Auto merge of #1923 - RalfJung:more-simd, r=RalfJung

More portable SIMD: rem, shl, shr

Also make sure we catch the potential UB in div, rem, shl, shr.

Author: bors

src/shims/intrinsics.rs

@@ -306,7 +306,14 @@ pub trait EvalContextExt<'mir, 'tcx: 'mir>: crate::MiriEvalContextExt<'mir, 'tcx
             }
 
             // SIMD operations
-            "simd_add" | "simd_sub" | "simd_mul" | "simd_div" => {
+            #[rustfmt::skip]
+            | "simd_add"
+            | "simd_sub"
+            | "simd_mul"
+            | "simd_div"
+            | "simd_rem"
+            | "simd_shl"
+            | "simd_shr" => {
                 let &[ref left, ref right] = check_arg_count(args)?;
                 let (left, left_len) = this.operand_to_simd(left)?;
                 let (right, right_len) = this.operand_to_simd(right)?;
@@ -320,14 +327,27 @@ pub trait EvalContextExt<'mir, 'tcx: 'mir>: crate::MiriEvalContextExt<'mir, 'tcx
                     "simd_sub" => mir::BinOp::Sub,
                     "simd_mul" => mir::BinOp::Mul,
                     "simd_div" => mir::BinOp::Div,
+                    "simd_rem" => mir::BinOp::Rem,
+                    "simd_shl" => mir::BinOp::Shl,
+                    "simd_shr" => mir::BinOp::Shr,
                     _ => unreachable!(),
                 };
 
                 for i in 0..dest_len {
                     let left = this.read_immediate(&this.mplace_index(&left, i)?.into())?;
                     let right = this.read_immediate(&this.mplace_index(&right, i)?.into())?;
-                    let dest = this.mplace_index(&dest, i)?.into();
-                    this.binop_ignore_overflow(op, &left, &right, &dest)?;
+                    let dest = this.mplace_index(&dest, i)?;
+                    let (val, overflowed, ty) = this.overflowing_binary_op(op, &left, &right)?;
+                    assert_eq!(ty, dest.layout.ty);
+                    if matches!(op, mir::BinOp::Shl | mir::BinOp::Shr) {
+                        // Shifts have extra UB as SIMD operations that the MIR binop does not have.
+                        // See <https://github.com/rust-lang/rust/issues/91237>.
+                        if overflowed {
+                            let r_val = right.to_scalar()?.to_bits(right.layout.size)?;
+                            throw_ub_format!("overflowing shift by {} in `{}` in SIMD lane {}", r_val, intrinsic_name, i);
+                        }
+                    }
+                    this.write_scalar(val, &dest.into())?;
                 }
             }
 

…compile-fail/intrinsics/div-by-zero-1.rs …s/compile-fail/intrinsics/div-by-zero.rstests/compile-fail/intrinsics/div-by-zero-1.rs renamed to tests/compile-fail/intrinsics/div-by-zero.rs tests/compile-fail/intrinsics/div-by-zero-1.rs renamed to tests/compile-fail/intrinsics/div-by-zero.rs

@@ -0,0 +1,15 @@
+#![feature(platform_intrinsics, repr_simd)]
+
+extern "platform-intrinsic" {
+    pub(crate) fn simd_div<T>(x: T, y: T) -> T;
+}
+
+#[repr(simd)]
+#[allow(non_camel_case_types)]
+struct i32x2(i32, i32);
+
+fn main() { unsafe {
+    let x = i32x2(1, 1);
+    let y = i32x2(1, 0);
+    simd_div(x, y); //~ERROR Undefined Behavior: dividing by zero
+} }

…compile-fail/intrinsics/div-by-zero-2.rs …s/compile-fail/intrinsics/rem-by-zero.rstests/compile-fail/intrinsics/div-by-zero-2.rs renamed to tests/compile-fail/intrinsics/rem-by-zero.rs tests/compile-fail/intrinsics/div-by-zero-2.rs renamed to tests/compile-fail/intrinsics/rem-by-zero.rs

@@ -0,0 +1,15 @@
+#![feature(platform_intrinsics, repr_simd)]
+
+extern "platform-intrinsic" {
+    pub(crate) fn simd_rem<T>(x: T, y: T) -> T;
+}
+
+#[repr(simd)]
+#[allow(non_camel_case_types)]
+struct i32x2(i32, i32);
+
+fn main() { unsafe {
+    let x = i32x2(1, 1);
+    let y = i32x2(1, 0);
+    simd_rem(x, y); //~ERROR Undefined Behavior: calculating the remainder with a divisor of zero
+} }

tests/compile-fail/intrinsics/simd-div-by-zero.rs

@@ -0,0 +1,15 @@
+#![feature(platform_intrinsics, repr_simd)]
+
+extern "platform-intrinsic" {
+    pub(crate) fn simd_shl<T>(x: T, y: T) -> T;
+}
+
+#[repr(simd)]
+#[allow(non_camel_case_types)]
+struct i32x2(i32, i32);
+
+fn main() { unsafe {
+    let x = i32x2(1, 1);
+    let y = i32x2(100, 0);
+    simd_shl(x, y); //~ERROR overflowing shift by 100 in `simd_shl` in SIMD lane 0
+} }

tests/compile-fail/intrinsics/simd-rem-by-zero.rs

@@ -0,0 +1,15 @@
+#![feature(platform_intrinsics, repr_simd)]
+
+extern "platform-intrinsic" {
+    pub(crate) fn simd_shr<T>(x: T, y: T) -> T;
+}
+
+#[repr(simd)]
+#[allow(non_camel_case_types)]
+struct i32x2(i32, i32);
+
+fn main() { unsafe {
+    let x = i32x2(1, 1);
+    let y = i32x2(20, 40);
+    simd_shr(x, y); //~ERROR overflowing shift by 40 in `simd_shr` in SIMD lane 1
+} }

tests/compile-fail/intrinsics/simd-shl-too-far.rs

@@ -9,6 +9,7 @@ fn simd_ops_f32() {
     assert_eq!(a * b, f32x4::from_array([10.0, 20.0, 30.0, 40.0]));
     assert_eq!(b / a, f32x4::from_array([0.1, 0.2, 0.3, 0.4]));
     assert_eq!(a / 2.0, f32x4::splat(5.0));
+    assert_eq!(a % b, f32x4::from_array([0.0, 0.0, 1.0, 2.0]));
 }
 
 fn simd_ops_i32() {
@@ -19,6 +20,9 @@ fn simd_ops_i32() {
     assert_eq!(a * b, i32x4::from_array([10, 20, 30, 40]));
     assert_eq!(a / b, i32x4::from_array([10, 5, 3, 2]));
     assert_eq!(a / 2, i32x4::splat(5));
+    assert_eq!(a % b, i32x4::from_array([0, 0, 1, 2]));
+    assert_eq!(b << 2, i32x4::from_array([4, 8, 12, 16]));
+    assert_eq!(b >> 1, i32x4::from_array([0, 1, 1, 2]));
 }
 
 fn main() {