Don't use usub.with.overflow intrinsic

The canonical form of a usub.with.overflow check in LLVM are
separate sub + icmp instructions, rather than a usub.with.overflow
intrinsic. Using usub.with.overflow will generally result in worse
optimization potential.

The backend will attempt to form usub.with.overflow when it comes
to actual instruction selection. This is not fully reliable, but
I believe this is a better tradeoff than using the intrinsic in
IR.

Fixes rust-lang#103285.

compiler/rustc_codegen_llvm/src/builder.rs

@@ -365,11 +365,14 @@ impl<'a, 'll, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'll, 'tcx> {
                 Int(I64) => "llvm.ssub.with.overflow.i64",
                 Int(I128) => "llvm.ssub.with.overflow.i128",
 
-                Uint(U8) => "llvm.usub.with.overflow.i8",
-                Uint(U16) => "llvm.usub.with.overflow.i16",
-                Uint(U32) => "llvm.usub.with.overflow.i32",
-                Uint(U64) => "llvm.usub.with.overflow.i64",
-                Uint(U128) => "llvm.usub.with.overflow.i128",
+                Uint(_) => {
+                    // Emit sub and icmp instead of llvm.usub.with.overflow. LLVM considers these
+                    // to be the canonical form. It will attempt to reform llvm.usub.with.overflow
+                    // in the backend if profitable.
+                    let sub = self.sub(lhs, rhs);
+                    let cmp = self.icmp(IntPredicate::IntULT, lhs, rhs);
+                    return (sub, cmp);
+                }
 
                 _ => unreachable!(),
             },

src/test/codegen/issue-103285-ptr-addr-overflow-check.rs

@@ -0,0 +1,16 @@
+// compile-flags: -O -C debug-assertions=yes
+
+#![crate_type = "lib"]
+#![feature(strict_provenance)]
+
+#[no_mangle]
+pub fn test(src: *const u8, dst: *const u8) -> usize {
+    // CHECK-LABEL: @test(
+    // CHECK-NOT: panic
+    let src_usize = src.addr();
+    let dst_usize = dst.addr();
+    if src_usize > dst_usize {
+        return src_usize - dst_usize;
+    }
+    return 0;
+}