loongarch: Use intrinsics::simd for vshuf4i

crates/core_arch/src/loongarch64/lasx/generated.rs

@@ -167,12 +167,6 @@ unsafe extern "unadjusted" {
     fn __lasx_xvbitsel_v(a: __v32u8, b: __v32u8, c: __v32u8) -> __v32u8;
     #[link_name = "llvm.loongarch.lasx.xvbitseli.b"]
     fn __lasx_xvbitseli_b(a: __v32u8, b: __v32u8, c: u32) -> __v32u8;
-    #[link_name = "llvm.loongarch.lasx.xvshuf4i.b"]
-    fn __lasx_xvshuf4i_b(a: __v32i8, b: u32) -> __v32i8;
-    #[link_name = "llvm.loongarch.lasx.xvshuf4i.h"]
-    fn __lasx_xvshuf4i_h(a: __v16i16, b: u32) -> __v16i16;
-    #[link_name = "llvm.loongarch.lasx.xvshuf4i.w"]
-    fn __lasx_xvshuf4i_w(a: __v8i32, b: u32) -> __v8i32;
     #[link_name = "llvm.loongarch.lasx.xvclo.b"]
     fn __lasx_xvclo_b(a: __v32i8) -> __v32i8;
     #[link_name = "llvm.loongarch.lasx.xvclo.h"]
@@ -1631,33 +1625,6 @@ pub fn lasx_xvbitseli_b<const IMM8: u32>(a: m256i, b: m256i) -> m256i {
     unsafe { transmute(__lasx_xvbitseli_b(transmute(a), transmute(b), IMM8)) }
 }
 
-#[inline]
-#[target_feature(enable = "lasx")]
-#[rustc_legacy_const_generics(1)]
-#[unstable(feature = "stdarch_loongarch", issue = "117427")]
-pub fn lasx_xvshuf4i_b<const IMM8: u32>(a: m256i) -> m256i {
-    static_assert_uimm_bits!(IMM8, 8);
-    unsafe { transmute(__lasx_xvshuf4i_b(transmute(a), IMM8)) }
-}
-
-#[inline]
-#[target_feature(enable = "lasx")]
-#[rustc_legacy_const_generics(1)]
-#[unstable(feature = "stdarch_loongarch", issue = "117427")]
-pub fn lasx_xvshuf4i_h<const IMM8: u32>(a: m256i) -> m256i {
-    static_assert_uimm_bits!(IMM8, 8);
-    unsafe { transmute(__lasx_xvshuf4i_h(transmute(a), IMM8)) }
-}
-
-#[inline]
-#[target_feature(enable = "lasx")]
-#[rustc_legacy_const_generics(1)]
-#[unstable(feature = "stdarch_loongarch", issue = "117427")]
-pub fn lasx_xvshuf4i_w<const IMM8: u32>(a: m256i) -> m256i {
-    static_assert_uimm_bits!(IMM8, 8);
-    unsafe { transmute(__lasx_xvshuf4i_w(transmute(a), IMM8)) }
-}
-
 #[inline]
 #[target_feature(enable = "lasx")]
 #[unstable(feature = "stdarch_loongarch", issue = "117427")]

crates/core_arch/src/loongarch64/lasx/portable.rs

@@ -267,6 +267,53 @@ pub(crate) const unsafe fn simd_packod_d<T: Copy>(a: T, b: T) -> T {
     simd_shuffle!(b, a, [1, 5, 3, 7])
 }
 
+#[inline(always)]
+#[rustc_const_unstable(feature = "stdarch_const_helpers", issue = "none")]
+const unsafe fn simd_shuf4i_b<const I: u32, T: Copy>(a: T) -> T {
+    simd_shuffle!(
+        a,
+        a,
+        [
+            ((I >> 0) & 3) + 0, ((I >> 2) & 3) + 0, ((I >> 4) & 3) + 0, ((I >> 6) & 3) + 0,
+            ((I >> 0) & 3) + 4, ((I >> 2) & 3) + 4, ((I >> 4) & 3) + 4, ((I >> 6) & 3) + 4,
+            ((I >> 0) & 3) + 8, ((I >> 2) & 3) + 8, ((I >> 4) & 3) + 8, ((I >> 6) & 3) + 8,
+            ((I >> 0) & 3) + 12, ((I >> 2) & 3) + 12, ((I >> 4) & 3) + 12, ((I >> 6) & 3) + 12,
+            ((I >> 0) & 3) + 16, ((I >> 2) & 3) + 16, ((I >> 4) & 3) + 16, ((I >> 6) & 3) + 16,
+            ((I >> 0) & 3) + 20, ((I >> 2) & 3) + 20, ((I >> 4) & 3) + 20, ((I >> 6) & 3) + 20,
+            ((I >> 0) & 3) + 24, ((I >> 2) & 3) + 24, ((I >> 4) & 3) + 24, ((I >> 6) & 3) + 24,
+            ((I >> 0) & 3) + 28, ((I >> 2) & 3) + 28, ((I >> 4) & 3) + 28, ((I >> 6) & 3) + 28
+        ]
+    )
+}
+
+#[inline(always)]
+#[rustc_const_unstable(feature = "stdarch_const_helpers", issue = "none")]
+const unsafe fn simd_shuf4i_h<const I: u32, T: Copy>(a: T) -> T {
+    simd_shuffle!(
+        a,
+        a,
+        [
+            ((I >> 0) & 3) + 0, ((I >> 2) & 3) + 0, ((I >> 4) & 3) + 0, ((I >> 6) & 3) + 0,
+            ((I >> 0) & 3) + 4, ((I >> 2) & 3) + 4, ((I >> 4) & 3) + 4, ((I >> 6) & 3) + 4,
+            ((I >> 0) & 3) + 8, ((I >> 2) & 3) + 8, ((I >> 4) & 3) + 8, ((I >> 6) & 3) + 8,
+            ((I >> 0) & 3) + 12, ((I >> 2) & 3) + 12, ((I >> 4) & 3) + 12, ((I >> 6) & 3) + 12,
+        ]
+    )
+}
+
+#[inline(always)]
+#[rustc_const_unstable(feature = "stdarch_const_helpers", issue = "none")]
+const unsafe fn simd_shuf4i_w<const I: u32, T: Copy>(a: T) -> T {
+    simd_shuffle!(
+        a,
+        a,
+        [
+            ((I >> 0) & 3) + 0, ((I >> 2) & 3) + 0, ((I >> 4) & 3) + 0, ((I >> 6) & 3) + 0,
+            ((I >> 0) & 3) + 4, ((I >> 2) & 3) + 4, ((I >> 4) & 3) + 4, ((I >> 6) & 3) + 4
+        ]
+    )
+}
+
 impl_vv!("lasx", lasx_xvpcnt_b, is::simd_ctpop, m256i, i8x32);
 impl_vv!("lasx", lasx_xvpcnt_h, is::simd_ctpop, m256i, i16x16);
 impl_vv!("lasx", lasx_xvpcnt_w, is::simd_ctpop, m256i, i32x8);
@@ -488,6 +535,9 @@ impl_vuv!("lasx", lasx_xvrepl128vei_b, simd_replvei_b, m256i, i8x32, 4, const);
 impl_vuv!("lasx", lasx_xvrepl128vei_h, simd_replvei_h, m256i, i16x16, 3, const);
 impl_vuv!("lasx", lasx_xvrepl128vei_w, simd_replvei_w, m256i, i32x8, 2, const);
 impl_vuv!("lasx", lasx_xvrepl128vei_d, simd_replvei_d, m256i, i64x4, 1, const);
+impl_vuv!("lasx", lasx_xvshuf4i_b, simd_shuf4i_b, m256i, i8x32, 8, const);
+impl_vuv!("lasx", lasx_xvshuf4i_h, simd_shuf4i_h, m256i, i16x16, 8, const);
+impl_vuv!("lasx", lasx_xvshuf4i_w, simd_shuf4i_w, m256i, i32x8, 8, const);
 
 impl_vug!("lasx", lasx_xvpickve2gr_w, is::simd_extract, m256i, i32x8, i32, 3);
 impl_vug!("lasx", lasx_xvpickve2gr_d, is::simd_extract, m256i, i64x4, i64, 2);

crates/core_arch/src/loongarch64/lsx/generated.rs

@@ -173,12 +173,6 @@ unsafe extern "unadjusted" {
     fn __lsx_vbitsel_v(a: __v16u8, b: __v16u8, c: __v16u8) -> __v16u8;
     #[link_name = "llvm.loongarch.lsx.vbitseli.b"]
     fn __lsx_vbitseli_b(a: __v16u8, b: __v16u8, c: u32) -> __v16u8;
-    #[link_name = "llvm.loongarch.lsx.vshuf4i.b"]
-    fn __lsx_vshuf4i_b(a: __v16i8, b: u32) -> __v16i8;
-    #[link_name = "llvm.loongarch.lsx.vshuf4i.h"]
-    fn __lsx_vshuf4i_h(a: __v8i16, b: u32) -> __v8i16;
-    #[link_name = "llvm.loongarch.lsx.vshuf4i.w"]
-    fn __lsx_vshuf4i_w(a: __v4i32, b: u32) -> __v4i32;
     #[link_name = "llvm.loongarch.lsx.vclo.b"]
     fn __lsx_vclo_b(a: __v16i8) -> __v16i8;
     #[link_name = "llvm.loongarch.lsx.vclo.h"]
@@ -363,8 +357,6 @@ unsafe extern "unadjusted" {
     fn __lsx_vfrstp_b(a: __v16i8, b: __v16i8, c: __v16i8) -> __v16i8;
     #[link_name = "llvm.loongarch.lsx.vfrstp.h"]
     fn __lsx_vfrstp_h(a: __v8i16, b: __v8i16, c: __v8i16) -> __v8i16;
-    #[link_name = "llvm.loongarch.lsx.vshuf4i.d"]
-    fn __lsx_vshuf4i_d(a: __v2i64, b: __v2i64, c: u32) -> __v2i64;
     #[link_name = "llvm.loongarch.lsx.vbsrl.v"]
     fn __lsx_vbsrl_v(a: __v16i8, b: u32) -> __v16i8;
     #[link_name = "llvm.loongarch.lsx.vbsll.v"]
@@ -1574,33 +1566,6 @@ pub fn lsx_vbitseli_b<const IMM8: u32>(a: m128i, b: m128i) -> m128i {
     unsafe { transmute(__lsx_vbitseli_b(transmute(a), transmute(b), IMM8)) }
 }
 
-#[inline]
-#[target_feature(enable = "lsx")]
-#[rustc_legacy_const_generics(1)]
-#[unstable(feature = "stdarch_loongarch", issue = "117427")]
-pub fn lsx_vshuf4i_b<const IMM8: u32>(a: m128i) -> m128i {
-    static_assert_uimm_bits!(IMM8, 8);
-    unsafe { transmute(__lsx_vshuf4i_b(transmute(a), IMM8)) }
-}
-
-#[inline]
-#[target_feature(enable = "lsx")]
-#[rustc_legacy_const_generics(1)]
-#[unstable(feature = "stdarch_loongarch", issue = "117427")]
-pub fn lsx_vshuf4i_h<const IMM8: u32>(a: m128i) -> m128i {
-    static_assert_uimm_bits!(IMM8, 8);
-    unsafe { transmute(__lsx_vshuf4i_h(transmute(a), IMM8)) }
-}
-
-#[inline]
-#[target_feature(enable = "lsx")]
-#[rustc_legacy_const_generics(1)]
-#[unstable(feature = "stdarch_loongarch", issue = "117427")]
-pub fn lsx_vshuf4i_w<const IMM8: u32>(a: m128i) -> m128i {
-    static_assert_uimm_bits!(IMM8, 8);
-    unsafe { transmute(__lsx_vshuf4i_w(transmute(a), IMM8)) }
-}
-
 #[inline]
 #[target_feature(enable = "lsx")]
 #[unstable(feature = "stdarch_loongarch", issue = "117427")]
@@ -2261,15 +2226,6 @@ pub fn lsx_vfrstp_h(a: m128i, b: m128i, c: m128i) -> m128i {
     unsafe { transmute(__lsx_vfrstp_h(transmute(a), transmute(b), transmute(c))) }
 }
 
-#[inline]
-#[target_feature(enable = "lsx")]
-#[rustc_legacy_const_generics(2)]
-#[unstable(feature = "stdarch_loongarch", issue = "117427")]
-pub fn lsx_vshuf4i_d<const IMM8: u32>(a: m128i, b: m128i) -> m128i {
-    static_assert_uimm_bits!(IMM8, 8);
-    unsafe { transmute(__lsx_vshuf4i_d(transmute(a), transmute(b), IMM8)) }
-}
-
 #[inline]
 #[target_feature(enable = "lsx")]
 #[rustc_legacy_const_generics(1)]

crates/core_arch/src/loongarch64/lsx/portable.rs

@@ -173,6 +173,46 @@ pub(crate) const unsafe fn simd_packod_d<T: Copy>(a: T, b: T) -> T {
     simd_shuffle!(b, a, [1, 3])
 }
 
+#[inline(always)]
+#[rustc_const_unstable(feature = "stdarch_const_helpers", issue = "none")]
+const unsafe fn simd_shuf4i_b<const I: u32, T: Copy>(a: T) -> T {
+    simd_shuffle!(
+        a,
+        a,
+        [
+            ((I >> 0) & 3) + 0, ((I >> 2) & 3) + 0, ((I >> 4) & 3) + 0, ((I >> 6) & 3) + 0,
+            ((I >> 0) & 3) + 4, ((I >> 2) & 3) + 4, ((I >> 4) & 3) + 4, ((I >> 6) & 3) + 4,
+            ((I >> 0) & 3) + 8, ((I >> 2) & 3) + 8, ((I >> 4) & 3) + 8, ((I >> 6) & 3) + 8,
+            ((I >> 0) & 3) + 12, ((I >> 2) & 3) + 12, ((I >> 4) & 3) + 12, ((I >> 6) & 3) + 12
+        ]
+    )
+}
+
+#[inline(always)]
+#[rustc_const_unstable(feature = "stdarch_const_helpers", issue = "none")]
+const unsafe fn simd_shuf4i_h<const I: u32, T: Copy>(a: T) -> T {
+    simd_shuffle!(
+        a,
+        a,
+        [
+            ((I >> 0) & 3) + 0, ((I >> 2) & 3) + 0, ((I >> 4) & 3) + 0, ((I >> 6) & 3) + 0,
+            ((I >> 0) & 3) + 4, ((I >> 2) & 3) + 4, ((I >> 4) & 3) + 4, ((I >> 6) & 3) + 4
+        ]
+    )
+}
+
+#[inline(always)]
+#[rustc_const_unstable(feature = "stdarch_const_helpers", issue = "none")]
+const unsafe fn simd_shuf4i_w<const I: u32, T: Copy>(a: T) -> T {
+    simd_shuffle!(a, a, [((I >> 0) & 3), ((I >> 2) & 3), ((I >> 4) & 3), ((I >> 6) & 3)])
+}
+
+#[inline(always)]
+#[rustc_const_unstable(feature = "stdarch_const_helpers", issue = "none")]
+const unsafe fn simd_shuf4i_d<const I: u32, T: Copy>(a: T, b: T) -> T {
+    simd_shuffle!(a, b, [((I >> 0) & 3), ((I >> 2) & 3)])
+}
+
 impl_vv!("lsx", lsx_vpcnt_b, is::simd_ctpop, m128i, i8x16);
 impl_vv!("lsx", lsx_vpcnt_h, is::simd_ctpop, m128i, i16x8);
 impl_vv!("lsx", lsx_vpcnt_w, is::simd_ctpop, m128i, i32x4);
@@ -389,6 +429,9 @@ impl_vuv!("lsx", lsx_vreplvei_b, simd_replvei_b, m128i, i8x16, 4, const);
 impl_vuv!("lsx", lsx_vreplvei_h, simd_replvei_h, m128i, i16x8, 3, const);
 impl_vuv!("lsx", lsx_vreplvei_w, simd_replvei_w, m128i, i32x4, 2, const);
 impl_vuv!("lsx", lsx_vreplvei_d, simd_replvei_d, m128i, i64x2, 1, const);
+impl_vuv!("lsx", lsx_vshuf4i_b, simd_shuf4i_b, m128i, i8x16, 8, const);
+impl_vuv!("lsx", lsx_vshuf4i_h, simd_shuf4i_h, m128i, i16x8, 8, const);
+impl_vuv!("lsx", lsx_vshuf4i_w, simd_shuf4i_w, m128i, i32x4, 8, const);
 
 impl_vug!("lsx", lsx_vpickve2gr_b, is::simd_extract, m128i, i8x16, i32, 4);
 impl_vug!("lsx", lsx_vpickve2gr_h, is::simd_extract, m128i, i16x8, i32, 3);
@@ -437,6 +480,8 @@ impl_vvvv!("lsx", lsx_vfnmadd_d, ls::simd_fnmadd, m128d, f64x2);
 impl_vvvv!("lsx", lsx_vfnmsub_s, ls::simd_fnmsub, m128, f32x4);
 impl_vvvv!("lsx", lsx_vfnmsub_d, ls::simd_fnmsub, m128d, f64x2);
 
+impl_vvuv!("lsx", lsx_vshuf4i_d, simd_shuf4i_d, m128i, i64x2, 8, const);
+
 impl_vugv!("lsx", lsx_vinsgr2vr_b, is::simd_insert, m128i, i8x16, i32, 4);
 impl_vugv!("lsx", lsx_vinsgr2vr_h, is::simd_insert, m128i, i16x8, i32, 3);
 impl_vugv!("lsx", lsx_vinsgr2vr_w, is::simd_insert, m128i, i32x4, i32, 2);

crates/core_arch/src/loongarch64/simd.rs

@@ -334,6 +334,26 @@ macro_rules! impl_vvvv {
 
 pub(super) use impl_vvvv;
 
+macro_rules! impl_vvuv {
+    ($ft:literal, $name:ident, $op:ident, $oty:ty, $ity:ident, $ibs:expr, const) => {
+        #[inline]
+        #[target_feature(enable = $ft)]
+        #[rustc_legacy_const_generics(2)]
+        #[unstable(feature = "stdarch_loongarch", issue = "117427")]
+        pub fn $name<const IMM: u32>(a: $oty, b: $oty) -> $oty {
+            static_assert_uimm_bits!(IMM, $ibs);
+            unsafe {
+                let a: $ity = transmute(a);
+                let b: $ity = transmute(b);
+                let r: $ity = $op::<IMM, _>(a, b);
+                transmute(r)
+            }
+        }
+    };
+}
+
+pub(super) use impl_vvuv;
+
 macro_rules! impl_vugv {
     ($ft:literal, $name:ident, $op:path, $oty:ty, $ity:ident, $gty:ty, $ibs:expr) => {
         #[inline]

crates/stdarch-gen-loongarch/lasx.spec

@@ -1642,16 +1642,19 @@ asm-fmts = xd, xj, ui8
 data-types = UV32QI, UV32QI, UV32QI, USI
 
 /// lasx_xvshuf4i_b
+impl = portable
 name = lasx_xvshuf4i_b
 asm-fmts = xd, xj, ui8
 data-types = V32QI, V32QI, USI
 
 /// lasx_xvshuf4i_h
+impl = portable
 name = lasx_xvshuf4i_h
 asm-fmts = xd, xj, ui8
 data-types = V16HI, V16HI, USI
 
 /// lasx_xvshuf4i_w
+impl = portable
 name = lasx_xvshuf4i_w
 asm-fmts = xd, xj, ui8
 data-types = V8SI, V8SI, USI

crates/stdarch-gen-loongarch/lsx.spec

@@ -1657,16 +1657,19 @@ asm-fmts = vd, vj, ui8
 data-types = UV16QI, UV16QI, UV16QI, USI
 
 /// lsx_vshuf4i_b
+impl = portable
 name = lsx_vshuf4i_b
 asm-fmts = vd, vj, ui8
 data-types = V16QI, V16QI, USI
 
 /// lsx_vshuf4i_h
+impl = portable
 name = lsx_vshuf4i_h
 asm-fmts = vd, vj, ui8
 data-types = V8HI, V8HI, USI
 
 /// lsx_vshuf4i_w
+impl = portable
 name = lsx_vshuf4i_w
 asm-fmts = vd, vj, ui8
 data-types = V4SI, V4SI, USI
@@ -2366,6 +2369,7 @@ asm-fmts = vd, vj, vk
 data-types = V8HI, V8HI, V8HI, V8HI
 
 /// lsx_vshuf4i_d
+impl = portable
 name = lsx_vshuf4i_d
 asm-fmts = vd, vj, ui8
 data-types = V2DI, V2DI, V2DI, USI

crates/stdarch-gen-loongarch/src/portable-intrinsics.txt

@@ -259,6 +259,10 @@ lsx_vpackod_b
 lsx_vpackod_h
 lsx_vpackod_w
 lsx_vpackod_d
+lsx_vshuf4i_b
+lsx_vshuf4i_h
+lsx_vshuf4i_w
+lsx_vshuf4i_d
 
 # LASX intrinsics
 lasx_xvsll_b
@@ -520,3 +524,6 @@ lasx_xvpackod_b
 lasx_xvpackod_h
 lasx_xvpackod_w
 lasx_xvpackod_d
+lasx_xvshuf4i_b
+lasx_xvshuf4i_h
+lasx_xvshuf4i_w