From 839e9a53245a8752529aee22f69e9fae06260d58 Mon Sep 17 00:00:00 2001
From: "Sergey \"Shnatsel\" Davidoff" <shnatsel@gmail.com>
Date: Thu, 30 Jul 2026 16:01:50 +0100
Subject: [PATCH] Use a faster formulation of swizzle_dyn_precise on AVX-512
 VBMI.

                                                                   Doubles the throughput on both Ice Lake and Zen4.
                                                                   Improves latency on Ice Lake from 23 to 13 cycles, on Zen4 from 17 to 15 cycles.

                                                                   Improves encoding performance of a toy base64 implementation on Zen4 by -20% (time) +25% (throughput), no change on decoding benchmark
---
 crates/core_simd/src/swizzle_dyn.rs | 22 ++++++++++------------
 1 file changed, 10 insertions(+), 12 deletions(-)

diff --git a/crates/core_simd/src/swizzle_dyn.rs b/crates/core_simd/src/swizzle_dyn.rs
index f0253c284e3..da91e01801c 100644
--- a/crates/core_simd/src/swizzle_dyn.rs
+++ b/crates/core_simd/src/swizzle_dyn.rs
@@ -73,13 +73,12 @@ impl<const N: usize> Simd<u8, N> {
                 16 => transize(x86::_mm_shuffle_epi8, self, zeroing_idxs(idxs)),
                 #[cfg(all(target_feature = "avx512vl", target_feature = "avx512vbmi"))]
                 32 => {
-                    // Unlike vpshufb, vpermb doesn't zero out values in the result based on the index high bit
                     let swizzler = |bytes, idxs| {
-                        let mask = x86::_mm256_cmp_epu8_mask::<{ x86::_MM_CMPINT_LT }>(
-                            idxs,
-                            Simd::<u8, 32>::splat(N as u8).into(),
-                        );
-                        x86::_mm256_maskz_permutexvar_epi8(mask, idxs, bytes)
+                        // Clamp out-of-range indices to the first byte of a
+                        // second, all-zero table.
+                        let idxs =
+                            x86::_mm256_min_epu8(idxs, Simd::<u8, 32>::splat(N as u8).into());
+                        x86::_mm256_permutex2var_epi8(bytes, idxs, x86::_mm256_setzero_si256())
                     };
                     transize(swizzler, self, idxs)
                 }
@@ -90,13 +89,12 @@ impl<const N: usize> Simd<u8, N> {
                 // Notable absence: avx512bw pshufb shuffle
                 #[cfg(all(target_feature = "avx512vl", target_feature = "avx512vbmi"))]
                 64 => {
-                    // Unlike vpshufb, vpermb doesn't zero out values in the result based on the index high bit
                     let swizzler = |bytes, idxs| {
-                        let mask = x86::_mm512_cmp_epu8_mask::<{ x86::_MM_CMPINT_LT }>(
-                            idxs,
-                            Simd::<u8, 64>::splat(N as u8).into(),
-                        );
-                        x86::_mm512_maskz_permutexvar_epi8(mask, idxs, bytes)
+                        // Clamp out-of-range indices to the first byte of a
+                        // second, all-zero table.
+                        let idxs =
+                            x86::_mm512_min_epu8(idxs, Simd::<u8, 64>::splat(N as u8).into());
+                        x86::_mm512_permutex2var_epi8(bytes, idxs, x86::_mm512_setzero_si512())
                     };
                     transize(swizzler, self, idxs)
                 }