Auto merge of #1978 - RalfJung:simd_eq, r=RalfJung

implement simd_eq and simd_reduce_any

This lets us re-enable the division and modulo tests, since those operations now internally use simd_eq and simd_reduce_any.

However, I am not sure what exactly the rules are for simd_reduce_any. `@workingjubilee` for now I made it UB to call those with inputs that are not all-0 or all-1, but that might be taking it too far?

Author: bors

rust-version

@@ -1 +1 @@
-78450d2d602b06d9b94349aaf8cece1a4acaf3a8
+b8967b0d52a2ba5f0c9da0da03e78ccba5534e4a

src/shims/intrinsics.rs

@@ -324,7 +324,8 @@ pub trait EvalContextExt<'mir, 'tcx: 'mir>: crate::MiriEvalContextExt<'mir, 'tcx
             | "simd_shl"
             | "simd_shr"
             | "simd_and"
-            | "simd_or" => {
+            | "simd_or"
+            | "simd_eq" => {
                 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)?;
@@ -343,6 +344,7 @@ pub trait EvalContextExt<'mir, 'tcx: 'mir>: crate::MiriEvalContextExt<'mir, 'tcx
                     "simd_shr" => mir::BinOp::Shr,
                     "simd_and" => mir::BinOp::BitAnd,
                     "simd_or" => mir::BinOp::BitOr,
+                    "simd_eq" => mir::BinOp::Eq,
                     _ => unreachable!(),
                 };
 
@@ -351,7 +353,6 @@ pub trait EvalContextExt<'mir, 'tcx: 'mir>: crate::MiriEvalContextExt<'mir, 'tcx
                     let right = this.read_immediate(&this.mplace_index(&right, i)?.into())?;
                     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>.
@@ -360,8 +361,40 @@ pub trait EvalContextExt<'mir, 'tcx: 'mir>: crate::MiriEvalContextExt<'mir, 'tcx
                             throw_ub_format!("overflowing shift by {} in `{}` in SIMD lane {}", r_val, intrinsic_name, i);
                         }
                     }
-                    this.write_scalar(val, &dest.into())?;
+                    if matches!(op, mir::BinOp::Eq) {
+                        // Special handling for boolean-returning operations
+                        assert_eq!(ty, this.tcx.types.bool);
+                        let val = val.to_bool().unwrap();
+                        let val = if val { -1 } else { 0 }; // SIMD uses all-1 as pattern for "true"
+                        let val = Scalar::from_int(val, dest.layout.size);
+                        this.write_scalar(val, &dest.into())?;
+                    } else {
+                        assert_eq!(ty, dest.layout.ty);
+                        this.write_scalar(val, &dest.into())?;
+                    }
+                }
+            }
+            "simd_reduce_any" => {
+                let &[ref arg] = check_arg_count(args)?;
+                let (arg, arg_len) = this.operand_to_simd(arg)?;
+
+                let mut res = false; // the neutral element
+                for i in 0..arg_len {
+                    let op = this.read_immediate(&this.mplace_index(&arg, i)?.into())?;
+                    // We convert it to a *signed* integer and expect either 0 or -1 (the latter means all bits were set).
+                    let val = op.to_scalar()?.to_int(op.layout.size)?;
+                    let val = match val {
+                        0 => false,
+                        -1 => true,
+                        _ =>
+                            throw_ub_format!(
+                                "each element of a simd_reduce_any operand must be all-0-bits or all-1-bits"
+                            ),
+                    };
+                    res = res | val;
                 }
+
+                this.write_scalar(Scalar::from_bool(res), dest)?;
             }
 
             // Atomic operations

tests/compile-fail/intrinsics/simd-reduce-invalid-bool.rs

@@ -0,0 +1,14 @@
+#![feature(platform_intrinsics, repr_simd)]
+
+extern "platform-intrinsic" {
+    pub(crate) fn simd_reduce_any<T>(x: T) -> bool;
+}
+
+#[repr(simd)]
+#[allow(non_camel_case_types)]
+struct i32x2(i32, i32);
+
+fn main() { unsafe {
+    let x = i32x2(0, 1);
+    simd_reduce_any(x); //~ERROR must be all-0-bits or all-1-bits
+} }

tests/run-pass/portable-simd.rs

@@ -1,4 +1,4 @@
-#![feature(portable_simd)]
+#![feature(portable_simd, platform_intrinsics)]
 use std::simd::*;
 
 fn simd_ops_f32() {
@@ -18,16 +18,35 @@ fn simd_ops_i32() {
     assert_eq!(a + b, i32x4::from_array([11, 12, 13, 14]));
     assert_eq!(a - b, i32x4::from_array([9, 8, 7, 6]));
     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 / i32x4::splat(2), i32x4::splat(5));
-    //assert_eq!(a % b, i32x4::from_array([0, 0, 1, 2]));
+    assert_eq!(a / b, i32x4::from_array([10, 5, 3, 2]));
+    assert_eq!(a / i32x4::splat(2), i32x4::splat(5));
+    assert_eq!(a % b, i32x4::from_array([0, 0, 1, 2]));
     assert_eq!(b << i32x4::splat(2), i32x4::from_array([4, 8, 12, 16]));
     assert_eq!(b >> i32x4::splat(1), i32x4::from_array([0, 1, 1, 2]));
     assert_eq!(b & i32x4::splat(2), i32x4::from_array([0, 2, 2, 0]));
     assert_eq!(b | i32x4::splat(2), i32x4::from_array([3, 2, 3, 6]));
 }
 
+fn simd_intrinsics() {
+    extern "platform-intrinsic" {
+        pub(crate) fn simd_eq<T, U>(x: T, y: T) -> U;
+        pub(crate) fn simd_reduce_any<T>(x: T) -> bool;
+    }
+
+    // Make sure simd_eq returns all-1 for `true`
+    let a = i32x4::splat(10);
+    let b = i32x4::from_array([1, 2, 10, 4]);
+    let c: i32x4 = unsafe { simd_eq(a, b) };
+    assert_eq!(c, i32x4::from_array([0, 0, -1, 0]));
+
+    unsafe {
+        assert!(!simd_reduce_any(i32x4::splat(0)));
+        assert!(simd_reduce_any(i32x4::splat(-1)));
+    }
+}
+
 fn main() {
     simd_ops_f32();
     simd_ops_i32();
+    simd_intrinsics();
 }