Use partial_cmp to implement tuple lt/le/ge/gt

rust-lang · Feb 17, 2023 · 680e216 · 680e216
1 parent dc37e37
commit 680e216
Show file tree

Hide file tree

Showing 2 changed files with 40 additions and 15 deletions.
diff --git a/library/core/src/tuple.rs b/library/core/src/tuple.rs
@@ -1,7 +1,7 @@
 // See src/libstd/primitive_docs.rs for documentation.
 
-use crate::cmp::Ordering::*;
-use crate::cmp::*;
+use crate::cmp::Ordering::{self, *};
+use crate::mem::transmute;
 
 // Recursive macro for implementing n-ary tuple functions and operations
 //
@@ -61,19 +61,19 @@ macro_rules! tuple_impls {
                 }
                 #[inline]
                 fn lt(&self, other: &($($T,)+)) -> bool {
-                    lexical_ord!(lt, $( ${ignore(T)} self.${index()}, other.${index()} ),+)
+                    lexical_ord!(lt, Less, $( ${ignore(T)} self.${index()}, other.${index()} ),+)
                 }
                 #[inline]
                 fn le(&self, other: &($($T,)+)) -> bool {
-                    lexical_ord!(le, $( ${ignore(T)} self.${index()}, other.${index()} ),+)
+                    lexical_ord!(le, Less, $( ${ignore(T)} self.${index()}, other.${index()} ),+)
                 }
                 #[inline]
                 fn ge(&self, other: &($($T,)+)) -> bool {
-                    lexical_ord!(ge, $( ${ignore(T)} self.${index()}, other.${index()} ),+)
+                    lexical_ord!(ge, Greater, $( ${ignore(T)} self.${index()}, other.${index()} ),+)
                 }
                 #[inline]
                 fn gt(&self, other: &($($T,)+)) -> bool {
-                    lexical_ord!(gt, $( ${ignore(T)} self.${index()}, other.${index()} ),+)
+                    lexical_ord!(gt, Greater, $( ${ignore(T)} self.${index()}, other.${index()} ),+)
                 }
             }
         }
@@ -123,16 +123,38 @@ macro_rules! maybe_tuple_doc {
     };
 }
 
-// Constructs an expression that performs a lexical ordering using method $rel.
+#[inline]
+const fn ordering_is_some(c: Option<Ordering>, x: Ordering) -> bool {
+    // FIXME: Just use `==` once that's const-stable on `Option`s.
+    // This isn't using `match` because that optimizes worse due to
+    // making a two-step check (`Some` *then* the inner value).
+
+    // SAFETY: There's no public guarantee for `Option<Ordering>`,
+    // but we're core so we know that it's definitely a byte.
+    unsafe {
+        let c: i8 = transmute(c);
+        let x: i8 = transmute(Some(x));
+        c == x
+    }
+}
+
+// Constructs an expression that performs a lexical ordering using method `$rel`.
 // The values are interleaved, so the macro invocation for
-// `(a1, a2, a3) < (b1, b2, b3)` would be `lexical_ord!(lt, a1, b1, a2, b2,
-// a3, b3)` (and similarly for `lexical_cmp`)
+// `(a1, a2, a3) < (b1, b2, b3)` would be `lexical_ord!(lt, opt_is_lt, a1, b1,
+// a2, b2, a3, b3)` (and similarly for `lexical_cmp`)
+//
+// `$ne_rel` is only used to determine the result after checking that they're
+// not equal, so `lt` and `le` can both just use `Less`.
 macro_rules! lexical_ord {
-    ($rel: ident, $a:expr, $b:expr, $($rest_a:expr, $rest_b:expr),+) => {
-        if $a != $b { lexical_ord!($rel, $a, $b) }
-        else { lexical_ord!($rel, $($rest_a, $rest_b),+) }
+    ($rel: ident, $ne_rel: ident, $a:expr, $b:expr, $($rest_a:expr, $rest_b:expr),+) => {{
+        let c = PartialOrd::partial_cmp(&$a, &$b);
+        if !ordering_is_some(c, Equal) { ordering_is_some(c, $ne_rel) }
+        else { lexical_ord!($rel, $ne_rel, $($rest_a, $rest_b),+) }
+    }};
+    ($rel: ident, $ne_rel: ident, $a:expr, $b:expr) => {
+        // Use the specific method for the last element
+        PartialOrd::$rel(&$a, &$b)
     };
-    ($rel: ident, $a:expr, $b:expr) => { ($a) . $rel (& $b) };
 }
 
 macro_rules! lexical_partial_cmp {

diff --git a/tests/codegen/comparison-operators-2-tuple.rs b/tests/codegen/comparison-operators-2-tuple.rs
@@ -11,6 +11,9 @@ type TwoTuple = (i16, u16);
 //
 // The operators are all overridden directly, so should optimize easily.
 //
+// Yes, the `s[lg]t` is correct for the `[lg]e` version because it's only used
+// in the side of the select where we know the values are *not* equal.
+//
 
 // CHECK-LABEL: @check_lt_direct
 // CHECK-SAME: (i16 noundef %[[A0:.+]], i16 noundef %[[A1:.+]], i16 noundef %[[B0:.+]], i16 noundef %[[B1:.+]])
@@ -29,7 +32,7 @@ pub fn check_lt_direct(a: TwoTuple, b: TwoTuple) -> bool {
 #[no_mangle]
 pub fn check_le_direct(a: TwoTuple, b: TwoTuple) -> bool {
     // CHECK-DAG: %[[EQ:.+]] = icmp eq i16 %[[A0]], %[[B0]]
-    // CHECK-DAG: %[[CMP0:.+]] = icmp sle i16 %[[A0]], %[[B0]]
+    // CHECK-DAG: %[[CMP0:.+]] = icmp slt i16 %[[A0]], %[[B0]]
     // CHECK-DAG: %[[CMP1:.+]] = icmp ule i16 %[[A1]], %[[B1]]
     // CHECK: %[[R:.+]] = select i1 %[[EQ]], i1 %[[CMP1]], i1 %[[CMP0]]
     // CHECK: ret i1 %[[R]]
@@ -53,7 +56,7 @@ pub fn check_gt_direct(a: TwoTuple, b: TwoTuple) -> bool {
 #[no_mangle]
 pub fn check_ge_direct(a: TwoTuple, b: TwoTuple) -> bool {
     // CHECK-DAG: %[[EQ:.+]] = icmp eq i16 %[[A0]], %[[B0]]
-    // CHECK-DAG: %[[CMP0:.+]] = icmp sge i16 %[[A0]], %[[B0]]
+    // CHECK-DAG: %[[CMP0:.+]] = icmp sgt i16 %[[A0]], %[[B0]]
     // CHECK-DAG: %[[CMP1:.+]] = icmp uge i16 %[[A1]], %[[B1]]
     // CHECK: %[[R:.+]] = select i1 %[[EQ]], i1 %[[CMP1]], i1 %[[CMP0]]
     // CHECK: ret i1 %[[R]]