Auto merge of #118310 - scottmcm:three-way-compare, r=davidtwco

Add `Ord::cmp` for primitives as a `BinOp` in MIR

Update: most of this OP was written months ago.  See #118310 (comment) below for where we got to recently that made it ready for review.

---

There are dozens of reasonable ways to implement `Ord::cmp` for integers using comparison, bit-ops, and branches.  Those differences are irrelevant at the rust level, however, so we can make things better by adding `BinOp::Cmp` at the MIR level:

1. Exactly how to implement it is left up to the backends, so LLVM can use whatever pattern its optimizer best recognizes and cranelift can use whichever pattern codegens the fastest.
2. By not inlining those details for every use of `cmp`, we drastically reduce the amount of MIR generated for `derive`d `PartialOrd`, while also making it more amenable to MIR-level optimizations.

Having extremely careful `if` ordering to μoptimize resource usage on broadwell (#63767) is great, but it really feels to me like libcore is the wrong place to put that logic.  Similarly, using subtraction [tricks](https://graphics.stanford.edu/~seander/bithacks.html#CopyIntegerSign) (#105840) is arguably even nicer, but depends on the optimizer understanding it (llvm/llvm-project#73417) to be practical.  Or maybe [bitor is better than add](https://discourse.llvm.org/t/representing-in-ir/67369/2?u=scottmcm)?  But maybe only on a future version that [has `or disjoint` support](https://discourse.llvm.org/t/rfc-add-or-disjoint-flag/75036?u=scottmcm)?  And just because one of those forms happens to be good for LLVM, there's no guarantee that it'd be the same form that GCC or Cranelift would rather see -- especially given their very different optimizers.  Not to mention that if LLVM gets a spaceship intrinsic -- [which it should](https://rust-lang.zulipchat.com/#narrow/stream/131828-t-compiler/topic/Suboptimal.20inlining.20in.20std.20function.20.60binary_search.60/near/404250586) -- we'll need at least a rustc intrinsic to be able to call it.

As for simplifying it in Rust, we now regularly inline `{integer}::partial_cmp`, but it's quite a large amount of IR.  The best way to see that is with 8811efa#diff-d134c32d028fbe2bf835fef2df9aca9d13332dd82284ff21ee7ebf717bfa4765R113 -- I added a new pre-codegen MIR test for a simple 3-tuple struct, and this PR change it from 36 locals and 26 basic blocks down to 24 locals and 8 basic blocks.  Even better, as soon as the construct-`Some`-then-match-it-in-same-BB noise is cleaned up, this'll expose the `Cmp == 0` branches clearly in MIR, so that an InstCombine (#105808) can simplify that to just a `BinOp::Eq` and thus fix some of our generated code perf issues.  (Tracking that through today's `if a < b { Less } else if a == b { Equal } else { Greater }` would be *much* harder.)

---

r? `@ghost`
But first I should check that perf is ok with this
~~...and my true nemesis, tidy.~~

compiler/rustc_codegen_cranelift/src/codegen_i128.rs

@@ -68,7 +68,7 @@ pub(crate) fn maybe_codegen<'tcx>(
                 Some(CValue::by_val(ret_val, lhs.layout()))
             }
         }
-        BinOp::Lt | BinOp::Le | BinOp::Eq | BinOp::Ge | BinOp::Gt | BinOp::Ne => None,
+        BinOp::Lt | BinOp::Le | BinOp::Eq | BinOp::Ge | BinOp::Gt | BinOp::Ne | BinOp::Cmp => None,
         BinOp::Shl | BinOp::ShlUnchecked | BinOp::Shr | BinOp::ShrUnchecked => None,
     }
 }
@@ -134,6 +134,7 @@ pub(crate) fn maybe_codegen_checked<'tcx>(
         BinOp::AddUnchecked | BinOp::SubUnchecked | BinOp::MulUnchecked => unreachable!(),
         BinOp::Offset => unreachable!("offset should only be used on pointers, not 128bit ints"),
         BinOp::Div | BinOp::Rem => unreachable!(),
+        BinOp::Cmp => unreachable!(),
         BinOp::Lt | BinOp::Le | BinOp::Eq | BinOp::Ge | BinOp::Gt | BinOp::Ne => unreachable!(),
         BinOp::Shl | BinOp::ShlUnchecked | BinOp::Shr | BinOp::ShrUnchecked => unreachable!(),
     }

compiler/rustc_codegen_cranelift/src/num.rs

@@ -40,13 +40,33 @@ pub(crate) fn bin_op_to_intcc(bin_op: BinOp, signed: bool) -> Option<IntCC> {
     })
 }
 
+fn codegen_three_way_compare<'tcx>(
+    fx: &mut FunctionCx<'_, '_, 'tcx>,
+    signed: bool,
+    lhs: Value,
+    rhs: Value,
+) -> CValue<'tcx> {
+    // This emits `(lhs > rhs) - (lhs < rhs)`, which is cranelift's preferred form per
+    // <https://github.com/bytecodealliance/wasmtime/blob/8052bb9e3b792503b225f2a5b2ba3bc023bff462/cranelift/codegen/src/prelude_opt.isle#L41-L47>
+    let gt_cc = crate::num::bin_op_to_intcc(BinOp::Gt, signed).unwrap();
+    let lt_cc = crate::num::bin_op_to_intcc(BinOp::Lt, signed).unwrap();
+    let gt = fx.bcx.ins().icmp(gt_cc, lhs, rhs);
+    let lt = fx.bcx.ins().icmp(lt_cc, lhs, rhs);
+    let val = fx.bcx.ins().isub(gt, lt);
+    CValue::by_val(val, fx.layout_of(fx.tcx.ty_ordering_enum(Some(fx.mir.span))))
+}
+
 fn codegen_compare_bin_op<'tcx>(
     fx: &mut FunctionCx<'_, '_, 'tcx>,
     bin_op: BinOp,
     signed: bool,
     lhs: Value,
     rhs: Value,
 ) -> CValue<'tcx> {
+    if bin_op == BinOp::Cmp {
+        return codegen_three_way_compare(fx, signed, lhs, rhs);
+    }
+
     let intcc = crate::num::bin_op_to_intcc(bin_op, signed).unwrap();
     let val = fx.bcx.ins().icmp(intcc, lhs, rhs);
     CValue::by_val(val, fx.layout_of(fx.tcx.types.bool))
@@ -59,7 +79,7 @@ pub(crate) fn codegen_binop<'tcx>(
     in_rhs: CValue<'tcx>,
 ) -> CValue<'tcx> {
     match bin_op {
-        BinOp::Eq | BinOp::Lt | BinOp::Le | BinOp::Ne | BinOp::Ge | BinOp::Gt => {
+        BinOp::Eq | BinOp::Lt | BinOp::Le | BinOp::Ne | BinOp::Ge | BinOp::Gt | BinOp::Cmp => {
             match in_lhs.layout().ty.kind() {
                 ty::Bool | ty::Uint(_) | ty::Int(_) | ty::Char => {
                     let signed = type_sign(in_lhs.layout().ty);
@@ -160,7 +180,7 @@ pub(crate) fn codegen_int_binop<'tcx>(
         }
         BinOp::Offset => unreachable!("Offset is not an integer operation"),
         // Compare binops handles by `codegen_binop`.
-        BinOp::Eq | BinOp::Ne | BinOp::Lt | BinOp::Le | BinOp::Gt | BinOp::Ge => {
+        BinOp::Eq | BinOp::Ne | BinOp::Lt | BinOp::Le | BinOp::Gt | BinOp::Ge | BinOp::Cmp => {
             unreachable!("{:?}({:?}, {:?})", bin_op, in_lhs.layout().ty, in_rhs.layout().ty);
         }
     };

compiler/rustc_codegen_gcc/src/common.rs

@@ -94,6 +94,10 @@ impl<'gcc, 'tcx> ConstMethods<'tcx> for CodegenCx<'gcc, 'tcx> {
         self.const_int(self.type_i32(), i as i64)
     }
 
+    fn const_i8(&self, i: i8) -> RValue<'gcc> {
+        self.const_int(self.type_i8(), i as i64)
+    }
+
     fn const_u32(&self, i: u32) -> RValue<'gcc> {
         self.const_uint(self.type_u32(), i as u64)
     }

compiler/rustc_codegen_llvm/src/common.rs

@@ -160,6 +160,10 @@ impl<'ll, 'tcx> ConstMethods<'tcx> for CodegenCx<'ll, 'tcx> {
         self.const_int(self.type_i32(), i as i64)
     }
 
+    fn const_i8(&self, i: i8) -> &'ll Value {
+        self.const_int(self.type_i8(), i as i64)
+    }
+
     fn const_u32(&self, i: u32) -> &'ll Value {
         self.const_uint(self.type_i32(), i as u64)
     }

compiler/rustc_codegen_ssa/src/mir/rvalue.rs

@@ -7,6 +7,7 @@ use crate::common::{self, IntPredicate};
 use crate::traits::*;
 use crate::MemFlags;
 
+use rustc_hir as hir;
 use rustc_middle::mir;
 use rustc_middle::mir::Operand;
 use rustc_middle::ty::cast::{CastTy, IntTy};
@@ -882,6 +883,35 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
                     bx.icmp(base::bin_op_to_icmp_predicate(op.to_hir_binop(), is_signed), lhs, rhs)
                 }
             }
+            mir::BinOp::Cmp => {
+                use std::cmp::Ordering;
+                debug_assert!(!is_float);
+                let pred = |op| base::bin_op_to_icmp_predicate(op, is_signed);
+                if bx.cx().tcx().sess.opts.optimize == OptLevel::No {
+                    // FIXME: This actually generates tighter assembly, and is a classic trick
+                    // <https://graphics.stanford.edu/~seander/bithacks.html#CopyIntegerSign>
+                    // However, as of 2023-11 it optimizes worse in things like derived
+                    // `PartialOrd`, so only use it in debug for now. Once LLVM can handle it
+                    // better (see <https://github.com/llvm/llvm-project/issues/73417>), it'll
+                    // be worth trying it in optimized builds as well.
+                    let is_gt = bx.icmp(pred(hir::BinOpKind::Gt), lhs, rhs);
+                    let gtext = bx.zext(is_gt, bx.type_i8());
+                    let is_lt = bx.icmp(pred(hir::BinOpKind::Lt), lhs, rhs);
+                    let ltext = bx.zext(is_lt, bx.type_i8());
+                    bx.unchecked_ssub(gtext, ltext)
+                } else {
+                    // These operations are those expected by `tests/codegen/integer-cmp.rs`,
+                    // from <https://github.com/rust-lang/rust/pull/63767>.
+                    let is_lt = bx.icmp(pred(hir::BinOpKind::Lt), lhs, rhs);
+                    let is_ne = bx.icmp(pred(hir::BinOpKind::Ne), lhs, rhs);
+                    let ge = bx.select(
+                        is_ne,
+                        bx.cx().const_i8(Ordering::Greater as i8),
+                        bx.cx().const_i8(Ordering::Equal as i8),
+                    );
+                    bx.select(is_lt, bx.cx().const_i8(Ordering::Less as i8), ge)
+                }
+            }
         }
     }
 

compiler/rustc_codegen_ssa/src/traits/consts.rs

@@ -19,6 +19,7 @@ pub trait ConstMethods<'tcx>: BackendTypes {
     fn const_bool(&self, val: bool) -> Self::Value;
     fn const_i16(&self, i: i16) -> Self::Value;
     fn const_i32(&self, i: i32) -> Self::Value;
+    fn const_i8(&self, i: i8) -> Self::Value;
     fn const_u32(&self, i: u32) -> Self::Value;
     fn const_u64(&self, i: u64) -> Self::Value;
     fn const_u128(&self, i: u128) -> Self::Value;

compiler/rustc_const_eval/src/interpret/operand.rs

@@ -235,6 +235,13 @@ impl<'tcx, Prov: Provenance> ImmTy<'tcx, Prov> {
         Self::from_scalar(Scalar::from_bool(b), layout)
     }
 
+    #[inline]
+    pub fn from_ordering(c: std::cmp::Ordering, tcx: TyCtxt<'tcx>) -> Self {
+        let ty = tcx.ty_ordering_enum(None);
+        let layout = tcx.layout_of(ty::ParamEnv::reveal_all().and(ty)).unwrap();
+        Self::from_scalar(Scalar::from_i8(c as i8), layout)
+    }
+
     #[inline]
     pub fn to_const_int(self) -> ConstInt {
         assert!(self.layout.ty.is_integral());

compiler/rustc_const_eval/src/interpret/operator.rs

@@ -61,6 +61,11 @@ impl<'mir, 'tcx: 'mir, M: Machine<'mir, 'tcx>> InterpCx<'mir, 'tcx, M> {
 }
 
 impl<'mir, 'tcx: 'mir, M: Machine<'mir, 'tcx>> InterpCx<'mir, 'tcx, M> {
+    fn three_way_compare<T: Ord>(&self, lhs: T, rhs: T) -> (ImmTy<'tcx, M::Provenance>, bool) {
+        let res = Ord::cmp(&lhs, &rhs);
+        return (ImmTy::from_ordering(res, *self.tcx), false);
+    }
+
     fn binary_char_op(
         &self,
         bin_op: mir::BinOp,
@@ -69,6 +74,10 @@ impl<'mir, 'tcx: 'mir, M: Machine<'mir, 'tcx>> InterpCx<'mir, 'tcx, M> {
     ) -> (ImmTy<'tcx, M::Provenance>, bool) {
         use rustc_middle::mir::BinOp::*;
 
+        if bin_op == Cmp {
+            return self.three_way_compare(l, r);
+        }
+
         let res = match bin_op {
             Eq => l == r,
             Ne => l != r,
@@ -231,6 +240,11 @@ impl<'mir, 'tcx: 'mir, M: Machine<'mir, 'tcx>> InterpCx<'mir, 'tcx, M> {
                 let r = self.sign_extend(r, right_layout) as i128;
                 return Ok((ImmTy::from_bool(op(&l, &r), *self.tcx), false));
             }
+            if bin_op == Cmp {
+                let l = self.sign_extend(l, left_layout) as i128;
+                let r = self.sign_extend(r, right_layout) as i128;
+                return Ok(self.three_way_compare(l, r));
+            }
             let op: Option<fn(i128, i128) -> (i128, bool)> = match bin_op {
                 Div if r == 0 => throw_ub!(DivisionByZero),
                 Rem if r == 0 => throw_ub!(RemainderByZero),
@@ -270,6 +284,10 @@ impl<'mir, 'tcx: 'mir, M: Machine<'mir, 'tcx>> InterpCx<'mir, 'tcx, M> {
             }
         }
 
+        if bin_op == Cmp {
+            return Ok(self.three_way_compare(l, r));
+        }
+
         let val = match bin_op {
             Eq => ImmTy::from_bool(l == r, *self.tcx),
             Ne => ImmTy::from_bool(l != r, *self.tcx),

compiler/rustc_const_eval/src/transform/validate.rs

@@ -986,6 +986,15 @@ impl<'a, 'tcx> Visitor<'tcx> for TypeChecker<'a, 'tcx> {
                             )
                         }
                     }
+                    Cmp => {
+                        for x in [a, b] {
+                            check_kinds!(
+                                x,
+                                "Cannot three-way compare non-integer type {:?}",
+                                ty::Char | ty::Uint(..) | ty::Int(..)
+                            )
+                        }
+                    }
                     AddUnchecked | SubUnchecked | MulUnchecked | Shl | ShlUnchecked | Shr
                     | ShrUnchecked => {
                         for x in [a, b] {

compiler/rustc_const_eval/src/util/mod.rs

@@ -19,7 +19,7 @@ pub fn binop_left_homogeneous(op: mir::BinOp) -> bool {
     match op {
         Add | AddUnchecked | Sub | SubUnchecked | Mul | MulUnchecked | Div | Rem | BitXor
         | BitAnd | BitOr | Offset | Shl | ShlUnchecked | Shr | ShrUnchecked => true,
-        Eq | Ne | Lt | Le | Gt | Ge => false,
+        Eq | Ne | Lt | Le | Gt | Ge | Cmp => false,
     }
 }
 
@@ -30,7 +30,7 @@ pub fn binop_right_homogeneous(op: mir::BinOp) -> bool {
     use rustc_middle::mir::BinOp::*;
     match op {
         Add | AddUnchecked | Sub | SubUnchecked | Mul | MulUnchecked | Div | Rem | BitXor
-        | BitAnd | BitOr | Eq | Ne | Lt | Le | Gt | Ge => true,
+        | BitAnd | BitOr | Eq | Ne | Lt | Le | Gt | Ge | Cmp => true,
         Offset | Shl | ShlUnchecked | Shr | ShrUnchecked => false,
     }
 }

compiler/rustc_hir/src/lang_items.rs

@@ -226,6 +226,7 @@ language_item_table! {
     Unpin,                   sym::unpin,               unpin_trait,                Target::Trait,          GenericRequirement::None;
     Pin,                     sym::pin,                 pin_type,                   Target::Struct,         GenericRequirement::None;
 
+    OrderingEnum,            sym::Ordering,            ordering_enum,              Target::Enum,           GenericRequirement::Exact(0);
     PartialEq,               sym::eq,                  eq_trait,                   Target::Trait,          GenericRequirement::Exact(1);
     PartialOrd,              sym::partial_ord,         partial_ord_trait,          Target::Trait,          GenericRequirement::Exact(1);
     CVoid,                   sym::c_void,              c_void,                     Target::Enum,           GenericRequirement::None;

compiler/rustc_hir_analysis/src/check/intrinsic.rs

@@ -107,6 +107,7 @@ pub fn intrinsic_operation_unsafety(tcx: TyCtxt<'_>, intrinsic_id: LocalDefId) -
         | sym::cttz
         | sym::bswap
         | sym::bitreverse
+        | sym::three_way_compare
         | sym::discriminant_value
         | sym::type_id
         | sym::likely
@@ -418,6 +419,10 @@ pub fn check_intrinsic_type(
             | sym::bswap
             | sym::bitreverse => (1, 0, vec![param(0)], param(0)),
 
+            sym::three_way_compare => {
+                (1, 0, vec![param(0), param(0)], tcx.ty_ordering_enum(Some(span)))
+            }
+
             sym::add_with_overflow | sym::sub_with_overflow | sym::mul_with_overflow => {
                 (1, 0, vec![param(0), param(0)], Ty::new_tup(tcx, &[param(0), tcx.types.bool]))
             }

compiler/rustc_middle/src/mir/syntax.rs

@@ -1438,6 +1438,16 @@ pub enum BinOp {
     Ge,
     /// The `>` operator (greater than)
     Gt,
+    /// The `<=>` operator (three-way comparison, like `Ord::cmp`)
+    ///
+    /// This is supported only on the integer types and `char`, always returning
+    /// [`rustc_hir::LangItem::OrderingEnum`] (aka [`std::cmp::Ordering`]).
+    ///
+    /// [`Rvalue::BinaryOp`]`(BinOp::Cmp, A, B)` returns
+    /// - `Ordering::Less` (`-1_i8`, as a Scalar) if `A < B`
+    /// - `Ordering::Equal` (`0_i8`, as a Scalar) if `A == B`
+    /// - `Ordering::Greater` (`+1_i8`, as a Scalar) if `A > B`
+    Cmp,
     /// The `ptr.offset` operator
     Offset,
 }

compiler/rustc_middle/src/mir/tcx.rs

@@ -276,6 +276,11 @@ impl<'tcx> BinOp {
             &BinOp::Eq | &BinOp::Lt | &BinOp::Le | &BinOp::Ne | &BinOp::Ge | &BinOp::Gt => {
                 tcx.types.bool
             }
+            &BinOp::Cmp => {
+                // these should be integer-like types of the same size.
+                assert_eq!(lhs_ty, rhs_ty);
+                tcx.ty_ordering_enum(None)
+            }
         }
     }
 }
@@ -312,7 +317,8 @@ impl BinOp {
             BinOp::Gt => hir::BinOpKind::Gt,
             BinOp::Le => hir::BinOpKind::Le,
             BinOp::Ge => hir::BinOpKind::Ge,
-            BinOp::AddUnchecked
+            BinOp::Cmp
+            | BinOp::AddUnchecked
             | BinOp::SubUnchecked
             | BinOp::MulUnchecked
             | BinOp::ShlUnchecked

compiler/rustc_middle/src/ty/context.rs

@@ -956,6 +956,13 @@ impl<'tcx> TyCtxt<'tcx> {
         self.get_lang_items(())
     }
 
+    /// Gets a `Ty` representing the [`LangItem::OrderingEnum`]
+    #[track_caller]
+    pub fn ty_ordering_enum(self, span: Option<Span>) -> Ty<'tcx> {
+        let ordering_enum = self.require_lang_item(hir::LangItem::OrderingEnum, span);
+        self.type_of(ordering_enum).no_bound_vars().unwrap()
+    }
+
     /// Obtain the given diagnostic item's `DefId`. Use `is_diagnostic_item` if you just want to
     /// compare against another `DefId`, since `is_diagnostic_item` is cheaper.
     pub fn get_diagnostic_item(self, name: Symbol) -> Option<DefId> {

compiler/rustc_mir_transform/src/lower_intrinsics.rs

@@ -90,6 +90,7 @@ impl<'tcx> MirPass<'tcx> for LowerIntrinsics {
                     sym::wrapping_add
                     | sym::wrapping_sub
                     | sym::wrapping_mul
+                    | sym::three_way_compare
                     | sym::unchecked_add
                     | sym::unchecked_sub
                     | sym::unchecked_mul
@@ -109,6 +110,7 @@ impl<'tcx> MirPass<'tcx> for LowerIntrinsics {
                             sym::wrapping_add => BinOp::Add,
                             sym::wrapping_sub => BinOp::Sub,
                             sym::wrapping_mul => BinOp::Mul,
+                            sym::three_way_compare => BinOp::Cmp,
                             sym::unchecked_add => BinOp::AddUnchecked,
                             sym::unchecked_sub => BinOp::SubUnchecked,
                             sym::unchecked_mul => BinOp::MulUnchecked,

compiler/rustc_mir_transform/src/promote_consts.rs

@@ -525,6 +525,7 @@ impl<'tcx> Validator<'_, 'tcx> {
                     | BinOp::Lt
                     | BinOp::Ge
                     | BinOp::Gt
+                    | BinOp::Cmp
                     | BinOp::Offset
                     | BinOp::Add
                     | BinOp::AddUnchecked

compiler/rustc_smir/src/rustc_smir/convert/mir.rs

@@ -493,6 +493,7 @@ impl<'tcx> Stable<'tcx> for mir::BinOp {
             BinOp::Ne => stable_mir::mir::BinOp::Ne,
             BinOp::Ge => stable_mir::mir::BinOp::Ge,
             BinOp::Gt => stable_mir::mir::BinOp::Gt,
+            BinOp::Cmp => stable_mir::mir::BinOp::Cmp,
             BinOp::Offset => stable_mir::mir::BinOp::Offset,
         }
     }

compiler/rustc_span/src/symbol.rs

@@ -1813,6 +1813,7 @@ symbols! {
         thread,
         thread_local,
         thread_local_macro,
+        three_way_compare,
         thumb2,
         thumb_mode: "thumb-mode",
         tmm_reg,

compiler/rustc_ty_utils/src/consts.rs

@@ -82,7 +82,7 @@ fn check_binop(op: mir::BinOp) -> bool {
     match op {
         Add | AddUnchecked | Sub | SubUnchecked | Mul | MulUnchecked | Div | Rem | BitXor
         | BitAnd | BitOr | Shl | ShlUnchecked | Shr | ShrUnchecked | Eq | Lt | Le | Ne | Ge
-        | Gt => true,
+        | Gt | Cmp => true,
         Offset => false,
     }
 }

compiler/stable_mir/src/mir/body.rs

@@ -329,6 +329,7 @@ pub enum BinOp {
     Ne,
     Ge,
     Gt,
+    Cmp,
     Offset,
 }
 
@@ -368,6 +369,9 @@ impl BinOp {
                 assert!(lhs_kind.is_primitive() || lhs_kind.is_raw_ptr() || lhs_kind.is_fn_ptr());
                 Ty::bool_ty()
             }
+            BinOp::Cmp => {
+                unimplemented!("Should cmp::Ordering be a RigidTy?");
+            }
         }
     }
 }