From 0ee67124999e876128361154ea218e9d426a4657 Mon Sep 17 00:00:00 2001
From: Stenzek <stenzek@gmail.com>
Date: Sun, 1 Oct 2023 15:03:56 +1000
Subject: [PATCH] Common: Add MemsetPtrs()

---
 src/common/intrin.h | 41 +++++++++++++++++++++++++++++++++++++++++
 1 file changed, 41 insertions(+)

diff --git a/src/common/intrin.h b/src/common/intrin.h
index 61595d297..2c730f058 100644
--- a/src/common/intrin.h
+++ b/src/common/intrin.h
@@ -7,6 +7,8 @@
 
 #include "types.h"
 
+#include <type_traits>
+
 #if defined(CPU_ARCH_X86) || defined(CPU_ARCH_X64)
 #define CPU_ARCH_SSE 1
 #include <emmintrin.h>
@@ -24,3 +26,42 @@
 #else
 #include <malloc.h> // alloca
 #endif
+
+template<typename T>
+static inline void MemsetPtrs(T* ptr, T value, u32 count)
+{
+  static_assert(std::is_pointer_v<T>, "T is pointer type");
+  static_assert(sizeof(T) == sizeof(void*), "T isn't a fat pointer");
+  T* dest = ptr;
+
+#if defined(CPU_ARCH_SSE) || defined(CPU_ARCH_NEON)
+  static constexpr u32 PTRS_PER_VECTOR = (16 / sizeof(T));
+  const u32 aligned_count = count / PTRS_PER_VECTOR;
+  const u32 remaining_count = count % PTRS_PER_VECTOR;
+
+#if defined(CPU_ARCH_SSE)
+  const __m128i svalue = _mm_set1_epi64x(reinterpret_cast<intptr_t>(value));
+#elif defined(CPU_ARCH_NEON)
+  const uint64x2_t svalue = vdupq_n_u64(reinterpret_cast<uintptr_t>(value));
+#endif
+
+  // Clang gets way too eager and tries to unroll these, emitting thousands of instructions.
+#ifdef __clang__
+#pragma clang loop unroll(disable)
+#endif
+  for (u32 i = 0; i < aligned_count; i++)
+  {
+#if defined(CPU_ARCH_SSE)
+    _mm_store_si128(reinterpret_cast<__m128i*>(dest), svalue);
+#elif defined(CPU_ARCH_NEON)
+    vst1q_u64(reinterpret_cast<u64*>(dest), svalue);
+#endif
+    dest += PTRS_PER_VECTOR;
+  }
+#else
+  const u32 remaining_count = count;
+#endif
+
+  for (u32 i = 0; i < remaining_count; i++)
+    *(dest++) = value;
+}