Auto merge of #1991 - RalfJung:rustup, r=RalfJung

Rustup: simd_select

Cc rust-lang/rust#94474

Author: bors

rust-version

@@ -1 +1 @@
-6a705566166debf5eff88c57140df607fa409aaa
+f0c4da49983aa699f715caf681e3154b445fb60b

src/helpers.rs

@@ -758,3 +758,18 @@ pub fn immty_from_uint_checked<'tcx>(
         err_unsup_format!("unsigned value {:#x} does not fit in {} bits", int, layout.size.bits())
     })?)
 }
+
+pub fn bool_to_simd_element(b: bool, size: Size) -> Scalar<Tag> {
+    // SIMD uses all-1 as pattern for "true"
+    let val = if b { -1 } else { 0 };
+    Scalar::from_int(val, size)
+}
+
+pub fn simd_element_to_bool<'tcx>(elem: ImmTy<'tcx, Tag>) -> InterpResult<'tcx, bool> {
+    let val = elem.to_scalar()?.to_int(elem.layout.size)?;
+    Ok(match val {
+        0 => false,
+        -1 => true,
+        _ => throw_ub_format!("each element of a SIMD mask must be all-0-bits or all-1-bits"),
+    })
+}

src/shims/intrinsics.rs

@@ -8,7 +8,7 @@ use rustc_middle::{mir, mir::BinOp, ty, ty::FloatTy};
 use rustc_target::abi::{Align, Integer};
 
 use crate::*;
-use helpers::check_arg_count;
+use helpers::{bool_to_simd_element, check_arg_count, simd_element_to_bool};
 
 pub enum AtomicOp {
     MirOp(mir::BinOp, bool),
@@ -365,8 +365,7 @@ pub trait EvalContextExt<'mir, 'tcx: 'mir>: crate::MiriEvalContextExt<'mir, 'tcx
                         // 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);
+                        let val = bool_to_simd_element(val, dest.layout.size);
                         this.write_scalar(val, &dest.into())?;
                     } else {
                         assert_eq!(ty, dest.layout.ty);
@@ -381,21 +380,34 @@ pub trait EvalContextExt<'mir, 'tcx: 'mir>: crate::MiriEvalContextExt<'mir, 'tcx
                 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"
-                            ),
-                    };
+                    let val = simd_element_to_bool(op)?;
                     res = res | val;
                 }
 
                 this.write_scalar(Scalar::from_bool(res), dest)?;
             }
+            "simd_select" => {
+                let &[ref mask, ref yes, ref no] = check_arg_count(args)?;
+                let (mask, mask_len) = this.operand_to_simd(mask)?;
+                let (yes, yes_len) = this.operand_to_simd(yes)?;
+                let (no, no_len) = this.operand_to_simd(no)?;
+                let (dest, dest_len) = this.place_to_simd(dest)?;
+
+                assert_eq!(dest_len, mask_len);
+                assert_eq!(dest_len, yes_len);
+                assert_eq!(dest_len, no_len);
+
+                for i in 0..dest_len {
+                    let mask = this.read_immediate(&this.mplace_index(&mask, i)?.into())?;
+                    let yes = this.read_immediate(&this.mplace_index(&yes, i)?.into())?;
+                    let no = this.read_immediate(&this.mplace_index(&no, i)?.into())?;
+                    let dest = this.mplace_index(&dest, i)?;
+
+                    let mask = simd_element_to_bool(mask)?;
+                    let val = if mask { yes } else { no };
+                    this.write_immediate(*val, &dest.into())?;
+                }
+            }
 
             // Atomic operations
             "atomic_load" => this.atomic_load(args, dest, AtomicReadOp::SeqCst)?,

tests/run-pass/portable-simd.rs

@@ -20,7 +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 / i32x4::splat(2), i32x4::splat(5));
+    assert_eq!(i32x2::splat(i32::MIN) / i32x2::splat(-1), i32x2::splat(i32::MIN));
     assert_eq!(a % b, i32x4::from_array([0, 0, 1, 2]));
+    assert_eq!(i32x2::splat(i32::MIN) % i32x2::splat(-1), i32x2::splat(0));
     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]));
@@ -29,19 +31,22 @@ fn simd_ops_i32() {
 
 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;
+        fn simd_eq<T, U>(x: T, y: T) -> U;
+        fn simd_reduce_any<T>(x: T) -> bool;
+        fn simd_select<M, T>(m: M, yes: T, no: T) -> T;
     }
-
-    // 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 {
+        // 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 = simd_eq(a, b);
+        assert_eq!(c, i32x4::from_array([0, 0, -1, 0]));
+
         assert!(!simd_reduce_any(i32x4::splat(0)));
         assert!(simd_reduce_any(i32x4::splat(-1)));
+
+        assert_eq!(simd_select(i8x4::from_array([0, -1, -1, 0]), a, b), i32x4::from_array([1, 10, 10, 4]));
+        assert_eq!(simd_select(i8x4::from_array([0, -1, -1, 0]), b, a), i32x4::from_array([10, 2, 10, 10]));
     }
 }