"testnzc_ps" => p!("avx.vtestnzc.ps.256", (f32x8, f32x8) -> i32),
"testz_pd" => p!("avx.vtestz.pd.256", (f64x4, f64x4) -> i32),
"testz_ps" => p!("avx.vtestz.ps.256", (f32x8, f32x8) -> i32),
+
+ "abs_epi16" => p!("avx2.pabs.w", (i16x16) -> i16x16),
+ "abs_epi32" => p!("avx2.pabs.d", (i32x8) -> i32x8),
+ "abs_epi8" => p!("avx2.pabs.b", (i8x32) -> i8x32),
+ "adds_epi16" => p!("avx2.padds.w", (i16x16, i16x16) -> i16x16),
+ "adds_epi8" => p!("avx2.padds.b", (i8x32, i8x32) -> i8x32),
+ "adds_epu16" => p!("avx2.paddus.w", (i16x16, i16x16) -> i16x16),
+ "adds_epu8" => p!("avx2.paddus.b", (i8x32, i8x32) -> i8x32),
+ "avg_epu16" => p!("avx2.pavg.w", (i16x16, i16x16) -> i16x16),
+ "avg_epu8" => p!("avx2.pavg.b", (i8x32, i8x32) -> i8x32),
+ "hadd_epi16" => p!("avx2.phadd.w", (i16x16, i16x16) -> i16x16),
+ "hadd_epi32" => p!("avx2.phadd.d", (i32x8, i32x8) -> i32x8),
+ "hadds_epi16" => p!("avx2.phadd.sw", (i16x16, i16x16) -> i16x16),
+ "hsub_epi16" => p!("avx2.phsub.w", (i16x16, i16x16) -> i16x16),
+ "hsub_epi32" => p!("avx2.phsub.d", (i32x8, i32x8) -> i32x8),
+ "hsubs_epi16" => p!("avx2.phsub.sw", (i16x16, i16x16) -> i16x16),
+ "madd_epi16" => p!("avx2.pmadd.wd", (i16x16, i16x16) -> i32x8),
+ "maddubs_epi16" => p!("avx2.pmadd.ub.sw", (i8x32, i8x32) -> i16x16),
+ "max_epi16" => p!("avx2.pmaxs.w", (i16x16, i16x16) -> i16x16),
+ "max_epi32" => p!("avx2.pmaxs.d", (i32x8, i32x8) -> i32x8),
+ "max_epi8" => p!("avx2.pmaxs.b", (i8x32, i8x32) -> i8x32),
+ "max_epu16" => p!("avx2.pmaxu.w", (i16x16, i16x16) -> i16x16),
+ "max_epu32" => p!("avx2.pmaxu.d", (i32x8, i32x8) -> i32x8),
+ "max_epu8" => p!("avx2.pmaxu.b", (i8x32, i8x32) -> i8x32),
+ "min_epi16" => p!("avx2.pmins.w", (i16x16, i16x16) -> i16x16),
+ "min_epi32" => p!("avx2.pmins.d", (i32x8, i32x8) -> i32x8),
+ "min_epi8" => p!("avx2.pmins.b", (i8x32, i8x32) -> i8x32),
+ "min_epu16" => p!("avx2.pminu.w", (i16x16, i16x16) -> i16x16),
+ "min_epu32" => p!("avx2.pminu.d", (i32x8, i32x8) -> i32x8),
+ "min_epu8" => p!("avx2.pminu.b", (i8x32, i8x32) -> i8x32),
+ "mul_epi32" => p!("avx2.mul.dq", (i32x8, i32x8) -> i64x4),
+ "mul_epu32" => p!("avx2.mulu.dq", (i32x8, i32x8) -> i64x4),
+ "mulhi_epi16" => p!("avx2.pmulh.w", (i8x32, i8x32) -> i8x32),
+ "mulhi_epu16" => p!("avx2.pmulhu.w", (i8x32, i8x32) -> i8x32),
+ "mulhrs_epi16" => p!("avx2.pmul.hr.sw", (i16x16, i16x16) -> i16x16),
+ "packs_epi16" => p!("avx2.packsswb", (i16x16, i16x16) -> i8x32),
+ "packs_epi32" => p!("avx2.packssdw", (i32x8, i32x8) -> i16x16),
+ "packus_epi16" => p!("avx2.packuswb", (i16x16, i16x16) -> i8x32),
+ "packus_epi32" => p!("avx2.packusdw", (i32x8, i32x8) -> i16x16),
+ "permutevar8x32_epi32" => p!("avx2.permd", (i32x8, i32x8) -> i32x8),
+ "permutevar8x32_ps" => p!("avx2.permps", (f32x8, i32x8) -> i32x8),
+ "sad_epu8" => p!("avx2.psad.bw", (i8x32, i8x32) -> i64x4),
+ "shuffle_epi8" => p!("avx2.pshuf.b", (i8x32, i8x32) -> i8x32),
+ "sign_epi16" => p!("avx2.psign.w", (i16x16, i16x16) -> i16x16),
+ "sign_epi32" => p!("avx2.psign.d", (i32x8, i32x8) -> i32x8),
+ "sign_epi8" => p!("avx2.psign.b", (i8x32, i8x32) -> i8x32),
+ "subs_epi16" => p!("avx2.psubs.w", (i16x16, i16x16) -> i16x16),
+ "subs_epi8" => p!("avx2.psubs.b", (i8x32, i8x32) -> i8x32),
+ "subs_epu16" => p!("avx2.psubus.w", (i16x16, i16x16) -> i16x16),
+ "subs_epu8" => p!("avx2.psubus.b", (i8x32, i8x32) -> i8x32),
+
_ => return None,
})
} else {