Auto merge of #2895 - RalfJung:simd_bitmask, r=RalfJung

support array return types in simd_bitmask

Fixes #2734

As usual I am stomped by the simd_bitmask behavior wrt endianess. I confirmed that for little endian, Miri matches what rustc currently does, but I can't test rustc on big endian. `@workingjubilee` `@calebzulawski` could you have a look whether those new tests make sense?

src/shims/intrinsics/simd.rs

@@ -563,7 +563,11 @@ pub trait EvalContextExt<'mir, 'tcx: 'mir>: crate::MiriInterpCxExt<'mir, 'tcx> {
                 let (op, op_len) = this.operand_to_simd(op)?;
                 let bitmask_len = op_len.max(8);
 
-                assert!(dest.layout.ty.is_integral());
+                // Returns either an unsigned integer or array of `u8`.
+                assert!(
+                    dest.layout.ty.is_integral()
+                        || matches!(dest.layout.ty.kind(), ty::Array(elemty, _) if elemty == &this.tcx.types.u8)
+                );
                 assert!(bitmask_len <= 64);
                 assert_eq!(bitmask_len, dest.layout.size.bits());
                 let op_len = u32::try_from(op_len).unwrap();
@@ -577,7 +581,10 @@ pub trait EvalContextExt<'mir, 'tcx: 'mir>: crate::MiriInterpCxExt<'mir, 'tcx> {
                             .unwrap();
                     }
                 }
-                this.write_int(res, dest)?;
+                // We have to force the place type to be an int so that we can write `res` into it.
+                let mut dest = this.force_allocation(dest)?;
+                dest.layout = this.machine.layouts.uint(dest.layout.size).unwrap();
+                this.write_int(res, &dest.into())?;
             }
 
             name => throw_unsup_format!("unimplemented intrinsic: `simd_{name}`"),

tests/pass/portable-simd.rs

@@ -2,6 +2,10 @@
 #![feature(portable_simd, platform_intrinsics)]
 use std::simd::*;
 
+extern "platform-intrinsic" {
+    pub(crate) fn simd_bitmask<T, U>(x: T) -> U;
+}
+
 fn simd_ops_f32() {
     let a = f32x4::splat(10.0);
     let b = f32x4::from_array([1.0, 2.0, 3.0, -4.0]);
@@ -208,11 +212,40 @@ fn simd_mask() {
     assert_eq!(bitmask, 0b1010001101001001);
     assert_eq!(Mask::<i64, 16>::from_bitmask(bitmask), mask);
 
+    // Also directly call intrinsic, to test both kinds of return types.
+    unsafe {
+        let bitmask1: u16 = simd_bitmask(mask.to_int());
+        let bitmask2: [u8; 2] = simd_bitmask(mask.to_int());
+        if cfg!(target_endian = "little") {
+            assert_eq!(bitmask1, 0b1010001101001001);
+            assert_eq!(bitmask2, [0b01001001, 0b10100011]);
+        } else {
+            // All the bitstrings are reversed compared to above, but the array elements are in the
+            // same order.
+            assert_eq!(bitmask1, 0b1001001011000101);
+            assert_eq!(bitmask2, [0b10010010, 0b11000101]);
+        }
+    }
+
+    // Mask less than 8 bits long, which is a special case (padding with 0s).
     let values = [false, false, false, true];
     let mask = Mask::<i64, 4>::from_array(values);
     let bitmask = mask.to_bitmask();
     assert_eq!(bitmask, 0b1000);
     assert_eq!(Mask::<i64, 4>::from_bitmask(bitmask), mask);
+
+    // Also directly call intrinsic, to test both kinds of return types.
+    unsafe {
+        let bitmask1: u8 = simd_bitmask(mask.to_int());
+        let bitmask2: [u8; 1] = simd_bitmask(mask.to_int());
+        if cfg!(target_endian = "little") {
+            assert_eq!(bitmask1, 0b1000);
+            assert_eq!(bitmask2, [0b1000]);
+        } else {
+            assert_eq!(bitmask1, 0b0001);
+            assert_eq!(bitmask2, [0b0001]);
+        }
+    }
 }
 
 fn simd_cast() {