Revert "Stop generating allocas+memcmp for simple array equality"

This reverts commit 2456495.

Author: Urgau

compiler/rustc_codegen_llvm/src/context.rs

@@ -604,7 +604,6 @@ impl<'ll> CodegenCx<'ll, '_> {
         let t_i32 = self.type_i32();
         let t_i64 = self.type_i64();
         let t_i128 = self.type_i128();
-        let t_isize = self.type_isize();
         let t_f32 = self.type_f32();
         let t_f64 = self.type_f64();
 
@@ -816,10 +815,6 @@ impl<'ll> CodegenCx<'ll, '_> {
         ifn!("llvm.assume", fn(i1) -> void);
         ifn!("llvm.prefetch", fn(i8p, t_i32, t_i32, t_i32) -> void);
 
-        // This isn't an "LLVM intrinsic", but LLVM's optimization passes
-        // recognize it like one and we assume it exists in `core::slice::cmp`
-        ifn!("memcmp", fn(i8p, i8p, t_isize) -> t_i32);
-
         // variadic intrinsics
         ifn!("llvm.va_start", fn(i8p) -> void);
         ifn!("llvm.va_end", fn(i8p) -> void);

compiler/rustc_codegen_llvm/src/intrinsic.rs

@@ -297,43 +297,6 @@ impl<'ll, 'tcx> IntrinsicCallMethods<'tcx> for Builder<'_, 'll, 'tcx> {
                 }
             }
 
-            sym::raw_eq => {
-                use abi::Abi::*;
-                let tp_ty = substs.type_at(0);
-                let layout = self.layout_of(tp_ty).layout;
-                let use_integer_compare = match layout.abi {
-                    Scalar(_) | ScalarPair(_, _) => true,
-                    Uninhabited | Vector { .. } => false,
-                    Aggregate { .. } => {
-                        // For rusty ABIs, small aggregates are actually passed
-                        // as `RegKind::Integer` (see `FnAbi::adjust_for_abi`),
-                        // so we re-use that same threshold here.
-                        layout.size <= self.data_layout().pointer_size * 2
-                    }
-                };
-
-                let a = args[0].immediate();
-                let b = args[1].immediate();
-                if layout.size.bytes() == 0 {
-                    self.const_bool(true)
-                } else if use_integer_compare {
-                    let integer_ty = self.type_ix(layout.size.bits());
-                    let ptr_ty = self.type_ptr_to(integer_ty);
-                    let a_ptr = self.bitcast(a, ptr_ty);
-                    let a_val = self.load(integer_ty, a_ptr, layout.align.abi);
-                    let b_ptr = self.bitcast(b, ptr_ty);
-                    let b_val = self.load(integer_ty, b_ptr, layout.align.abi);
-                    self.icmp(IntPredicate::IntEQ, a_val, b_val)
-                } else {
-                    let i8p_ty = self.type_i8p();
-                    let a_ptr = self.bitcast(a, i8p_ty);
-                    let b_ptr = self.bitcast(b, i8p_ty);
-                    let n = self.const_usize(layout.size.bytes());
-                    let cmp = self.call_intrinsic("memcmp", &[a_ptr, b_ptr, n]);
-                    self.icmp(IntPredicate::IntEQ, cmp, self.const_i32(0))
-                }
-            }
-
             sym::black_box => {
                 args[0].val.store(self, result);
 

compiler/rustc_const_eval/src/interpret/intrinsics.rs

@@ -478,10 +478,6 @@ impl<'mir, 'tcx: 'mir, M: Machine<'mir, 'tcx>> InterpCx<'mir, 'tcx, M> {
                     throw_ub_format!("`assume` intrinsic called with `false`");
                 }
             }
-            sym::raw_eq => {
-                let result = self.raw_eq_intrinsic(&args[0], &args[1])?;
-                self.write_scalar(result, dest)?;
-            }
             _ => return Ok(false),
         }
 
@@ -590,19 +586,4 @@ impl<'mir, 'tcx: 'mir, M: Machine<'mir, 'tcx>> InterpCx<'mir, 'tcx, M> {
         let bytes = std::iter::repeat(byte).take(len.bytes_usize());
         self.memory.write_bytes(dst, bytes)
     }
-
-    pub(crate) fn raw_eq_intrinsic(
-        &mut self,
-        lhs: &OpTy<'tcx, <M as Machine<'mir, 'tcx>>::PointerTag>,
-        rhs: &OpTy<'tcx, <M as Machine<'mir, 'tcx>>::PointerTag>,
-    ) -> InterpResult<'tcx, Scalar<M::PointerTag>> {
-        let layout = self.layout_of(lhs.layout.ty.builtin_deref(true).unwrap().ty)?;
-        assert!(!layout.is_unsized());
-
-        let lhs = self.read_pointer(lhs)?;
-        let rhs = self.read_pointer(rhs)?;
-        let lhs_bytes = self.memory.read_bytes(lhs, layout.size)?;
-        let rhs_bytes = self.memory.read_bytes(rhs, layout.size)?;
-        Ok(Scalar::from_bool(lhs_bytes == rhs_bytes))
-    }
 }

compiler/rustc_span/src/symbol.rs

@@ -1069,7 +1069,6 @@ symbols! {
         quote,
         range_inclusive_new,
         raw_dylib,
-        raw_eq,
         raw_identifiers,
         raw_ref_op,
         re_rebalance_coherence,

compiler/rustc_typeck/src/check/intrinsic.rs

@@ -387,13 +387,6 @@ pub fn check_intrinsic_type(tcx: TyCtxt<'_>, it: &hir::ForeignItem<'_>) {
 
             sym::nontemporal_store => (1, vec![tcx.mk_mut_ptr(param(0)), param(0)], tcx.mk_unit()),
 
-            sym::raw_eq => {
-                let br = ty::BoundRegion { var: ty::BoundVar::from_u32(0), kind: ty::BrAnon(0) };
-                let param_ty =
-                    tcx.mk_imm_ref(tcx.mk_region(ty::ReLateBound(ty::INNERMOST, br)), param(0));
-                (1, vec![param_ty; 2], tcx.types.bool)
-            }
-
             sym::black_box => (1, vec![param(0)], param(0)),
 
             sym::const_eval_select => (4, vec![param(0), param(1), param(2)], param(3)),

library/core/src/array/equality.rs

@@ -1,6 +1,4 @@
 use crate::convert::TryInto;
-use crate::num::{NonZeroI128, NonZeroI16, NonZeroI32, NonZeroI64, NonZeroI8, NonZeroIsize};
-use crate::num::{NonZeroU128, NonZeroU16, NonZeroU32, NonZeroU64, NonZeroU8, NonZeroUsize};
 
 #[stable(feature = "rust1", since = "1.0.0")]
 impl<A, B, const N: usize> PartialEq<[B; N]> for [A; N]
@@ -9,11 +7,11 @@ where
 {
     #[inline]
     fn eq(&self, other: &[B; N]) -> bool {
-        SpecArrayEq::spec_eq(self, other)
+        self[..] == other[..]
     }
     #[inline]
     fn ne(&self, other: &[B; N]) -> bool {
-        SpecArrayEq::spec_ne(self, other)
+        self[..] != other[..]
     }
 }
 
@@ -129,88 +127,3 @@ where
 
 #[stable(feature = "rust1", since = "1.0.0")]
 impl<T: Eq, const N: usize> Eq for [T; N] {}
-
-trait SpecArrayEq<Other, const N: usize>: Sized {
-    fn spec_eq(a: &[Self; N], b: &[Other; N]) -> bool;
-    fn spec_ne(a: &[Self; N], b: &[Other; N]) -> bool;
-}
-
-impl<T: PartialEq<Other>, Other, const N: usize> SpecArrayEq<Other, N> for T {
-    default fn spec_eq(a: &[Self; N], b: &[Other; N]) -> bool {
-        a[..] == b[..]
-    }
-    default fn spec_ne(a: &[Self; N], b: &[Other; N]) -> bool {
-        a[..] != b[..]
-    }
-}
-
-impl<T: IsRawEqComparable<U>, U, const N: usize> SpecArrayEq<U, N> for T {
-    fn spec_eq(a: &[T; N], b: &[U; N]) -> bool {
-        // SAFETY: This is why `IsRawEqComparable` is an `unsafe trait`.
-        unsafe {
-            let b = &*b.as_ptr().cast::<[T; N]>();
-            crate::intrinsics::raw_eq(a, b)
-        }
-    }
-    fn spec_ne(a: &[T; N], b: &[U; N]) -> bool {
-        !Self::spec_eq(a, b)
-    }
-}
-
-/// `U` exists on here mostly because `min_specialization` didn't let me
-/// repeat the `T` type parameter in the above specialization, so instead
-/// the `T == U` constraint comes from the impls on this.
-/// # Safety
-/// - Neither `Self` nor `U` has any padding.
-/// - `Self` and `U` have the same layout.
-/// - `Self: PartialEq<U>` is byte-wise (this means no floats, among other things)
-#[rustc_specialization_trait]
-unsafe trait IsRawEqComparable<U>: PartialEq<U> {}
-
-macro_rules! is_raw_eq_comparable {
-    ($($t:ty),+ $(,)?) => {$(
-        unsafe impl IsRawEqComparable<$t> for $t {}
-    )+};
-}
-
-// SAFETY: All the ordinary integer types allow all bit patterns as distinct values
-is_raw_eq_comparable!(u8, u16, u32, u64, u128, usize, i8, i16, i32, i64, i128, isize);
-
-// SAFETY: bool and char have *niches*, but no *padding*, so this is sound
-is_raw_eq_comparable!(bool, char);
-
-// SAFETY: Similarly, the non-zero types have a niche, but no undef,
-// and they compare like their underlying numeric type.
-is_raw_eq_comparable!(
-    NonZeroU8,
-    NonZeroU16,
-    NonZeroU32,
-    NonZeroU64,
-    NonZeroU128,
-    NonZeroUsize,
-    NonZeroI8,
-    NonZeroI16,
-    NonZeroI32,
-    NonZeroI64,
-    NonZeroI128,
-    NonZeroIsize,
-);
-
-// SAFETY: The NonZero types have the "null" optimization guaranteed, and thus
-// are also safe to equality-compare bitwise inside an `Option`.
-// The way `PartialOrd` is defined for `Option` means that this wouldn't work
-// for `<` or `>` on the signed types, but since we only do `==` it's fine.
-is_raw_eq_comparable!(
-    Option<NonZeroU8>,
-    Option<NonZeroU16>,
-    Option<NonZeroU32>,
-    Option<NonZeroU64>,
-    Option<NonZeroU128>,
-    Option<NonZeroUsize>,
-    Option<NonZeroI8>,
-    Option<NonZeroI16>,
-    Option<NonZeroI32>,
-    Option<NonZeroI64>,
-    Option<NonZeroI128>,
-    Option<NonZeroIsize>,
-);

library/core/src/intrinsics.rs

@@ -1939,25 +1939,6 @@ extern "rust-intrinsic" {
     #[cfg(not(bootstrap))]
     pub fn const_deallocate(ptr: *mut u8, size: usize, align: usize);
 
-    /// Determines whether the raw bytes of the two values are equal.
-    ///
-    /// This is particularly handy for arrays, since it allows things like just
-    /// comparing `i96`s instead of forcing `alloca`s for `[6 x i16]`.
-    ///
-    /// Above some backend-decided threshold this will emit calls to `memcmp`,
-    /// like slice equality does, instead of causing massive code size.
-    ///
-    /// # Safety
-    ///
-    /// It's UB to call this if any of the *bytes* in `*a` or `*b` are uninitialized.
-    /// Note that this is a stricter criterion than just the *values* being
-    /// fully-initialized: if `T` has padding, it's UB to call this intrinsic.
-    ///
-    /// (The implementation is allowed to branch on the results of comparisons,
-    /// which is UB if any of their inputs are `undef`.)
-    #[rustc_const_unstable(feature = "const_intrinsic_raw_eq", issue = "none")]
-    pub fn raw_eq<T>(a: &T, b: &T) -> bool;
-
     /// See documentation of [`std::hint::black_box`] for details.
     ///
     /// [`std::hint::black_box`]: crate::hint::black_box

src/test/codegen/slice-ref-equality.rs

@@ -2,25 +2,6 @@
 
 #![crate_type = "lib"]
 
-// #71602 reported a simple array comparison just generating a loop.
-// This was originally fixed by ensuring it generates a single bcmp,
-// but we now generate it as a load+icmp instead. `is_zero_slice` was
-// tweaked to still test the case of comparison against a slice,
-// and `is_zero_array` tests the new array-specific behaviour.
-// The optimization was then extended to short slice-to-array comparisons,
-// so the first test here now has a long slice to still get the bcmp.
-
-// CHECK-LABEL: @is_zero_slice_long
-#[no_mangle]
-pub fn is_zero_slice_long(data: &[u8; 456]) -> bool {
-    // CHECK: :
-    // CHECK-NEXT: %{{.+}} = getelementptr {{.+}}
-    // CHECK-NEXT: %[[BCMP:.+]] = tail call i32 @{{bcmp|memcmp}}({{.+}})
-    // CHECK-NEXT: %[[EQ:.+]] = icmp eq i32 %[[BCMP]], 0
-    // CHECK-NEXT: ret i1 %[[EQ]]
-    &data[..] == [0; 456]
-}
-
 // CHECK-LABEL: @is_zero_slice_short
 #[no_mangle]
 pub fn is_zero_slice_short(data: &[u8; 4]) -> bool {