]> git.lizzy.rs Git - rust.git/blobdiff - src/shims/intrinsics.rs
Auto merge of #2162 - RalfJung:rustup, r=RalfJung
[rust.git] / src / shims / intrinsics.rs
index 49c9c0fb0d9657beb25df8017ee959b128d13145..1f06971a3e70db0128c06876f4c3dc95f451bbe4 100644 (file)
@@ -3,12 +3,12 @@
 use log::trace;
 
 use rustc_apfloat::{Float, Round};
-use rustc_middle::ty::layout::{IntegerExt, LayoutOf};
+use rustc_middle::ty::layout::{HasParamEnv, IntegerExt, LayoutOf};
 use rustc_middle::{mir, mir::BinOp, ty, ty::FloatTy};
-use rustc_target::abi::{Align, Integer};
+use rustc_target::abi::{Align, Endian, HasDataLayout, Integer, Size};
 
 use crate::*;
-use helpers::{bool_to_simd_element, check_arg_count, simd_element_to_bool};
+use helpers::check_arg_count;
 
 pub enum AtomicOp {
     MirOp(mir::BinOp, bool),
@@ -22,19 +22,20 @@ fn call_intrinsic(
         &mut self,
         instance: ty::Instance<'tcx>,
         args: &[OpTy<'tcx, Tag>],
-        ret: Option<(&PlaceTy<'tcx, Tag>, mir::BasicBlock)>,
+        dest: &PlaceTy<'tcx, Tag>,
+        ret: Option<mir::BasicBlock>,
         _unwind: StackPopUnwind,
     ) -> InterpResult<'tcx> {
         let this = self.eval_context_mut();
 
-        if this.emulate_intrinsic(instance, args, ret)? {
+        if this.emulate_intrinsic(instance, args, dest, ret)? {
             return Ok(());
         }
 
         // All supported intrinsics have a return place.
         let intrinsic_name = this.tcx.item_name(instance.def_id());
         let intrinsic_name = intrinsic_name.as_str();
-        let (dest, ret) = match ret {
+        let ret = match ret {
             None => throw_unsup_format!("unimplemented (diverging) intrinsic: {}", intrinsic_name),
             Some(p) => p,
         };
@@ -43,13 +44,13 @@ fn call_intrinsic(
         match intrinsic_name {
             // Miri overwriting CTFE intrinsics.
             "ptr_guaranteed_eq" => {
-                let &[ref left, ref right] = check_arg_count(args)?;
+                let [left, right] = check_arg_count(args)?;
                 let left = this.read_immediate(left)?;
                 let right = this.read_immediate(right)?;
                 this.binop_ignore_overflow(mir::BinOp::Eq, &left, &right, dest)?;
             }
             "ptr_guaranteed_ne" => {
-                let &[ref left, ref right] = check_arg_count(args)?;
+                let [left, right] = check_arg_count(args)?;
                 let left = this.read_immediate(left)?;
                 let right = this.read_immediate(right)?;
                 this.binop_ignore_overflow(mir::BinOp::Ne, &left, &right, dest)?;
@@ -65,34 +66,49 @@ fn call_intrinsic(
 
             // Raw memory accesses
             "volatile_load" => {
-                let &[ref place] = check_arg_count(args)?;
+                let [place] = check_arg_count(args)?;
                 let place = this.deref_operand(place)?;
                 this.copy_op(&place.into(), dest)?;
             }
             "volatile_store" => {
-                let &[ref place, ref dest] = check_arg_count(args)?;
+                let [place, dest] = check_arg_count(args)?;
                 let place = this.deref_operand(place)?;
                 this.copy_op(dest, &place.into())?;
             }
 
             "write_bytes" | "volatile_set_memory" => {
-                let &[ref ptr, ref val_byte, ref count] = check_arg_count(args)?;
+                let [ptr, val_byte, count] = check_arg_count(args)?;
                 let ty = instance.substs.type_at(0);
                 let ty_layout = this.layout_of(ty)?;
                 let val_byte = this.read_scalar(val_byte)?.to_u8()?;
                 let ptr = this.read_pointer(ptr)?;
                 let count = this.read_scalar(count)?.to_machine_usize(this)?;
+                // `checked_mul` enforces a too small bound (the correct one would probably be machine_isize_max),
+                // but no actual allocation can be big enough for the difference to be noticeable.
                 let byte_count = ty_layout.size.checked_mul(count, this).ok_or_else(|| {
                     err_ub_format!("overflow computing total size of `{}`", intrinsic_name)
                 })?;
-                this.memory
-                    .write_bytes(ptr, iter::repeat(val_byte).take(byte_count.bytes() as usize))?;
+                this.write_bytes_ptr(
+                    ptr,
+                    iter::repeat(val_byte).take(byte_count.bytes() as usize),
+                )?;
             }
 
             // Floating-point operations
+            "fabsf32" => {
+                let [f] = check_arg_count(args)?;
+                let f = this.read_scalar(f)?.to_f32()?;
+                // Can be implemented in soft-floats.
+                this.write_scalar(Scalar::from_f32(f.abs()), dest)?;
+            }
+            "fabsf64" => {
+                let [f] = check_arg_count(args)?;
+                let f = this.read_scalar(f)?.to_f64()?;
+                // Can be implemented in soft-floats.
+                this.write_scalar(Scalar::from_f64(f.abs()), dest)?;
+            }
             #[rustfmt::skip]
             | "sinf32"
-            | "fabsf32"
             | "cosf32"
             | "sqrtf32"
             | "expf32"
@@ -105,12 +121,11 @@ fn call_intrinsic(
             | "truncf32"
             | "roundf32"
             => {
-                let &[ref f] = check_arg_count(args)?;
+                let [f] = check_arg_count(args)?;
                 // FIXME: Using host floats.
                 let f = f32::from_bits(this.read_scalar(f)?.to_u32()?);
                 let f = match intrinsic_name {
                     "sinf32" => f.sin(),
-                    "fabsf32" => f.abs(),
                     "cosf32" => f.cos(),
                     "sqrtf32" => f.sqrt(),
                     "expf32" => f.exp(),
@@ -129,7 +144,6 @@ fn call_intrinsic(
 
             #[rustfmt::skip]
             | "sinf64"
-            | "fabsf64"
             | "cosf64"
             | "sqrtf64"
             | "expf64"
@@ -142,12 +156,11 @@ fn call_intrinsic(
             | "truncf64"
             | "roundf64"
             => {
-                let &[ref f] = check_arg_count(args)?;
+                let [f] = check_arg_count(args)?;
                 // FIXME: Using host floats.
                 let f = f64::from_bits(this.read_scalar(f)?.to_u64()?);
                 let f = match intrinsic_name {
                     "sinf64" => f.sin(),
-                    "fabsf64" => f.abs(),
                     "cosf64" => f.cos(),
                     "sqrtf64" => f.sqrt(),
                     "expf64" => f.exp(),
@@ -171,7 +184,7 @@ fn call_intrinsic(
             | "fdiv_fast"
             | "frem_fast"
             => {
-                let &[ref a, ref b] = check_arg_count(args)?;
+                let [a, b] = check_arg_count(args)?;
                 let a = this.read_immediate(a)?;
                 let b = this.read_immediate(b)?;
                 let op = match intrinsic_name {
@@ -216,7 +229,7 @@ fn call_intrinsic(
             | "maxnumf32"
             | "copysignf32"
             => {
-                let &[ref a, ref b] = check_arg_count(args)?;
+                let [a, b] = check_arg_count(args)?;
                 let a = this.read_scalar(a)?.to_f32()?;
                 let b = this.read_scalar(b)?.to_f32()?;
                 let res = match intrinsic_name {
@@ -233,7 +246,7 @@ fn call_intrinsic(
             | "maxnumf64"
             | "copysignf64"
             => {
-                let &[ref a, ref b] = check_arg_count(args)?;
+                let [a, b] = check_arg_count(args)?;
                 let a = this.read_scalar(a)?.to_f64()?;
                 let b = this.read_scalar(b)?.to_f64()?;
                 let res = match intrinsic_name {
@@ -246,7 +259,7 @@ fn call_intrinsic(
             }
 
             "powf32" => {
-                let &[ref f, ref f2] = check_arg_count(args)?;
+                let [f, f2] = check_arg_count(args)?;
                 // FIXME: Using host floats.
                 let f = f32::from_bits(this.read_scalar(f)?.to_u32()?);
                 let f2 = f32::from_bits(this.read_scalar(f2)?.to_u32()?);
@@ -254,7 +267,7 @@ fn call_intrinsic(
             }
 
             "powf64" => {
-                let &[ref f, ref f2] = check_arg_count(args)?;
+                let [f, f2] = check_arg_count(args)?;
                 // FIXME: Using host floats.
                 let f = f64::from_bits(this.read_scalar(f)?.to_u64()?);
                 let f2 = f64::from_bits(this.read_scalar(f2)?.to_u64()?);
@@ -262,7 +275,7 @@ fn call_intrinsic(
             }
 
             "fmaf32" => {
-                let &[ref a, ref b, ref c] = check_arg_count(args)?;
+                let [a, b, c] = check_arg_count(args)?;
                 let a = this.read_scalar(a)?.to_f32()?;
                 let b = this.read_scalar(b)?.to_f32()?;
                 let c = this.read_scalar(c)?.to_f32()?;
@@ -271,7 +284,7 @@ fn call_intrinsic(
             }
 
             "fmaf64" => {
-                let &[ref a, ref b, ref c] = check_arg_count(args)?;
+                let [a, b, c] = check_arg_count(args)?;
                 let a = this.read_scalar(a)?.to_f64()?;
                 let b = this.read_scalar(b)?.to_f64()?;
                 let c = this.read_scalar(c)?.to_f64()?;
@@ -280,7 +293,7 @@ fn call_intrinsic(
             }
 
             "powif32" => {
-                let &[ref f, ref i] = check_arg_count(args)?;
+                let [f, i] = check_arg_count(args)?;
                 // FIXME: Using host floats.
                 let f = f32::from_bits(this.read_scalar(f)?.to_u32()?);
                 let i = this.read_scalar(i)?.to_i32()?;
@@ -288,7 +301,7 @@ fn call_intrinsic(
             }
 
             "powif64" => {
-                let &[ref f, ref i] = check_arg_count(args)?;
+                let [f, i] = check_arg_count(args)?;
                 // FIXME: Using host floats.
                 let f = f64::from_bits(this.read_scalar(f)?.to_u64()?);
                 let i = this.read_scalar(i)?.to_i32()?;
@@ -296,7 +309,7 @@ fn call_intrinsic(
             }
 
             "float_to_int_unchecked" => {
-                let &[ref val] = check_arg_count(args)?;
+                let [val] = check_arg_count(args)?;
                 let val = this.read_immediate(val)?;
 
                 let res = match val.layout.ty.kind() {
@@ -317,20 +330,40 @@ fn call_intrinsic(
             // SIMD operations
             #[rustfmt::skip]
             | "simd_neg"
-            | "simd_fabs" => {
-                let &[ref op] = check_arg_count(args)?;
+            | "simd_fabs"
+            | "simd_ceil"
+            | "simd_floor"
+            | "simd_round"
+            | "simd_trunc"
+            | "simd_fsqrt" => {
+                let [op] = check_arg_count(args)?;
                 let (op, op_len) = this.operand_to_simd(op)?;
                 let (dest, dest_len) = this.place_to_simd(dest)?;
 
                 assert_eq!(dest_len, op_len);
 
+                #[derive(Copy, Clone)]
+                enum HostFloatOp {
+                    Ceil,
+                    Floor,
+                    Round,
+                    Trunc,
+                    Sqrt,
+                }
+                #[derive(Copy, Clone)]
                 enum Op {
                     MirOp(mir::UnOp),
                     Abs,
+                    HostOp(HostFloatOp),
                 }
                 let which = match intrinsic_name {
                     "simd_neg" => Op::MirOp(mir::UnOp::Neg),
                     "simd_fabs" => Op::Abs,
+                    "simd_ceil" => Op::HostOp(HostFloatOp::Ceil),
+                    "simd_floor" => Op::HostOp(HostFloatOp::Floor),
+                    "simd_round" => Op::HostOp(HostFloatOp::Round),
+                    "simd_trunc" => Op::HostOp(HostFloatOp::Trunc),
+                    "simd_fsqrt" => Op::HostOp(HostFloatOp::Sqrt),
                     _ => unreachable!(),
                 };
 
@@ -342,7 +375,7 @@ enum Op {
                         Op::Abs => {
                             // Works for f32 and f64.
                             let ty::Float(float_ty) = op.layout.ty.kind() else {
-                                bug!("simd_fabs operand is not a float")
+                                bug!("{} operand is not a float", intrinsic_name)
                             };
                             let op = op.to_scalar()?;
                             match float_ty {
@@ -350,6 +383,37 @@ enum Op {
                                 FloatTy::F64 => Scalar::from_f64(op.to_f64()?.abs()),
                             }
                         }
+                        Op::HostOp(host_op) => {
+                            let ty::Float(float_ty) = op.layout.ty.kind() else {
+                                bug!("{} operand is not a float", intrinsic_name)
+                            };
+                            // FIXME using host floats
+                            match float_ty {
+                                FloatTy::F32 => {
+                                    let f = f32::from_bits(op.to_scalar()?.to_u32()?);
+                                    let res = match host_op {
+                                        HostFloatOp::Ceil => f.ceil(),
+                                        HostFloatOp::Floor => f.floor(),
+                                        HostFloatOp::Round => f.round(),
+                                        HostFloatOp::Trunc => f.trunc(),
+                                        HostFloatOp::Sqrt => f.sqrt(),
+                                    };
+                                    Scalar::from_u32(res.to_bits())
+                                }
+                                FloatTy::F64 => {
+                                    let f = f64::from_bits(op.to_scalar()?.to_u64()?);
+                                    let res = match host_op {
+                                        HostFloatOp::Ceil => f.ceil(),
+                                        HostFloatOp::Floor => f.floor(),
+                                        HostFloatOp::Round => f.round(),
+                                        HostFloatOp::Trunc => f.trunc(),
+                                        HostFloatOp::Sqrt => f.sqrt(),
+                                    };
+                                    Scalar::from_u64(res.to_bits())
+                                }
+                            }
+
+                        }
                     };
                     this.write_scalar(val, &dest.into())?;
                 }
@@ -374,10 +438,11 @@ enum Op {
             | "simd_fmax"
             | "simd_fmin"
             | "simd_saturating_add"
-            | "simd_saturating_sub" => {
+            | "simd_saturating_sub"
+            | "simd_arith_offset" => {
                 use mir::BinOp;
 
-                let &[ref left, ref right] = check_arg_count(args)?;
+                let [left, right] = check_arg_count(args)?;
                 let (left, left_len) = this.operand_to_simd(left)?;
                 let (right, right_len) = this.operand_to_simd(right)?;
                 let (dest, dest_len) = this.place_to_simd(dest)?;
@@ -390,6 +455,7 @@ enum Op {
                     SaturatingOp(BinOp),
                     FMax,
                     FMin,
+                    WrappingOffset,
                 }
                 let which = match intrinsic_name {
                     "simd_add" => Op::MirOp(BinOp::Add),
@@ -412,6 +478,7 @@ enum Op {
                     "simd_fmin" => Op::FMin,
                     "simd_saturating_add" => Op::SaturatingOp(BinOp::Add),
                     "simd_saturating_sub" => Op::SaturatingOp(BinOp::Sub),
+                    "simd_arith_offset" => Op::WrappingOffset,
                     _ => unreachable!(),
                 };
 
@@ -441,15 +508,55 @@ enum Op {
                                 val
                             }
                         }
+                        Op::SaturatingOp(mir_op) => {
+                            this.saturating_arith(mir_op, &left, &right)?
+                        }
+                        Op::WrappingOffset => {
+                            let ptr = this.scalar_to_ptr(left.to_scalar()?)?;
+                            let offset_count = right.to_scalar()?.to_machine_isize(this)?;
+                            let pointee_ty = left.layout.ty.builtin_deref(true).unwrap().ty;
+
+                            let pointee_size = i64::try_from(this.layout_of(pointee_ty)?.size.bytes()).unwrap();
+                            let offset_bytes = offset_count.wrapping_mul(pointee_size);
+                            let offset_ptr = ptr.wrapping_signed_offset(offset_bytes, this);
+                            Scalar::from_maybe_pointer(offset_ptr, this)
+                        }
                         Op::FMax => {
                             fmax_op(&left, &right)?
                         }
                         Op::FMin => {
                             fmin_op(&left, &right)?
                         }
-                        Op::SaturatingOp(mir_op) => {
-                            this.saturating_arith(mir_op, &left, &right)?
-                        }
+                    };
+                    this.write_scalar(val, &dest.into())?;
+                }
+            }
+            "simd_fma" => {
+                let [a, b, c] = check_arg_count(args)?;
+                let (a, a_len) = this.operand_to_simd(a)?;
+                let (b, b_len) = this.operand_to_simd(b)?;
+                let (c, c_len) = this.operand_to_simd(c)?;
+                let (dest, dest_len) = this.place_to_simd(dest)?;
+
+                assert_eq!(dest_len, a_len);
+                assert_eq!(dest_len, b_len);
+                assert_eq!(dest_len, c_len);
+
+                for i in 0..dest_len {
+                    let a = this.read_immediate(&this.mplace_index(&a, i)?.into())?.to_scalar()?;
+                    let b = this.read_immediate(&this.mplace_index(&b, i)?.into())?.to_scalar()?;
+                    let c = this.read_immediate(&this.mplace_index(&c, i)?.into())?.to_scalar()?;
+                    let dest = this.mplace_index(&dest, i)?;
+
+                    // Works for f32 and f64.
+                    let ty::Float(float_ty) = dest.layout.ty.kind() else {
+                        bug!("{} operand is not a float", intrinsic_name)
+                    };
+                    let val = match float_ty {
+                        FloatTy::F32 =>
+                            Scalar::from_f32(a.to_f32()?.mul_add(b.to_f32()?, c.to_f32()?).value),
+                        FloatTy::F64 =>
+                            Scalar::from_f64(a.to_f64()?.mul_add(b.to_f64()?, c.to_f64()?).value),
                     };
                     this.write_scalar(val, &dest.into())?;
                 }
@@ -464,7 +571,7 @@ enum Op {
             | "simd_reduce_min" => {
                 use mir::BinOp;
 
-                let &[ref op] = check_arg_count(args)?;
+                let [op] = check_arg_count(args)?;
                 let (op, op_len) = this.operand_to_simd(op)?;
 
                 let imm_from_bool =
@@ -536,7 +643,7 @@ enum Op {
             | "simd_reduce_mul_ordered" => {
                 use mir::BinOp;
 
-                let &[ref op, ref init] = check_arg_count(args)?;
+                let [op, init] = check_arg_count(args)?;
                 let (op, op_len) = this.operand_to_simd(op)?;
                 let init = this.read_immediate(init)?;
 
@@ -554,7 +661,7 @@ enum Op {
                 this.write_immediate(*res, dest)?;
             }
             "simd_select" => {
-                let &[ref mask, ref yes, ref no] = check_arg_count(args)?;
+                let [mask, yes, no] = check_arg_count(args)?;
                 let (mask, mask_len) = this.operand_to_simd(mask)?;
                 let (yes, yes_len) = this.operand_to_simd(yes)?;
                 let (no, no_len) = this.operand_to_simd(no)?;
@@ -570,14 +677,52 @@ enum Op {
                     let no = this.read_immediate(&this.mplace_index(&no, i)?.into())?;
                     let dest = this.mplace_index(&dest, i)?;
 
-                    let mask = simd_element_to_bool(mask)?;
-                    let val = if mask { yes } else { no };
+                    let val = if simd_element_to_bool(mask)? { yes } else { no };
+                    this.write_immediate(*val, &dest.into())?;
+                }
+            }
+            "simd_select_bitmask" => {
+                let [mask, yes, no] = check_arg_count(args)?;
+                let (yes, yes_len) = this.operand_to_simd(yes)?;
+                let (no, no_len) = this.operand_to_simd(no)?;
+                let (dest, dest_len) = this.place_to_simd(dest)?;
+                let bitmask_len = dest_len.max(8);
+
+                assert!(mask.layout.ty.is_integral());
+                assert!(bitmask_len <= 64);
+                assert_eq!(bitmask_len, mask.layout.size.bits());
+                assert_eq!(dest_len, yes_len);
+                assert_eq!(dest_len, no_len);
+
+                let mask: u64 = this
+                    .read_scalar(mask)?
+                    .check_init()?
+                    .to_bits(mask.layout.size)?
+                    .try_into()
+                    .unwrap();
+                for i in 0..dest_len {
+                    let mask =
+                        mask & (1 << simd_bitmask_index(i, dest_len, this.data_layout().endian));
+                    let yes = this.read_immediate(&this.mplace_index(&yes, i)?.into())?;
+                    let no = this.read_immediate(&this.mplace_index(&no, i)?.into())?;
+                    let dest = this.mplace_index(&dest, i)?;
+
+                    let val = if mask != 0 { yes } else { no };
                     this.write_immediate(*val, &dest.into())?;
                 }
+                for i in dest_len..bitmask_len {
+                    // If the mask is "padded", ensure that padding is all-zero.
+                    let mask = mask & (1 << i);
+                    if mask != 0 {
+                        throw_ub_format!(
+                            "a SIMD bitmask less than 8 bits long must be filled with 0s for the remaining bits"
+                        );
+                    }
+                }
             }
             #[rustfmt::skip]
             "simd_cast" | "simd_as" => {
-                let &[ref op] = check_arg_count(args)?;
+                let [op] = check_arg_count(args)?;
                 let (op, op_len) = this.operand_to_simd(op)?;
                 let (dest, dest_len) = this.place_to_simd(dest)?;
 
@@ -614,6 +759,109 @@ enum Op {
                     this.write_immediate(val, &dest.into())?;
                 }
             }
+            "simd_shuffle" => {
+                let [left, right, index] = check_arg_count(args)?;
+                let (left, left_len) = this.operand_to_simd(left)?;
+                let (right, right_len) = this.operand_to_simd(right)?;
+                let (dest, dest_len) = this.place_to_simd(dest)?;
+
+                // `index` is an array, not a SIMD type
+                let ty::Array(_, index_len) = index.layout.ty.kind() else {
+                    bug!("simd_shuffle index argument has non-array type {}", index.layout.ty)
+                };
+                let index_len = index_len.eval_usize(*this.tcx, this.param_env());
+
+                assert_eq!(left_len, right_len);
+                assert_eq!(index_len, dest_len);
+
+                for i in 0..dest_len {
+                    let src_index: u64 = this
+                        .read_immediate(&this.operand_index(index, i)?)?
+                        .to_scalar()?
+                        .to_u32()?
+                        .into();
+                    let dest = this.mplace_index(&dest, i)?;
+
+                    let val = if src_index < left_len {
+                        this.read_immediate(&this.mplace_index(&left, src_index)?.into())?
+                    } else if src_index < left_len.checked_add(right_len).unwrap() {
+                        this.read_immediate(
+                            &this.mplace_index(&right, src_index - left_len)?.into(),
+                        )?
+                    } else {
+                        bug!(
+                            "simd_shuffle index {} is out of bounds for 2 vectors of size {}",
+                            src_index,
+                            left_len
+                        );
+                    };
+                    this.write_immediate(*val, &dest.into())?;
+                }
+            }
+            "simd_gather" => {
+                let [passthru, ptrs, mask] = check_arg_count(args)?;
+                let (passthru, passthru_len) = this.operand_to_simd(passthru)?;
+                let (ptrs, ptrs_len) = this.operand_to_simd(ptrs)?;
+                let (mask, mask_len) = this.operand_to_simd(mask)?;
+                let (dest, dest_len) = this.place_to_simd(dest)?;
+
+                assert_eq!(dest_len, passthru_len);
+                assert_eq!(dest_len, ptrs_len);
+                assert_eq!(dest_len, mask_len);
+
+                for i in 0..dest_len {
+                    let passthru = this.read_immediate(&this.mplace_index(&passthru, i)?.into())?;
+                    let ptr = this.read_immediate(&this.mplace_index(&ptrs, i)?.into())?;
+                    let mask = this.read_immediate(&this.mplace_index(&mask, i)?.into())?;
+                    let dest = this.mplace_index(&dest, i)?;
+
+                    let val = if simd_element_to_bool(mask)? {
+                        let place = this.deref_operand(&ptr.into())?;
+                        this.read_immediate(&place.into())?
+                    } else {
+                        passthru
+                    };
+                    this.write_immediate(*val, &dest.into())?;
+                }
+            }
+            "simd_scatter" => {
+                let [value, ptrs, mask] = check_arg_count(args)?;
+                let (value, value_len) = this.operand_to_simd(value)?;
+                let (ptrs, ptrs_len) = this.operand_to_simd(ptrs)?;
+                let (mask, mask_len) = this.operand_to_simd(mask)?;
+
+                assert_eq!(ptrs_len, value_len);
+                assert_eq!(ptrs_len, mask_len);
+
+                for i in 0..ptrs_len {
+                    let value = this.read_immediate(&this.mplace_index(&value, i)?.into())?;
+                    let ptr = this.read_immediate(&this.mplace_index(&ptrs, i)?.into())?;
+                    let mask = this.read_immediate(&this.mplace_index(&mask, i)?.into())?;
+
+                    if simd_element_to_bool(mask)? {
+                        let place = this.deref_operand(&ptr.into())?;
+                        this.write_immediate(*value, &place.into())?;
+                    }
+                }
+            }
+            "simd_bitmask" => {
+                let [op] = check_arg_count(args)?;
+                let (op, op_len) = this.operand_to_simd(op)?;
+                let bitmask_len = op_len.max(8);
+
+                assert!(dest.layout.ty.is_integral());
+                assert!(bitmask_len <= 64);
+                assert_eq!(bitmask_len, dest.layout.size.bits());
+
+                let mut res = 0u64;
+                for i in 0..op_len {
+                    let op = this.read_immediate(&this.mplace_index(&op, i)?.into())?;
+                    if simd_element_to_bool(op)? {
+                        res |= 1 << simd_bitmask_index(i, op_len, this.data_layout().endian);
+                    }
+                }
+                this.write_int(res, dest)?;
+            }
 
             // Atomic operations
             "atomic_load" => this.atomic_load(args, dest, AtomicReadOp::SeqCst)?,
@@ -816,14 +1064,14 @@ enum Op {
 
             // Other
             "exact_div" => {
-                let &[ref num, ref denom] = check_arg_count(args)?;
+                let [num, denom] = check_arg_count(args)?;
                 this.exact_div(&this.read_immediate(num)?, &this.read_immediate(denom)?, dest)?;
             }
 
             "try" => return this.handle_try(args, dest, ret),
 
             "breakpoint" => {
-                let &[] = check_arg_count(args)?;
+                let [] = check_arg_count(args)?;
                 // normally this would raise a SIGTRAP, which aborts if no debugger is connected
                 throw_machine_stop!(TerminationInfo::Abort("Trace/breakpoint trap".to_string()))
             }
@@ -844,7 +1092,7 @@ fn atomic_load(
     ) -> InterpResult<'tcx> {
         let this = self.eval_context_mut();
 
-        let &[ref place] = check_arg_count(args)?;
+        let [place] = check_arg_count(args)?;
         let place = this.deref_operand(place)?;
 
         // make sure it fits into a scalar; otherwise it cannot be atomic
@@ -854,7 +1102,7 @@ fn atomic_load(
         // even if the type they wrap would be less aligned (e.g. AtomicU64 on 32bit must
         // be 8-aligned).
         let align = Align::from_bytes(place.layout.size.bytes()).unwrap();
-        this.memory.check_ptr_access_align(
+        this.check_ptr_access_align(
             place.ptr,
             place.layout.size,
             align,
@@ -872,7 +1120,7 @@ fn atomic_store(
     ) -> InterpResult<'tcx> {
         let this = self.eval_context_mut();
 
-        let &[ref place, ref val] = check_arg_count(args)?;
+        let [place, val] = check_arg_count(args)?;
         let place = this.deref_operand(place)?;
         let val = this.read_scalar(val)?; // make sure it fits into a scalar; otherwise it cannot be atomic
 
@@ -880,7 +1128,7 @@ fn atomic_store(
         // even if the type they wrap would be less aligned (e.g. AtomicU64 on 32bit must
         // be 8-aligned).
         let align = Align::from_bytes(place.layout.size.bytes()).unwrap();
-        this.memory.check_ptr_access_align(
+        this.check_ptr_access_align(
             place.ptr,
             place.layout.size,
             align,
@@ -897,7 +1145,7 @@ fn compiler_fence(
         args: &[OpTy<'tcx, Tag>],
         atomic: AtomicFenceOp,
     ) -> InterpResult<'tcx> {
-        let &[] = check_arg_count(args)?;
+        let [] = check_arg_count(args)?;
         let _ = atomic;
         //FIXME: compiler fences are currently ignored
         Ok(())
@@ -909,7 +1157,7 @@ fn atomic_fence(
         atomic: AtomicFenceOp,
     ) -> InterpResult<'tcx> {
         let this = self.eval_context_mut();
-        let &[] = check_arg_count(args)?;
+        let [] = check_arg_count(args)?;
         this.validate_atomic_fence(atomic)?;
         Ok(())
     }
@@ -923,7 +1171,7 @@ fn atomic_op(
     ) -> InterpResult<'tcx> {
         let this = self.eval_context_mut();
 
-        let &[ref place, ref rhs] = check_arg_count(args)?;
+        let [place, rhs] = check_arg_count(args)?;
         let place = this.deref_operand(place)?;
 
         if !place.layout.ty.is_integral() {
@@ -935,7 +1183,7 @@ fn atomic_op(
         // even if the type they wrap would be less aligned (e.g. AtomicU64 on 32bit must
         // be 8-aligned).
         let align = Align::from_bytes(place.layout.size.bytes()).unwrap();
-        this.memory.check_ptr_access_align(
+        this.check_ptr_access_align(
             place.ptr,
             place.layout.size,
             align,
@@ -945,12 +1193,12 @@ fn atomic_op(
         match atomic_op {
             AtomicOp::Min => {
                 let old = this.atomic_min_max_scalar(&place, rhs, true, atomic)?;
-                this.write_immediate(*old, &dest)?; // old value is returned
+                this.write_immediate(*old, dest)?; // old value is returned
                 Ok(())
             }
             AtomicOp::Max => {
                 let old = this.atomic_min_max_scalar(&place, rhs, false, atomic)?;
-                this.write_immediate(*old, &dest)?; // old value is returned
+                this.write_immediate(*old, dest)?; // old value is returned
                 Ok(())
             }
             AtomicOp::MirOp(op, neg) => {
@@ -969,7 +1217,7 @@ fn atomic_exchange(
     ) -> InterpResult<'tcx> {
         let this = self.eval_context_mut();
 
-        let &[ref place, ref new] = check_arg_count(args)?;
+        let [place, new] = check_arg_count(args)?;
         let place = this.deref_operand(place)?;
         let new = this.read_scalar(new)?;
 
@@ -977,7 +1225,7 @@ fn atomic_exchange(
         // even if the type they wrap would be less aligned (e.g. AtomicU64 on 32bit must
         // be 8-aligned).
         let align = Align::from_bytes(place.layout.size.bytes()).unwrap();
-        this.memory.check_ptr_access_align(
+        this.check_ptr_access_align(
             place.ptr,
             place.layout.size,
             align,
@@ -999,7 +1247,7 @@ fn atomic_compare_exchange_impl(
     ) -> InterpResult<'tcx> {
         let this = self.eval_context_mut();
 
-        let &[ref place, ref expect_old, ref new] = check_arg_count(args)?;
+        let [place, expect_old, new] = check_arg_count(args)?;
         let place = this.deref_operand(place)?;
         let expect_old = this.read_immediate(expect_old)?; // read as immediate for the sake of `binary_op()`
         let new = this.read_scalar(new)?;
@@ -1008,7 +1256,7 @@ fn atomic_compare_exchange_impl(
         // even if the type they wrap would be less aligned (e.g. AtomicU64 on 32bit must
         // be 8-aligned).
         let align = Align::from_bytes(place.layout.size.bytes()).unwrap();
-        this.memory.check_ptr_access_align(
+        this.check_ptr_access_align(
             place.ptr,
             place.layout.size,
             align,
@@ -1134,3 +1382,26 @@ fn fmin_op<'tcx>(
         FloatTy::F64 => Scalar::from_f64(left.to_f64()?.min(right.to_f64()?)),
     })
 }
+
+fn bool_to_simd_element(b: bool, size: Size) -> Scalar<Tag> {
+    // SIMD uses all-1 as pattern for "true"
+    let val = if b { -1 } else { 0 };
+    Scalar::from_int(val, size)
+}
+
+fn simd_element_to_bool<'tcx>(elem: ImmTy<'tcx, Tag>) -> InterpResult<'tcx, bool> {
+    let val = elem.to_scalar()?.to_int(elem.layout.size)?;
+    Ok(match val {
+        0 => false,
+        -1 => true,
+        _ => throw_ub_format!("each element of a SIMD mask must be all-0-bits or all-1-bits"),
+    })
+}
+
+fn simd_bitmask_index(idx: u64, vec_len: u64, endianess: Endian) -> u64 {
+    assert!(idx < vec_len);
+    match endianess {
+        Endian::Little => idx,
+        Endian::Big => vec_len - 1 - idx, // reverse order of bits
+    }
+}