, CValue<'tcx>) -> CValue<'tcx>,
96 assert_eq!(x.layout(), y.layout());
97 let layout = x.layout();
99 let (lane_count, _lane_ty) = layout.ty.simd_size_and_type(fx.tcx);
100 let (ret_lane_count, _ret_lane_ty) = ret.layout().ty.simd_size_and_type(fx.tcx);
101 assert_eq!(lane_count, ret_lane_count);
103 for lane_idx in 0..lane_count {
104 let x_lane = x.value_lane(fx, lane_idx);
105 let y_lane = y.value_lane(fx, lane_idx);
107 let res_lane = f(fx, x_lane, y_lane);
109 ret.place_lane(fx, lane_idx).write_cvalue(fx, res_lane);
113 fn simd_pair_for_each_lane<'tcx>(
114 fx: &mut FunctionCx<'_, '_, 'tcx>,
118 f: &dyn Fn(&mut FunctionCx<'_, '_, 'tcx>, Ty<'tcx>, Ty<'tcx>, Value, Value) -> Value,
120 assert_eq!(x.layout(), y.layout());
121 let layout = x.layout();
123 let (lane_count, lane_ty) = layout.ty.simd_size_and_type(fx.tcx);
124 let lane_layout = fx.layout_of(lane_ty);
125 let (ret_lane_count, ret_lane_ty) = ret.layout().ty.simd_size_and_type(fx.tcx);
126 let ret_lane_layout = fx.layout_of(ret_lane_ty);
127 assert_eq!(lane_count, ret_lane_count);
129 for lane_idx in 0..lane_count {
130 let x_lane = x.value_lane(fx, lane_idx).load_scalar(fx);
131 let y_lane = y.value_lane(fx, lane_idx).load_scalar(fx);
133 let res_lane = f(fx, lane_layout.ty, ret_lane_layout.ty, x_lane, y_lane);
134 let res_lane = CValue::by_val(res_lane, ret_lane_layout);
136 ret.place_lane(fx, lane_idx).write_cvalue(fx, res_lane);
140 fn simd_reduce<'tcx>(
141 fx: &mut FunctionCx<'_, '_, 'tcx>,
145 f: &dyn Fn(&mut FunctionCx<'_, '_, 'tcx>, Ty<'tcx>, Value, Value) -> Value,
147 let (lane_count, lane_ty) = val.layout().ty.simd_size_and_type(fx.tcx);
148 let lane_layout = fx.layout_of(lane_ty);
149 assert_eq!(lane_layout, ret.layout());
151 let (mut res_val, start_lane) =
152 if let Some(acc) = acc { (acc, 0) } else { (val.value_lane(fx, 0).load_scalar(fx), 1) };
153 for lane_idx in start_lane..lane_count {
154 let lane = val.value_lane(fx, lane_idx).load_scalar(fx);
155 res_val = f(fx, lane_layout.ty, res_val, lane);
157 let res = CValue::by_val(res_val, lane_layout);
158 ret.write_cvalue(fx, res);
161 // FIXME move all uses to `simd_reduce`
162 fn simd_reduce_bool<'tcx>(
163 fx: &mut FunctionCx<'_, '_, 'tcx>,
166 f: &dyn Fn(&mut FunctionCx<'_, '_, 'tcx>, Value, Value) -> Value,
168 let (lane_count, _lane_ty) = val.layout().ty.simd_size_and_type(fx.tcx);
169 assert!(ret.layout().ty.is_bool());
171 let res_val = val.value_lane(fx, 0).load_scalar(fx);
172 let mut res_val = fx.bcx.ins().band_imm(res_val, 1); // mask to boolean
173 for lane_idx in 1..lane_count {
174 let lane = val.value_lane(fx, lane_idx).load_scalar(fx);
175 let lane = fx.bcx.ins().band_imm(lane, 1); // mask to boolean
176 res_val = f(fx, res_val, lane);
178 let res_val = if fx.bcx.func.dfg.value_type(res_val) != types::I8 {
179 fx.bcx.ins().ireduce(types::I8, res_val)
183 let res = CValue::by_val(res_val, ret.layout());
184 ret.write_cvalue(fx, res);
187 fn bool_to_zero_or_max_uint<'tcx>(
188 fx: &mut FunctionCx<'_, '_, 'tcx>,
192 let ty = fx.clif_type(ty).unwrap();
194 let int_ty = match ty {
195 types::F32 => types::I32,
196 types::F64 => types::I64,
200 let mut res = fx.bcx.ins().bmask(int_ty, val);
203 res = fx.bcx.ins().bitcast(ty, res);
209 pub(crate) fn codegen_intrinsic_call<'tcx>(
210 fx: &mut FunctionCx<'_, '_, 'tcx>,
211 instance: Instance<'tcx>,
212 args: &[mir::Operand<'tcx>],
213 destination: CPlace<'tcx>,
214 target: Option<BasicBlock>,
215 source_info: mir::SourceInfo,
217 let intrinsic = fx.tcx.item_name(instance.def_id());
218 let substs = instance.substs;
220 let target = if let Some(target) = target {
223 // Insert non returning intrinsics here
226 fx.bcx.ins().trap(TrapCode::User(0));
229 crate::base::codegen_panic(fx, "Transmuting to uninhabited type.", source_info);
231 _ => unimplemented!("unsupported intrinsic {}", intrinsic),
236 if intrinsic.as_str().starts_with("simd_") {
237 self::simd::codegen_simd_intrinsic_call(
245 let ret_block = fx.get_block(target);
246 fx.bcx.ins().jump(ret_block, &[]);
247 } else if codegen_float_intrinsic_call(fx, intrinsic, args, destination) {
248 let ret_block = fx.get_block(target);
249 fx.bcx.ins().jump(ret_block, &[]);
251 codegen_regular_intrinsic_call(
264 fn codegen_float_intrinsic_call<'tcx>(
265 fx: &mut FunctionCx<'_, '_, 'tcx>,
267 args: &[mir::Operand<'tcx>],
270 let (name, arg_count, ty) = match intrinsic {
271 sym::expf32 => ("expf", 1, fx.tcx.types.f32),
272 sym::expf64 => ("exp", 1, fx.tcx.types.f64),
273 sym::exp2f32 => ("exp2f", 1, fx.tcx.types.f32),
274 sym::exp2f64 => ("exp2", 1, fx.tcx.types.f64),
275 sym::sqrtf32 => ("sqrtf", 1, fx.tcx.types.f32),
276 sym::sqrtf64 => ("sqrt", 1, fx.tcx.types.f64),
277 sym::powif32 => ("__powisf2", 2, fx.tcx.types.f32), // compiler-builtins
278 sym::powif64 => ("__powidf2", 2, fx.tcx.types.f64), // compiler-builtins
279 sym::powf32 => ("powf", 2, fx.tcx.types.f32),
280 sym::powf64 => ("pow", 2, fx.tcx.types.f64),
281 sym::logf32 => ("logf", 1, fx.tcx.types.f32),
282 sym::logf64 => ("log", 1, fx.tcx.types.f64),
283 sym::log2f32 => ("log2f", 1, fx.tcx.types.f32),
284 sym::log2f64 => ("log2", 1, fx.tcx.types.f64),
285 sym::log10f32 => ("log10f", 1, fx.tcx.types.f32),
286 sym::log10f64 => ("log10", 1, fx.tcx.types.f64),
287 sym::fabsf32 => ("fabsf", 1, fx.tcx.types.f32),
288 sym::fabsf64 => ("fabs", 1, fx.tcx.types.f64),
289 sym::fmaf32 => ("fmaf", 3, fx.tcx.types.f32),
290 sym::fmaf64 => ("fma", 3, fx.tcx.types.f64),
291 sym::copysignf32 => ("copysignf", 2, fx.tcx.types.f32),
292 sym::copysignf64 => ("copysign", 2, fx.tcx.types.f64),
293 sym::floorf32 => ("floorf", 1, fx.tcx.types.f32),
294 sym::floorf64 => ("floor", 1, fx.tcx.types.f64),
295 sym::ceilf32 => ("ceilf", 1, fx.tcx.types.f32),
296 sym::ceilf64 => ("ceil", 1, fx.tcx.types.f64),
297 sym::truncf32 => ("truncf", 1, fx.tcx.types.f32),
298 sym::truncf64 => ("trunc", 1, fx.tcx.types.f64),
299 sym::roundf32 => ("roundf", 1, fx.tcx.types.f32),
300 sym::roundf64 => ("round", 1, fx.tcx.types.f64),
301 sym::sinf32 => ("sinf", 1, fx.tcx.types.f32),
302 sym::sinf64 => ("sin", 1, fx.tcx.types.f64),
303 sym::cosf32 => ("cosf", 1, fx.tcx.types.f32),
304 sym::cosf64 => ("cos", 1, fx.tcx.types.f64),
308 if args.len() != arg_count {
309 bug!("wrong number of args for intrinsic {:?}", intrinsic);
313 let args = match args {
315 a = [codegen_operand(fx, x)];
319 b = [codegen_operand(fx, x), codegen_operand(fx, y)];
323 c = [codegen_operand(fx, x), codegen_operand(fx, y), codegen_operand(fx, z)];
329 let layout = fx.layout_of(ty);
330 let res = match intrinsic {
331 sym::fmaf32 | sym::fmaf64 => {
332 let a = args[0].load_scalar(fx);
333 let b = args[1].load_scalar(fx);
334 let c = args[2].load_scalar(fx);
335 CValue::by_val(fx.bcx.ins().fma(a, b, c), layout)
337 sym::copysignf32 | sym::copysignf64 => {
338 let a = args[0].load_scalar(fx);
339 let b = args[1].load_scalar(fx);
340 CValue::by_val(fx.bcx.ins().fcopysign(a, b), layout)
350 let a = args[0].load_scalar(fx);
352 let val = match intrinsic {
353 sym::fabsf32 | sym::fabsf64 => fx.bcx.ins().fabs(a),
354 sym::floorf32 | sym::floorf64 => fx.bcx.ins().floor(a),
355 sym::ceilf32 | sym::ceilf64 => fx.bcx.ins().ceil(a),
356 sym::truncf32 | sym::truncf64 => fx.bcx.ins().trunc(a),
360 CValue::by_val(val, layout)
362 // These intrinsics aren't supported natively by Cranelift.
363 // Lower them to a libcall.
364 _ => fx.easy_call(name, &args, ty),
367 ret.write_cvalue(fx, res);
372 fn codegen_regular_intrinsic_call<'tcx>(
373 fx: &mut FunctionCx<'_, '_, 'tcx>,
374 instance: Instance<'tcx>,
376 substs: SubstsRef<'tcx>,
377 args: &[mir::Operand<'tcx>],
379 destination: Option<BasicBlock>,
380 source_info: mir::SourceInfo,
382 let usize_layout = fx.layout_of(fx.tcx.types.usize);
385 sym::likely | sym::unlikely => {
386 intrinsic_args!(fx, args => (a); intrinsic);
388 ret.write_cvalue(fx, a);
391 intrinsic_args!(fx, args => (); intrinsic);
393 fx.bcx.ins().debugtrap();
395 sym::copy | sym::copy_nonoverlapping => {
396 intrinsic_args!(fx, args => (src, dst, count); intrinsic);
397 let src = src.load_scalar(fx);
398 let dst = dst.load_scalar(fx);
399 let count = count.load_scalar(fx);
401 let elem_ty = substs.type_at(0);
402 let elem_size: u64 = fx.layout_of(elem_ty).size.bytes();
403 assert_eq!(args.len(), 3);
405 if elem_size != 1 { fx.bcx.ins().imul_imm(count, elem_size as i64) } else { count };
407 if intrinsic == sym::copy_nonoverlapping {
408 // FIXME emit_small_memcpy
409 fx.bcx.call_memcpy(fx.target_config, dst, src, byte_amount);
411 // FIXME emit_small_memmove
412 fx.bcx.call_memmove(fx.target_config, dst, src, byte_amount);
415 sym::volatile_copy_memory | sym::volatile_copy_nonoverlapping_memory => {
416 // NOTE: the volatile variants have src and dst swapped
417 intrinsic_args!(fx, args => (dst, src, count); intrinsic);
418 let dst = dst.load_scalar(fx);
419 let src = src.load_scalar(fx);
420 let count = count.load_scalar(fx);
422 let elem_ty = substs.type_at(0);
423 let elem_size: u64 = fx.layout_of(elem_ty).size.bytes();
424 assert_eq!(args.len(), 3);
426 if elem_size != 1 { fx.bcx.ins().imul_imm(count, elem_size as i64) } else { count };
428 // FIXME make the copy actually volatile when using emit_small_mem{cpy,move}
429 if intrinsic == sym::volatile_copy_nonoverlapping_memory {
430 // FIXME emit_small_memcpy
431 fx.bcx.call_memcpy(fx.target_config, dst, src, byte_amount);
433 // FIXME emit_small_memmove
434 fx.bcx.call_memmove(fx.target_config, dst, src, byte_amount);
437 sym::size_of_val => {
438 intrinsic_args!(fx, args => (ptr); intrinsic);
440 let layout = fx.layout_of(substs.type_at(0));
441 // Note: Can't use is_unsized here as truly unsized types need to take the fixed size
443 let size = if let Abi::ScalarPair(_, _) = ptr.layout().abi {
444 let (_ptr, info) = ptr.load_scalar_pair(fx);
445 let (size, _align) = crate::unsize::size_and_align_of_dst(fx, layout, info);
448 fx.bcx.ins().iconst(fx.pointer_type, layout.size.bytes() as i64)
450 ret.write_cvalue(fx, CValue::by_val(size, usize_layout));
452 sym::min_align_of_val => {
453 intrinsic_args!(fx, args => (ptr); intrinsic);
455 let layout = fx.layout_of(substs.type_at(0));
456 // Note: Can't use is_unsized here as truly unsized types need to take the fixed size
458 let align = if let Abi::ScalarPair(_, _) = ptr.layout().abi {
459 let (_ptr, info) = ptr.load_scalar_pair(fx);
460 let (_size, align) = crate::unsize::size_and_align_of_dst(fx, layout, info);
463 fx.bcx.ins().iconst(fx.pointer_type, layout.align.abi.bytes() as i64)
465 ret.write_cvalue(fx, CValue::by_val(align, usize_layout));
468 sym::vtable_size => {
469 intrinsic_args!(fx, args => (vtable); intrinsic);
470 let vtable = vtable.load_scalar(fx);
472 let size = crate::vtable::size_of_obj(fx, vtable);
473 ret.write_cvalue(fx, CValue::by_val(size, usize_layout));
476 sym::vtable_align => {
477 intrinsic_args!(fx, args => (vtable); intrinsic);
478 let vtable = vtable.load_scalar(fx);
480 let align = crate::vtable::min_align_of_obj(fx, vtable);
481 ret.write_cvalue(fx, CValue::by_val(align, usize_layout));
491 | sym::unchecked_shr => {
492 intrinsic_args!(fx, args => (x, y); intrinsic);
494 // FIXME trap on overflow
495 let bin_op = match intrinsic {
496 sym::unchecked_add => BinOp::Add,
497 sym::unchecked_sub => BinOp::Sub,
498 sym::unchecked_mul => BinOp::Mul,
499 sym::unchecked_div | sym::exact_div => BinOp::Div,
500 sym::unchecked_rem => BinOp::Rem,
501 sym::unchecked_shl => BinOp::Shl,
502 sym::unchecked_shr => BinOp::Shr,
505 let res = crate::num::codegen_int_binop(fx, bin_op, x, y);
506 ret.write_cvalue(fx, res);
508 sym::add_with_overflow | sym::sub_with_overflow | sym::mul_with_overflow => {
509 intrinsic_args!(fx, args => (x, y); intrinsic);
511 assert_eq!(x.layout().ty, y.layout().ty);
512 let bin_op = match intrinsic {
513 sym::add_with_overflow => BinOp::Add,
514 sym::sub_with_overflow => BinOp::Sub,
515 sym::mul_with_overflow => BinOp::Mul,
519 let res = crate::num::codegen_checked_int_binop(fx, bin_op, x, y);
520 ret.write_cvalue(fx, res);
522 sym::saturating_add | sym::saturating_sub => {
523 intrinsic_args!(fx, args => (lhs, rhs); intrinsic);
525 assert_eq!(lhs.layout().ty, rhs.layout().ty);
526 let bin_op = match intrinsic {
527 sym::saturating_add => BinOp::Add,
528 sym::saturating_sub => BinOp::Sub,
532 let res = crate::num::codegen_saturating_int_binop(fx, bin_op, lhs, rhs);
533 ret.write_cvalue(fx, res);
535 sym::rotate_left => {
536 intrinsic_args!(fx, args => (x, y); intrinsic);
537 let y = y.load_scalar(fx);
539 let layout = x.layout();
540 let x = x.load_scalar(fx);
541 let res = fx.bcx.ins().rotl(x, y);
542 ret.write_cvalue(fx, CValue::by_val(res, layout));
544 sym::rotate_right => {
545 intrinsic_args!(fx, args => (x, y); intrinsic);
546 let y = y.load_scalar(fx);
548 let layout = x.layout();
549 let x = x.load_scalar(fx);
550 let res = fx.bcx.ins().rotr(x, y);
551 ret.write_cvalue(fx, CValue::by_val(res, layout));
554 // The only difference between offset and arith_offset is regarding UB. Because Cranelift
555 // doesn't have UB both are codegen'ed the same way
556 sym::offset | sym::arith_offset => {
557 intrinsic_args!(fx, args => (base, offset); intrinsic);
558 let offset = offset.load_scalar(fx);
560 let pointee_ty = base.layout().ty.builtin_deref(true).unwrap().ty;
561 let pointee_size = fx.layout_of(pointee_ty).size.bytes();
562 let ptr_diff = if pointee_size != 1 {
563 fx.bcx.ins().imul_imm(offset, pointee_size as i64)
567 let base_val = base.load_scalar(fx);
568 let res = fx.bcx.ins().iadd(base_val, ptr_diff);
569 ret.write_cvalue(fx, CValue::by_val(res, base.layout()));
573 intrinsic_args!(fx, args => (ptr, mask); intrinsic);
574 let ptr = ptr.load_scalar(fx);
575 let mask = mask.load_scalar(fx);
576 fx.bcx.ins().band(ptr, mask);
580 intrinsic_args!(fx, args => (from); intrinsic);
582 ret.write_cvalue_transmute(fx, from);
584 sym::write_bytes | sym::volatile_set_memory => {
585 intrinsic_args!(fx, args => (dst, val, count); intrinsic);
586 let val = val.load_scalar(fx);
587 let count = count.load_scalar(fx);
589 let pointee_ty = dst.layout().ty.builtin_deref(true).unwrap().ty;
590 let pointee_size = fx.layout_of(pointee_ty).size.bytes();
591 let count = if pointee_size != 1 {
592 fx.bcx.ins().imul_imm(count, pointee_size as i64)
596 let dst_ptr = dst.load_scalar(fx);
597 // FIXME make the memset actually volatile when switching to emit_small_memset
598 // FIXME use emit_small_memset
599 fx.bcx.call_memset(fx.target_config, dst_ptr, val, count);
601 sym::ctlz | sym::ctlz_nonzero => {
602 intrinsic_args!(fx, args => (arg); intrinsic);
603 let val = arg.load_scalar(fx);
605 // FIXME trap on `ctlz_nonzero` with zero arg.
606 let res = fx.bcx.ins().clz(val);
607 let res = CValue::by_val(res, arg.layout());
608 ret.write_cvalue(fx, res);
610 sym::cttz | sym::cttz_nonzero => {
611 intrinsic_args!(fx, args => (arg); intrinsic);
612 let val = arg.load_scalar(fx);
614 // FIXME trap on `cttz_nonzero` with zero arg.
615 let res = fx.bcx.ins().ctz(val);
616 let res = CValue::by_val(res, arg.layout());
617 ret.write_cvalue(fx, res);
620 intrinsic_args!(fx, args => (arg); intrinsic);
621 let val = arg.load_scalar(fx);
623 let res = fx.bcx.ins().popcnt(val);
624 let res = CValue::by_val(res, arg.layout());
625 ret.write_cvalue(fx, res);
628 intrinsic_args!(fx, args => (arg); intrinsic);
629 let val = arg.load_scalar(fx);
631 let res = fx.bcx.ins().bitrev(val);
632 let res = CValue::by_val(res, arg.layout());
633 ret.write_cvalue(fx, res);
636 intrinsic_args!(fx, args => (arg); intrinsic);
637 let val = arg.load_scalar(fx);
639 let res = if fx.bcx.func.dfg.value_type(val) == types::I8 {
642 fx.bcx.ins().bswap(val)
644 let res = CValue::by_val(res, arg.layout());
645 ret.write_cvalue(fx, res);
647 sym::assert_inhabited | sym::assert_zero_valid | sym::assert_mem_uninitialized_valid => {
648 intrinsic_args!(fx, args => (); intrinsic);
650 let layout = fx.layout_of(substs.type_at(0));
651 if layout.abi.is_uninhabited() {
652 with_no_trimmed_paths!({
653 crate::base::codegen_panic(
655 &format!("attempted to instantiate uninhabited type `{}`", layout.ty),
662 if intrinsic == sym::assert_zero_valid && !fx.tcx.permits_zero_init(layout) {
663 with_no_trimmed_paths!({
664 crate::base::codegen_panic(
667 "attempted to zero-initialize type `{}`, which is invalid",
676 if intrinsic == sym::assert_mem_uninitialized_valid
677 && !fx.tcx.permits_uninit_init(layout)
679 with_no_trimmed_paths!({
680 crate::base::codegen_panic(
683 "attempted to leave type `{}` uninitialized, which is invalid",
693 sym::volatile_load | sym::unaligned_volatile_load => {
694 intrinsic_args!(fx, args => (ptr); intrinsic);
696 // Cranelift treats loads as volatile by default
697 // FIXME correctly handle unaligned_volatile_load
698 let inner_layout = fx.layout_of(ptr.layout().ty.builtin_deref(true).unwrap().ty);
699 let val = CValue::by_ref(Pointer::new(ptr.load_scalar(fx)), inner_layout);
700 ret.write_cvalue(fx, val);
702 sym::volatile_store | sym::unaligned_volatile_store => {
703 intrinsic_args!(fx, args => (ptr, val); intrinsic);
704 let ptr = ptr.load_scalar(fx);
706 // Cranelift treats stores as volatile by default
707 // FIXME correctly handle unaligned_volatile_store
708 let dest = CPlace::for_ptr(Pointer::new(ptr), val.layout());
709 dest.write_cvalue(fx, val);
716 | sym::variant_count => {
717 intrinsic_args!(fx, args => (); intrinsic);
720 fx.tcx.const_eval_instance(ParamEnv::reveal_all(), instance, None).unwrap();
721 let val = crate::constant::codegen_const_value(fx, const_val, ret.layout().ty);
722 ret.write_cvalue(fx, val);
725 sym::ptr_offset_from | sym::ptr_offset_from_unsigned => {
726 intrinsic_args!(fx, args => (ptr, base); intrinsic);
727 let ptr = ptr.load_scalar(fx);
728 let base = base.load_scalar(fx);
729 let ty = substs.type_at(0);
731 let pointee_size: u64 = fx.layout_of(ty).size.bytes();
732 let diff_bytes = fx.bcx.ins().isub(ptr, base);
733 // FIXME this can be an exact division.
734 let val = if intrinsic == sym::ptr_offset_from_unsigned {
735 let usize_layout = fx.layout_of(fx.tcx.types.usize);
736 // Because diff_bytes ULE isize::MAX, this would be fine as signed,
737 // but unsigned is slightly easier to codegen, so might as well.
738 CValue::by_val(fx.bcx.ins().udiv_imm(diff_bytes, pointee_size as i64), usize_layout)
740 let isize_layout = fx.layout_of(fx.tcx.types.isize);
741 CValue::by_val(fx.bcx.ins().sdiv_imm(diff_bytes, pointee_size as i64), isize_layout)
743 ret.write_cvalue(fx, val);
746 sym::ptr_guaranteed_cmp => {
747 intrinsic_args!(fx, args => (a, b); intrinsic);
749 let val = crate::num::codegen_ptr_binop(fx, BinOp::Eq, a, b).load_scalar(fx);
750 ret.write_cvalue(fx, CValue::by_val(val, fx.layout_of(fx.tcx.types.u8)));
753 sym::caller_location => {
754 intrinsic_args!(fx, args => (); intrinsic);
756 let caller_location = fx.get_caller_location(source_info);
757 ret.write_cvalue(fx, caller_location);
760 _ if intrinsic.as_str().starts_with("atomic_fence") => {
761 intrinsic_args!(fx, args => (); intrinsic);
763 fx.bcx.ins().fence();
765 _ if intrinsic.as_str().starts_with("atomic_singlethreadfence") => {
766 intrinsic_args!(fx, args => (); intrinsic);
768 // FIXME use a compiler fence once Cranelift supports it
769 fx.bcx.ins().fence();
771 _ if intrinsic.as_str().starts_with("atomic_load") => {
772 intrinsic_args!(fx, args => (ptr); intrinsic);
773 let ptr = ptr.load_scalar(fx);
775 let ty = substs.type_at(0);
777 ty::Uint(UintTy::U128) | ty::Int(IntTy::I128) => {
778 // FIXME implement 128bit atomics
779 if fx.tcx.is_compiler_builtins(LOCAL_CRATE) {
780 // special case for compiler-builtins to avoid having to patch it
781 crate::trap::trap_unimplemented(fx, "128bit atomics not yet supported");
786 .span_fatal(source_info.span, "128bit atomics not yet supported");
789 ty::Uint(_) | ty::Int(_) | ty::RawPtr(..) => {}
791 report_atomic_type_validation_error(fx, intrinsic, source_info.span, ty);
795 let clif_ty = fx.clif_type(ty).unwrap();
797 let val = fx.bcx.ins().atomic_load(clif_ty, MemFlags::trusted(), ptr);
799 let val = CValue::by_val(val, fx.layout_of(ty));
800 ret.write_cvalue(fx, val);
802 _ if intrinsic.as_str().starts_with("atomic_store") => {
803 intrinsic_args!(fx, args => (ptr, val); intrinsic);
804 let ptr = ptr.load_scalar(fx);
806 let ty = substs.type_at(0);
808 ty::Uint(UintTy::U128) | ty::Int(IntTy::I128) => {
809 // FIXME implement 128bit atomics
810 if fx.tcx.is_compiler_builtins(LOCAL_CRATE) {
811 // special case for compiler-builtins to avoid having to patch it
812 crate::trap::trap_unimplemented(fx, "128bit atomics not yet supported");
817 .span_fatal(source_info.span, "128bit atomics not yet supported");
820 ty::Uint(_) | ty::Int(_) | ty::RawPtr(..) => {}
822 report_atomic_type_validation_error(fx, intrinsic, source_info.span, ty);
827 let val = val.load_scalar(fx);
829 fx.bcx.ins().atomic_store(MemFlags::trusted(), val, ptr);
831 _ if intrinsic.as_str().starts_with("atomic_xchg") => {
832 intrinsic_args!(fx, args => (ptr, new); intrinsic);
833 let ptr = ptr.load_scalar(fx);
835 let layout = new.layout();
836 match layout.ty.kind() {
837 ty::Uint(_) | ty::Int(_) | ty::RawPtr(..) => {}
839 report_atomic_type_validation_error(fx, intrinsic, source_info.span, layout.ty);
843 let ty = fx.clif_type(layout.ty).unwrap();
845 let new = new.load_scalar(fx);
847 let old = fx.bcx.ins().atomic_rmw(ty, MemFlags::trusted(), AtomicRmwOp::Xchg, ptr, new);
849 let old = CValue::by_val(old, layout);
850 ret.write_cvalue(fx, old);
852 _ if intrinsic.as_str().starts_with("atomic_cxchg") => {
853 // both atomic_cxchg_* and atomic_cxchgweak_*
854 intrinsic_args!(fx, args => (ptr, test_old, new); intrinsic);
855 let ptr = ptr.load_scalar(fx);
857 let layout = new.layout();
858 match layout.ty.kind() {
859 ty::Uint(_) | ty::Int(_) | ty::RawPtr(..) => {}
861 report_atomic_type_validation_error(fx, intrinsic, source_info.span, layout.ty);
866 let test_old = test_old.load_scalar(fx);
867 let new = new.load_scalar(fx);
869 let old = fx.bcx.ins().atomic_cas(MemFlags::trusted(), ptr, test_old, new);
870 let is_eq = fx.bcx.ins().icmp(IntCC::Equal, old, test_old);
872 let ret_val = CValue::by_val_pair(old, is_eq, ret.layout());
873 ret.write_cvalue(fx, ret_val)
876 _ if intrinsic.as_str().starts_with("atomic_xadd") => {
877 intrinsic_args!(fx, args => (ptr, amount); intrinsic);
878 let ptr = ptr.load_scalar(fx);
880 let layout = amount.layout();
881 match layout.ty.kind() {
882 ty::Uint(_) | ty::Int(_) | ty::RawPtr(..) => {}
884 report_atomic_type_validation_error(fx, intrinsic, source_info.span, layout.ty);
888 let ty = fx.clif_type(layout.ty).unwrap();
890 let amount = amount.load_scalar(fx);
893 fx.bcx.ins().atomic_rmw(ty, MemFlags::trusted(), AtomicRmwOp::Add, ptr, amount);
895 let old = CValue::by_val(old, layout);
896 ret.write_cvalue(fx, old);
898 _ if intrinsic.as_str().starts_with("atomic_xsub") => {
899 intrinsic_args!(fx, args => (ptr, amount); intrinsic);
900 let ptr = ptr.load_scalar(fx);
902 let layout = amount.layout();
903 match layout.ty.kind() {
904 ty::Uint(_) | ty::Int(_) | ty::RawPtr(..) => {}
906 report_atomic_type_validation_error(fx, intrinsic, source_info.span, layout.ty);
910 let ty = fx.clif_type(layout.ty).unwrap();
912 let amount = amount.load_scalar(fx);
915 fx.bcx.ins().atomic_rmw(ty, MemFlags::trusted(), AtomicRmwOp::Sub, ptr, amount);
917 let old = CValue::by_val(old, layout);
918 ret.write_cvalue(fx, old);
920 _ if intrinsic.as_str().starts_with("atomic_and") => {
921 intrinsic_args!(fx, args => (ptr, src); intrinsic);
922 let ptr = ptr.load_scalar(fx);
924 let layout = src.layout();
925 match layout.ty.kind() {
926 ty::Uint(_) | ty::Int(_) | ty::RawPtr(..) => {}
928 report_atomic_type_validation_error(fx, intrinsic, source_info.span, layout.ty);
932 let ty = fx.clif_type(layout.ty).unwrap();
934 let src = src.load_scalar(fx);
936 let old = fx.bcx.ins().atomic_rmw(ty, MemFlags::trusted(), AtomicRmwOp::And, ptr, src);
938 let old = CValue::by_val(old, layout);
939 ret.write_cvalue(fx, old);
941 _ if intrinsic.as_str().starts_with("atomic_or") => {
942 intrinsic_args!(fx, args => (ptr, src); intrinsic);
943 let ptr = ptr.load_scalar(fx);
945 let layout = src.layout();
946 match layout.ty.kind() {
947 ty::Uint(_) | ty::Int(_) | ty::RawPtr(..) => {}
949 report_atomic_type_validation_error(fx, intrinsic, source_info.span, layout.ty);
953 let ty = fx.clif_type(layout.ty).unwrap();
955 let src = src.load_scalar(fx);
957 let old = fx.bcx.ins().atomic_rmw(ty, MemFlags::trusted(), AtomicRmwOp::Or, ptr, src);
959 let old = CValue::by_val(old, layout);
960 ret.write_cvalue(fx, old);
962 _ if intrinsic.as_str().starts_with("atomic_xor") => {
963 intrinsic_args!(fx, args => (ptr, src); intrinsic);
964 let ptr = ptr.load_scalar(fx);
966 let layout = src.layout();
967 match layout.ty.kind() {
968 ty::Uint(_) | ty::Int(_) | ty::RawPtr(..) => {}
970 report_atomic_type_validation_error(fx, intrinsic, source_info.span, layout.ty);
974 let ty = fx.clif_type(layout.ty).unwrap();
976 let src = src.load_scalar(fx);
978 let old = fx.bcx.ins().atomic_rmw(ty, MemFlags::trusted(), AtomicRmwOp::Xor, ptr, src);
980 let old = CValue::by_val(old, layout);
981 ret.write_cvalue(fx, old);
983 _ if intrinsic.as_str().starts_with("atomic_nand") => {
984 intrinsic_args!(fx, args => (ptr, src); intrinsic);
985 let ptr = ptr.load_scalar(fx);
987 let layout = src.layout();
988 match layout.ty.kind() {
989 ty::Uint(_) | ty::Int(_) | ty::RawPtr(..) => {}
991 report_atomic_type_validation_error(fx, intrinsic, source_info.span, layout.ty);
995 let ty = fx.clif_type(layout.ty).unwrap();
997 let src = src.load_scalar(fx);
999 let old = fx.bcx.ins().atomic_rmw(ty, MemFlags::trusted(), AtomicRmwOp::Nand, ptr, src);
1001 let old = CValue::by_val(old, layout);
1002 ret.write_cvalue(fx, old);
1004 _ if intrinsic.as_str().starts_with("atomic_max") => {
1005 intrinsic_args!(fx, args => (ptr, src); intrinsic);
1006 let ptr = ptr.load_scalar(fx);
1008 let layout = src.layout();
1009 match layout.ty.kind() {
1010 ty::Uint(_) | ty::Int(_) | ty::RawPtr(..) => {}
1012 report_atomic_type_validation_error(fx, intrinsic, source_info.span, layout.ty);
1016 let ty = fx.clif_type(layout.ty).unwrap();
1018 let src = src.load_scalar(fx);
1020 let old = fx.bcx.ins().atomic_rmw(ty, MemFlags::trusted(), AtomicRmwOp::Smax, ptr, src);
1022 let old = CValue::by_val(old, layout);
1023 ret.write_cvalue(fx, old);
1025 _ if intrinsic.as_str().starts_with("atomic_umax") => {
1026 intrinsic_args!(fx, args => (ptr, src); intrinsic);
1027 let ptr = ptr.load_scalar(fx);
1029 let layout = src.layout();
1030 match layout.ty.kind() {
1031 ty::Uint(_) | ty::Int(_) | ty::RawPtr(..) => {}
1033 report_atomic_type_validation_error(fx, intrinsic, source_info.span, layout.ty);
1037 let ty = fx.clif_type(layout.ty).unwrap();
1039 let src = src.load_scalar(fx);
1041 let old = fx.bcx.ins().atomic_rmw(ty, MemFlags::trusted(), AtomicRmwOp::Umax, ptr, src);
1043 let old = CValue::by_val(old, layout);
1044 ret.write_cvalue(fx, old);
1046 _ if intrinsic.as_str().starts_with("atomic_min") => {
1047 intrinsic_args!(fx, args => (ptr, src); intrinsic);
1048 let ptr = ptr.load_scalar(fx);
1050 let layout = src.layout();
1051 match layout.ty.kind() {
1052 ty::Uint(_) | ty::Int(_) | ty::RawPtr(..) => {}
1054 report_atomic_type_validation_error(fx, intrinsic, source_info.span, layout.ty);
1058 let ty = fx.clif_type(layout.ty).unwrap();
1060 let src = src.load_scalar(fx);
1062 let old = fx.bcx.ins().atomic_rmw(ty, MemFlags::trusted(), AtomicRmwOp::Smin, ptr, src);
1064 let old = CValue::by_val(old, layout);
1065 ret.write_cvalue(fx, old);
1067 _ if intrinsic.as_str().starts_with("atomic_umin") => {
1068 intrinsic_args!(fx, args => (ptr, src); intrinsic);
1069 let ptr = ptr.load_scalar(fx);
1071 let layout = src.layout();
1072 match layout.ty.kind() {
1073 ty::Uint(_) | ty::Int(_) | ty::RawPtr(..) => {}
1075 report_atomic_type_validation_error(fx, intrinsic, source_info.span, layout.ty);
1079 let ty = fx.clif_type(layout.ty).unwrap();
1081 let src = src.load_scalar(fx);
1083 let old = fx.bcx.ins().atomic_rmw(ty, MemFlags::trusted(), AtomicRmwOp::Umin, ptr, src);
1085 let old = CValue::by_val(old, layout);
1086 ret.write_cvalue(fx, old);
1090 intrinsic_args!(fx, args => (a, b); intrinsic);
1091 let a = a.load_scalar(fx);
1092 let b = b.load_scalar(fx);
1094 let val = crate::num::codegen_float_min(fx, a, b);
1095 let val = CValue::by_val(val, fx.layout_of(fx.tcx.types.f32));
1096 ret.write_cvalue(fx, val);
1099 intrinsic_args!(fx, args => (a, b); intrinsic);
1100 let a = a.load_scalar(fx);
1101 let b = b.load_scalar(fx);
1103 let val = crate::num::codegen_float_min(fx, a, b);
1104 let val = CValue::by_val(val, fx.layout_of(fx.tcx.types.f64));
1105 ret.write_cvalue(fx, val);
1108 intrinsic_args!(fx, args => (a, b); intrinsic);
1109 let a = a.load_scalar(fx);
1110 let b = b.load_scalar(fx);
1112 let val = crate::num::codegen_float_max(fx, a, b);
1113 let val = CValue::by_val(val, fx.layout_of(fx.tcx.types.f32));
1114 ret.write_cvalue(fx, val);
1117 intrinsic_args!(fx, args => (a, b); intrinsic);
1118 let a = a.load_scalar(fx);
1119 let b = b.load_scalar(fx);
1121 let val = crate::num::codegen_float_max(fx, a, b);
1122 let val = CValue::by_val(val, fx.layout_of(fx.tcx.types.f64));
1123 ret.write_cvalue(fx, val);
1127 intrinsic_args!(fx, args => (f, data, catch_fn); intrinsic);
1128 let f = f.load_scalar(fx);
1129 let data = data.load_scalar(fx);
1130 let _catch_fn = catch_fn.load_scalar(fx);
1132 // FIXME once unwinding is supported, change this to actually catch panics
1133 let f_sig = fx.bcx.func.import_signature(Signature {
1134 call_conv: fx.target_config.default_call_conv,
1135 params: vec![AbiParam::new(pointer_ty(fx.tcx))],
1139 fx.bcx.ins().call_indirect(f_sig, f, &[data]);
1141 let layout = ret.layout();
1142 let ret_val = CValue::const_val(fx, layout, ty::ScalarInt::null(layout.size));
1143 ret.write_cvalue(fx, ret_val);
1146 sym::fadd_fast | sym::fsub_fast | sym::fmul_fast | sym::fdiv_fast | sym::frem_fast => {
1147 intrinsic_args!(fx, args => (x, y); intrinsic);
1149 let res = crate::num::codegen_float_binop(
1152 sym::fadd_fast => BinOp::Add,
1153 sym::fsub_fast => BinOp::Sub,
1154 sym::fmul_fast => BinOp::Mul,
1155 sym::fdiv_fast => BinOp::Div,
1156 sym::frem_fast => BinOp::Rem,
1157 _ => unreachable!(),
1162 ret.write_cvalue(fx, res);
1164 sym::float_to_int_unchecked => {
1165 intrinsic_args!(fx, args => (f); intrinsic);
1166 let f = f.load_scalar(fx);
1168 let res = crate::cast::clif_int_or_float_cast(
1172 fx.clif_type(ret.layout().ty).unwrap(),
1173 type_sign(ret.layout().ty),
1175 ret.write_cvalue(fx, CValue::by_val(res, ret.layout()));
1179 intrinsic_args!(fx, args => (lhs_ref, rhs_ref); intrinsic);
1180 let lhs_ref = lhs_ref.load_scalar(fx);
1181 let rhs_ref = rhs_ref.load_scalar(fx);
1183 let size = fx.layout_of(substs.type_at(0)).layout.size();
1184 // FIXME add and use emit_small_memcmp
1185 let is_eq_value = if size == Size::ZERO {
1186 // No bytes means they're trivially equal
1187 fx.bcx.ins().iconst(types::I8, 1)
1188 } else if let Some(clty) = size.bits().try_into().ok().and_then(Type::int) {
1189 // Can't use `trusted` for these loads; they could be unaligned.
1190 let mut flags = MemFlags::new();
1192 let lhs_val = fx.bcx.ins().load(clty, flags, lhs_ref, 0);
1193 let rhs_val = fx.bcx.ins().load(clty, flags, rhs_ref, 0);
1194 fx.bcx.ins().icmp(IntCC::Equal, lhs_val, rhs_val)
1196 // Just call `memcmp` (like slices do in core) when the
1197 // size is too large or it's not a power-of-two.
1198 let signed_bytes = i64::try_from(size.bytes()).unwrap();
1199 let bytes_val = fx.bcx.ins().iconst(fx.pointer_type, signed_bytes);
1200 let params = vec![AbiParam::new(fx.pointer_type); 3];
1201 let returns = vec![AbiParam::new(types::I32)];
1202 let args = &[lhs_ref, rhs_ref, bytes_val];
1203 let cmp = fx.lib_call("memcmp", params, returns, args)[0];
1204 fx.bcx.ins().icmp_imm(IntCC::Equal, cmp, 0)
1206 ret.write_cvalue(fx, CValue::by_val(is_eq_value, ret.layout()));
1209 sym::const_allocate => {
1210 intrinsic_args!(fx, args => (_size, _align); intrinsic);
1212 // returns a null pointer at runtime.
1213 let null = fx.bcx.ins().iconst(fx.pointer_type, 0);
1214 ret.write_cvalue(fx, CValue::by_val(null, ret.layout()));
1217 sym::const_deallocate => {
1218 intrinsic_args!(fx, args => (_ptr, _size, _align); intrinsic);
1223 intrinsic_args!(fx, args => (a); intrinsic);
1225 // FIXME implement black_box semantics
1226 ret.write_cvalue(fx, a);
1229 // FIXME implement variadics in cranelift
1230 sym::va_copy | sym::va_arg | sym::va_end => {
1231 fx.tcx.sess.span_fatal(
1233 "Defining variadic functions is not yet supported by Cranelift",
1240 .span_fatal(source_info.span, &format!("unsupported intrinsic {}", intrinsic));
1244 let ret_block = fx.get_block(destination.unwrap());
1245 fx.bcx.ins().jump(ret_block, &[]);