pq_crypto 0.4.2 → 0.5.0

data/ext/pqcrypto/vendor/mldsa-native/mldsa/src/native/aarch64/src/mld_polyvecl_pointwise_acc_montgomery_l4.S

@@ -0,0 +1,129 @@
+/* Copyright (c) The mldsa-native project authors
+ * SPDX-License-Identifier: Apache-2.0 OR ISC OR MIT
+ */
+
+#include "../../../common.h"
+#if defined(MLD_ARITH_BACKEND_AARCH64) && !defined(MLD_CONFIG_MULTILEVEL_NO_SHARED) && \
+           (defined(MLD_CONFIG_MULTILEVEL_WITH_SHARED) || MLDSA_L == 4)
+
+/*
+ * WARNING: This file is auto-derived from the mldsa-native source file
+ *   dev/aarch64_opt/src/mld_polyvecl_pointwise_acc_montgomery_l4.S using scripts/simpasm. Do not modify it directly.
+ */
+
+#if defined(__ELF__)
+.section .note.GNU-stack,"",@progbits
+#endif
+
+.text
+.balign 4
+.global MLD_ASM_NAMESPACE(polyvecl_pointwise_acc_montgomery_l4_asm)
+MLD_ASM_FN_SYMBOL(polyvecl_pointwise_acc_montgomery_l4_asm)
+
+        .cfi_startproc
+        mov	w3, #0xe001             // =57345
+        movk	w3, #0x7f, lsl #16
+        dup	v0.4s, w3
+        mov	w3, #0x2001             // =8193
+        movk	w3, #0x380, lsl #16
+        dup	v1.4s, w3
+        mov	x3, #0x40               // =64
+
+Lpolyvecl_pointwise_acc_montgomery_l4_loop_start:
+        ldr	q17, [x1, #0x10]
+        ldr	q18, [x1, #0x20]
+        ldr	q19, [x1, #0x30]
+        ldr	q16, [x1], #0x40
+        ldr	q21, [x2, #0x10]
+        ldr	q22, [x2, #0x20]
+        ldr	q23, [x2, #0x30]
+        ldr	q20, [x2], #0x40
+        smull	v24.2d, v16.2s, v20.2s
+        smull2	v25.2d, v16.4s, v20.4s
+        smull	v26.2d, v17.2s, v21.2s
+        smull2	v27.2d, v17.4s, v21.4s
+        smull	v28.2d, v18.2s, v22.2s
+        smull2	v29.2d, v18.4s, v22.4s
+        smull	v30.2d, v19.2s, v23.2s
+        smull2	v31.2d, v19.4s, v23.4s
+        ldr	q16, [x1, #0x3c0]
+        ldr	q17, [x1, #0x3d0]
+        ldr	q18, [x1, #0x3e0]
+        ldr	q19, [x1, #0x3f0]
+        ldr	q20, [x2, #0x3c0]
+        ldr	q21, [x2, #0x3d0]
+        ldr	q22, [x2, #0x3e0]
+        ldr	q23, [x2, #0x3f0]
+        smlal	v24.2d, v16.2s, v20.2s
+        smlal2	v25.2d, v16.4s, v20.4s
+        smlal	v26.2d, v17.2s, v21.2s
+        smlal2	v27.2d, v17.4s, v21.4s
+        smlal	v28.2d, v18.2s, v22.2s
+        smlal2	v29.2d, v18.4s, v22.4s
+        smlal	v30.2d, v19.2s, v23.2s
+        smlal2	v31.2d, v19.4s, v23.4s
+        ldr	q16, [x1, #0x7c0]
+        ldr	q17, [x1, #0x7d0]
+        ldr	q18, [x1, #0x7e0]
+        ldr	q19, [x1, #0x7f0]
+        ldr	q20, [x2, #0x7c0]
+        ldr	q21, [x2, #0x7d0]
+        ldr	q22, [x2, #0x7e0]
+        ldr	q23, [x2, #0x7f0]
+        smlal	v24.2d, v16.2s, v20.2s
+        smlal2	v25.2d, v16.4s, v20.4s
+        smlal	v26.2d, v17.2s, v21.2s
+        smlal2	v27.2d, v17.4s, v21.4s
+        smlal	v28.2d, v18.2s, v22.2s
+        smlal2	v29.2d, v18.4s, v22.4s
+        smlal	v30.2d, v19.2s, v23.2s
+        smlal2	v31.2d, v19.4s, v23.4s
+        ldr	q16, [x1, #0xbc0]
+        ldr	q17, [x1, #0xbd0]
+        ldr	q18, [x1, #0xbe0]
+        ldr	q19, [x1, #0xbf0]
+        ldr	q20, [x2, #0xbc0]
+        ldr	q21, [x2, #0xbd0]
+        ldr	q22, [x2, #0xbe0]
+        ldr	q23, [x2, #0xbf0]
+        smlal	v24.2d, v16.2s, v20.2s
+        smlal2	v25.2d, v16.4s, v20.4s
+        smlal	v26.2d, v17.2s, v21.2s
+        smlal2	v27.2d, v17.4s, v21.4s
+        smlal	v28.2d, v18.2s, v22.2s
+        smlal2	v29.2d, v18.4s, v22.4s
+        smlal	v30.2d, v19.2s, v23.2s
+        smlal2	v31.2d, v19.4s, v23.4s
+        uzp1	v16.4s, v24.4s, v25.4s
+        mul	v16.4s, v16.4s, v1.4s
+        smlsl	v24.2d, v16.2s, v0.2s
+        smlsl2	v25.2d, v16.4s, v0.4s
+        uzp2	v16.4s, v24.4s, v25.4s
+        uzp1	v17.4s, v26.4s, v27.4s
+        mul	v17.4s, v17.4s, v1.4s
+        smlsl	v26.2d, v17.2s, v0.2s
+        smlsl2	v27.2d, v17.4s, v0.4s
+        uzp2	v17.4s, v26.4s, v27.4s
+        uzp1	v18.4s, v28.4s, v29.4s
+        mul	v18.4s, v18.4s, v1.4s
+        smlsl	v28.2d, v18.2s, v0.2s
+        smlsl2	v29.2d, v18.4s, v0.4s
+        uzp2	v18.4s, v28.4s, v29.4s
+        uzp1	v19.4s, v30.4s, v31.4s
+        mul	v19.4s, v19.4s, v1.4s
+        smlsl	v30.2d, v19.2s, v0.2s
+        smlsl2	v31.2d, v19.4s, v0.4s
+        uzp2	v19.4s, v30.4s, v31.4s
+        str	q17, [x0, #0x10]
+        str	q18, [x0, #0x20]
+        str	q19, [x0, #0x30]
+        str	q16, [x0], #0x40
+        subs	x3, x3, #0x4
+        cbnz	x3, Lpolyvecl_pointwise_acc_montgomery_l4_loop_start
+        ret
+        .cfi_endproc
+
+MLD_ASM_FN_SIZE(polyvecl_pointwise_acc_montgomery_l4_asm)
+
+#endif /* MLD_ARITH_BACKEND_AARCH64 && !MLD_CONFIG_MULTILEVEL_NO_SHARED && \
+          (MLD_CONFIG_MULTILEVEL_WITH_SHARED || MLDSA_L == 4) */

data/ext/pqcrypto/vendor/mldsa-native/mldsa/src/native/aarch64/src/mld_polyvecl_pointwise_acc_montgomery_l5.S

@@ -0,0 +1,145 @@
+/* Copyright (c) The mldsa-native project authors
+ * SPDX-License-Identifier: Apache-2.0 OR ISC OR MIT
+ */
+
+#include "../../../common.h"
+#if defined(MLD_ARITH_BACKEND_AARCH64) && !defined(MLD_CONFIG_MULTILEVEL_NO_SHARED) && \
+           (defined(MLD_CONFIG_MULTILEVEL_WITH_SHARED) || MLDSA_L == 5)
+
+/*
+ * WARNING: This file is auto-derived from the mldsa-native source file
+ *   dev/aarch64_opt/src/mld_polyvecl_pointwise_acc_montgomery_l5.S using scripts/simpasm. Do not modify it directly.
+ */
+
+#if defined(__ELF__)
+.section .note.GNU-stack,"",@progbits
+#endif
+
+.text
+.balign 4
+.global MLD_ASM_NAMESPACE(polyvecl_pointwise_acc_montgomery_l5_asm)
+MLD_ASM_FN_SYMBOL(polyvecl_pointwise_acc_montgomery_l5_asm)
+
+        .cfi_startproc
+        mov	w3, #0xe001             // =57345
+        movk	w3, #0x7f, lsl #16
+        dup	v0.4s, w3
+        mov	w3, #0x2001             // =8193
+        movk	w3, #0x380, lsl #16
+        dup	v1.4s, w3
+        mov	x3, #0x40               // =64
+
+Lpolyvecl_pointwise_acc_montgomery_l5_loop_start:
+        ldr	q17, [x1, #0x10]
+        ldr	q18, [x1, #0x20]
+        ldr	q19, [x1, #0x30]
+        ldr	q16, [x1], #0x40
+        ldr	q21, [x2, #0x10]
+        ldr	q22, [x2, #0x20]
+        ldr	q23, [x2, #0x30]
+        ldr	q20, [x2], #0x40
+        smull	v24.2d, v16.2s, v20.2s
+        smull2	v25.2d, v16.4s, v20.4s
+        smull	v26.2d, v17.2s, v21.2s
+        smull2	v27.2d, v17.4s, v21.4s
+        smull	v28.2d, v18.2s, v22.2s
+        smull2	v29.2d, v18.4s, v22.4s
+        smull	v30.2d, v19.2s, v23.2s
+        smull2	v31.2d, v19.4s, v23.4s
+        ldr	q16, [x1, #0x3c0]
+        ldr	q17, [x1, #0x3d0]
+        ldr	q18, [x1, #0x3e0]
+        ldr	q19, [x1, #0x3f0]
+        ldr	q20, [x2, #0x3c0]
+        ldr	q21, [x2, #0x3d0]
+        ldr	q22, [x2, #0x3e0]
+        ldr	q23, [x2, #0x3f0]
+        smlal	v24.2d, v16.2s, v20.2s
+        smlal2	v25.2d, v16.4s, v20.4s
+        smlal	v26.2d, v17.2s, v21.2s
+        smlal2	v27.2d, v17.4s, v21.4s
+        smlal	v28.2d, v18.2s, v22.2s
+        smlal2	v29.2d, v18.4s, v22.4s
+        smlal	v30.2d, v19.2s, v23.2s
+        smlal2	v31.2d, v19.4s, v23.4s
+        ldr	q16, [x1, #0x7c0]
+        ldr	q17, [x1, #0x7d0]
+        ldr	q18, [x1, #0x7e0]
+        ldr	q19, [x1, #0x7f0]
+        ldr	q20, [x2, #0x7c0]
+        ldr	q21, [x2, #0x7d0]
+        ldr	q22, [x2, #0x7e0]
+        ldr	q23, [x2, #0x7f0]
+        smlal	v24.2d, v16.2s, v20.2s
+        smlal2	v25.2d, v16.4s, v20.4s
+        smlal	v26.2d, v17.2s, v21.2s
+        smlal2	v27.2d, v17.4s, v21.4s
+        smlal	v28.2d, v18.2s, v22.2s
+        smlal2	v29.2d, v18.4s, v22.4s
+        smlal	v30.2d, v19.2s, v23.2s
+        smlal2	v31.2d, v19.4s, v23.4s
+        ldr	q16, [x1, #0xbc0]
+        ldr	q17, [x1, #0xbd0]
+        ldr	q18, [x1, #0xbe0]
+        ldr	q19, [x1, #0xbf0]
+        ldr	q20, [x2, #0xbc0]
+        ldr	q21, [x2, #0xbd0]
+        ldr	q22, [x2, #0xbe0]
+        ldr	q23, [x2, #0xbf0]
+        smlal	v24.2d, v16.2s, v20.2s
+        smlal2	v25.2d, v16.4s, v20.4s
+        smlal	v26.2d, v17.2s, v21.2s
+        smlal2	v27.2d, v17.4s, v21.4s
+        smlal	v28.2d, v18.2s, v22.2s
+        smlal2	v29.2d, v18.4s, v22.4s
+        smlal	v30.2d, v19.2s, v23.2s
+        smlal2	v31.2d, v19.4s, v23.4s
+        ldr	q16, [x1, #0xfc0]
+        ldr	q17, [x1, #0xfd0]
+        ldr	q18, [x1, #0xfe0]
+        ldr	q19, [x1, #0xff0]
+        ldr	q20, [x2, #0xfc0]
+        ldr	q21, [x2, #0xfd0]
+        ldr	q22, [x2, #0xfe0]
+        ldr	q23, [x2, #0xff0]
+        smlal	v24.2d, v16.2s, v20.2s
+        smlal2	v25.2d, v16.4s, v20.4s
+        smlal	v26.2d, v17.2s, v21.2s
+        smlal2	v27.2d, v17.4s, v21.4s
+        smlal	v28.2d, v18.2s, v22.2s
+        smlal2	v29.2d, v18.4s, v22.4s
+        smlal	v30.2d, v19.2s, v23.2s
+        smlal2	v31.2d, v19.4s, v23.4s
+        uzp1	v16.4s, v24.4s, v25.4s
+        mul	v16.4s, v16.4s, v1.4s
+        smlsl	v24.2d, v16.2s, v0.2s
+        smlsl2	v25.2d, v16.4s, v0.4s
+        uzp2	v16.4s, v24.4s, v25.4s
+        uzp1	v17.4s, v26.4s, v27.4s
+        mul	v17.4s, v17.4s, v1.4s
+        smlsl	v26.2d, v17.2s, v0.2s
+        smlsl2	v27.2d, v17.4s, v0.4s
+        uzp2	v17.4s, v26.4s, v27.4s
+        uzp1	v18.4s, v28.4s, v29.4s
+        mul	v18.4s, v18.4s, v1.4s
+        smlsl	v28.2d, v18.2s, v0.2s
+        smlsl2	v29.2d, v18.4s, v0.4s
+        uzp2	v18.4s, v28.4s, v29.4s
+        uzp1	v19.4s, v30.4s, v31.4s
+        mul	v19.4s, v19.4s, v1.4s
+        smlsl	v30.2d, v19.2s, v0.2s
+        smlsl2	v31.2d, v19.4s, v0.4s
+        uzp2	v19.4s, v30.4s, v31.4s
+        str	q17, [x0, #0x10]
+        str	q18, [x0, #0x20]
+        str	q19, [x0, #0x30]
+        str	q16, [x0], #0x40
+        subs	x3, x3, #0x4
+        cbnz	x3, Lpolyvecl_pointwise_acc_montgomery_l5_loop_start
+        ret
+        .cfi_endproc
+
+MLD_ASM_FN_SIZE(polyvecl_pointwise_acc_montgomery_l5_asm)
+
+#endif /* MLD_ARITH_BACKEND_AARCH64 && !MLD_CONFIG_MULTILEVEL_NO_SHARED && \
+          (MLD_CONFIG_MULTILEVEL_WITH_SHARED || MLDSA_L == 5) */

data/ext/pqcrypto/vendor/mldsa-native/mldsa/src/native/aarch64/src/mld_polyvecl_pointwise_acc_montgomery_l7.S

@@ -0,0 +1,177 @@
+/* Copyright (c) The mldsa-native project authors
+ * SPDX-License-Identifier: Apache-2.0 OR ISC OR MIT
+ */
+
+#include "../../../common.h"
+#if defined(MLD_ARITH_BACKEND_AARCH64) && !defined(MLD_CONFIG_MULTILEVEL_NO_SHARED) && \
+           (defined(MLD_CONFIG_MULTILEVEL_WITH_SHARED) || MLDSA_L == 7)
+
+/*
+ * WARNING: This file is auto-derived from the mldsa-native source file
+ *   dev/aarch64_opt/src/mld_polyvecl_pointwise_acc_montgomery_l7.S using scripts/simpasm. Do not modify it directly.
+ */
+
+#if defined(__ELF__)
+.section .note.GNU-stack,"",@progbits
+#endif
+
+.text
+.balign 4
+.global MLD_ASM_NAMESPACE(polyvecl_pointwise_acc_montgomery_l7_asm)
+MLD_ASM_FN_SYMBOL(polyvecl_pointwise_acc_montgomery_l7_asm)
+
+        .cfi_startproc
+        mov	w3, #0xe001             // =57345
+        movk	w3, #0x7f, lsl #16
+        dup	v0.4s, w3
+        mov	w3, #0x2001             // =8193
+        movk	w3, #0x380, lsl #16
+        dup	v1.4s, w3
+        mov	x3, #0x40               // =64
+
+Lpolyvecl_pointwise_acc_montgomery_l7_loop_start:
+        ldr	q17, [x1, #0x10]
+        ldr	q18, [x1, #0x20]
+        ldr	q19, [x1, #0x30]
+        ldr	q16, [x1], #0x40
+        ldr	q21, [x2, #0x10]
+        ldr	q22, [x2, #0x20]
+        ldr	q23, [x2, #0x30]
+        ldr	q20, [x2], #0x40
+        smull	v24.2d, v16.2s, v20.2s
+        smull2	v25.2d, v16.4s, v20.4s
+        smull	v26.2d, v17.2s, v21.2s
+        smull2	v27.2d, v17.4s, v21.4s
+        smull	v28.2d, v18.2s, v22.2s
+        smull2	v29.2d, v18.4s, v22.4s
+        smull	v30.2d, v19.2s, v23.2s
+        smull2	v31.2d, v19.4s, v23.4s
+        ldr	q16, [x1, #0x3c0]
+        ldr	q17, [x1, #0x3d0]
+        ldr	q18, [x1, #0x3e0]
+        ldr	q19, [x1, #0x3f0]
+        ldr	q20, [x2, #0x3c0]
+        ldr	q21, [x2, #0x3d0]
+        ldr	q22, [x2, #0x3e0]
+        ldr	q23, [x2, #0x3f0]
+        smlal	v24.2d, v16.2s, v20.2s
+        smlal2	v25.2d, v16.4s, v20.4s
+        smlal	v26.2d, v17.2s, v21.2s
+        smlal2	v27.2d, v17.4s, v21.4s
+        smlal	v28.2d, v18.2s, v22.2s
+        smlal2	v29.2d, v18.4s, v22.4s
+        smlal	v30.2d, v19.2s, v23.2s
+        smlal2	v31.2d, v19.4s, v23.4s
+        ldr	q16, [x1, #0x7c0]
+        ldr	q17, [x1, #0x7d0]
+        ldr	q18, [x1, #0x7e0]
+        ldr	q19, [x1, #0x7f0]
+        ldr	q20, [x2, #0x7c0]
+        ldr	q21, [x2, #0x7d0]
+        ldr	q22, [x2, #0x7e0]
+        ldr	q23, [x2, #0x7f0]
+        smlal	v24.2d, v16.2s, v20.2s
+        smlal2	v25.2d, v16.4s, v20.4s
+        smlal	v26.2d, v17.2s, v21.2s
+        smlal2	v27.2d, v17.4s, v21.4s
+        smlal	v28.2d, v18.2s, v22.2s
+        smlal2	v29.2d, v18.4s, v22.4s
+        smlal	v30.2d, v19.2s, v23.2s
+        smlal2	v31.2d, v19.4s, v23.4s
+        ldr	q16, [x1, #0xbc0]
+        ldr	q17, [x1, #0xbd0]
+        ldr	q18, [x1, #0xbe0]
+        ldr	q19, [x1, #0xbf0]
+        ldr	q20, [x2, #0xbc0]
+        ldr	q21, [x2, #0xbd0]
+        ldr	q22, [x2, #0xbe0]
+        ldr	q23, [x2, #0xbf0]
+        smlal	v24.2d, v16.2s, v20.2s
+        smlal2	v25.2d, v16.4s, v20.4s
+        smlal	v26.2d, v17.2s, v21.2s
+        smlal2	v27.2d, v17.4s, v21.4s
+        smlal	v28.2d, v18.2s, v22.2s
+        smlal2	v29.2d, v18.4s, v22.4s
+        smlal	v30.2d, v19.2s, v23.2s
+        smlal2	v31.2d, v19.4s, v23.4s
+        ldr	q16, [x1, #0xfc0]
+        ldr	q17, [x1, #0xfd0]
+        ldr	q18, [x1, #0xfe0]
+        ldr	q19, [x1, #0xff0]
+        ldr	q20, [x2, #0xfc0]
+        ldr	q21, [x2, #0xfd0]
+        ldr	q22, [x2, #0xfe0]
+        ldr	q23, [x2, #0xff0]
+        smlal	v24.2d, v16.2s, v20.2s
+        smlal2	v25.2d, v16.4s, v20.4s
+        smlal	v26.2d, v17.2s, v21.2s
+        smlal2	v27.2d, v17.4s, v21.4s
+        smlal	v28.2d, v18.2s, v22.2s
+        smlal2	v29.2d, v18.4s, v22.4s
+        smlal	v30.2d, v19.2s, v23.2s
+        smlal2	v31.2d, v19.4s, v23.4s
+        ldr	q16, [x1, #0x13c0]
+        ldr	q17, [x1, #0x13d0]
+        ldr	q18, [x1, #0x13e0]
+        ldr	q19, [x1, #0x13f0]
+        ldr	q20, [x2, #0x13c0]
+        ldr	q21, [x2, #0x13d0]
+        ldr	q22, [x2, #0x13e0]
+        ldr	q23, [x2, #0x13f0]
+        smlal	v24.2d, v16.2s, v20.2s
+        smlal2	v25.2d, v16.4s, v20.4s
+        smlal	v26.2d, v17.2s, v21.2s
+        smlal2	v27.2d, v17.4s, v21.4s
+        smlal	v28.2d, v18.2s, v22.2s
+        smlal2	v29.2d, v18.4s, v22.4s
+        smlal	v30.2d, v19.2s, v23.2s
+        smlal2	v31.2d, v19.4s, v23.4s
+        ldr	q16, [x1, #0x17c0]
+        ldr	q17, [x1, #0x17d0]
+        ldr	q18, [x1, #0x17e0]
+        ldr	q19, [x1, #0x17f0]
+        ldr	q20, [x2, #0x17c0]
+        ldr	q21, [x2, #0x17d0]
+        ldr	q22, [x2, #0x17e0]
+        ldr	q23, [x2, #0x17f0]
+        smlal	v24.2d, v16.2s, v20.2s
+        smlal2	v25.2d, v16.4s, v20.4s
+        smlal	v26.2d, v17.2s, v21.2s
+        smlal2	v27.2d, v17.4s, v21.4s
+        smlal	v28.2d, v18.2s, v22.2s
+        smlal2	v29.2d, v18.4s, v22.4s
+        smlal	v30.2d, v19.2s, v23.2s
+        smlal2	v31.2d, v19.4s, v23.4s
+        uzp1	v16.4s, v24.4s, v25.4s
+        mul	v16.4s, v16.4s, v1.4s
+        smlsl	v24.2d, v16.2s, v0.2s
+        smlsl2	v25.2d, v16.4s, v0.4s
+        uzp2	v16.4s, v24.4s, v25.4s
+        uzp1	v17.4s, v26.4s, v27.4s
+        mul	v17.4s, v17.4s, v1.4s
+        smlsl	v26.2d, v17.2s, v0.2s
+        smlsl2	v27.2d, v17.4s, v0.4s
+        uzp2	v17.4s, v26.4s, v27.4s
+        uzp1	v18.4s, v28.4s, v29.4s
+        mul	v18.4s, v18.4s, v1.4s
+        smlsl	v28.2d, v18.2s, v0.2s
+        smlsl2	v29.2d, v18.4s, v0.4s
+        uzp2	v18.4s, v28.4s, v29.4s
+        uzp1	v19.4s, v30.4s, v31.4s
+        mul	v19.4s, v19.4s, v1.4s
+        smlsl	v30.2d, v19.2s, v0.2s
+        smlsl2	v31.2d, v19.4s, v0.4s
+        uzp2	v19.4s, v30.4s, v31.4s
+        str	q17, [x0, #0x10]
+        str	q18, [x0, #0x20]
+        str	q19, [x0, #0x30]
+        str	q16, [x0], #0x40
+        subs	x3, x3, #0x4
+        cbnz	x3, Lpolyvecl_pointwise_acc_montgomery_l7_loop_start
+        ret
+        .cfi_endproc
+
+MLD_ASM_FN_SIZE(polyvecl_pointwise_acc_montgomery_l7_asm)
+
+#endif /* MLD_ARITH_BACKEND_AARCH64 && !MLD_CONFIG_MULTILEVEL_NO_SHARED && \
+          (MLD_CONFIG_MULTILEVEL_WITH_SHARED || MLDSA_L == 7) */