From c87d3e2bf711b03ac3eb9b7e00b328a6afaa94f9 Mon Sep 17 00:00:00 2001 From: WANG Rui Date: Wed, 6 May 2026 16:27:37 +0800 Subject: [PATCH] loongarch: Use `intrinsics::simd` for vadda --- .../src/loongarch64/lasx/generated.rs | 36 ------------------- .../src/loongarch64/lasx/portable.rs | 4 +++ .../src/loongarch64/lsx/generated.rs | 36 ------------------- .../core_arch/src/loongarch64/lsx/portable.rs | 4 +++ crates/core_arch/src/loongarch64/simd.rs | 13 +++++++ crates/stdarch-gen-loongarch/lasx.spec | 4 +++ crates/stdarch-gen-loongarch/lsx.spec | 4 +++ .../src/portable-intrinsics.txt | 8 +++++ 8 files changed, 37 insertions(+), 72 deletions(-) diff --git a/crates/core_arch/src/loongarch64/lasx/generated.rs b/crates/core_arch/src/loongarch64/lasx/generated.rs index 387a3e564c..5ef7132637 100644 --- a/crates/core_arch/src/loongarch64/lasx/generated.rs +++ b/crates/core_arch/src/loongarch64/lasx/generated.rs @@ -91,14 +91,6 @@ unsafe extern "unadjusted" { fn __lasx_xvsat_wu(a: __v8u32, b: u32) -> __v8u32; #[link_name = "llvm.loongarch.lasx.xvsat.du"] fn __lasx_xvsat_du(a: __v4u64, b: u32) -> __v4u64; - #[link_name = "llvm.loongarch.lasx.xvadda.b"] - fn __lasx_xvadda_b(a: __v32i8, b: __v32i8) -> __v32i8; - #[link_name = "llvm.loongarch.lasx.xvadda.h"] - fn __lasx_xvadda_h(a: __v16i16, b: __v16i16) -> __v16i16; - #[link_name = "llvm.loongarch.lasx.xvadda.w"] - fn __lasx_xvadda_w(a: __v8i32, b: __v8i32) -> __v8i32; - #[link_name = "llvm.loongarch.lasx.xvadda.d"] - fn __lasx_xvadda_d(a: __v4i64, b: __v4i64) -> __v4i64; #[link_name = "llvm.loongarch.lasx.xvavg.b"] fn __lasx_xvavg_b(a: __v32i8, b: __v32i8) -> __v32i8; #[link_name = "llvm.loongarch.lasx.xvavg.h"] @@ -1445,34 +1437,6 @@ pub fn lasx_xvsat_du(a: m256i) -> m256i { unsafe { transmute(__lasx_xvsat_du(transmute(a), IMM6)) } } -#[inline] -#[target_feature(enable = "lasx")] -#[unstable(feature = "stdarch_loongarch", issue = "117427")] -pub fn lasx_xvadda_b(a: m256i, b: m256i) -> m256i { - unsafe { transmute(__lasx_xvadda_b(transmute(a), transmute(b))) } -} - -#[inline] -#[target_feature(enable = "lasx")] -#[unstable(feature = "stdarch_loongarch", issue = "117427")] -pub fn lasx_xvadda_h(a: m256i, b: m256i) -> m256i { - unsafe { transmute(__lasx_xvadda_h(transmute(a), transmute(b))) } -} - -#[inline] -#[target_feature(enable = "lasx")] -#[unstable(feature = "stdarch_loongarch", issue = "117427")] -pub fn lasx_xvadda_w(a: m256i, b: m256i) -> m256i { - unsafe { transmute(__lasx_xvadda_w(transmute(a), transmute(b))) } -} - -#[inline] -#[target_feature(enable = "lasx")] -#[unstable(feature = "stdarch_loongarch", issue = "117427")] -pub fn lasx_xvadda_d(a: m256i, b: m256i) -> m256i { - unsafe { transmute(__lasx_xvadda_d(transmute(a), transmute(b))) } -} - #[inline] #[target_feature(enable = "lasx")] #[unstable(feature = "stdarch_loongarch", issue = "117427")] diff --git a/crates/core_arch/src/loongarch64/lasx/portable.rs b/crates/core_arch/src/loongarch64/lasx/portable.rs index 147d94fe21..366deda0c1 100644 --- a/crates/core_arch/src/loongarch64/lasx/portable.rs +++ b/crates/core_arch/src/loongarch64/lasx/portable.rs @@ -148,6 +148,10 @@ impl_vvv!("lasx", lasx_xvssub_bu, is::simd_saturating_sub, m256i, u8x32); impl_vvv!("lasx", lasx_xvssub_hu, is::simd_saturating_sub, m256i, u16x16); impl_vvv!("lasx", lasx_xvssub_wu, is::simd_saturating_sub, m256i, u32x8); impl_vvv!("lasx", lasx_xvssub_du, is::simd_saturating_sub, m256i, u64x4); +impl_vvv!("lasx", lasx_xvadda_b, ls::simd_adda, m256i, i8x32); +impl_vvv!("lasx", lasx_xvadda_h, ls::simd_adda, m256i, i16x16); +impl_vvv!("lasx", lasx_xvadda_w, ls::simd_adda, m256i, i32x8); +impl_vvv!("lasx", lasx_xvadda_d, ls::simd_adda, m256i, i64x4); impl_vuv!("lasx", lasx_xvslli_b, is::simd_shl, m256i, i8x32); impl_vuv!("lasx", lasx_xvslli_h, is::simd_shl, m256i, i16x16); diff --git a/crates/core_arch/src/loongarch64/lsx/generated.rs b/crates/core_arch/src/loongarch64/lsx/generated.rs index 4f9a022c1d..1df46a43fe 100644 --- a/crates/core_arch/src/loongarch64/lsx/generated.rs +++ b/crates/core_arch/src/loongarch64/lsx/generated.rs @@ -91,14 +91,6 @@ unsafe extern "unadjusted" { fn __lsx_vsat_wu(a: __v4u32, b: u32) -> __v4u32; #[link_name = "llvm.loongarch.lsx.vsat.du"] fn __lsx_vsat_du(a: __v2u64, b: u32) -> __v2u64; - #[link_name = "llvm.loongarch.lsx.vadda.b"] - fn __lsx_vadda_b(a: __v16i8, b: __v16i8) -> __v16i8; - #[link_name = "llvm.loongarch.lsx.vadda.h"] - fn __lsx_vadda_h(a: __v8i16, b: __v8i16) -> __v8i16; - #[link_name = "llvm.loongarch.lsx.vadda.w"] - fn __lsx_vadda_w(a: __v4i32, b: __v4i32) -> __v4i32; - #[link_name = "llvm.loongarch.lsx.vadda.d"] - fn __lsx_vadda_d(a: __v2i64, b: __v2i64) -> __v2i64; #[link_name = "llvm.loongarch.lsx.vavg.b"] fn __lsx_vavg_b(a: __v16i8, b: __v16i8) -> __v16i8; #[link_name = "llvm.loongarch.lsx.vavg.h"] @@ -1357,34 +1349,6 @@ pub fn lsx_vsat_du(a: m128i) -> m128i { unsafe { transmute(__lsx_vsat_du(transmute(a), IMM6)) } } -#[inline] -#[target_feature(enable = "lsx")] -#[unstable(feature = "stdarch_loongarch", issue = "117427")] -pub fn lsx_vadda_b(a: m128i, b: m128i) -> m128i { - unsafe { transmute(__lsx_vadda_b(transmute(a), transmute(b))) } -} - -#[inline] -#[target_feature(enable = "lsx")] -#[unstable(feature = "stdarch_loongarch", issue = "117427")] -pub fn lsx_vadda_h(a: m128i, b: m128i) -> m128i { - unsafe { transmute(__lsx_vadda_h(transmute(a), transmute(b))) } -} - -#[inline] -#[target_feature(enable = "lsx")] -#[unstable(feature = "stdarch_loongarch", issue = "117427")] -pub fn lsx_vadda_w(a: m128i, b: m128i) -> m128i { - unsafe { transmute(__lsx_vadda_w(transmute(a), transmute(b))) } -} - -#[inline] -#[target_feature(enable = "lsx")] -#[unstable(feature = "stdarch_loongarch", issue = "117427")] -pub fn lsx_vadda_d(a: m128i, b: m128i) -> m128i { - unsafe { transmute(__lsx_vadda_d(transmute(a), transmute(b))) } -} - #[inline] #[target_feature(enable = "lsx")] #[unstable(feature = "stdarch_loongarch", issue = "117427")] diff --git a/crates/core_arch/src/loongarch64/lsx/portable.rs b/crates/core_arch/src/loongarch64/lsx/portable.rs index e98dcf050f..55d349142e 100644 --- a/crates/core_arch/src/loongarch64/lsx/portable.rs +++ b/crates/core_arch/src/loongarch64/lsx/portable.rs @@ -148,6 +148,10 @@ impl_vvv!("lsx", lsx_vssub_bu, is::simd_saturating_sub, m128i, u8x16); impl_vvv!("lsx", lsx_vssub_hu, is::simd_saturating_sub, m128i, u16x8); impl_vvv!("lsx", lsx_vssub_wu, is::simd_saturating_sub, m128i, u32x4); impl_vvv!("lsx", lsx_vssub_du, is::simd_saturating_sub, m128i, u64x2); +impl_vvv!("lsx", lsx_vadda_b, ls::simd_adda, m128i, i8x16); +impl_vvv!("lsx", lsx_vadda_h, ls::simd_adda, m128i, i16x8); +impl_vvv!("lsx", lsx_vadda_w, ls::simd_adda, m128i, i32x4); +impl_vvv!("lsx", lsx_vadda_d, ls::simd_adda, m128i, i64x2); impl_vuv!("lsx", lsx_vslli_b, is::simd_shl, m128i, i8x16); impl_vuv!("lsx", lsx_vslli_h, is::simd_shl, m128i, i16x8); diff --git a/crates/core_arch/src/loongarch64/simd.rs b/crates/core_arch/src/loongarch64/simd.rs index 6f573f56d6..ccad4b8c48 100644 --- a/crates/core_arch/src/loongarch64/simd.rs +++ b/crates/core_arch/src/loongarch64/simd.rs @@ -45,6 +45,19 @@ impl_simd_ext!(i64x4, i64); impl_simd_ext!(u64x2, u64); impl_simd_ext!(u64x4, u64); +#[inline(always)] +#[rustc_const_unstable(feature = "stdarch_const_helpers", issue = "none")] +pub(crate) const unsafe fn simd_abs(a: T) -> T { + let m: T = is::simd_lt(a, ls::simd_splat(0)); + is::simd_select(m, is::simd_neg(a), a) +} + +#[inline(always)] +#[rustc_const_unstable(feature = "stdarch_const_helpers", issue = "none")] +pub(crate) const unsafe fn simd_adda(a: T, b: T) -> T { + is::simd_add(ls::simd_abs(a), ls::simd_abs(b)) +} + #[inline(always)] #[rustc_const_unstable(feature = "stdarch_const_helpers", issue = "none")] pub(super) const unsafe fn simd_andn(a: T, b: T) -> T { diff --git a/crates/stdarch-gen-loongarch/lasx.spec b/crates/stdarch-gen-loongarch/lasx.spec index 27362d3259..07f2ed49d3 100644 --- a/crates/stdarch-gen-loongarch/lasx.spec +++ b/crates/stdarch-gen-loongarch/lasx.spec @@ -924,21 +924,25 @@ asm-fmts = xd, xj, ui6 data-types = UV4DI, UV4DI, UQI /// lasx_xvadda_b +impl = portable name = lasx_xvadda_b asm-fmts = xd, xj, xk data-types = V32QI, V32QI, V32QI /// lasx_xvadda_h +impl = portable name = lasx_xvadda_h asm-fmts = xd, xj, xk data-types = V16HI, V16HI, V16HI /// lasx_xvadda_w +impl = portable name = lasx_xvadda_w asm-fmts = xd, xj, xk data-types = V8SI, V8SI, V8SI /// lasx_xvadda_d +impl = portable name = lasx_xvadda_d asm-fmts = xd, xj, xk data-types = V4DI, V4DI, V4DI diff --git a/crates/stdarch-gen-loongarch/lsx.spec b/crates/stdarch-gen-loongarch/lsx.spec index 9b1b7cc245..0b32dd83c7 100644 --- a/crates/stdarch-gen-loongarch/lsx.spec +++ b/crates/stdarch-gen-loongarch/lsx.spec @@ -924,21 +924,25 @@ asm-fmts = vd, vj, ui6 data-types = UV2DI, UV2DI, UQI /// lsx_vadda_b +impl = portable name = lsx_vadda_b asm-fmts = vd, vj, vk data-types = V16QI, V16QI, V16QI /// lsx_vadda_h +impl = portable name = lsx_vadda_h asm-fmts = vd, vj, vk data-types = V8HI, V8HI, V8HI /// lsx_vadda_w +impl = portable name = lsx_vadda_w asm-fmts = vd, vj, vk data-types = V4SI, V4SI, V4SI /// lsx_vadda_d +impl = portable name = lsx_vadda_d asm-fmts = vd, vj, vk data-types = V2DI, V2DI, V2DI diff --git a/crates/stdarch-gen-loongarch/src/portable-intrinsics.txt b/crates/stdarch-gen-loongarch/src/portable-intrinsics.txt index 423af4cd71..6e3e5fe32e 100644 --- a/crates/stdarch-gen-loongarch/src/portable-intrinsics.txt +++ b/crates/stdarch-gen-loongarch/src/portable-intrinsics.txt @@ -219,6 +219,10 @@ lsx_vssub_bu lsx_vssub_hu lsx_vssub_wu lsx_vssub_du +lsx_vadda_b +lsx_vadda_h +lsx_vadda_w +lsx_vadda_d # LASX intrinsics lasx_xvsll_b @@ -435,3 +439,7 @@ lasx_xvssub_bu lasx_xvssub_hu lasx_xvssub_wu lasx_xvssub_du +lasx_xvadda_b +lasx_xvadda_h +lasx_xvadda_w +lasx_xvadda_d