Add 'neozip/' from commit 'c2712b8a345191f6ed79558c089777df94590087'

git-subtree-dir: neozip
git-subtree-mainline: b1e34e861b5d732afe828d58aad2c638135061fd
git-subtree-split: c2712b8a345191f6ed79558c089777df94590087

1 files changed, 79 insertions, 0 deletions

diff --git a/neozip/arch/arm/neon_intrins.h b/neozip/arch/arm/neon_intrins.h
new file mode 100644
index 0000000000..449916e0b7
--- /dev/null
+++ b/neozip/arch/arm/neon_intrins.h
@@ -0,0 +1,79 @@
+#ifndef ARM_NEON_INTRINS_H
+#define ARM_NEON_INTRINS_H
+
+#if defined(_MSC_VER) && defined(ARCH_ARM) && defined(ARCH_64BIT)
+/* arm64_neon.h is MSVC specific */
+#  include <arm64_neon.h>
+#else
+#  include <arm_neon.h>
+#endif
+
+#if defined(ARM_NEON) && defined(ARCH_ARM) && defined(ARCH_32BIT)
+/* Compatibility shim for the _high family of functions */
+#define vmull_high_u8(a, b) vmull_u8(vget_high_u8(a), vget_high_u8(b))
+#define vmlal_high_u8(a, b, c) vmlal_u8(a, vget_high_u8(b), vget_high_u8(c))
+#define vmlal_high_u16(a, b, c) vmlal_u16(a, vget_high_u16(b), vget_high_u16(c))
+#define vaddw_high_u8(a, b) vaddw_u8(a, vget_high_u8(b))
+#endif
+
+#ifdef ARM_NEON
+
+#define vqsubq_u16_x4_x1(out, a, b) do { \
+    out.val[0] = vqsubq_u16(a.val[0], b); \
+    out.val[1] = vqsubq_u16(a.val[1], b); \
+    out.val[2] = vqsubq_u16(a.val[2], b); \
+    out.val[3] = vqsubq_u16(a.val[3], b); \
+} while (0)
+
+#  if defined(ARCH_ARM) && defined(ARCH_32BIT) && defined(__clang__) && \
+    (!defined(__clang_major__) || __clang_major__ < 20)
+/* Clang versions before 20 have too strict of an
+ * alignment requirement (:256) for x4 NEON intrinsics */
+#    undef ARM_NEON_HASLD4
+#    undef vld1q_u16_x4
+#    undef vld1q_u8_x4
+#    undef vst1q_u16_x4
+#  endif
+
+#  ifndef ARM_NEON_HASLD4
+
+static inline uint16x8x4_t vld1q_u16_x4(uint16_t const *a) {
+    uint16x8x4_t ret;
+    ret.val[0] = vld1q_u16(a);
+    ret.val[1] = vld1q_u16(a+8);
+    ret.val[2] = vld1q_u16(a+16);
+    ret.val[3] = vld1q_u16(a+24);
+    return ret;
+}
+
+static inline uint8x16x4_t vld1q_u8_x4(uint8_t const *a) {
+    uint8x16x4_t ret;
+    ret.val[0] = vld1q_u8(a);
+    ret.val[1] = vld1q_u8(a+16);
+    ret.val[2] = vld1q_u8(a+32);
+    ret.val[3] = vld1q_u8(a+48);
+    return ret;
+}
+
+static inline void vst1q_u16_x4(uint16_t *p, uint16x8x4_t a) {
+    vst1q_u16(p, a.val[0]);
+    vst1q_u16(p + 8, a.val[1]);
+    vst1q_u16(p + 16, a.val[2]);
+    vst1q_u16(p + 24, a.val[3]);
+}
+#  endif // HASLD4 check
+
+#  ifndef _MSC_VER
+#    define vld1_u8_ex(p, align) vld1_u8(HINT_ALIGNED((p), (align)/8))
+#    define vld1q_u8_ex(p, align) vld1q_u8(HINT_ALIGNED((p), (align)/8))
+#    define vld1q_u64_ex(p, align) vld1q_u64(HINT_ALIGNED((p), (align)/8))
+#  endif
+#  if !defined(_MSC_VER) || !defined(ARM_NEON_HASLD4)
+#    define vld1q_u8_x4_ex(p, align) vld1q_u8_x4(HINT_ALIGNED((p), (align)/8))
+#    define vld1q_u16_x4_ex(p, align) vld1q_u16_x4(HINT_ALIGNED((p), (align)/8))
+#    define vst1q_u16_x4_ex(p, a, align) vst1q_u16_x4(HINT_ALIGNED((p), (align)/8), a)
+#  endif
+
+#endif
+
+#endif // include guard ARM_NEON_INTRINS_H