mldsa_gh 0.1.0

data/ext/mldsa_gh_native/vendor/mldsa-native/mldsa/src/native/x86_64/src/mldsa_poly_use_hint_88_avx2_asm.S

@@ -0,0 +1,125 @@
+/*
+ * Copyright (c) The mldsa-native project authors
+ * SPDX-License-Identifier: Apache-2.0 OR ISC OR MIT
+ */
+
+/* References
+ * ==========
+ *
+ * - [REF_AVX2]
+ *   CRYSTALS-Dilithium optimized AVX2 implementation
+ *   Bai, Ducas, Kiltz, Lepoint, Lyubashevsky, Schwabe, Seiler, Stehlé
+ *   https://github.com/pq-crystals/dilithium/tree/master/avx2
+ */
+
+/*
+ * This file is derived from the public domain
+ * AVX2 Dilithium implementation @[REF_AVX2].
+ */
+
+
+/*yaml
+  Name: poly_use_hint_88_avx2_asm
+  Description: x86_64 AVX2 hint application (alpha = (Q-1)/88).
+    Use the hint polynomial h to correct the high bits of the polynomial a,
+    in place. Variant for parameter set ML-DSA-44 (GAMMA2 = (Q-1)/88).
+  Signature: void mld_poly_use_hint_88_avx2_asm(int32_t *a, const int32_t *h)
+  ABI:
+    Architecture: x86_64
+    CallingConvention: SysV
+    Features: [AVX2]
+    rdi:
+      type: buffer
+      size_bytes: 1024
+      permissions: read/write
+      c_parameter: int32_t *a
+      description: Input/output polynomial (256 x int32_t)
+    rsi:
+      type: buffer
+      size_bytes: 1024
+      permissions: read-only
+      c_parameter: const int32_t *h
+      description: Hint polynomial (256 x int32_t)
+*/
+
+#include "../../../common.h"
+
+#if defined(MLD_ARITH_BACKEND_X86_64_DEFAULT) &&   \
+    !defined(MLD_CONFIG_NO_VERIFY_API) &&          \
+    !defined(MLD_CONFIG_MULTILEVEL_NO_SHARED) &&   \
+    (defined(MLD_CONFIG_MULTILEVEL_WITH_SHARED) || \
+    MLD_CONFIG_PARAMETER_SET == 44)
+
+/*
+ * WARNING: This file is auto-derived from the mldsa-native source file
+ *   dev/x86_64/src/mldsa_poly_use_hint_88_avx2_asm.S using scripts/simpasm. Do not modify it directly.
+ */
+
+.text
+.balign 4
+.global MLD_ASM_NAMESPACE(poly_use_hint_88_avx2_asm)
+MLD_ASM_FN_SYMBOL(poly_use_hint_88_avx2_asm)
+
+        .cfi_startproc
+        movl $0x7f, %ecx
+        xorl %eax, %eax
+        vpxor %xmm5, %xmm5, %xmm5
+        movl $0x2c0b, %r8d           # imm = 0x2C0B
+        vmovd %r8d, %xmm8
+        vpbroadcastd %xmm8, %ymm8
+        vmovd %ecx, %xmm4
+        movl $0x7e6c00, %ecx         # imm = 0x7E6C00
+        movl $0x80, %r9d
+        vmovd %r9d, %xmm7
+        vpbroadcastd %xmm7, %ymm7
+        movl $0x2b, %r10d
+        vmovd %r10d, %xmm6
+        vpbroadcastd %xmm6, %ymm6
+        vmovd %ecx, %xmm3
+        vpbroadcastd %xmm4, %ymm4
+        vpbroadcastd %xmm3, %ymm3
+
+Lmld_poly_use_hint_88_avx2_asm_loop:
+        vmovdqa (%rdi), %ymm0
+        vmovdqa (%rsi), %ymm1
+        vpaddd %ymm0, %ymm4, %ymm10
+        vpsrld $0x7, %ymm10, %ymm10
+        vpmulhuw %ymm8, %ymm10, %ymm10
+        vpmulhrsw %ymm7, %ymm10, %ymm10
+        vpcmpgtd %ymm3, %ymm0, %ymm11
+        vpandn %ymm10, %ymm11, %ymm9
+        vpslld $0x1, %ymm10, %ymm12
+        vpaddd %ymm10, %ymm12, %ymm12
+        vpslld $0x5, %ymm12, %ymm10
+        vpsubd %ymm12, %ymm10, %ymm10
+        vpslld $0xb, %ymm10, %ymm10
+        vpsubd %ymm10, %ymm0, %ymm0
+        vpaddd %ymm11, %ymm0, %ymm0
+        vpcmpgtd %ymm5, %ymm0, %ymm0
+        vpandn %ymm1, %ymm0, %ymm0
+        vpslld $0x1, %ymm0, %ymm0
+        vpsubd %ymm0, %ymm1, %ymm0
+        vpaddd %ymm9, %ymm0, %ymm0
+        vpblendvb %ymm0, %ymm6, %ymm0, %ymm0
+        vpcmpgtd %ymm6, %ymm0, %ymm1
+        vpandn %ymm0, %ymm1, %ymm0
+        vmovdqa %ymm0, (%rdi)
+        addq $0x20, %rdi
+        addq $0x20, %rsi
+        addq $0x20, %rax
+        cmpq $0x400, %rax            # imm = 0x400
+        jne Lmld_poly_use_hint_88_avx2_asm_loop
+        retq
+        .cfi_endproc
+
+MLD_ASM_FN_SIZE(poly_use_hint_88_avx2_asm)
+
+
+#endif /* MLD_ARITH_BACKEND_X86_64_DEFAULT && !MLD_CONFIG_NO_VERIFY_API && \
+          !MLD_CONFIG_MULTILEVEL_NO_SHARED && \
+          (MLD_CONFIG_MULTILEVEL_WITH_SHARED || MLD_CONFIG_PARAMETER_SET == 44) \
+        */
+
+#if defined(__ELF__)
+.section .note.GNU-stack,"",%progbits
+#endif

data/ext/mldsa_gh_native/vendor/mldsa-native/mldsa/src/native/x86_64/src/mldsa_polyz_unpack_17_avx2_asm.S

@@ -0,0 +1,355 @@
+/*
+ * Copyright (c) The mldsa-native project authors
+ * SPDX-License-Identifier: Apache-2.0 OR ISC OR MIT
+ */
+
+/* References
+ * ==========
+ *
+ * - [REF_AVX2]
+ *   CRYSTALS-Dilithium optimized AVX2 implementation
+ *   Bai, Ducas, Kiltz, Lepoint, Lyubashevsky, Schwabe, Seiler, Stehlé
+ *   https://github.com/pq-crystals/dilithium/tree/master/avx2
+ */
+
+/*
+ * This file is derived from the public domain
+ * AVX2 Dilithium implementation @[REF_AVX2].
+ */
+
+
+/*yaml
+  Name: polyz_unpack_17_avx2_asm
+  Description: x86_64 AVX2 unpacking of 17-bit packed coefficients.
+    Unpack polynomial z with 18-bit packed coefficients (GAMMA1 = 2^17),
+    mapping packed [0, 2^18-1] to signed [-(2^17-1), 2^17] via GAMMA1 - x.
+  Signature: void mld_polyz_unpack_17_avx2_asm(int32_t *r, const uint8_t *a)
+  ABI:
+    Architecture: x86_64
+    CallingConvention: SysV
+    Features: [AVX2]
+    rdi:
+      type: buffer
+      size_bytes: 1024
+      permissions: write-only
+      c_parameter: int32_t *r
+      description: Output polynomial (256 x int32_t)
+    rsi:
+      type: buffer
+      size_bytes: 576
+      permissions: read-only
+      c_parameter: const uint8_t *a
+      description: Packed input bytes
+*/
+
+#include "../../../common.h"
+
+#if defined(MLD_ARITH_BACKEND_X86_64_DEFAULT) &&   \
+    (!defined(MLD_CONFIG_NO_SIGN_API) ||           \
+     !defined(MLD_CONFIG_NO_VERIFY_API)) &&        \
+    !defined(MLD_CONFIG_MULTILEVEL_NO_SHARED) &&   \
+    (defined(MLD_CONFIG_MULTILEVEL_WITH_SHARED) || \
+     MLD_CONFIG_PARAMETER_SET == 44)
+
+
+/*
+ * WARNING: This file is auto-derived from the mldsa-native source file
+ *   dev/x86_64/src/mldsa_polyz_unpack_17_avx2_asm.S using scripts/simpasm. Do not modify it directly.
+ */
+
+.text
+.balign 4
+.global MLD_ASM_NAMESPACE(polyz_unpack_17_avx2_asm)
+MLD_ASM_FN_SYMBOL(polyz_unpack_17_avx2_asm)
+
+        .cfi_startproc
+        movabsq $-0xfbfcfd00fdff00, %rax # imm = 0xFF040302FF020100
+        vmovq %rax, %xmm1
+        movabsq $-0xf7f8f900f9fafc, %rax # imm = 0xFF080706FF060504
+        vpinsrq $0x1, %rax, %xmm1, %xmm1
+        movabsq $-0xe4e5e600e6e7e9, %rax # imm = 0xFF1B1A19FF191817
+        vmovq %rax, %xmm5
+        movabsq $-0xe0e1e200e2e3e5, %rax # imm = 0xFF1F1E1DFF1D1C1B
+        vpinsrq $0x1, %rax, %xmm5, %xmm5
+        vinserti128 $0x1, %xmm5, %ymm1, %ymm1
+        movabsq $0x200000000, %rax      # imm = 0x200000000
+        vmovq %rax, %xmm2
+        movabsq $0x600000004, %rax      # imm = 0x600000004
+        vpinsrq $0x1, %rax, %xmm2, %xmm2
+        vinserti128 $0x1, %xmm2, %ymm2, %ymm2
+        movl $0x3ffff, %eax          # imm = 0x3FFFF
+        vmovd %eax, %xmm3
+        vpbroadcastd %xmm3, %ymm3
+        movl $0x20000, %eax          # imm = 0x20000
+        vmovd %eax, %xmm4
+        vpbroadcastd %xmm4, %ymm4
+        vmovdqu (%rsi), %xmm0
+        vmovdqu 0x2(%rsi), %xmm5
+        vinserti128 $0x1, %xmm5, %ymm0, %ymm0
+        vpshufb %ymm1, %ymm0, %ymm0
+        vpsrlvd %ymm2, %ymm0, %ymm0
+        vpand %ymm3, %ymm0, %ymm0
+        vpsubd %ymm0, %ymm4, %ymm0
+        vmovdqa %ymm0, (%rdi)
+        vmovdqu 0x12(%rsi), %xmm0
+        vmovdqu 0x14(%rsi), %xmm5
+        vinserti128 $0x1, %xmm5, %ymm0, %ymm0
+        vpshufb %ymm1, %ymm0, %ymm0
+        vpsrlvd %ymm2, %ymm0, %ymm0
+        vpand %ymm3, %ymm0, %ymm0
+        vpsubd %ymm0, %ymm4, %ymm0
+        vmovdqa %ymm0, 0x20(%rdi)
+        vmovdqu 0x24(%rsi), %xmm0
+        vmovdqu 0x26(%rsi), %xmm5
+        vinserti128 $0x1, %xmm5, %ymm0, %ymm0
+        vpshufb %ymm1, %ymm0, %ymm0
+        vpsrlvd %ymm2, %ymm0, %ymm0
+        vpand %ymm3, %ymm0, %ymm0
+        vpsubd %ymm0, %ymm4, %ymm0
+        vmovdqa %ymm0, 0x40(%rdi)
+        vmovdqu 0x36(%rsi), %xmm0
+        vmovdqu 0x38(%rsi), %xmm5
+        vinserti128 $0x1, %xmm5, %ymm0, %ymm0
+        vpshufb %ymm1, %ymm0, %ymm0
+        vpsrlvd %ymm2, %ymm0, %ymm0
+        vpand %ymm3, %ymm0, %ymm0
+        vpsubd %ymm0, %ymm4, %ymm0
+        vmovdqa %ymm0, 0x60(%rdi)
+        vmovdqu 0x48(%rsi), %xmm0
+        vmovdqu 0x4a(%rsi), %xmm5
+        vinserti128 $0x1, %xmm5, %ymm0, %ymm0
+        vpshufb %ymm1, %ymm0, %ymm0
+        vpsrlvd %ymm2, %ymm0, %ymm0
+        vpand %ymm3, %ymm0, %ymm0
+        vpsubd %ymm0, %ymm4, %ymm0
+        vmovdqa %ymm0, 0x80(%rdi)
+        vmovdqu 0x5a(%rsi), %xmm0
+        vmovdqu 0x5c(%rsi), %xmm5
+        vinserti128 $0x1, %xmm5, %ymm0, %ymm0
+        vpshufb %ymm1, %ymm0, %ymm0
+        vpsrlvd %ymm2, %ymm0, %ymm0
+        vpand %ymm3, %ymm0, %ymm0
+        vpsubd %ymm0, %ymm4, %ymm0
+        vmovdqa %ymm0, 0xa0(%rdi)
+        vmovdqu 0x6c(%rsi), %xmm0
+        vmovdqu 0x6e(%rsi), %xmm5
+        vinserti128 $0x1, %xmm5, %ymm0, %ymm0
+        vpshufb %ymm1, %ymm0, %ymm0
+        vpsrlvd %ymm2, %ymm0, %ymm0
+        vpand %ymm3, %ymm0, %ymm0
+        vpsubd %ymm0, %ymm4, %ymm0
+        vmovdqa %ymm0, 0xc0(%rdi)
+        vmovdqu 0x7e(%rsi), %xmm0
+        vmovdqu 0x80(%rsi), %xmm5
+        vinserti128 $0x1, %xmm5, %ymm0, %ymm0
+        vpshufb %ymm1, %ymm0, %ymm0
+        vpsrlvd %ymm2, %ymm0, %ymm0
+        vpand %ymm3, %ymm0, %ymm0
+        vpsubd %ymm0, %ymm4, %ymm0
+        vmovdqa %ymm0, 0xe0(%rdi)
+        vmovdqu 0x90(%rsi), %xmm0
+        vmovdqu 0x92(%rsi), %xmm5
+        vinserti128 $0x1, %xmm5, %ymm0, %ymm0
+        vpshufb %ymm1, %ymm0, %ymm0
+        vpsrlvd %ymm2, %ymm0, %ymm0
+        vpand %ymm3, %ymm0, %ymm0
+        vpsubd %ymm0, %ymm4, %ymm0
+        vmovdqa %ymm0, 0x100(%rdi)
+        vmovdqu 0xa2(%rsi), %xmm0
+        vmovdqu 0xa4(%rsi), %xmm5
+        vinserti128 $0x1, %xmm5, %ymm0, %ymm0
+        vpshufb %ymm1, %ymm0, %ymm0
+        vpsrlvd %ymm2, %ymm0, %ymm0
+        vpand %ymm3, %ymm0, %ymm0
+        vpsubd %ymm0, %ymm4, %ymm0
+        vmovdqa %ymm0, 0x120(%rdi)
+        vmovdqu 0xb4(%rsi), %xmm0
+        vmovdqu 0xb6(%rsi), %xmm5
+        vinserti128 $0x1, %xmm5, %ymm0, %ymm0
+        vpshufb %ymm1, %ymm0, %ymm0
+        vpsrlvd %ymm2, %ymm0, %ymm0
+        vpand %ymm3, %ymm0, %ymm0
+        vpsubd %ymm0, %ymm4, %ymm0
+        vmovdqa %ymm0, 0x140(%rdi)
+        vmovdqu 0xc6(%rsi), %xmm0
+        vmovdqu 0xc8(%rsi), %xmm5
+        vinserti128 $0x1, %xmm5, %ymm0, %ymm0
+        vpshufb %ymm1, %ymm0, %ymm0
+        vpsrlvd %ymm2, %ymm0, %ymm0
+        vpand %ymm3, %ymm0, %ymm0
+        vpsubd %ymm0, %ymm4, %ymm0
+        vmovdqa %ymm0, 0x160(%rdi)
+        vmovdqu 0xd8(%rsi), %xmm0
+        vmovdqu 0xda(%rsi), %xmm5
+        vinserti128 $0x1, %xmm5, %ymm0, %ymm0
+        vpshufb %ymm1, %ymm0, %ymm0
+        vpsrlvd %ymm2, %ymm0, %ymm0
+        vpand %ymm3, %ymm0, %ymm0
+        vpsubd %ymm0, %ymm4, %ymm0
+        vmovdqa %ymm0, 0x180(%rdi)
+        vmovdqu 0xea(%rsi), %xmm0
+        vmovdqu 0xec(%rsi), %xmm5
+        vinserti128 $0x1, %xmm5, %ymm0, %ymm0
+        vpshufb %ymm1, %ymm0, %ymm0
+        vpsrlvd %ymm2, %ymm0, %ymm0
+        vpand %ymm3, %ymm0, %ymm0
+        vpsubd %ymm0, %ymm4, %ymm0
+        vmovdqa %ymm0, 0x1a0(%rdi)
+        vmovdqu 0xfc(%rsi), %xmm0
+        vmovdqu 0xfe(%rsi), %xmm5
+        vinserti128 $0x1, %xmm5, %ymm0, %ymm0
+        vpshufb %ymm1, %ymm0, %ymm0
+        vpsrlvd %ymm2, %ymm0, %ymm0
+        vpand %ymm3, %ymm0, %ymm0
+        vpsubd %ymm0, %ymm4, %ymm0
+        vmovdqa %ymm0, 0x1c0(%rdi)
+        vmovdqu 0x10e(%rsi), %xmm0
+        vmovdqu 0x110(%rsi), %xmm5
+        vinserti128 $0x1, %xmm5, %ymm0, %ymm0
+        vpshufb %ymm1, %ymm0, %ymm0
+        vpsrlvd %ymm2, %ymm0, %ymm0
+        vpand %ymm3, %ymm0, %ymm0
+        vpsubd %ymm0, %ymm4, %ymm0
+        vmovdqa %ymm0, 0x1e0(%rdi)
+        vmovdqu 0x120(%rsi), %xmm0
+        vmovdqu 0x122(%rsi), %xmm5
+        vinserti128 $0x1, %xmm5, %ymm0, %ymm0
+        vpshufb %ymm1, %ymm0, %ymm0
+        vpsrlvd %ymm2, %ymm0, %ymm0
+        vpand %ymm3, %ymm0, %ymm0
+        vpsubd %ymm0, %ymm4, %ymm0
+        vmovdqa %ymm0, 0x200(%rdi)
+        vmovdqu 0x132(%rsi), %xmm0
+        vmovdqu 0x134(%rsi), %xmm5
+        vinserti128 $0x1, %xmm5, %ymm0, %ymm0
+        vpshufb %ymm1, %ymm0, %ymm0
+        vpsrlvd %ymm2, %ymm0, %ymm0
+        vpand %ymm3, %ymm0, %ymm0
+        vpsubd %ymm0, %ymm4, %ymm0
+        vmovdqa %ymm0, 0x220(%rdi)
+        vmovdqu 0x144(%rsi), %xmm0
+        vmovdqu 0x146(%rsi), %xmm5
+        vinserti128 $0x1, %xmm5, %ymm0, %ymm0
+        vpshufb %ymm1, %ymm0, %ymm0
+        vpsrlvd %ymm2, %ymm0, %ymm0
+        vpand %ymm3, %ymm0, %ymm0
+        vpsubd %ymm0, %ymm4, %ymm0
+        vmovdqa %ymm0, 0x240(%rdi)
+        vmovdqu 0x156(%rsi), %xmm0
+        vmovdqu 0x158(%rsi), %xmm5
+        vinserti128 $0x1, %xmm5, %ymm0, %ymm0
+        vpshufb %ymm1, %ymm0, %ymm0
+        vpsrlvd %ymm2, %ymm0, %ymm0
+        vpand %ymm3, %ymm0, %ymm0
+        vpsubd %ymm0, %ymm4, %ymm0
+        vmovdqa %ymm0, 0x260(%rdi)
+        vmovdqu 0x168(%rsi), %xmm0
+        vmovdqu 0x16a(%rsi), %xmm5
+        vinserti128 $0x1, %xmm5, %ymm0, %ymm0
+        vpshufb %ymm1, %ymm0, %ymm0
+        vpsrlvd %ymm2, %ymm0, %ymm0
+        vpand %ymm3, %ymm0, %ymm0
+        vpsubd %ymm0, %ymm4, %ymm0
+        vmovdqa %ymm0, 0x280(%rdi)
+        vmovdqu 0x17a(%rsi), %xmm0
+        vmovdqu 0x17c(%rsi), %xmm5
+        vinserti128 $0x1, %xmm5, %ymm0, %ymm0
+        vpshufb %ymm1, %ymm0, %ymm0
+        vpsrlvd %ymm2, %ymm0, %ymm0
+        vpand %ymm3, %ymm0, %ymm0
+        vpsubd %ymm0, %ymm4, %ymm0
+        vmovdqa %ymm0, 0x2a0(%rdi)
+        vmovdqu 0x18c(%rsi), %xmm0
+        vmovdqu 0x18e(%rsi), %xmm5
+        vinserti128 $0x1, %xmm5, %ymm0, %ymm0
+        vpshufb %ymm1, %ymm0, %ymm0
+        vpsrlvd %ymm2, %ymm0, %ymm0
+        vpand %ymm3, %ymm0, %ymm0
+        vpsubd %ymm0, %ymm4, %ymm0
+        vmovdqa %ymm0, 0x2c0(%rdi)
+        vmovdqu 0x19e(%rsi), %xmm0
+        vmovdqu 0x1a0(%rsi), %xmm5
+        vinserti128 $0x1, %xmm5, %ymm0, %ymm0
+        vpshufb %ymm1, %ymm0, %ymm0
+        vpsrlvd %ymm2, %ymm0, %ymm0
+        vpand %ymm3, %ymm0, %ymm0
+        vpsubd %ymm0, %ymm4, %ymm0
+        vmovdqa %ymm0, 0x2e0(%rdi)
+        vmovdqu 0x1b0(%rsi), %xmm0
+        vmovdqu 0x1b2(%rsi), %xmm5
+        vinserti128 $0x1, %xmm5, %ymm0, %ymm0
+        vpshufb %ymm1, %ymm0, %ymm0
+        vpsrlvd %ymm2, %ymm0, %ymm0
+        vpand %ymm3, %ymm0, %ymm0
+        vpsubd %ymm0, %ymm4, %ymm0
+        vmovdqa %ymm0, 0x300(%rdi)
+        vmovdqu 0x1c2(%rsi), %xmm0
+        vmovdqu 0x1c4(%rsi), %xmm5
+        vinserti128 $0x1, %xmm5, %ymm0, %ymm0
+        vpshufb %ymm1, %ymm0, %ymm0
+        vpsrlvd %ymm2, %ymm0, %ymm0
+        vpand %ymm3, %ymm0, %ymm0
+        vpsubd %ymm0, %ymm4, %ymm0
+        vmovdqa %ymm0, 0x320(%rdi)
+        vmovdqu 0x1d4(%rsi), %xmm0
+        vmovdqu 0x1d6(%rsi), %xmm5
+        vinserti128 $0x1, %xmm5, %ymm0, %ymm0
+        vpshufb %ymm1, %ymm0, %ymm0
+        vpsrlvd %ymm2, %ymm0, %ymm0
+        vpand %ymm3, %ymm0, %ymm0
+        vpsubd %ymm0, %ymm4, %ymm0
+        vmovdqa %ymm0, 0x340(%rdi)
+        vmovdqu 0x1e6(%rsi), %xmm0
+        vmovdqu 0x1e8(%rsi), %xmm5
+        vinserti128 $0x1, %xmm5, %ymm0, %ymm0
+        vpshufb %ymm1, %ymm0, %ymm0
+        vpsrlvd %ymm2, %ymm0, %ymm0
+        vpand %ymm3, %ymm0, %ymm0
+        vpsubd %ymm0, %ymm4, %ymm0
+        vmovdqa %ymm0, 0x360(%rdi)
+        vmovdqu 0x1f8(%rsi), %xmm0
+        vmovdqu 0x1fa(%rsi), %xmm5
+        vinserti128 $0x1, %xmm5, %ymm0, %ymm0
+        vpshufb %ymm1, %ymm0, %ymm0
+        vpsrlvd %ymm2, %ymm0, %ymm0
+        vpand %ymm3, %ymm0, %ymm0
+        vpsubd %ymm0, %ymm4, %ymm0
+        vmovdqa %ymm0, 0x380(%rdi)
+        vmovdqu 0x20a(%rsi), %xmm0
+        vmovdqu 0x20c(%rsi), %xmm5
+        vinserti128 $0x1, %xmm5, %ymm0, %ymm0
+        vpshufb %ymm1, %ymm0, %ymm0
+        vpsrlvd %ymm2, %ymm0, %ymm0
+        vpand %ymm3, %ymm0, %ymm0
+        vpsubd %ymm0, %ymm4, %ymm0
+        vmovdqa %ymm0, 0x3a0(%rdi)
+        vmovdqu 0x21c(%rsi), %xmm0
+        vmovdqu 0x21e(%rsi), %xmm5
+        vinserti128 $0x1, %xmm5, %ymm0, %ymm0
+        vpshufb %ymm1, %ymm0, %ymm0
+        vpsrlvd %ymm2, %ymm0, %ymm0
+        vpand %ymm3, %ymm0, %ymm0
+        vpsubd %ymm0, %ymm4, %ymm0
+        vmovdqa %ymm0, 0x3c0(%rdi)
+        vmovdqu 0x22e(%rsi), %xmm0
+        vmovdqu 0x230(%rsi), %xmm5
+        vinserti128 $0x1, %xmm5, %ymm0, %ymm0
+        vpshufb %ymm1, %ymm0, %ymm0
+        vpsrlvd %ymm2, %ymm0, %ymm0
+        vpand %ymm3, %ymm0, %ymm0
+        vpsubd %ymm0, %ymm4, %ymm0
+        vmovdqa %ymm0, 0x3e0(%rdi)
+        retq
+        .cfi_endproc
+
+MLD_ASM_FN_SIZE(polyz_unpack_17_avx2_asm)
+
+
+#endif /* MLD_ARITH_BACKEND_X86_64_DEFAULT && (!MLD_CONFIG_NO_SIGN_API || \
+          !MLD_CONFIG_NO_VERIFY_API) && !MLD_CONFIG_MULTILEVEL_NO_SHARED && \
+          (MLD_CONFIG_MULTILEVEL_WITH_SHARED || MLD_CONFIG_PARAMETER_SET == 44) \
+        */
+
+#if defined(__ELF__)
+.section .note.GNU-stack,"",%progbits
+#endif