core-arch: Add NEON fp16 intrinsics

Author: mrkajetanp

Diff for: crates/core_arch/src/aarch64/neon/generated.rs

@@ -4783,6 +4783,24 @@ pub unsafe fn vbsl_u64(a: uint64x1_t, b: uint64x1_t, c: uint64x1_t) -> uint64x1_
     simd_or(simd_and(a, b), simd_and(simd_xor(a, transmute(not)), c))
 }
 
+/// Bitwise Select.
+#[inline]
+#[target_feature(enable = "neon,fp16")]
+#[cfg_attr(target_arch = "arm", target_feature(enable = "v7"))]
+#[cfg_attr(all(test, target_arch = "arm"), assert_instr(vbsl))]
+#[cfg_attr(
+    all(test, any(target_arch = "aarch64", target_arch = "arm64ec")),
+    assert_instr(bsl)
+)]
+#[unstable(feature = "stdarch_neon_f16", issue = "136306")]
+pub unsafe fn vbsl_f16(a: uint16x4_t, b: float16x4_t, c: float16x4_t) -> float16x4_t {
+    let not = int16x4_t::splat(-1);
+    transmute(simd_or(
+        simd_and(a, transmute(b)),
+        simd_and(simd_xor(a, transmute(not)), transmute(c)),
+    ))
+}
+
 /// Bitwise Select.
 #[inline]
 #[target_feature(enable = "neon")]
@@ -5096,6 +5114,24 @@ pub unsafe fn vbslq_p16(a: uint16x8_t, b: poly16x8_t, c: poly16x8_t) -> poly16x8
     ))
 }
 
+/// Bitwise Select.
+#[inline]
+#[target_feature(enable = "neon,fp16")]
+#[cfg_attr(target_arch = "arm", target_feature(enable = "v7"))]
+#[cfg_attr(all(test, target_arch = "arm"), assert_instr(vbsl))]
+#[cfg_attr(
+    all(test, any(target_arch = "aarch64", target_arch = "arm64ec")),
+    assert_instr(bsl)
+)]
+#[unstable(feature = "stdarch_neon_f16", issue = "136306")]
+pub unsafe fn vbslq_f16(a: uint16x8_t, b: float16x8_t, c: float16x8_t) -> float16x8_t {
+    let not = int16x8_t::splat(-1);
+    transmute(simd_or(
+        simd_and(a, transmute(b)),
+        simd_and(simd_xor(a, transmute(not)), transmute(c)),
+    ))
+}
+
 /// Bitwise Select. (128-bit)
 #[inline]
 #[target_feature(enable = "neon")]

Diff for: crates/core_arch/src/arm_shared/neon/generated.rs

@@ -108,6 +108,74 @@ vsri_n_p64
 vsriq_n_p64
 vtst_p64
 vtstq_p64
+vaddh_f16
+vsubh_f16
+vabsh_f16
+vdivh_f16
+vmulh_f16
+vfmsh_f16
+vfmah_f16
+vminnmh_f16
+vmaxnmh_f16
+vrndh_f16
+vrndnh_f16
+vrndih_f16
+vrndah_f16
+vrndph_f16
+vrndmh_f16
+vrndxh_f16
+vsqrth_f16
+vnegh_f16
+vcvth_f16_s32
+vcvth_s32_f16
+vcvth_n_f16_s32
+vcvth_n_s32_f16
+vcvth_f16_u32
+vcvth_u32_f16
+vcvth_n_f16_u32
+vcvth_n_u32_f16
+vcvtah_s32_f16
+vcvtah_u32_f16
+vcvtmh_s32_f16
+vcvtmh_u32_f16
+vcvtpq_s16_f16
+vcvtpq_u16_f16
+vcvtp_s16_f16
+vcvtp_u16_f16
+vcvtph_s32_f16
+vcvtph_u32_f16
+vcvtnh_u32_f16
+vcvtnh_s32_f16
+vfmlsl_low_f16
+vfmlslq_low_f16
+vfmlsl_high_f16
+vfmlslq_high_f16
+vfmlsl_lane_high_f16
+vfmlsl_laneq_high_f16
+vfmlslq_lane_high_f16
+vfmlslq_laneq_high_f16
+vfmlsl_lane_low_f16
+vfmlsl_laneq_low_f16
+vfmlslq_lane_low_f16
+vfmlslq_laneq_low_f16
+vfmlal_low_f16
+vfmlalq_low_f16
+vfmlal_high_f16
+vfmlalq_high_f16
+vfmlal_lane_low_f16
+vfmlal_laneq_low_f16
+vfmlalq_lane_low_f16
+vfmlalq_laneq_low_f16
+vfmlal_lane_high_f16
+vfmlal_laneq_high_f16
+vfmlalq_lane_high_f16
+vfmlalq_laneq_high_f16
+vreinterpret_f16_p64
+vreinterpretq_f16_p64
+vreinterpret_p64_f16
+vreinterpretq_p64_f16
+vreinterpret_p128_f16
+vreinterpretq_p128_f16
 
 # Present in Clang header but triggers an ICE due to lack of backend support.
 vcmla_f32
@@ -134,6 +202,31 @@ vcmlaq_rot270_laneq_f32
 vcmlaq_rot90_f32
 vcmlaq_rot90_lane_f32
 vcmlaq_rot90_laneq_f32
+vcmla_f16
+vcmlaq_f16
+vcmla_laneq_f16
+vcmla_lane_f16
+vcmla_laneq_f16
+vcmlaq_lane_f16
+vcmlaq_laneq_f16
+vcmla_rot90_f16
+vcmlaq_rot90_f16
+vcmla_rot180_f16
+vcmlaq_rot180_f16
+vcmla_rot270_f16
+vcmlaq_rot270_f16
+vcmla_rot90_lane_f16
+vcmla_rot90_laneq_f16
+vcmlaq_rot90_lane_f16
+vcmlaq_rot90_laneq_f16
+vcmla_rot180_lane_f16
+vcmla_rot180_laneq_f16
+vcmlaq_rot180_lane_f16
+vcmlaq_rot180_laneq_f16
+vcmla_rot270_lane_f16
+vcmla_rot270_laneq_f16
+vcmlaq_rot270_lane_f16
+vcmlaq_rot270_laneq_f16
 
 # Implemented in stdarch for A64 only, Clang support both A32/A64
 vadd_s64
@@ -182,4 +275,46 @@ vrndpq_f32
 vrndq_f32
 vrndq_f32
 vrndx_f32
-vrndxq_f32
+vrndxq_f32
+vrnda_f16
+vrnda_f16
+vrndaq_f16
+vrndaq_f16
+vrnd_f16
+vrnd_f16
+vrndi_f16
+vrndi_f16
+vrndiq_f16
+vrndiq_f16
+vrndm_f16
+vrndm_f16
+vrndmq_f16
+vrndmq_f16
+vrndns_f16
+vrndp_f16
+vrndpq_f16
+vrndq_f16
+vrndx_f16
+vrndxq_f16
+vpmin_f16
+vpmax_f16
+vcaddq_rot270_f16
+vcaddq_rot90_f16
+vcadd_rot270_f16
+vcadd_rot90_f16
+vcvtm_s16_f16
+vcvtmq_s16_f16
+vcvtm_u16_f16
+vcvtmq_u16_f16
+vcvtaq_s16_f16
+vcvtaq_u16_f16
+vcvtnq_s16_f16
+vcvtnq_u16_f16
+vcvtn_s16_f16
+vcvtn_u16_f16
+vcvtaq_s16_f16
+vcvtaq_u16_f16
+vcvta_s16_f16
+vcvta_u16_f16
+vceqz_f16
+vceqzq_f16

Diff for: crates/core_arch/src/arm_shared/neon/mod.rs

@@ -194,7 +194,7 @@ fn generate_rust_program(notices: &str, intrinsic: &Intrinsic, target: &str) ->
 #![cfg_attr(any(target_arch = "aarch64", target_arch = "arm64ec"), feature(stdarch_neon_sha3))]
 #![cfg_attr(any(target_arch = "aarch64", target_arch = "arm64ec"), feature(stdarch_neon_sm4))]
 #![cfg_attr(any(target_arch = "aarch64", target_arch = "arm64ec"), feature(stdarch_neon_ftts))]
-#![cfg_attr(any(target_arch = "aarch64", target_arch = "arm64ec"), feature(stdarch_neon_f16))]
+#![feature(stdarch_neon_f16)]
 #![allow(non_upper_case_globals)]
 use core_arch::arch::{target_arch}::*;