fix simd_bitmask return type for non-power-of-two inputs, and add tests

Author: RalfJung

compiler/rustc_codegen_llvm/src/intrinsic.rs

@@ -1120,7 +1120,7 @@ fn generic_simd_intrinsic<'ll, 'tcx>(
         let (len, _) = require_simd!(arg_tys[1], SimdArgument);
 
         let expected_int_bits = (len.max(8) - 1).next_power_of_two();
-        let expected_bytes = len / 8 + ((len % 8 > 0) as u64);
+        let expected_bytes = len.div_ceil(8);
 
         let mask_ty = arg_tys[0];
         let mask = match mask_ty.kind() {
@@ -1386,8 +1386,8 @@ fn generic_simd_intrinsic<'ll, 'tcx>(
         //
         // The bit order of the result depends on the byte endianness, LSB-first for little
         // endian and MSB-first for big endian.
-        let expected_int_bits = in_len.max(8);
-        let expected_bytes = expected_int_bits / 8 + ((expected_int_bits % 8 > 0) as u64);
+        let expected_int_bits = (in_len.max(8) - 1).next_power_of_two();
+        let expected_bytes = in_len.div_ceil(8);
 
         // Integer vector <i{in_bitwidth} x in_len>:
         let (i_xn, in_elem_bitwidth) = match in_elem.kind() {
@@ -1408,6 +1408,7 @@ fn generic_simd_intrinsic<'ll, 'tcx>(
         };
 
         // Shift the MSB to the right by "in_elem_bitwidth - 1" into the first bit position.
+        // FIXME: the docs say that valid inputs are `0` or `!0`, so the shift seems unnecessary.
         let shift_indices =
             vec![
                 bx.cx.const_int(bx.type_ix(in_elem_bitwidth), (in_elem_bitwidth - 1) as _);

tests/ui/simd/simd-bitmask.rs

@@ -1,5 +1,4 @@
 //@run-pass
-//@ignore-endian-big behavior of simd_select_bitmask is endian-specific
 #![feature(repr_simd, intrinsics)]
 
 extern "rust-intrinsic" {
@@ -17,36 +16,97 @@ fn main() {
         let i: u8 = simd_bitmask(v);
         let a: [u8; 1] = simd_bitmask(v);
 
-        assert_eq!(i, 0b0101);
-        assert_eq!(a, [0b0101]);
+        if cfg!(target_endian = "little") {
+            assert_eq!(i, 0b0101);
+            assert_eq!(a, [0b0101]);
+        } else {
+            assert_eq!(i, 0b1010);
+            assert_eq!(a, [0b1010]);
+        }
 
         let v = Simd::<i8, 16>([0, 0, -1, -1, 0, 0, 0, 0, 0, 0, 0, 0, -1, 0, -1, 0]);
         let i: u16 = simd_bitmask(v);
         let a: [u8; 2] = simd_bitmask(v);
 
-        assert_eq!(i, 0b0101000000001100);
-        assert_eq!(a, [0b1100, 0b01010000]);
+        if cfg!(target_endian = "little") {
+            assert_eq!(i, 0b0101000000001100);
+            assert_eq!(a, [0b00001100, 0b01010000]);
+        } else {
+            assert_eq!(i, 0b0011000000001010);
+            assert_eq!(a, [0b00110000, 0b00001010]);
+        }
     }
 
     unsafe {
-        let a = Simd::<i32, 8>([0, 1, 2, 3, 4, 5, 6, 7]);
-        let b = Simd::<i32, 8>([8, 9, 10, 11, 12, 13, 14, 15]);
-        let e = [0, 9, 2, 11, 12, 13, 14, 15];
+        let a = Simd::<i32, 4>([0, 1, 2, 3]);
+        let b = Simd::<i32, 4>([8, 9, 10, 11]);
+        let e = [0, 9, 2, 11];
 
-        let r = simd_select_bitmask(0b0101u8, a, b);
+        let mask = if cfg!(target_endian = "little") { 0b0101u8 } else { 0b1010u8 };
+        let r = simd_select_bitmask(mask, a, b);
         assert_eq!(r.0, e);
 
-        let r = simd_select_bitmask([0b0101u8], a, b);
+        let mask = if cfg!(target_endian = "little") { [0b0101u8] } else { [0b1010u8] };
+        let r = simd_select_bitmask(mask, a, b);
         assert_eq!(r.0, e);
 
         let a = Simd::<i32, 16>([0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15]);
         let b = Simd::<i32, 16>([16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31]);
         let e = [16, 17, 2, 3, 20, 21, 22, 23, 24, 25, 26, 27, 12, 29, 14, 31];
 
-        let r = simd_select_bitmask(0b0101000000001100u16, a, b);
+        let mask = if cfg!(target_endian = "little") {
+            0b0101000000001100u16
+        } else {
+            0b0011000000001010u16
+        };
+        let r = simd_select_bitmask(mask, a, b);
         assert_eq!(r.0, e);
 
-        let r = simd_select_bitmask([0b1100u8, 0b01010000u8], a, b);
+        let mask = if cfg!(target_endian = "little") {
+            [0b00001100u8, 0b01010000u8]
+        } else {
+            [0b00110000u8, 0b00001010u8]
+        };
+        let r = simd_select_bitmask(mask, a, b);
         assert_eq!(r.0, e);
     }
+
+    non_pow2();
+}
+
+fn non_pow2() {
+    // Non-power-of-2 multi-byte mask.
+    #[repr(simd, packed)]
+    #[allow(non_camel_case_types)]
+    #[derive(Copy, Clone, Debug, PartialEq)]
+    struct i32x10([i32; 10]);
+    impl i32x10 {
+        fn splat(x: i32) -> Self {
+            Self([x; 10])
+        }
+    }
+    unsafe {
+        let mask = i32x10([!0, !0, 0, !0, 0, 0, !0, 0, !0, 0]);
+        let mask_bits = if cfg!(target_endian = "little") { 0b0101001011 } else { 0b1101001010 };
+        let mask_bytes =
+            if cfg!(target_endian = "little") { [0b01001011, 0b01] } else { [0b11, 0b01001010] };
+
+        let bitmask1: u16 = simd_bitmask(mask);
+        let bitmask2: [u8; 2] = simd_bitmask(mask);
+        assert_eq!(bitmask1, mask_bits);
+        assert_eq!(bitmask2, mask_bytes);
+
+        let selected1 = simd_select_bitmask::<u16, _>(
+            mask_bits,
+            i32x10::splat(!0), // yes
+            i32x10::splat(0),  // no
+        );
+        let selected2 = simd_select_bitmask::<[u8; 2], _>(
+            mask_bytes,
+            i32x10::splat(!0), // yes
+            i32x10::splat(0),  // no
+        );
+        assert_eq!(selected1, mask);
+        assert_eq!(selected2, mask);
+    }
 }