From: Nathan Moinvaziri <nathan@nathanm.com>
Date: Sun, 11 Jan 2026 19:32:44 +0000 (-0800)
Subject: Eliminate extra vmovdqu instruction folding xmm into zmm.
X-Git-Url: http://git.ipfire.org/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=26ab3c5e29afe63a2435fa0b1064449614d45bb7;p=thirdparty%2Fzlib-ng.git

Eliminate extra vmovdqu instruction folding xmm into zmm.

Fixed by using _mm512_castsi128_si512() and removing redundant insert.
---

diff --git a/arch/x86/crc32_pclmulqdq_tpl.h b/arch/x86/crc32_pclmulqdq_tpl.h
index e951c0579..a7b8edfdd 100644
--- a/arch/x86/crc32_pclmulqdq_tpl.h
+++ b/arch/x86/crc32_pclmulqdq_tpl.h
@@ -264,7 +264,7 @@ Z_FORCEINLINE static uint32_t crc32_copy_impl(uint32_t crc, uint8_t *dst, const
         }
 
         // Fold existing xmm state into first 64 bytes
-        zmm_t0 = _mm512_inserti32x4(_mm512_setzero_si512(), xmm_crc0, 0);
+        zmm_t0 = _mm512_castsi128_si512(xmm_crc0);
         zmm_t0 = _mm512_inserti32x4(zmm_t0, xmm_crc1, 1);
         zmm_t0 = _mm512_inserti32x4(zmm_t0, xmm_crc2, 2);
         zmm_t0 = _mm512_inserti32x4(zmm_t0, xmm_crc3, 3);