diff --git a/src/ds/defines.h b/src/ds/defines.h
index b0aab14..3019493 100644
--- a/src/ds/defines.h
+++ b/src/ds/defines.h
@@ -1,12 +1,20 @@
 #pragma once
 
 #if defined(_MSC_VER) && !defined(__clang__)
+// 28 is FAST_FAIL_INVALID_BUFFER_ACCESS.  Not using the symbolic constant to
+// avoid depending on winnt.h
+#  define SNMALLOC_FAST_FAIL() __fastfail(28)
 #  define ALWAYSINLINE __forceinline
 #  define NOINLINE __declspec(noinline)
 #  define likely(x) !!(x)
 #  define unlikely(x) !!(x)
 #  define SNMALLOC_SLOW_PATH NOINLINE
 #  define SNMALLOC_FAST_PATH ALWAYSINLINE
+/**
+ * Fast path with inline linkage.  MSVC assumes that `__forceinline` implies
+ * `inline` and complains if you specify `SNMALLOC_FAST_PATH` and `inline`.
+ */
+#  define SNMALLOC_FAST_PATH_INLINE ALWAYSINLINE
 #  if _MSC_VER >= 1927 && !defined(SNMALLOC_USE_CXX17)
 #    define SNMALLOC_FAST_PATH_LAMBDA [[msvc::forceinline]]
 #  else
@@ -17,12 +25,23 @@
 #  define SNMALLOC_REQUIRE_CONSTINIT
 #  define SNMALLOC_UNUSED_FUNCTION
 #else
+#  define SNMALLOC_FAST_FAIL() __builtin_trap()
 #  define likely(x) __builtin_expect(!!(x), 1)
 #  define unlikely(x) __builtin_expect(!!(x), 0)
 #  define ALWAYSINLINE __attribute__((always_inline))
 #  define NOINLINE __attribute__((noinline))
 #  define SNMALLOC_SLOW_PATH NOINLINE
 #  define SNMALLOC_FAST_PATH ALWAYSINLINE
+/**
+ * Fast path with inline linkage.  GCC assumes that
+ * `__attribute__((always_inline))` is orthogonal to `inline` and complains if
+ * you specify `SNMALLOC_FAST_PATH` and don't specify `inline` in places where
+ * `inline` would be required for the one definition rule.  The error message
+ * in this case is confusing: always-inline function may not be inlined.  If
+ * you see this error message when using `SNMALLOC_FAST_PATH` then switch to
+ * `SNMALLOC_FAST_PATH_INLINE`.
+ */
+#  define SNMALLOC_FAST_PATH_INLINE ALWAYSINLINE inline
 #  define SNMALLOC_FAST_PATH_LAMBDA SNMALLOC_FAST_PATH
 #  define SNMALLOC_PURE __attribute__((const))
 #  define SNMALLOC_COLD __attribute__((cold))
diff --git a/src/override/memcpy.cc b/src/override/memcpy.cc
index 4d41bf3..82bbf9a 100644
--- a/src/override/memcpy.cc
+++ b/src/override/memcpy.cc
@@ -16,7 +16,7 @@ using namespace snmalloc;
 // Windows has it with an underscore prefix
 #elif defined(_MSC_VER)
 #  define snprintf_l(buf, size, loc, msg, ...) \
-    _snprintf_l(buf, size, msg, loc, __VA_ARGS__)
+    _snprintf_s_l(buf, size, _TRUNCATE, msg, loc, __VA_ARGS__)
 #endif
 
 namespace
@@ -79,7 +79,7 @@ namespace
    * expands to a single load and store.
    */
   template<size_t Size>
-  SNMALLOC_FAST_PATH inline void copy_one(void* dst, const void* src)
+  SNMALLOC_FAST_PATH_INLINE void copy_one(void* dst, const void* src)
   {
 #if __has_builtin(__builtin_memcpy_inline)
     __builtin_memcpy_inline(dst, src, Size);
@@ -128,7 +128,7 @@ namespace
    * function is a no-op when `CheckReads` is false.
    */
   template<bool IsRead = false>
-  SNMALLOC_FAST_PATH inline void
+  SNMALLOC_FAST_PATH_INLINE void
   check_bounds(const void* ptr, size_t len, const char* msg = "")
   {
     if constexpr (!IsRead || CheckReads)
@@ -144,7 +144,7 @@ namespace
           UNUSED(ptr);
           UNUSED(len);
           UNUSED(msg);
-          __builtin_trap();
+          SNMALLOC_FAST_FAIL();
         }
         else
         {
@@ -165,7 +165,7 @@ namespace
    * chunks of size `Size` as are possible from `len`.
    */
   template<size_t Size>
-  SNMALLOC_FAST_PATH inline void
+  SNMALLOC_FAST_PATH_INLINE void
   block_copy(void* dst, const void* src, size_t len)
   {
     for (size_t i = 0; (i + Size) <= len; i += Size)
@@ -180,7 +180,7 @@ namespace
    * This may overlap other bits of the copy.
    */
   template<size_t Size>
-  SNMALLOC_FAST_PATH inline void
+  SNMALLOC_FAST_PATH_INLINE void
   copy_end(void* dst, const void* src, size_t len)
   {
     copy_one<Size>(