Move utils from rustc_middle to rustc_ty_utils

compiler/rustc_ty_utils/src/abi.rs

@@ -0,0 +1,518 @@
+use rustc_hir as hir;
+use rustc_hir::lang_items::LangItem;
+use rustc_middle::ty::layout::{
+    fn_can_unwind, FnAbiError, HasParamEnv, HasTyCtxt, LayoutCx, LayoutOf, TyAndLayout,
+};
+use rustc_middle::ty::{self, Ty, TyCtxt};
+use rustc_span::def_id::DefId;
+use rustc_target::abi::call::{
+    ArgAbi, ArgAttribute, ArgAttributes, ArgExtension, Conv, FnAbi, PassMode, Reg, RegKind,
+};
+use rustc_target::abi::*;
+use rustc_target::spec::abi::Abi as SpecAbi;
+
+use std::iter;
+
+pub fn provide(providers: &mut ty::query::Providers) {
+    *providers = ty::query::Providers { fn_abi_of_fn_ptr, fn_abi_of_instance, ..*providers };
+}
+
+// NOTE(eddyb) this is private to avoid using it from outside of
+// `fn_abi_of_instance` - any other uses are either too high-level
+// for `Instance` (e.g. typeck would use `Ty::fn_sig` instead),
+// or should go through `FnAbi` instead, to avoid losing any
+// adjustments `fn_abi_of_instance` might be performing.
+#[tracing::instrument(level = "debug", skip(tcx, param_env))]
+fn fn_sig_for_fn_abi<'tcx>(
+    tcx: TyCtxt<'tcx>,
+    instance: ty::Instance<'tcx>,
+    param_env: ty::ParamEnv<'tcx>,
+) -> ty::PolyFnSig<'tcx> {
+    let ty = instance.ty(tcx, param_env);
+    match *ty.kind() {
+        ty::FnDef(..) => {
+            // HACK(davidtwco,eddyb): This is a workaround for polymorphization considering
+            // parameters unused if they show up in the signature, but not in the `mir::Body`
+            // (i.e. due to being inside a projection that got normalized, see
+            // `src/test/ui/polymorphization/normalized_sig_types.rs`), and codegen not keeping
+            // track of a polymorphization `ParamEnv` to allow normalizing later.
+            //
+            // We normalize the `fn_sig` again after substituting at a later point.
+            let mut sig = match *ty.kind() {
+                ty::FnDef(def_id, substs) => tcx
+                    .bound_fn_sig(def_id)
+                    .map_bound(|fn_sig| {
+                        tcx.normalize_erasing_regions(tcx.param_env(def_id), fn_sig)
+                    })
+                    .subst(tcx, substs),
+                _ => unreachable!(),
+            };
+
+            if let ty::InstanceDef::VTableShim(..) = instance.def {
+                // Modify `fn(self, ...)` to `fn(self: *mut Self, ...)`.
+                sig = sig.map_bound(|mut sig| {
+                    let mut inputs_and_output = sig.inputs_and_output.to_vec();
+                    inputs_and_output[0] = tcx.mk_mut_ptr(inputs_and_output[0]);
+                    sig.inputs_and_output = tcx.intern_type_list(&inputs_and_output);
+                    sig
+                });
+            }
+            sig
+        }
+        ty::Closure(def_id, substs) => {
+            let sig = substs.as_closure().sig();
+
+            let bound_vars = tcx.mk_bound_variable_kinds(
+                sig.bound_vars().iter().chain(iter::once(ty::BoundVariableKind::Region(ty::BrEnv))),
+            );
+            let br = ty::BoundRegion {
+                var: ty::BoundVar::from_usize(bound_vars.len() - 1),
+                kind: ty::BoundRegionKind::BrEnv,
+            };
+            let env_region = ty::ReLateBound(ty::INNERMOST, br);
+            let env_ty = tcx.closure_env_ty(def_id, substs, env_region).unwrap();
+
+            let sig = sig.skip_binder();
+            ty::Binder::bind_with_vars(
+                tcx.mk_fn_sig(
+                    iter::once(env_ty).chain(sig.inputs().iter().cloned()),
+                    sig.output(),
+                    sig.c_variadic,
+                    sig.unsafety,
+                    sig.abi,
+                ),
+                bound_vars,
+            )
+        }
+        ty::Generator(_, substs, _) => {
+            let sig = substs.as_generator().poly_sig();
+
+            let bound_vars = tcx.mk_bound_variable_kinds(
+                sig.bound_vars().iter().chain(iter::once(ty::BoundVariableKind::Region(ty::BrEnv))),
+            );
+            let br = ty::BoundRegion {
+                var: ty::BoundVar::from_usize(bound_vars.len() - 1),
+                kind: ty::BoundRegionKind::BrEnv,
+            };
+            let env_region = ty::ReLateBound(ty::INNERMOST, br);
+            let env_ty = tcx.mk_mut_ref(tcx.mk_region(env_region), ty);
+
+            let pin_did = tcx.require_lang_item(LangItem::Pin, None);
+            let pin_adt_ref = tcx.adt_def(pin_did);
+            let pin_substs = tcx.intern_substs(&[env_ty.into()]);
+            let env_ty = tcx.mk_adt(pin_adt_ref, pin_substs);
+
+            let sig = sig.skip_binder();
+            let state_did = tcx.require_lang_item(LangItem::GeneratorState, None);
+            let state_adt_ref = tcx.adt_def(state_did);
+            let state_substs = tcx.intern_substs(&[sig.yield_ty.into(), sig.return_ty.into()]);
+            let ret_ty = tcx.mk_adt(state_adt_ref, state_substs);
+            ty::Binder::bind_with_vars(
+                tcx.mk_fn_sig(
+                    [env_ty, sig.resume_ty].iter(),
+                    &ret_ty,
+                    false,
+                    hir::Unsafety::Normal,
+                    rustc_target::spec::abi::Abi::Rust,
+                ),
+                bound_vars,
+            )
+        }
+        _ => bug!("unexpected type {:?} in Instance::fn_sig", ty),
+    }
+}
+
+#[inline]
+fn conv_from_spec_abi(tcx: TyCtxt<'_>, abi: SpecAbi) -> Conv {
+    use rustc_target::spec::abi::Abi::*;
+    match tcx.sess.target.adjust_abi(abi) {
+        RustIntrinsic | PlatformIntrinsic | Rust | RustCall => Conv::Rust,
+        RustCold => Conv::RustCold,
+
+        // It's the ABI's job to select this, not ours.
+        System { .. } => bug!("system abi should be selected elsewhere"),
+        EfiApi => bug!("eficall abi should be selected elsewhere"),
+
+        Stdcall { .. } => Conv::X86Stdcall,
+        Fastcall { .. } => Conv::X86Fastcall,
+        Vectorcall { .. } => Conv::X86VectorCall,
+        Thiscall { .. } => Conv::X86ThisCall,
+        C { .. } => Conv::C,
+        Unadjusted => Conv::C,
+        Win64 { .. } => Conv::X86_64Win64,
+        SysV64 { .. } => Conv::X86_64SysV,
+        Aapcs { .. } => Conv::ArmAapcs,
+        CCmseNonSecureCall => Conv::CCmseNonSecureCall,
+        PtxKernel => Conv::PtxKernel,
+        Msp430Interrupt => Conv::Msp430Intr,
+        X86Interrupt => Conv::X86Intr,
+        AmdGpuKernel => Conv::AmdGpuKernel,
+        AvrInterrupt => Conv::AvrInterrupt,
+        AvrNonBlockingInterrupt => Conv::AvrNonBlockingInterrupt,
+        Wasm => Conv::C,
+
+        // These API constants ought to be more specific...
+        Cdecl { .. } => Conv::C,
+    }
+}
+
+fn fn_abi_of_fn_ptr<'tcx>(
+    tcx: TyCtxt<'tcx>,
+    query: ty::ParamEnvAnd<'tcx, (ty::PolyFnSig<'tcx>, &'tcx ty::List<Ty<'tcx>>)>,
+) -> Result<&'tcx FnAbi<'tcx, Ty<'tcx>>, FnAbiError<'tcx>> {
+    let (param_env, (sig, extra_args)) = query.into_parts();
+
+    let cx = LayoutCx { tcx, param_env };
+    fn_abi_new_uncached(&cx, sig, extra_args, None, None, false)
+}
+
+fn fn_abi_of_instance<'tcx>(
+    tcx: TyCtxt<'tcx>,
+    query: ty::ParamEnvAnd<'tcx, (ty::Instance<'tcx>, &'tcx ty::List<Ty<'tcx>>)>,
+) -> Result<&'tcx FnAbi<'tcx, Ty<'tcx>>, FnAbiError<'tcx>> {
+    let (param_env, (instance, extra_args)) = query.into_parts();
+
+    let sig = fn_sig_for_fn_abi(tcx, instance, param_env);
+
+    let caller_location = if instance.def.requires_caller_location(tcx) {
+        Some(tcx.caller_location_ty())
+    } else {
+        None
+    };
+
+    fn_abi_new_uncached(
+        &LayoutCx { tcx, param_env },
+        sig,
+        extra_args,
+        caller_location,
+        Some(instance.def_id()),
+        matches!(instance.def, ty::InstanceDef::Virtual(..)),
+    )
+}
+
+// Handle safe Rust thin and fat pointers.
+fn adjust_for_rust_scalar<'tcx>(
+    cx: LayoutCx<'tcx, TyCtxt<'tcx>>,
+    attrs: &mut ArgAttributes,
+    scalar: Scalar,
+    layout: TyAndLayout<'tcx>,
+    offset: Size,
+    is_return: bool,
+) {
+    // Booleans are always a noundef i1 that needs to be zero-extended.
+    if scalar.is_bool() {
+        attrs.ext(ArgExtension::Zext);
+        attrs.set(ArgAttribute::NoUndef);
+        return;
+    }
+
+    // Scalars which have invalid values cannot be undef.
+    if !scalar.is_always_valid(&cx) {
+        attrs.set(ArgAttribute::NoUndef);
+    }
+
+    // Only pointer types handled below.
+    let Scalar::Initialized { value: Pointer, valid_range} = scalar else { return };
+
+    if !valid_range.contains(0) {
+        attrs.set(ArgAttribute::NonNull);
+    }
+
+    if let Some(pointee) = layout.pointee_info_at(&cx, offset) {
+        if let Some(kind) = pointee.safe {
+            attrs.pointee_align = Some(pointee.align);
+
+            // `Box` (`UniqueBorrowed`) are not necessarily dereferenceable
+            // for the entire duration of the function as they can be deallocated
+            // at any time. Same for shared mutable references. If LLVM had a
+            // way to say "dereferenceable on entry" we could use it here.
+            attrs.pointee_size = match kind {
+                PointerKind::UniqueBorrowed
+                | PointerKind::UniqueBorrowedPinned
+                | PointerKind::Frozen => pointee.size,
+                PointerKind::SharedMutable | PointerKind::UniqueOwned => Size::ZERO,
+            };
+
+            // `Box`, `&T`, and `&mut T` cannot be undef.
+            // Note that this only applies to the value of the pointer itself;
+            // this attribute doesn't make it UB for the pointed-to data to be undef.
+            attrs.set(ArgAttribute::NoUndef);
+
+            // The aliasing rules for `Box<T>` are still not decided, but currently we emit
+            // `noalias` for it. This can be turned off using an unstable flag.
+            // See https://github.com/rust-lang/unsafe-code-guidelines/issues/326
+            let noalias_for_box = cx.tcx.sess.opts.unstable_opts.box_noalias.unwrap_or(true);
+
+            // `&mut` pointer parameters never alias other parameters,
+            // or mutable global data
+            //
+            // `&T` where `T` contains no `UnsafeCell<U>` is immutable,
+            // and can be marked as both `readonly` and `noalias`, as
+            // LLVM's definition of `noalias` is based solely on memory
+            // dependencies rather than pointer equality
+            //
+            // Due to past miscompiles in LLVM, we apply a separate NoAliasMutRef attribute
+            // for UniqueBorrowed arguments, so that the codegen backend can decide whether
+            // or not to actually emit the attribute. It can also be controlled with the
+            // `-Zmutable-noalias` debugging option.
+            let no_alias = match kind {
+                PointerKind::SharedMutable
+                | PointerKind::UniqueBorrowed
+                | PointerKind::UniqueBorrowedPinned => false,
+                PointerKind::UniqueOwned => noalias_for_box,
+                PointerKind::Frozen => !is_return,
+            };
+            if no_alias {
+                attrs.set(ArgAttribute::NoAlias);
+            }
+
+            if kind == PointerKind::Frozen && !is_return {
+                attrs.set(ArgAttribute::ReadOnly);
+            }
+
+            if kind == PointerKind::UniqueBorrowed && !is_return {
+                attrs.set(ArgAttribute::NoAliasMutRef);
+            }
+        }
+    }
+}
+
+// FIXME(eddyb) perhaps group the signature/type-containing (or all of them?)
+// arguments of this method, into a separate `struct`.
+#[tracing::instrument(level = "debug", skip(cx, caller_location, fn_def_id, force_thin_self_ptr))]
+fn fn_abi_new_uncached<'tcx>(
+    cx: &LayoutCx<'tcx, TyCtxt<'tcx>>,
+    sig: ty::PolyFnSig<'tcx>,
+    extra_args: &[Ty<'tcx>],
+    caller_location: Option<Ty<'tcx>>,
+    fn_def_id: Option<DefId>,
+    // FIXME(eddyb) replace this with something typed, like an `enum`.
+    force_thin_self_ptr: bool,
+) -> Result<&'tcx FnAbi<'tcx, Ty<'tcx>>, FnAbiError<'tcx>> {
+    let sig = cx.tcx.normalize_erasing_late_bound_regions(cx.param_env, sig);
+
+    let conv = conv_from_spec_abi(cx.tcx(), sig.abi);
+
+    let mut inputs = sig.inputs();
+    let extra_args = if sig.abi == RustCall {
+        assert!(!sig.c_variadic && extra_args.is_empty());
+
+        if let Some(input) = sig.inputs().last() {
+            if let ty::Tuple(tupled_arguments) = input.kind() {
+                inputs = &sig.inputs()[0..sig.inputs().len() - 1];
+                tupled_arguments
+            } else {
+                bug!(
+                    "argument to function with \"rust-call\" ABI \
+                        is not a tuple"
+                );
+            }
+        } else {
+            bug!(
+                "argument to function with \"rust-call\" ABI \
+                    is not a tuple"
+            );
+        }
+    } else {
+        assert!(sig.c_variadic || extra_args.is_empty());
+        extra_args
+    };
+
+    let target = &cx.tcx.sess.target;
+    let target_env_gnu_like = matches!(&target.env[..], "gnu" | "musl" | "uclibc");
+    let win_x64_gnu = target.os == "windows" && target.arch == "x86_64" && target.env == "gnu";
+    let linux_s390x_gnu_like =
+        target.os == "linux" && target.arch == "s390x" && target_env_gnu_like;
+    let linux_sparc64_gnu_like =
+        target.os == "linux" && target.arch == "sparc64" && target_env_gnu_like;
+    let linux_powerpc_gnu_like =
+        target.os == "linux" && target.arch == "powerpc" && target_env_gnu_like;
+    use SpecAbi::*;
+    let rust_abi = matches!(sig.abi, RustIntrinsic | PlatformIntrinsic | Rust | RustCall);
+
+    let arg_of = |ty: Ty<'tcx>, arg_idx: Option<usize>| -> Result<_, FnAbiError<'tcx>> {
+        let span = tracing::debug_span!("arg_of");
+        let _entered = span.enter();
+        let is_return = arg_idx.is_none();
+
+        let layout = cx.layout_of(ty)?;
+        let layout = if force_thin_self_ptr && arg_idx == Some(0) {
+            // Don't pass the vtable, it's not an argument of the virtual fn.
+            // Instead, pass just the data pointer, but give it the type `*const/mut dyn Trait`
+            // or `&/&mut dyn Trait` because this is special-cased elsewhere in codegen
+            make_thin_self_ptr(cx, layout)
+        } else {
+            layout
+        };
+
+        let mut arg = ArgAbi::new(cx, layout, |layout, scalar, offset| {
+            let mut attrs = ArgAttributes::new();
+            adjust_for_rust_scalar(*cx, &mut attrs, scalar, *layout, offset, is_return);
+            attrs
+        });
+
+        if arg.layout.is_zst() {
+            // For some forsaken reason, x86_64-pc-windows-gnu
+            // doesn't ignore zero-sized struct arguments.
+            // The same is true for {s390x,sparc64,powerpc}-unknown-linux-{gnu,musl,uclibc}.
+            if is_return
+                || rust_abi
+                || (!win_x64_gnu
+                    && !linux_s390x_gnu_like
+                    && !linux_sparc64_gnu_like
+                    && !linux_powerpc_gnu_like)
+            {
+                arg.mode = PassMode::Ignore;
+            }
+        }
+
+        Ok(arg)
+    };
+
+    let mut fn_abi = FnAbi {
+        ret: arg_of(sig.output(), None)?,
+        args: inputs
+            .iter()
+            .copied()
+            .chain(extra_args.iter().copied())
+            .chain(caller_location)
+            .enumerate()
+            .map(|(i, ty)| arg_of(ty, Some(i)))
+            .collect::<Result<_, _>>()?,
+        c_variadic: sig.c_variadic,
+        fixed_count: inputs.len() as u32,
+        conv,
+        can_unwind: fn_can_unwind(cx.tcx(), fn_def_id, sig.abi),
+    };
+    fn_abi_adjust_for_abi(cx, &mut fn_abi, sig.abi)?;
+    debug!("fn_abi_new_uncached = {:?}", fn_abi);
+    Ok(cx.tcx.arena.alloc(fn_abi))
+}
+
+#[tracing::instrument(level = "trace", skip(cx))]
+fn fn_abi_adjust_for_abi<'tcx>(
+    cx: &LayoutCx<'tcx, TyCtxt<'tcx>>,
+    fn_abi: &mut FnAbi<'tcx, Ty<'tcx>>,
+    abi: SpecAbi,
+) -> Result<(), FnAbiError<'tcx>> {
+    if abi == SpecAbi::Unadjusted {
+        return Ok(());
+    }
+
+    if abi == SpecAbi::Rust
+        || abi == SpecAbi::RustCall
+        || abi == SpecAbi::RustIntrinsic
+        || abi == SpecAbi::PlatformIntrinsic
+    {
+        let fixup = |arg: &mut ArgAbi<'tcx, Ty<'tcx>>| {
+            if arg.is_ignore() {
+                return;
+            }
+
+            match arg.layout.abi {
+                Abi::Aggregate { .. } => {}
+
+                // This is a fun case! The gist of what this is doing is
+                // that we want callers and callees to always agree on the
+                // ABI of how they pass SIMD arguments. If we were to *not*
+                // make these arguments indirect then they'd be immediates
+                // in LLVM, which means that they'd used whatever the
+                // appropriate ABI is for the callee and the caller. That
+                // means, for example, if the caller doesn't have AVX
+                // enabled but the callee does, then passing an AVX argument
+                // across this boundary would cause corrupt data to show up.
+                //
+                // This problem is fixed by unconditionally passing SIMD
+                // arguments through memory between callers and callees
+                // which should get them all to agree on ABI regardless of
+                // target feature sets. Some more information about this
+                // issue can be found in #44367.
+                //
+                // Note that the platform intrinsic ABI is exempt here as
+                // that's how we connect up to LLVM and it's unstable
+                // anyway, we control all calls to it in libstd.
+                Abi::Vector { .. }
+                    if abi != SpecAbi::PlatformIntrinsic
+                        && cx.tcx.sess.target.simd_types_indirect =>
+                {
+                    arg.make_indirect();
+                    return;
+                }
+
+                _ => return,
+            }
+
+            let size = arg.layout.size;
+            if arg.layout.is_unsized() || size > Pointer.size(cx) {
+                arg.make_indirect();
+            } else {
+                // We want to pass small aggregates as immediates, but using
+                // a LLVM aggregate type for this leads to bad optimizations,
+                // so we pick an appropriately sized integer type instead.
+                arg.cast_to(Reg { kind: RegKind::Integer, size });
+            }
+        };
+        fixup(&mut fn_abi.ret);
+        for arg in fn_abi.args.iter_mut() {
+            fixup(arg);
+        }
+    } else {
+        fn_abi.adjust_for_foreign_abi(cx, abi)?;
+    }
+
+    Ok(())
+}
+
+#[tracing::instrument(level = "debug", skip(cx))]
+fn make_thin_self_ptr<'tcx>(
+    cx: &(impl HasTyCtxt<'tcx> + HasParamEnv<'tcx>),
+    layout: TyAndLayout<'tcx>,
+) -> TyAndLayout<'tcx> {
+    let tcx = cx.tcx();
+    let fat_pointer_ty = if layout.is_unsized() {
+        // unsized `self` is passed as a pointer to `self`
+        // FIXME (mikeyhew) change this to use &own if it is ever added to the language
+        tcx.mk_mut_ptr(layout.ty)
+    } else {
+        match layout.abi {
+            Abi::ScalarPair(..) | Abi::Scalar(..) => (),
+            _ => bug!("receiver type has unsupported layout: {:?}", layout),
+        }
+
+        // In the case of Rc<Self>, we need to explicitly pass a *mut RcBox<Self>
+        // with a Scalar (not ScalarPair) ABI. This is a hack that is understood
+        // elsewhere in the compiler as a method on a `dyn Trait`.
+        // To get the type `*mut RcBox<Self>`, we just keep unwrapping newtypes until we
+        // get a built-in pointer type
+        let mut fat_pointer_layout = layout;
+        'descend_newtypes: while !fat_pointer_layout.ty.is_unsafe_ptr()
+            && !fat_pointer_layout.ty.is_region_ptr()
+        {
+            for i in 0..fat_pointer_layout.fields.count() {
+                let field_layout = fat_pointer_layout.field(cx, i);
+
+                if !field_layout.is_zst() {
+                    fat_pointer_layout = field_layout;
+                    continue 'descend_newtypes;
+                }
+            }
+
+            bug!("receiver has no non-zero-sized fields {:?}", fat_pointer_layout);
+        }
+
+        fat_pointer_layout.ty
+    };
+
+    // we now have a type like `*mut RcBox<dyn Trait>`
+    // change its layout to that of `*mut ()`, a thin pointer, but keep the same type
+    // this is understood as a special case elsewhere in the compiler
+    let unit_ptr_ty = tcx.mk_mut_ptr(tcx.mk_unit());
+
+    TyAndLayout {
+        ty: fat_pointer_ty,
+
+        // NOTE(eddyb) using an empty `ParamEnv`, and `unwrap`-ing the `Result`
+        // should always work because the type is always `*mut ()`.
+        ..tcx.layout_of(ty::ParamEnv::reveal_all().and(unit_ptr_ty)).unwrap()
+    }
+}

compiler/rustc_ty_utils/src/layout_sanity_check.rs

@@ -0,0 +1,303 @@
+use rustc_middle::ty::{
+    layout::{LayoutCx, TyAndLayout},
+    TyCtxt,
+};
+use rustc_target::abi::*;
+
+use std::cmp;
+
+/// Enforce some basic invariants on layouts.
+pub(super) fn sanity_check_layout<'tcx>(
+    cx: &LayoutCx<'tcx, TyCtxt<'tcx>>,
+    layout: &TyAndLayout<'tcx>,
+) {
+    // Type-level uninhabitedness should always imply ABI uninhabitedness.
+    if cx.tcx.conservative_is_privately_uninhabited(cx.param_env.and(layout.ty)) {
+        assert!(layout.abi.is_uninhabited());
+    }
+
+    if layout.size.bytes() % layout.align.abi.bytes() != 0 {
+        bug!("size is not a multiple of align, in the following layout:\n{layout:#?}");
+    }
+
+    if cfg!(debug_assertions) {
+        /// Yields non-ZST fields of the type
+        fn non_zst_fields<'tcx, 'a>(
+            cx: &'a LayoutCx<'tcx, TyCtxt<'tcx>>,
+            layout: &'a TyAndLayout<'tcx>,
+        ) -> impl Iterator<Item = (Size, TyAndLayout<'tcx>)> + 'a {
+            (0..layout.layout.fields().count()).filter_map(|i| {
+                let field = layout.field(cx, i);
+                // Also checking `align == 1` here leads to test failures in
+                // `layout/zero-sized-array-union.rs`, where a type has a zero-size field with
+                // alignment 4 that still gets ignored during layout computation (which is okay
+                // since other fields already force alignment 4).
+                let zst = field.is_zst();
+                (!zst).then(|| (layout.fields.offset(i), field))
+            })
+        }
+
+        fn skip_newtypes<'tcx>(
+            cx: &LayoutCx<'tcx, TyCtxt<'tcx>>,
+            layout: &TyAndLayout<'tcx>,
+        ) -> TyAndLayout<'tcx> {
+            if matches!(layout.layout.variants(), Variants::Multiple { .. }) {
+                // Definitely not a newtype of anything.
+                return *layout;
+            }
+            let mut fields = non_zst_fields(cx, layout);
+            let Some(first) = fields.next() else {
+                // No fields here, so this could be a primitive or enum -- either way it's not a newtype around a thing
+                return *layout
+            };
+            if fields.next().is_none() {
+                let (offset, first) = first;
+                if offset == Size::ZERO && first.layout.size() == layout.size {
+                    // This is a newtype, so keep recursing.
+                    // FIXME(RalfJung): I don't think it would be correct to do any checks for
+                    // alignment here, so we don't. Is that correct?
+                    return skip_newtypes(cx, &first);
+                }
+            }
+            // No more newtypes here.
+            *layout
+        }
+
+        fn check_layout_abi<'tcx>(cx: &LayoutCx<'tcx, TyCtxt<'tcx>>, layout: &TyAndLayout<'tcx>) {
+            match layout.layout.abi() {
+                Abi::Scalar(scalar) => {
+                    // No padding in scalars.
+                    let size = scalar.size(cx);
+                    let align = scalar.align(cx).abi;
+                    assert_eq!(
+                        layout.layout.size(),
+                        size,
+                        "size mismatch between ABI and layout in {layout:#?}"
+                    );
+                    assert_eq!(
+                        layout.layout.align().abi,
+                        align,
+                        "alignment mismatch between ABI and layout in {layout:#?}"
+                    );
+                    // Check that this matches the underlying field.
+                    let inner = skip_newtypes(cx, layout);
+                    assert!(
+                        matches!(inner.layout.abi(), Abi::Scalar(_)),
+                        "`Scalar` type {} is newtype around non-`Scalar` type {}",
+                        layout.ty,
+                        inner.ty
+                    );
+                    match inner.layout.fields() {
+                        FieldsShape::Primitive => {
+                            // Fine.
+                        }
+                        FieldsShape::Union(..) => {
+                            // FIXME: I guess we could also check something here? Like, look at all fields?
+                            return;
+                        }
+                        FieldsShape::Arbitrary { .. } => {
+                            // Should be an enum, the only field is the discriminant.
+                            assert!(
+                                inner.ty.is_enum(),
+                                "`Scalar` layout for non-primitive non-enum type {}",
+                                inner.ty
+                            );
+                            assert_eq!(
+                                inner.layout.fields().count(),
+                                1,
+                                "`Scalar` layout for multiple-field type in {inner:#?}",
+                            );
+                            let offset = inner.layout.fields().offset(0);
+                            let field = inner.field(cx, 0);
+                            // The field should be at the right offset, and match the `scalar` layout.
+                            assert_eq!(
+                                offset,
+                                Size::ZERO,
+                                "`Scalar` field at non-0 offset in {inner:#?}",
+                            );
+                            assert_eq!(
+                                field.size, size,
+                                "`Scalar` field with bad size in {inner:#?}",
+                            );
+                            assert_eq!(
+                                field.align.abi, align,
+                                "`Scalar` field with bad align in {inner:#?}",
+                            );
+                            assert!(
+                                matches!(field.abi, Abi::Scalar(_)),
+                                "`Scalar` field with bad ABI in {inner:#?}",
+                            );
+                        }
+                        _ => {
+                            panic!("`Scalar` layout for non-primitive non-enum type {}", inner.ty);
+                        }
+                    }
+                }
+                Abi::ScalarPair(scalar1, scalar2) => {
+                    // Sanity-check scalar pairs. These are a bit more flexible and support
+                    // padding, but we can at least ensure both fields actually fit into the layout
+                    // and the alignment requirement has not been weakened.
+                    let size1 = scalar1.size(cx);
+                    let align1 = scalar1.align(cx).abi;
+                    let size2 = scalar2.size(cx);
+                    let align2 = scalar2.align(cx).abi;
+                    assert!(
+                        layout.layout.align().abi >= cmp::max(align1, align2),
+                        "alignment mismatch between ABI and layout in {layout:#?}",
+                    );
+                    let field2_offset = size1.align_to(align2);
+                    assert!(
+                        layout.layout.size() >= field2_offset + size2,
+                        "size mismatch between ABI and layout in {layout:#?}"
+                    );
+                    // Check that the underlying pair of fields matches.
+                    let inner = skip_newtypes(cx, layout);
+                    assert!(
+                        matches!(inner.layout.abi(), Abi::ScalarPair(..)),
+                        "`ScalarPair` type {} is newtype around non-`ScalarPair` type {}",
+                        layout.ty,
+                        inner.ty
+                    );
+                    if matches!(inner.layout.variants(), Variants::Multiple { .. }) {
+                        // FIXME: ScalarPair for enums is enormously complicated and it is very hard
+                        // to check anything about them.
+                        return;
+                    }
+                    match inner.layout.fields() {
+                        FieldsShape::Arbitrary { .. } => {
+                            // Checked below.
+                        }
+                        FieldsShape::Union(..) => {
+                            // FIXME: I guess we could also check something here? Like, look at all fields?
+                            return;
+                        }
+                        _ => {
+                            panic!("`ScalarPair` layout with unexpected field shape in {inner:#?}");
+                        }
+                    }
+                    let mut fields = non_zst_fields(cx, &inner);
+                    let (offset1, field1) = fields.next().unwrap_or_else(|| {
+                        panic!("`ScalarPair` layout for type with not even one non-ZST field: {inner:#?}")
+                    });
+                    let (offset2, field2) = fields.next().unwrap_or_else(|| {
+                        panic!("`ScalarPair` layout for type with less than two non-ZST fields: {inner:#?}")
+                    });
+                    assert!(
+                        fields.next().is_none(),
+                        "`ScalarPair` layout for type with at least three non-ZST fields: {inner:#?}"
+                    );
+                    // The fields might be in opposite order.
+                    let (offset1, field1, offset2, field2) = if offset1 <= offset2 {
+                        (offset1, field1, offset2, field2)
+                    } else {
+                        (offset2, field2, offset1, field1)
+                    };
+                    // The fields should be at the right offset, and match the `scalar` layout.
+                    assert_eq!(
+                        offset1,
+                        Size::ZERO,
+                        "`ScalarPair` first field at non-0 offset in {inner:#?}",
+                    );
+                    assert_eq!(
+                        field1.size, size1,
+                        "`ScalarPair` first field with bad size in {inner:#?}",
+                    );
+                    assert_eq!(
+                        field1.align.abi, align1,
+                        "`ScalarPair` first field with bad align in {inner:#?}",
+                    );
+                    assert!(
+                        matches!(field1.abi, Abi::Scalar(_)),
+                        "`ScalarPair` first field with bad ABI in {inner:#?}",
+                    );
+                    assert_eq!(
+                        offset2, field2_offset,
+                        "`ScalarPair` second field at bad offset in {inner:#?}",
+                    );
+                    assert_eq!(
+                        field2.size, size2,
+                        "`ScalarPair` second field with bad size in {inner:#?}",
+                    );
+                    assert_eq!(
+                        field2.align.abi, align2,
+                        "`ScalarPair` second field with bad align in {inner:#?}",
+                    );
+                    assert!(
+                        matches!(field2.abi, Abi::Scalar(_)),
+                        "`ScalarPair` second field with bad ABI in {inner:#?}",
+                    );
+                }
+                Abi::Vector { count, element } => {
+                    // No padding in vectors. Alignment can be strengthened, though.
+                    assert!(
+                        layout.layout.align().abi >= element.align(cx).abi,
+                        "alignment mismatch between ABI and layout in {layout:#?}"
+                    );
+                    let size = element.size(cx) * count;
+                    assert_eq!(
+                        layout.layout.size(),
+                        size.align_to(cx.data_layout().vector_align(size).abi),
+                        "size mismatch between ABI and layout in {layout:#?}"
+                    );
+                }
+                Abi::Uninhabited | Abi::Aggregate { .. } => {} // Nothing to check.
+            }
+        }
+
+        check_layout_abi(cx, layout);
+
+        if let Variants::Multiple { variants, .. } = &layout.variants {
+            for variant in variants.iter() {
+                // No nested "multiple".
+                assert!(matches!(variant.variants(), Variants::Single { .. }));
+                // Variants should have the same or a smaller size as the full thing,
+                // and same for alignment.
+                if variant.size() > layout.size {
+                    bug!(
+                        "Type with size {} bytes has variant with size {} bytes: {layout:#?}",
+                        layout.size.bytes(),
+                        variant.size().bytes(),
+                    )
+                }
+                if variant.align().abi > layout.align.abi {
+                    bug!(
+                        "Type with alignment {} bytes has variant with alignment {} bytes: {layout:#?}",
+                        layout.align.abi.bytes(),
+                        variant.align().abi.bytes(),
+                    )
+                }
+                // Skip empty variants.
+                if variant.size() == Size::ZERO
+                    || variant.fields().count() == 0
+                    || variant.abi().is_uninhabited()
+                {
+                    // These are never actually accessed anyway, so we can skip the coherence check
+                    // for them. They also fail that check, since they have
+                    // `Aggregate`/`Uninhbaited` ABI even when the main type is
+                    // `Scalar`/`ScalarPair`. (Note that sometimes, variants with fields have size
+                    // 0, and sometimes, variants without fields have non-0 size.)
+                    continue;
+                }
+                // The top-level ABI and the ABI of the variants should be coherent.
+                let scalar_coherent = |s1: Scalar, s2: Scalar| {
+                    s1.size(cx) == s2.size(cx) && s1.align(cx) == s2.align(cx)
+                };
+                let abi_coherent = match (layout.abi, variant.abi()) {
+                    (Abi::Scalar(s1), Abi::Scalar(s2)) => scalar_coherent(s1, s2),
+                    (Abi::ScalarPair(a1, b1), Abi::ScalarPair(a2, b2)) => {
+                        scalar_coherent(a1, a2) && scalar_coherent(b1, b2)
+                    }
+                    (Abi::Uninhabited, _) => true,
+                    (Abi::Aggregate { .. }, _) => true,
+                    _ => false,
+                };
+                if !abi_coherent {
+                    bug!(
+                        "Variant ABI is incompatible with top-level ABI:\nvariant={:#?}\nTop-level: {layout:#?}",
+                        variant
+                    );
+                }
+            }
+        }
+    }
+}

compiler/rustc_ty_utils/src/lib.rs

@@ -9,8 +9,6 @@
 #![feature(never_type)]
 #![feature(box_patterns)]
 #![recursion_limit = "256"]
-#![deny(rustc::untranslatable_diagnostic)]
-#![deny(rustc::diagnostic_outside_of_impl)]
 
 #[macro_use]
 extern crate rustc_middle;
@@ -19,21 +17,26 @@ extern crate tracing;
 
 use rustc_middle::ty::query::Providers;
 
+mod abi;
 mod assoc;
 mod common_traits;
 mod consts;
 mod errors;
 mod implied_bounds;
 pub mod instance;
+mod layout;
+mod layout_sanity_check;
 mod needs_drop;
 pub mod representability;
 mod ty;
 
 pub fn provide(providers: &mut Providers) {
+    abi::provide(providers);
     assoc::provide(providers);
     common_traits::provide(providers);
     consts::provide(providers);
     implied_bounds::provide(providers);
+    layout::provide(providers);
     needs_drop::provide(providers);
     ty::provide(providers);
     instance::provide(providers);