Auto merge of rust-lang#122629 - RalfJung:assert-unsafe-precondition, r=saethlin

refactor check_{lang,library}_ub: use a single intrinsic

This enacts the plan I laid out [here](rust-lang#122282 (comment)): use a single intrinsic, called `ub_checks` (in aniticpation of rust-lang/compiler-team#725), that just exposes the value of `debug_assertions` (consistently implemented in both codegen and the interpreter). Put the language vs library UB logic into the library.

This makes it easier to do something like rust-lang#122282 in the future: that just slightly alters the semantics of `ub_checks` (making it more approximating when crates built with different flags are mixed), but it no longer affects whether these checks can happen in Miri or compile-time.

The first commit just moves things around; I don't think these macros and functions belong into `intrinsics.rs` as they are not intrinsics.

r? `@saethlin`

Author: bors

core/src/char/convert.rs

@@ -4,9 +4,9 @@ use crate::char::TryFromCharError;
 use crate::convert::TryFrom;
 use crate::error::Error;
 use crate::fmt;
-use crate::intrinsics::assert_unsafe_precondition;
 use crate::mem::transmute;
 use crate::str::FromStr;
+use crate::ub_checks::assert_unsafe_precondition;
 
 /// Converts a `u32` to a `char`. See [`char::from_u32`].
 #[must_use]

core/src/hint.rs

@@ -4,6 +4,7 @@
 //! Hints may be compile time or runtime.
 
 use crate::intrinsics;
+use crate::ub_checks;
 
 /// Informs the compiler that the site which is calling this function is not
 /// reachable, possibly enabling further optimizations.
@@ -98,7 +99,7 @@ use crate::intrinsics;
 #[rustc_const_stable(feature = "const_unreachable_unchecked", since = "1.57.0")]
 #[cfg_attr(miri, track_caller)] // even without panics, this helps for Miri backtraces
 pub const unsafe fn unreachable_unchecked() -> ! {
-    intrinsics::assert_unsafe_precondition!(
+    ub_checks::assert_unsafe_precondition!(
         check_language_ub,
         "hint::unreachable_unchecked must never be reached",
         () => false
@@ -148,7 +149,7 @@ pub const unsafe fn unreachable_unchecked() -> ! {
 pub const unsafe fn assert_unchecked(cond: bool) {
     // SAFETY: The caller promised `cond` is true.
     unsafe {
-        intrinsics::assert_unsafe_precondition!(
+        ub_checks::assert_unsafe_precondition!(
             check_language_ub,
             "hint::assert_unchecked must never be called when the condition is false",
             (cond: bool = cond) => cond,

core/src/intrinsics.rs

@@ -67,6 +67,7 @@ use crate::marker::DiscriminantKind;
 use crate::marker::Tuple;
 use crate::mem::align_of;
 use crate::ptr;
+use crate::ub_checks;
 
 pub mod mir;
 pub mod simd;
@@ -2660,38 +2661,22 @@ pub const unsafe fn typed_swap<T>(x: *mut T, y: *mut T) {
     unsafe { ptr::swap_nonoverlapping(x, y, 1) };
 }
 
-/// Returns whether we should check for library UB. This evaluate to the value of `cfg!(debug_assertions)`
-/// during monomorphization.
-///
-/// This intrinsic is evaluated after monomorphization, and therefore branching on this value can
-/// be used to implement debug assertions that are included in the precompiled standard library,
-/// but can be optimized out by builds that monomorphize the standard library code with debug
-/// assertions disabled. This intrinsic is primarily used by [`assert_unsafe_precondition`].
-///
-/// We have separate intrinsics for library UB and language UB because checkers like the const-eval
-/// interpreter and Miri already implement checks for language UB. Since such checkers do not know
-/// about library preconditions, checks guarded by this intrinsic let them find more UB.
-#[rustc_const_unstable(feature = "ub_checks", issue = "none")]
-#[unstable(feature = "core_intrinsics", issue = "none")]
-#[inline(always)]
-#[rustc_intrinsic]
-pub(crate) const fn check_library_ub() -> bool {
-    cfg!(debug_assertions)
-}
-
-/// Returns whether we should check for language UB. This evaluate to the value of `cfg!(debug_assertions)`
-/// during monomorphization.
-///
-/// Since checks implemented at the source level must come strictly before the operation that
-/// executes UB, if we enabled language UB checks in const-eval/Miri we would miss out on the
-/// interpreter's improved diagnostics for the cases that our source-level checks catch.
-///
-/// See `check_library_ub` for more information.
-#[rustc_const_unstable(feature = "ub_checks", issue = "none")]
+/// Returns whether we should perform some UB-checking at runtime. This evaluate to the value of
+/// `cfg!(debug_assertions)` during monomorphization.
+///
+/// This intrinsic is evaluated after monomorphization, which is relevant when mixing crates
+/// compiled with and without debug_assertions. The common case here is a user program built with
+/// debug_assertions linked against the distributed sysroot which is built without debug_assertions.
+/// For code that gets monomorphized in the user crate (i.e., generic functions and functions with
+/// `#[inline]`), gating assertions on `ub_checks()` rather than `cfg!(debug_assertions)` means that
+/// assertions are enabled whenever the *user crate* has debug assertions enabled. However if the
+/// user has debug assertions disabled, the checks will still get optimized out. This intrinsic is
+/// primarily used by [`ub_checks::assert_unsafe_precondition`].
+#[rustc_const_unstable(feature = "const_ub_checks", issue = "none")]
 #[unstable(feature = "core_intrinsics", issue = "none")]
 #[inline(always)]
-#[rustc_intrinsic]
-pub(crate) const fn check_language_ub() -> bool {
+#[cfg_attr(not(bootstrap), rustc_intrinsic)] // just make it a regular fn in bootstrap
+pub(crate) const fn ub_checks() -> bool {
     cfg!(debug_assertions)
 }
 
@@ -2755,132 +2740,6 @@ pub unsafe fn vtable_align(_ptr: *const ()) -> usize {
 // (`transmute` also falls into this category, but it cannot be wrapped due to the
 // check that `T` and `U` have the same size.)
 
-/// Check that the preconditions of an unsafe function are followed. The check is enabled at
-/// runtime if debug assertions are enabled when the caller is monomorphized. In const-eval/Miri
-/// checks implemented with this macro for language UB are always ignored.
-///
-/// This macro should be called as
-/// `assert_unsafe_precondition!(check_{library,lang}_ub, "message", (ident: type = expr, ident: type = expr) => check_expr)`
-/// where each `expr` will be evaluated and passed in as function argument `ident: type`. Then all
-/// those arguments are passed to a function with the body `check_expr`.
-/// Pick `check_language_ub` when this is guarding a violation of language UB, i.e., immediate UB
-/// according to the Rust Abstract Machine. Pick `check_library_ub` when this is guarding a violation
-/// of a documented library precondition that does not *immediately* lead to language UB.
-///
-/// If `check_library_ub` is used but the check is actually guarding language UB, the check will
-/// slow down const-eval/Miri and we'll get the panic message instead of the interpreter's nice
-/// diagnostic, but our ability to detect UB is unchanged.
-/// But if `check_language_ub` is used when the check is actually for library UB, the check is
-/// omitted in const-eval/Miri and thus if we eventually execute language UB which relies on the
-/// library UB, the backtrace Miri reports may be far removed from original cause.
-///
-/// These checks are behind a condition which is evaluated at codegen time, not expansion time like
-/// [`debug_assert`]. This means that a standard library built with optimizations and debug
-/// assertions disabled will have these checks optimized out of its monomorphizations, but if a
-/// caller of the standard library has debug assertions enabled and monomorphizes an expansion of
-/// this macro, that monomorphization will contain the check.
-///
-/// Since these checks cannot be optimized out in MIR, some care must be taken in both call and
-/// implementation to mitigate their compile-time overhead. Calls to this macro always expand to
-/// this structure:
-/// ```ignore (pseudocode)
-/// if ::core::intrinsics::check_language_ub() {
-///     precondition_check(args)
-/// }
-/// ```
-/// where `precondition_check` is monomorphic with the attributes `#[rustc_nounwind]`, `#[inline]` and
-/// `#[rustc_no_mir_inline]`. This combination of attributes ensures that the actual check logic is
-/// compiled only once and generates a minimal amount of IR because the check cannot be inlined in
-/// MIR, but *can* be inlined and fully optimized by a codegen backend.
-///
-/// Callers should avoid introducing any other `let` bindings or any code outside this macro in
-/// order to call it. Since the precompiled standard library is built with full debuginfo and these
-/// variables cannot be optimized out in MIR, an innocent-looking `let` can produce enough
-/// debuginfo to have a measurable compile-time impact on debug builds.
-#[allow_internal_unstable(ub_checks)] // permit this to be called in stably-const fn
-macro_rules! assert_unsafe_precondition {
-    ($kind:ident, $message:expr, ($($name:ident:$ty:ty = $arg:expr),*$(,)?) => $e:expr $(,)?) => {
-        {
-            // This check is inlineable, but not by the MIR inliner.
-            // The reason for this is that the MIR inliner is in an exceptionally bad position
-            // to think about whether or not to inline this. In MIR, this call is gated behind `debug_assertions`,
-            // which will codegen to `false` in release builds. Inlining the check would be wasted work in that case and
-            // would be bad for compile times.
-            //
-            // LLVM on the other hand sees the constant branch, so if it's `false`, it can immediately delete it without
-            // inlining the check. If it's `true`, it can inline it and get significantly better performance.
-            #[rustc_no_mir_inline]
-            #[inline]
-            #[rustc_nounwind]
-            #[rustc_const_unstable(feature = "ub_checks", issue = "none")]
-            const fn precondition_check($($name:$ty),*) {
-                if !$e {
-                    ::core::panicking::panic_nounwind(
-                        concat!("unsafe precondition(s) violated: ", $message)
-                    );
-                }
-            }
-
-            if ::core::intrinsics::$kind() {
-                precondition_check($($arg,)*);
-            }
-        }
-    };
-}
-pub(crate) use assert_unsafe_precondition;
-
-/// Checks whether `ptr` is properly aligned with respect to
-/// `align_of::<T>()`.
-///
-/// In `const` this is approximate and can fail spuriously. It is primarily intended
-/// for `assert_unsafe_precondition!` with `check_language_ub`, in which case the
-/// check is anyway not executed in `const`.
-#[inline]
-pub(crate) const fn is_aligned_and_not_null(ptr: *const (), align: usize) -> bool {
-    !ptr.is_null() && ptr.is_aligned_to(align)
-}
-
-#[inline]
-pub(crate) const fn is_valid_allocation_size(size: usize, len: usize) -> bool {
-    let max_len = if size == 0 { usize::MAX } else { isize::MAX as usize / size };
-    len <= max_len
-}
-
-/// Checks whether the regions of memory starting at `src` and `dst` of size
-/// `count * size` do *not* overlap.
-///
-/// Note that in const-eval this function just returns `true` and therefore must
-/// only be used with `assert_unsafe_precondition!`, similar to `is_aligned_and_not_null`.
-#[inline]
-pub(crate) const fn is_nonoverlapping(
-    src: *const (),
-    dst: *const (),
-    size: usize,
-    count: usize,
-) -> bool {
-    #[inline]
-    fn runtime(src: *const (), dst: *const (), size: usize, count: usize) -> bool {
-        let src_usize = src.addr();
-        let dst_usize = dst.addr();
-        let Some(size) = size.checked_mul(count) else {
-            crate::panicking::panic_nounwind(
-                "is_nonoverlapping: `size_of::<T>() * count` overflows a usize",
-            )
-        };
-        let diff = src_usize.abs_diff(dst_usize);
-        // If the absolute distance between the ptrs is at least as big as the size of the buffer,
-        // they do not overlap.
-        diff >= size
-    }
-
-    #[inline]
-    const fn comptime(_: *const (), _: *const (), _: usize, _: usize) -> bool {
-        true
-    }
-
-    const_eval_select((src, dst, size, count), comptime, runtime)
-}
-
 /// Copies `count * size_of::<T>()` bytes from `src` to `dst`. The source
 /// and destination must *not* overlap.
 ///
@@ -2979,7 +2838,7 @@ pub const unsafe fn copy_nonoverlapping<T>(src: *const T, dst: *mut T, count: us
         pub fn copy_nonoverlapping<T>(src: *const T, dst: *mut T, count: usize);
     }
 
-    assert_unsafe_precondition!(
+    ub_checks::assert_unsafe_precondition!(
         check_language_ub,
         "ptr::copy_nonoverlapping requires that both pointer arguments are aligned and non-null \
         and the specified memory ranges do not overlap",
@@ -2990,9 +2849,9 @@ pub const unsafe fn copy_nonoverlapping<T>(src: *const T, dst: *mut T, count: us
             align: usize = align_of::<T>(),
             count: usize = count,
         ) =>
-        is_aligned_and_not_null(src, align)
-            && is_aligned_and_not_null(dst, align)
-            && is_nonoverlapping(src, dst, size, count)
+        ub_checks::is_aligned_and_not_null(src, align)
+            && ub_checks::is_aligned_and_not_null(dst, align)
+            && ub_checks::is_nonoverlapping(src, dst, size, count)
     );
 
     // SAFETY: the safety contract for `copy_nonoverlapping` must be
@@ -3083,7 +2942,7 @@ pub const unsafe fn copy<T>(src: *const T, dst: *mut T, count: usize) {
 
     // SAFETY: the safety contract for `copy` must be upheld by the caller.
     unsafe {
-        assert_unsafe_precondition!(
+        ub_checks::assert_unsafe_precondition!(
             check_language_ub,
             "ptr::copy_nonoverlapping requires that both pointer arguments are aligned and non-null \
             and the specified memory ranges do not overlap",
@@ -3092,8 +2951,8 @@ pub const unsafe fn copy<T>(src: *const T, dst: *mut T, count: usize) {
                 dst: *mut () = dst as *mut (),
                 align: usize = align_of::<T>(),
             ) =>
-            is_aligned_and_not_null(src, align)
-                && is_aligned_and_not_null(dst, align)
+            ub_checks::is_aligned_and_not_null(src, align)
+                && ub_checks::is_aligned_and_not_null(dst, align)
         );
         copy(src, dst, count)
     }
@@ -3164,13 +3023,13 @@ pub const unsafe fn write_bytes<T>(dst: *mut T, val: u8, count: usize) {
 
     // SAFETY: the safety contract for `write_bytes` must be upheld by the caller.
     unsafe {
-        assert_unsafe_precondition!(
+        ub_checks::assert_unsafe_precondition!(
             check_language_ub,
             "ptr::write_bytes requires that the destination pointer is aligned and non-null",
             (
                 addr: *const () = dst as *const (),
                 align: usize = align_of::<T>(),
-            ) => is_aligned_and_not_null(addr, align)
+            ) => ub_checks::is_aligned_and_not_null(addr, align)
         );
         write_bytes(dst, val, count)
     }

core/src/lib.rs

@@ -171,6 +171,7 @@
 #![feature(const_type_id)]
 #![feature(const_type_name)]
 #![feature(const_typed_swap)]
+#![feature(const_ub_checks)]
 #![feature(const_unicode_case_lookup)]
 #![feature(const_unsafecell_get_mut)]
 #![feature(const_waker)]
@@ -366,6 +367,7 @@ pub mod hint;
 pub mod intrinsics;
 pub mod mem;
 pub mod ptr;
+mod ub_checks;
 
 /* Core language traits */
 

core/src/num/nonzero.rs

@@ -9,6 +9,7 @@ use crate::ops::{BitOr, BitOrAssign, Div, DivAssign, Neg, Rem, RemAssign};
 use crate::panic::{RefUnwindSafe, UnwindSafe};
 use crate::ptr;
 use crate::str::FromStr;
+use crate::ub_checks;
 
 use super::from_str_radix;
 use super::{IntErrorKind, ParseIntError};
@@ -369,7 +370,7 @@ where
             None => {
                 // SAFETY: The caller guarantees that `n` is non-zero, so this is unreachable.
                 unsafe {
-                    intrinsics::assert_unsafe_precondition!(
+                    ub_checks::assert_unsafe_precondition!(
                         check_language_ub,
                         "NonZero::new_unchecked requires the argument to be non-zero",
                         () => false,
@@ -409,7 +410,7 @@ where
             None => {
                 // SAFETY: The caller guarantees that `n` references a value that is non-zero, so this is unreachable.
                 unsafe {
-                    intrinsics::assert_unsafe_precondition!(
+                    ub_checks::assert_unsafe_precondition!(
                         check_library_ub,
                         "NonZero::from_mut_unchecked requires the argument to dereference as non-zero",
                         () => false,

core/src/ops/index_range.rs

@@ -1,6 +1,7 @@
-use crate::intrinsics::{assert_unsafe_precondition, unchecked_add, unchecked_sub};
+use crate::intrinsics::{unchecked_add, unchecked_sub};
 use crate::iter::{FusedIterator, TrustedLen};
 use crate::num::NonZero;
+use crate::ub_checks;
 
 /// Like a `Range<usize>`, but with a safety invariant that `start <= end`.
 ///
@@ -19,7 +20,7 @@ impl IndexRange {
     /// - `start <= end`
     #[inline]
     pub const unsafe fn new_unchecked(start: usize, end: usize) -> Self {
-        assert_unsafe_precondition!(
+        ub_checks::assert_unsafe_precondition!(
             check_library_ub,
             "IndexRange::new_unchecked requires `start <= end`",
             (start: usize = start, end: usize = end) => start <= end,

core/src/ptr/alignment.rs

@@ -1,7 +1,7 @@
 use crate::convert::{TryFrom, TryInto};
-#[cfg(debug_assertions)]
-use crate::intrinsics::assert_unsafe_precondition;
 use crate::num::NonZero;
+#[cfg(debug_assertions)]
+use crate::ub_checks::assert_unsafe_precondition;
 use crate::{cmp, fmt, hash, mem, num};
 
 /// A type storing a `usize` which is a power of two, and thus

core/src/ptr/const_ptr.rs

@@ -818,7 +818,7 @@ impl<T: ?Sized> *const T {
             intrinsics::const_eval_select((this, origin), comptime, runtime)
         }
 
-        assert_unsafe_precondition!(
+        ub_checks::assert_unsafe_precondition!(
             check_language_ub,
             "ptr::sub_ptr requires `self >= origin`",
             (