diff --git a/compiler/rustc_codegen_ssa/src/mir/intrinsic.rs b/compiler/rustc_codegen_ssa/src/mir/intrinsic.rs index 1d1826d984474..f6be8f2ba5d37 100644 --- a/compiler/rustc_codegen_ssa/src/mir/intrinsic.rs +++ b/compiler/rustc_codegen_ssa/src/mir/intrinsic.rs @@ -76,6 +76,22 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> { let name = bx.tcx().item_name(def_id); let name_str = name.as_str(); + // If we're swapping something that's *not* an `OperandValue::Ref`, + // then we can do it directly and avoid the alloca. + // Otherwise, we'll let the fallback MIR body take care of it. + if let sym::typed_swap = name { + let pointee_ty = fn_args.type_at(0); + let pointee_layout = bx.layout_of(pointee_ty); + if !bx.is_backend_ref(pointee_layout) { + let x_place = PlaceRef::new_sized(args[0].immediate(), pointee_layout); + let y_place = PlaceRef::new_sized(args[1].immediate(), pointee_layout); + let temp = bx.load_operand(x_place); + bx.typed_place_copy(x_place, y_place); + temp.val.store(bx, y_place); + return Ok(()); + } + } + let llret_ty = bx.backend_type(bx.layout_of(ret_ty)); let result = PlaceRef::new_sized(llresult, fn_abi.ret.layout); diff --git a/compiler/rustc_codegen_ssa/src/traits/builder.rs b/compiler/rustc_codegen_ssa/src/traits/builder.rs index 36f37e3791bc5..36ece77b4ee5c 100644 --- a/compiler/rustc_codegen_ssa/src/traits/builder.rs +++ b/compiler/rustc_codegen_ssa/src/traits/builder.rs @@ -1,5 +1,6 @@ use super::abi::AbiBuilderMethods; use super::asm::AsmBuilderMethods; +use super::consts::ConstMethods; use super::coverageinfo::CoverageInfoBuilderMethods; use super::debuginfo::DebugInfoBuilderMethods; use super::intrinsic::IntrinsicCallMethods; @@ -267,6 +268,26 @@ pub trait BuilderMethods<'a, 'tcx>: flags: MemFlags, ); + /// *Typed* copy for non-overlapping places. + /// + /// Has a default implementation in terms of `memcpy`, but specific backends + /// can override to do something smarter if possible. + /// + /// (For example, typed load-stores with alias metadata.) + fn typed_place_copy( + &mut self, + dst: PlaceRef<'tcx, Self::Value>, + src: PlaceRef<'tcx, Self::Value>, + ) { + debug_assert!(src.llextra.is_none()); + debug_assert!(dst.llextra.is_none()); + debug_assert_eq!(dst.layout.size, src.layout.size); + if !dst.layout.is_zst() { + let bytes = self.const_usize(dst.layout.size.bytes()); + self.memcpy(dst.llval, dst.align, src.llval, src.align, bytes, MemFlags::empty()); + } + } + fn select( &mut self, cond: Self::Value, diff --git a/compiler/rustc_codegen_ssa/src/traits/type_.rs b/compiler/rustc_codegen_ssa/src/traits/type_.rs index 72cce43a3fac2..555833759ebce 100644 --- a/compiler/rustc_codegen_ssa/src/traits/type_.rs +++ b/compiler/rustc_codegen_ssa/src/traits/type_.rs @@ -120,6 +120,20 @@ pub trait LayoutTypeMethods<'tcx>: Backend<'tcx> { immediate: bool, ) -> Self::Type; + /// A type that produces an [`OperandValue::Ref`] when loaded. + /// + /// AKA one that's not a ZST, not `is_backend_immediate`, and + /// not `is_backend_scalar_pair`. For such a type, a + /// [`load_operand`] doesn't actually `load` anything. + /// + /// [`OperandValue::Ref`]: crate::mir::operand::OperandValue::Ref + /// [`load_operand`]: super::BuilderMethods::load_operand + fn is_backend_ref(&self, layout: TyAndLayout<'tcx>) -> bool { + !(layout.is_zst() + || self.is_backend_immediate(layout) + || self.is_backend_scalar_pair(layout)) + } + /// A type that can be used in a [`super::BuilderMethods::load`] + /// [`super::BuilderMethods::store`] pair to implement a *typed* copy, /// such as a MIR `*_0 = *_1`. diff --git a/compiler/rustc_const_eval/src/interpret/intrinsics.rs b/compiler/rustc_const_eval/src/interpret/intrinsics.rs index b68bfb4211d1e..fda7b954c7d16 100644 --- a/compiler/rustc_const_eval/src/interpret/intrinsics.rs +++ b/compiler/rustc_const_eval/src/interpret/intrinsics.rs @@ -21,8 +21,8 @@ use rustc_span::symbol::{sym, Symbol}; use rustc_target::abi::Size; use super::{ - util::ensure_monomorphic_enough, CheckInAllocMsg, ImmTy, InterpCx, MPlaceTy, Machine, OpTy, - Pointer, + memory::MemoryKind, util::ensure_monomorphic_enough, CheckInAllocMsg, ImmTy, InterpCx, + MPlaceTy, Machine, OpTy, Pointer, }; use crate::fluent_generated as fluent; @@ -415,6 +415,9 @@ impl<'mir, 'tcx: 'mir, M: Machine<'mir, 'tcx>> InterpCx<'mir, 'tcx, M> { let result = self.raw_eq_intrinsic(&args[0], &args[1])?; self.write_scalar(result, dest)?; } + sym::typed_swap => { + self.typed_swap_intrinsic(&args[0], &args[1])?; + } sym::vtable_size => { let ptr = self.read_pointer(&args[0])?; @@ -608,6 +611,24 @@ impl<'mir, 'tcx: 'mir, M: Machine<'mir, 'tcx>> InterpCx<'mir, 'tcx, M> { self.mem_copy(src, dst, size, nonoverlapping) } + /// Does a *typed* swap of `*left` and `*right`. + pub(crate) fn typed_swap_intrinsic( + &mut self, + left: &OpTy<'tcx, >::Provenance>, + right: &OpTy<'tcx, >::Provenance>, + ) -> InterpResult<'tcx> { + let left = self.deref_pointer(left)?; + let right = self.deref_pointer(right)?; + debug_assert_eq!(left.layout, right.layout); + let kind = MemoryKind::Stack; + let temp = self.allocate(left.layout, kind)?; + self.copy_op(&left, &temp)?; + self.copy_op(&right, &left)?; + self.copy_op(&temp, &right)?; + self.deallocate_ptr(temp.ptr(), None, kind)?; + Ok(()) + } + pub(crate) fn write_bytes_intrinsic( &mut self, dst: &OpTy<'tcx, >::Provenance>, diff --git a/compiler/rustc_hir_analysis/src/check/intrinsic.rs b/compiler/rustc_hir_analysis/src/check/intrinsic.rs index bf5838143d945..9869d59eb3b09 100644 --- a/compiler/rustc_hir_analysis/src/check/intrinsic.rs +++ b/compiler/rustc_hir_analysis/src/check/intrinsic.rs @@ -500,6 +500,8 @@ pub fn check_intrinsic_type( (1, 0, vec![Ty::new_mut_ptr(tcx, param(0)), param(0)], Ty::new_unit(tcx)) } + sym::typed_swap => (1, 1, vec![Ty::new_mut_ptr(tcx, param(0)); 2], Ty::new_unit(tcx)), + sym::discriminant_value => { let assoc_items = tcx.associated_item_def_ids( tcx.require_lang_item(hir::LangItem::DiscriminantKind, None), diff --git a/compiler/rustc_span/src/symbol.rs b/compiler/rustc_span/src/symbol.rs index e43c9533382e1..21bea6d1f4cf3 100644 --- a/compiler/rustc_span/src/symbol.rs +++ b/compiler/rustc_span/src/symbol.rs @@ -1835,6 +1835,7 @@ symbols! { type_macros, type_name, type_privacy_lints, + typed_swap, u128, u128_legacy_const_max, u128_legacy_const_min, diff --git a/library/core/src/mem/mod.rs b/library/core/src/mem/mod.rs index e247b9ed83df5..098c731ed110f 100644 --- a/library/core/src/mem/mod.rs +++ b/library/core/src/mem/mod.rs @@ -726,31 +726,19 @@ pub unsafe fn uninitialized() -> T { #[rustc_const_unstable(feature = "const_swap", issue = "83163")] #[rustc_diagnostic_item = "mem_swap"] pub const fn swap(x: &mut T, y: &mut T) { - // NOTE(eddyb) SPIR-V's Logical addressing model doesn't allow for arbitrary - // reinterpretation of values as (chunkable) byte arrays, and the loop in the - // block optimization in `swap_slice` is hard to rewrite back - // into the (unoptimized) direct swapping implementation, so we disable it. - #[cfg(not(any(target_arch = "spirv")))] - { - // For types that are larger multiples of their alignment, the simple way - // tends to copy the whole thing to stack rather than doing it one part - // at a time, so instead treat them as one-element slices and piggy-back - // the slice optimizations that will split up the swaps. - if const { size_of::() / align_of::() > 2 } { - // SAFETY: exclusive references always point to one non-overlapping - // element and are non-null and properly aligned. - return unsafe { ptr::swap_nonoverlapping(x, y, 1) }; + cfg_if! { + // NOTE(eddyb) SPIR-V's Logical addressing model doesn't allow for arbitrary + // reinterpretation of values as (chunkable) byte arrays, and the loop in the + // block optimization in `ptr::swap_nonoverlapping` is hard to rewrite back + // into the (unoptimized) direct swapping implementation, so we disable it. + if #[cfg(any(target_arch = "spirv"))] { + swap_simple(x, y) + } else { + // SAFETY: `&mut` guarantees these are typed readable and writable + // as well as non-overlapping. + unsafe { ptr::typed_swap(x, y) } } } - - // If a scalar consists of just a small number of alignment units, let - // the codegen just swap those pieces directly, as it's likely just a - // few instructions and anything else is probably overcomplicated. - // - // Most importantly, this covers primitives and simd types that tend to - // have size=align where doing anything else can be a pessimization. - // (This will also be used for ZSTs, though any solution works for them.) - swap_simple(x, y); } /// Same as [`swap`] semantically, but always uses the simple implementation. diff --git a/library/core/src/ptr/mod.rs b/library/core/src/ptr/mod.rs index 30af752236846..08ebf999bedda 100644 --- a/library/core/src/ptr/mod.rs +++ b/library/core/src/ptr/mod.rs @@ -954,6 +954,23 @@ pub const unsafe fn swap(x: *mut T, y: *mut T) { } } +/// Non-overlapping *typed* swap of a single value. +/// +/// The codegen backends will replace this with a better implementation when +/// `T` is a simple type that can be an immediate. +/// +/// # Safety +/// +/// `x` and `y` are readable and writable as `T`, and non-overlapping. +#[rustc_nounwind] +#[inline] +#[cfg_attr(not(bootstrap), rustc_intrinsic)] +pub(crate) const unsafe fn typed_swap(x: *mut T, y: *mut T) { + // SAFETY: The caller provided single non-overlapping items behind + // pointers, so swapping them with `count: 1` is fine. + unsafe { swap_nonoverlapping(x, y, 1) }; +} + /// Swaps `count * size_of::()` bytes between the two regions of memory /// beginning at `x` and `y`. The two regions must *not* overlap. /// diff --git a/src/tools/miri/tests/fail/intrinsics/typed-swap-invalid-array.rs b/src/tools/miri/tests/fail/intrinsics/typed-swap-invalid-array.rs new file mode 100644 index 0000000000000..488e431187f73 --- /dev/null +++ b/src/tools/miri/tests/fail/intrinsics/typed-swap-invalid-array.rs @@ -0,0 +1,22 @@ +#![feature(rustc_attrs)] + +use std::ptr::addr_of_mut; + +#[rustc_intrinsic] +const unsafe fn typed_swap(_: *mut T, _: *mut T) { + unreachable!() +} + +fn invalid_array() { + let mut a = [1_u8; 100]; + let mut b = [2_u8; 100]; + unsafe { + let a = &mut *addr_of_mut!(a).cast::<[bool; 100]>(); + let b = &mut *addr_of_mut!(b).cast::<[bool; 100]>(); + typed_swap::<_, 0>(a, b); //~ERROR: constructing invalid value + } +} + +fn main() { + invalid_array(); +} diff --git a/src/tools/miri/tests/fail/intrinsics/typed-swap-invalid-array.stderr b/src/tools/miri/tests/fail/intrinsics/typed-swap-invalid-array.stderr new file mode 100644 index 0000000000000..d008265797fe2 --- /dev/null +++ b/src/tools/miri/tests/fail/intrinsics/typed-swap-invalid-array.stderr @@ -0,0 +1,20 @@ +error: Undefined Behavior: constructing invalid value at [0]: encountered 0x02, but expected a boolean + --> $DIR/typed-swap-invalid-array.rs:LL:CC + | +LL | typed_swap::<_, 0>(a, b); + | ^^^^^^^^^^^^^^^^^^^^^^^^ constructing invalid value at [0]: encountered 0x02, but expected a boolean + | + = help: this indicates a bug in the program: it performed an invalid operation, and caused Undefined Behavior + = help: see https://doc.rust-lang.org/nightly/reference/behavior-considered-undefined.html for further information + = note: BACKTRACE: + = note: inside `invalid_array` at $DIR/typed-swap-invalid-array.rs:LL:CC +note: inside `main` + --> $DIR/typed-swap-invalid-array.rs:LL:CC + | +LL | invalid_array(); + | ^^^^^^^^^^^^^^^ + +note: some details are omitted, run with `MIRIFLAGS=-Zmiri-backtrace=full` for a verbose backtrace + +error: aborting due to 1 previous error + diff --git a/src/tools/miri/tests/fail/intrinsics/typed-swap-invalid-scalar.rs b/src/tools/miri/tests/fail/intrinsics/typed-swap-invalid-scalar.rs new file mode 100644 index 0000000000000..1f62a95e76901 --- /dev/null +++ b/src/tools/miri/tests/fail/intrinsics/typed-swap-invalid-scalar.rs @@ -0,0 +1,22 @@ +#![feature(rustc_attrs)] + +use std::ptr::addr_of_mut; + +#[rustc_intrinsic] +const unsafe fn typed_swap(_: *mut T, _: *mut T) { + unreachable!() +} + +fn invalid_scalar() { + let mut a = 1_u8; + let mut b = 2_u8; + unsafe { + let a = &mut *addr_of_mut!(a).cast::(); + let b = &mut *addr_of_mut!(b).cast::(); + typed_swap::<_, 0>(a, b); //~ERROR: constructing invalid value + } +} + +fn main() { + invalid_scalar(); +} diff --git a/src/tools/miri/tests/fail/intrinsics/typed-swap-invalid-scalar.stderr b/src/tools/miri/tests/fail/intrinsics/typed-swap-invalid-scalar.stderr new file mode 100644 index 0000000000000..9022348cb480a --- /dev/null +++ b/src/tools/miri/tests/fail/intrinsics/typed-swap-invalid-scalar.stderr @@ -0,0 +1,20 @@ +error: Undefined Behavior: constructing invalid value: encountered 0x02, but expected a boolean + --> $DIR/typed-swap-invalid-scalar.rs:LL:CC + | +LL | typed_swap::<_, 0>(a, b); + | ^^^^^^^^^^^^^^^^^^^^^^^^ constructing invalid value: encountered 0x02, but expected a boolean + | + = help: this indicates a bug in the program: it performed an invalid operation, and caused Undefined Behavior + = help: see https://doc.rust-lang.org/nightly/reference/behavior-considered-undefined.html for further information + = note: BACKTRACE: + = note: inside `invalid_scalar` at $DIR/typed-swap-invalid-scalar.rs:LL:CC +note: inside `main` + --> $DIR/typed-swap-invalid-scalar.rs:LL:CC + | +LL | invalid_scalar(); + | ^^^^^^^^^^^^^^^^ + +note: some details are omitted, run with `MIRIFLAGS=-Zmiri-backtrace=full` for a verbose backtrace + +error: aborting due to 1 previous error + diff --git a/tests/assembly/x86_64-typed-swap.rs b/tests/assembly/x86_64-typed-swap.rs new file mode 100644 index 0000000000000..dd2a9f6edae8b --- /dev/null +++ b/tests/assembly/x86_64-typed-swap.rs @@ -0,0 +1,50 @@ +//@ only-x86_64 +//@ assembly-output: emit-asm +//@ compile-flags: --crate-type=lib -O + +use std::arch::x86_64::__m128; +use std::mem::swap; + +// CHECK-LABEL: swap_i32: +#[no_mangle] +pub fn swap_i32(x: &mut i32, y: &mut i32) { + // CHECK: movl (%[[ARG1:.+]]), %[[T1:.+]] + // CHECK: movl (%[[ARG2:.+]]), %[[T2:.+]] + // CHECK: movl %[[T2]], (%[[ARG1]]) + // CHECK: movl %[[T1]], (%[[ARG2]]) + // CHECK: retq + swap(x, y) +} + +// CHECK-LABEL: swap_pair: +#[no_mangle] +pub fn swap_pair(x: &mut (i32, u32), y: &mut (i32, u32)) { + // CHECK: movq (%[[ARG1]]), %[[T1:.+]] + // CHECK: movq (%[[ARG2]]), %[[T2:.+]] + // CHECK: movq %[[T2]], (%[[ARG1]]) + // CHECK: movq %[[T1]], (%[[ARG2]]) + // CHECK: retq + swap(x, y) +} + +// CHECK-LABEL: swap_str: +#[no_mangle] +pub fn swap_str<'a>(x: &mut &'a str, y: &mut &'a str) { + // CHECK: movups (%[[ARG1]]), %[[T1:xmm.]] + // CHECK: movups (%[[ARG2]]), %[[T2:xmm.]] + // CHECK: movups %[[T2]], (%[[ARG1]]) + // CHECK: movups %[[T1]], (%[[ARG2]]) + // CHECK: retq + swap(x, y) +} + +// CHECK-LABEL: swap_simd: +#[no_mangle] +pub fn swap_simd(x: &mut __m128, y: &mut __m128) { + // CHECK: movaps (%[[ARG1]]), %[[T1:xmm.]] + // CHECK: movaps (%[[ARG2]]), %[[T2:xmm.]] + // CHECK: movaps %[[T2]], (%[[ARG1]]) + // CHECK: movaps %[[T1]], (%[[ARG2]]) + // CHECK: retq + swap(x, y) +} diff --git a/tests/codegen/intrinsics/typed_swap.rs b/tests/codegen/intrinsics/typed_swap.rs new file mode 100644 index 0000000000000..704cc6b3a1f83 --- /dev/null +++ b/tests/codegen/intrinsics/typed_swap.rs @@ -0,0 +1,58 @@ +//@ compile-flags: -O -C no-prepopulate-passes +//@ only-64bit (so I don't need to worry about usize) + +#![crate_type = "lib"] + +use std::mem::swap; + +// CHECK-LABEL: @swap_unit( +#[no_mangle] +pub fn swap_unit(x: &mut (), y: &mut ()) { + // CHECK: start + // CHECK-NEXT: ret void + swap(x, y) +} + +// CHECK-LABEL: @swap_i32( +#[no_mangle] +pub fn swap_i32(x: &mut i32, y: &mut i32) { + // CHECK-NOT: alloca + + // CHECK: %[[TEMP:.+]] = load i32, ptr %x, align 4 + // CHECK-SAME: !noundef + // CHECK: call void @llvm.memcpy.p0.p0.i64(ptr align 4 %x, ptr align 4 %y, i64 4, i1 false) + // CHECK: store i32 %0, ptr %y, align 4 + // CHECK: ret void + swap(x, y) +} + +// CHECK-LABEL: @swap_pair( +#[no_mangle] +pub fn swap_pair(x: &mut (i32, u32), y: &mut (i32, u32)) { + // CHECK-NOT: alloca + + // CHECK: load i32 + // CHECK-SAME: !noundef + // CHECK: load i32 + // CHECK-SAME: !noundef + // CHECK: call void @llvm.memcpy.p0.p0.i64(ptr align 4 %x, ptr align 4 %y, i64 8, i1 false) + // CHECK: store i32 + // CHECK: store i32 + swap(x, y) +} + +// CHECK-LABEL: @swap_str( +#[no_mangle] +pub fn swap_str<'a>(x: &mut &'a str, y: &mut &'a str) { + // CHECK-NOT: alloca + + // CHECK: load ptr + // CHECK-SAME: !nonnull + // CHECK-SAME: !noundef + // CHECK: load i64 + // CHECK-SAME: !noundef + // CHECK: call void @llvm.memcpy.p0.p0.i64(ptr align 8 %x, ptr align 8 %y, i64 16, i1 false) + // CHECK: store ptr + // CHECK: store i64 + swap(x, y) +} diff --git a/tests/codegen/swap-small-types.rs b/tests/codegen/swap-small-types.rs index 5fdf4a5804a9c..4dcfed2a53af9 100644 --- a/tests/codegen/swap-small-types.rs +++ b/tests/codegen/swap-small-types.rs @@ -70,10 +70,7 @@ pub fn swap_slices<'a>(x: &mut &'a [u32], y: &mut &'a [u32]) { // CHECK-NOT: alloca // CHECK: load ptr // CHECK: load i64 - // CHECK: load ptr - // CHECK: load i64 - // CHECK: store ptr - // CHECK: store i64 + // CHECK: call void @llvm.memcpy.p0.p0.i64({{.+}}, i64 16, i1 false) // CHECK: store ptr // CHECK: store i64 swap(x, y)