diff --git a/include/SDL3/SDL_stdinc.h b/include/SDL3/SDL_stdinc.h
index e8f41543a0..8e8a0f8888 100644
--- a/include/SDL3/SDL_stdinc.h
+++ b/include/SDL3/SDL_stdinc.h
@@ -2600,6 +2600,26 @@ extern SDL_DECLSPEC void * SDLCALL SDL_memmove(SDL_OUT_BYTECAP(len) void *dst, S
  */
 extern SDL_DECLSPEC void * SDLCALL SDL_memset(SDL_OUT_BYTECAP(len) void *dst, int c, size_t len);
 
+/**
+ * Initialize all 16-bit words of buffer of memory to a specific value.
+ *
+ * This function will set a buffer of `words` Uint16 values, pointed to by
+ * `dst`, to the value specified in `val`.
+ *
+ * Unlike SDL_memset, this sets 16-bit values, not bytes, so it's not limited
+ * to a range of 0-255.
+ *
+ * \param dst the destination memory region. Must not be NULL.
+ * \param val the Uint16 value to set.
+ * \param words the number of Uint16 values to set in `dst`.
+ * \returns `dst`.
+ *
+ * \threadsafety It is safe to call this function from any thread.
+ *
+ * \since This function is available since SDL 3.6.0.
+ */
+extern SDL_DECLSPEC void * SDLCALL SDL_memset2(void *dst, Uint16 val, size_t words);
+
 /**
  * Initialize all 32-bit words of buffer of memory to a specific value.
  *
diff --git a/src/dynapi/SDL_dynapi.exports b/src/dynapi/SDL_dynapi.exports
index 9864557071..b9fc41795a 100644
--- a/src/dynapi/SDL_dynapi.exports
+++ b/src/dynapi/SDL_dynapi.exports
@@ -1290,3 +1290,4 @@ _SDL_LoadJPG
 _SDL_HasSVE2
 _SDL_GamepadHasCapSense
 _SDL_GetGamepadCapSense
+_SDL_memset2
diff --git a/src/dynapi/SDL_dynapi.sym b/src/dynapi/SDL_dynapi.sym
index 3958a52aa6..9c791f39ef 100644
--- a/src/dynapi/SDL_dynapi.sym
+++ b/src/dynapi/SDL_dynapi.sym
@@ -1291,6 +1291,7 @@ SDL3_0.0.0 {
     SDL_HasSVE2;
     SDL_GamepadHasCapSense;
     SDL_GetGamepadCapSense;
+    SDL_memset2;
     # extra symbols go here (don't modify this line)
   local: *;
 };
diff --git a/src/dynapi/SDL_dynapi_overrides.h b/src/dynapi/SDL_dynapi_overrides.h
index b54d32ae6d..6ae05c394f 100644
--- a/src/dynapi/SDL_dynapi_overrides.h
+++ b/src/dynapi/SDL_dynapi_overrides.h
@@ -1317,3 +1317,4 @@
 #define SDL_HasSVE2 SDL_HasSVE2_REAL
 #define SDL_GamepadHasCapSense SDL_GamepadHasCapSense_REAL
 #define SDL_GetGamepadCapSense SDL_GetGamepadCapSense_REAL
+#define SDL_memset2 SDL_memset2_REAL
diff --git a/src/dynapi/SDL_dynapi_procs.h b/src/dynapi/SDL_dynapi_procs.h
index 4f8ac0ba0c..95092e7049 100644
--- a/src/dynapi/SDL_dynapi_procs.h
+++ b/src/dynapi/SDL_dynapi_procs.h
@@ -1325,3 +1325,4 @@ SDL_DYNAPI_PROC(SDL_Surface*,SDL_LoadJPG,(const char *a),(a),return)
 SDL_DYNAPI_PROC(bool,SDL_HasSVE2,(void),(),return)
 SDL_DYNAPI_PROC(bool,SDL_GamepadHasCapSense,(SDL_Gamepad *a,SDL_GamepadCapSenseType b),(a,b),return)
 SDL_DYNAPI_PROC(bool,SDL_GetGamepadCapSense,(SDL_Gamepad *a,SDL_GamepadCapSenseType b),(a,b),return)
+SDL_DYNAPI_PROC(void*,SDL_memset2,(void *a,Uint16 b,size_t c),(a,b,c),return)
diff --git a/src/render/software/SDL_drawline.c b/src/render/software/SDL_drawline.c
index 057af0186e..006c6fbf6c 100644
--- a/src/render/software/SDL_drawline.c
+++ b/src/render/software/SDL_drawline.c
@@ -31,15 +31,14 @@ static void SDL_DrawLine1(SDL_Surface *dst, int x1, int y1, int x2, int y2, Uint
 {
     if (y1 == y2) {
         int length;
-        int pitch = (dst->pitch / dst->fmt->bytes_per_pixel);
         Uint8 *pixels;
         if (x1 <= x2) {
-            pixels = (Uint8 *)dst->pixels + y1 * pitch + x1;
+            pixels = (Uint8 *)dst->pixels + y1 * dst->pitch + x1 * sizeof(Uint8);
             length = draw_end ? (x2 - x1 + 1) : (x2 - x1);
         } else {
-            pixels = (Uint8 *)dst->pixels + y1 * pitch + x2;
+            pixels = (Uint8 *)dst->pixels + y1 * dst->pitch + x2 * sizeof(Uint8);
             if (!draw_end) {
-                ++pixels;
+                pixels += sizeof(Uint8);
             }
             length = draw_end ? (x1 - x2 + 1) : (x1 - x2);
         }
@@ -57,7 +56,19 @@ static void SDL_DrawLine2(SDL_Surface *dst, int x1, int y1, int x2, int y2, Uint
                           bool draw_end)
 {
     if (y1 == y2) {
-        HLINE(Uint16, DRAW_FASTSETPIXEL2, draw_end);
+        int length;
+        Uint8 *pixels;
+        if (x1 <= x2) {
+            pixels = (Uint8 *)dst->pixels + y1 * dst->pitch + x1 * sizeof(Uint16);
+            length = draw_end ? (x2 - x1 + 1) : (x2 - x1);
+        } else {
+            pixels = (Uint8 *)dst->pixels + y1 * dst->pitch + x2 * sizeof(Uint16);
+            if (!draw_end) {
+                pixels += sizeof(Uint16);
+            }
+            length = draw_end ? (x1 - x2 + 1) : (x1 - x2);
+        }
+        SDL_memset2(pixels, color, length);
     } else if (x1 == x2) {
         VLINE(Uint16, DRAW_FASTSETPIXEL2, draw_end);
     } else if (ABS(x1 - x2) == ABS(y1 - y2)) {
@@ -86,7 +97,19 @@ static void SDL_DrawLine4(SDL_Surface *dst, int x1, int y1, int x2, int y2, Uint
                           bool draw_end)
 {
     if (y1 == y2) {
-        HLINE(Uint32, DRAW_FASTSETPIXEL4, draw_end);
+        int length;
+        Uint8 *pixels;
+        if (x1 <= x2) {
+            pixels = (Uint8 *)dst->pixels + y1 * dst->pitch + x1 * sizeof(Uint32);
+            length = draw_end ? (x2 - x1 + 1) : (x2 - x1);
+        } else {
+            pixels = (Uint8 *)dst->pixels + y1 * dst->pitch + x2 * sizeof(Uint32);
+            if (!draw_end) {
+                pixels += sizeof(Uint32);
+            }
+            length = draw_end ? (x1 - x2 + 1) : (x1 - x2);
+        }
+        SDL_memset4(pixels, color, length);
     } else if (x1 == x2) {
         VLINE(Uint32, DRAW_FASTSETPIXEL4, draw_end);
     } else if (ABS(x1 - x2) == ABS(y1 - y2)) {
diff --git a/src/stdlib/SDL_memset.c b/src/stdlib/SDL_memset.c
index d12df67cf1..c459d3ba41 100644
--- a/src/stdlib/SDL_memset.c
+++ b/src/stdlib/SDL_memset.c
@@ -59,9 +59,7 @@ void *SDL_memset(SDL_OUT_BYTECAP(len) void *dst, int c, size_t len)
     dstp4 = (Uint32 *)dstp1;
     left = (len % 4);
     len /= 4;
-    while (len--) {
-        *dstp4++ = value4;
-    }
+    SDL_memset4(dstp4, value4, len);
 
     dstp1 = (Uint8 *)dstp4;
     switch (left) {
@@ -79,6 +77,28 @@ void *SDL_memset(SDL_OUT_BYTECAP(len) void *dst, int c, size_t len)
 #endif // HAVE_MEMSET
 }
 
+// Note that memset() is a byte assignment and this is a 16-bit assignment, so they're not directly equivalent.
+void *SDL_memset2(void *dst, Uint16 val, size_t words)
+{
+    size_t _n = words;
+    Uint16 *_p = SDL_static_cast(Uint16 *, dst);
+    Uint16 _val = (val);
+
+    if (_n > 1) {
+        if ((uintptr_t)_p & 2) {
+            *_p++ = _val;
+            --_n;
+        }
+        SDL_memset4(_p, (_val << 16) | _val, (_n >> 1));
+    }
+
+    if (_n & 1) {
+        _p[_n - 1] = _val;
+    }
+
+    return dst;
+}
+
 // Note that memset() is a byte assignment and this is a 32-bit assignment, so they're not directly equivalent.
 void *SDL_memset4(void *dst, Uint32 val, size_t dwords)
 {
diff --git a/src/video/SDL_fillrect.c b/src/video/SDL_fillrect.c
index cd42b56ed6..e07272d155 100644
--- a/src/video/SDL_fillrect.c
+++ b/src/video/SDL_fillrect.c
@@ -198,65 +198,16 @@ DEFINE_LSX_FILLRECT(4, Uint32)
 
 static void SDL_FillSurfaceRect1(Uint8 *pixels, int pitch, Uint32 color, int w, int h)
 {
-    int n;
-    Uint8 *p = NULL;
-
     while (h--) {
-        n = w;
-        p = pixels;
-
-        if (n > 3) {
-            switch ((uintptr_t)p & 3) {
-            case 1:
-                *p++ = (Uint8)color;
-                --n;
-                SDL_FALLTHROUGH;
-            case 2:
-                *p++ = (Uint8)color;
-                --n;
-                SDL_FALLTHROUGH;
-            case 3:
-                *p++ = (Uint8)color;
-                --n;
-            }
-            SDL_memset4(p, color, (n >> 2));
-        }
-        if (n & 3) {
-            p += (n & ~3);
-            switch (n & 3) {
-            case 3:
-                *p++ = (Uint8)color;
-                SDL_FALLTHROUGH;
-            case 2:
-                *p++ = (Uint8)color;
-                SDL_FALLTHROUGH;
-            case 1:
-                *p++ = (Uint8)color;
-            }
-        }
+        SDL_memset(pixels, color, w);
         pixels += pitch;
     }
 }
 
 static void SDL_FillSurfaceRect2(Uint8 *pixels, int pitch, Uint32 color, int w, int h)
 {
-    int n;
-    Uint16 *p = NULL;
-
     while (h--) {
-        n = w;
-        p = (Uint16 *)pixels;
-
-        if (n > 1) {
-            if ((uintptr_t)p & 2) {
-                *p++ = (Uint16)color;
-                --n;
-            }
-            SDL_memset4(p, color, (n >> 1));
-        }
-        if (n & 1) {
-            p[n - 1] = (Uint16)color;
-        }
+        SDL_memset2(pixels, color, w);
         pixels += pitch;
     }
 }