diff --git a/src/video/SDL_blit_auto.c b/src/video/SDL_blit_auto.c
index 47511148f..efd561064 100644
--- a/src/video/SDL_blit_auto.c
+++ b/src/video/SDL_blit_auto.c
@@ -69,7 +69,7 @@ static void SDL_Blit_RGB888_RGB888_Blend(SDL_BlitInfo *info)
 {
     const int flags = info->flags;
     Uint32 srcpixel;
-    Uint32 srcR, srcG, srcB, srcA;
+    Uint32 srcR, srcG, srcB;
     Uint32 dstpixel;
     Uint32 dstR, dstG, dstB;
 
@@ -79,22 +79,14 @@ static void SDL_Blit_RGB888_RGB888_Blend(SDL_BlitInfo *info)
         int n = info->dst_w;
         while (n--) {
             srcpixel = *src;
-            srcR = (Uint8)(srcpixel >> 16); srcG = (Uint8)(srcpixel >> 8); srcB = (Uint8)srcpixel; srcA = 0xFF;
+            srcR = (Uint8)(srcpixel >> 16); srcG = (Uint8)(srcpixel >> 8); srcB = (Uint8)srcpixel;
             dstpixel = *dst;
             dstR = (Uint8)(dstpixel >> 16); dstG = (Uint8)(dstpixel >> 8); dstB = (Uint8)dstpixel;
-            if (flags & (SDL_COPY_BLEND|SDL_COPY_ADD)) {
-                /* This goes away if we ever use premultiplied alpha */
-                if (srcA < 255) {
-                    srcR = (srcR * srcA) / 255;
-                    srcG = (srcG * srcA) / 255;
-                    srcB = (srcB * srcA) / 255;
-                }
-            }
             switch (flags & (SDL_COPY_BLEND|SDL_COPY_ADD|SDL_COPY_MOD)) {
             case SDL_COPY_BLEND:
-                dstR = srcR + ((255 - srcA) * dstR) / 255;
-                dstG = srcG + ((255 - srcA) * dstG) / 255;
-                dstB = srcB + ((255 - srcA) * dstB) / 255;
+                dstR = srcR;
+                dstG = srcG;
+                dstB = srcB;
                 break;
             case SDL_COPY_ADD:
                 dstR = srcR + dstR; if (dstR > 255) dstR = 255;
@@ -121,7 +113,7 @@ static void SDL_Blit_RGB888_RGB888_Blend_Scale(SDL_BlitInfo *info)
 {
     const int flags = info->flags;
     Uint32 srcpixel;
-    Uint32 srcR, srcG, srcB, srcA;
+    Uint32 srcR, srcG, srcB;
     Uint32 dstpixel;
     Uint32 dstR, dstG, dstB;
     int srcy, srcx;
@@ -152,22 +144,14 @@ static void SDL_Blit_RGB888_RGB888_Blend_Scale(SDL_BlitInfo *info)
                 src = (Uint32 *)(info->src + (srcy * info->src_pitch) + (srcx * 4));
             }
             srcpixel = *src;
-            srcR = (Uint8)(srcpixel >> 16); srcG = (Uint8)(srcpixel >> 8); srcB = (Uint8)srcpixel; srcA = 0xFF;
+            srcR = (Uint8)(srcpixel >> 16); srcG = (Uint8)(srcpixel >> 8); srcB = (Uint8)srcpixel;
             dstpixel = *dst;
             dstR = (Uint8)(dstpixel >> 16); dstG = (Uint8)(dstpixel >> 8); dstB = (Uint8)dstpixel;
-            if (flags & (SDL_COPY_BLEND|SDL_COPY_ADD)) {
-                /* This goes away if we ever use premultiplied alpha */
-                if (srcA < 255) {
-                    srcR = (srcR * srcA) / 255;
-                    srcG = (srcG * srcA) / 255;
-                    srcB = (srcB * srcA) / 255;
-                }
-            }
             switch (flags & (SDL_COPY_BLEND|SDL_COPY_ADD|SDL_COPY_MOD)) {
             case SDL_COPY_BLEND:
-                dstR = srcR + ((255 - srcA) * dstR) / 255;
-                dstG = srcG + ((255 - srcA) * dstG) / 255;
-                dstB = srcB + ((255 - srcA) * dstB) / 255;
+                dstR = srcR;
+                dstG = srcG;
+                dstB = srcB;
                 break;
             case SDL_COPY_ADD:
                 dstR = srcR + dstR; if (dstR > 255) dstR = 255;
@@ -281,7 +265,8 @@ static void SDL_Blit_RGB888_RGB888_Modulate_Blend(SDL_BlitInfo *info)
     const Uint32 modulateB = info->b;
     const Uint32 modulateA = info->a;
     Uint32 srcpixel;
-    Uint32 srcR, srcG, srcB, srcA;
+    const Uint32 srcA = (flags & SDL_COPY_MODULATE_ALPHA) ? modulateA : 0xFF;
+    Uint32 srcR, srcG, srcB;
     Uint32 dstpixel;
     Uint32 dstR, dstG, dstB;
 
@@ -291,7 +276,7 @@ static void SDL_Blit_RGB888_RGB888_Modulate_Blend(SDL_BlitInfo *info)
         int n = info->dst_w;
         while (n--) {
             srcpixel = *src;
-            srcR = (Uint8)(srcpixel >> 16); srcG = (Uint8)(srcpixel >> 8); srcB = (Uint8)srcpixel; srcA = 0xFF;
+            srcR = (Uint8)(srcpixel >> 16); srcG = (Uint8)(srcpixel >> 8); srcB = (Uint8)srcpixel;
             dstpixel = *dst;
             dstR = (Uint8)(dstpixel >> 16); dstG = (Uint8)(dstpixel >> 8); dstB = (Uint8)dstpixel;
             if (flags & SDL_COPY_MODULATE_COLOR) {
@@ -299,9 +284,6 @@ static void SDL_Blit_RGB888_RGB888_Modulate_Blend(SDL_BlitInfo *info)
                 srcG = (srcG * modulateG) / 255;
                 srcB = (srcB * modulateB) / 255;
             }
-            if (flags & SDL_COPY_MODULATE_ALPHA) {
-                srcA = (srcA * modulateA) / 255;
-            }
             if (flags & (SDL_COPY_BLEND|SDL_COPY_ADD)) {
                 /* This goes away if we ever use premultiplied alpha */
                 if (srcA < 255) {
@@ -345,7 +327,8 @@ static void SDL_Blit_RGB888_RGB888_Modulate_Blend_Scale(SDL_BlitInfo *info)
     const Uint32 modulateB = info->b;
     const Uint32 modulateA = info->a;
     Uint32 srcpixel;
-    Uint32 srcR, srcG, srcB, srcA;
+    const Uint32 srcA = (flags & SDL_COPY_MODULATE_ALPHA) ? modulateA : 0xFF;
+    Uint32 srcR, srcG, srcB;
     Uint32 dstpixel;
     Uint32 dstR, dstG, dstB;
     int srcy, srcx;
@@ -376,7 +359,7 @@ static void SDL_Blit_RGB888_RGB888_Modulate_Blend_Scale(SDL_BlitInfo *info)
                 src = (Uint32 *)(info->src + (srcy * info->src_pitch) + (srcx * 4));
             }
             srcpixel = *src;
-            srcR = (Uint8)(srcpixel >> 16); srcG = (Uint8)(srcpixel >> 8); srcB = (Uint8)srcpixel; srcA = 0xFF;
+            srcR = (Uint8)(srcpixel >> 16); srcG = (Uint8)(srcpixel >> 8); srcB = (Uint8)srcpixel;
             dstpixel = *dst;
             dstR = (Uint8)(dstpixel >> 16); dstG = (Uint8)(dstpixel >> 8); dstB = (Uint8)dstpixel;
             if (flags & SDL_COPY_MODULATE_COLOR) {
@@ -384,9 +367,6 @@ static void SDL_Blit_RGB888_RGB888_Modulate_Blend_Scale(SDL_BlitInfo *info)
                 srcG = (srcG * modulateG) / 255;
                 srcB = (srcB * modulateB) / 255;
             }
-            if (flags & SDL_COPY_MODULATE_ALPHA) {
-                srcA = (srcA * modulateA) / 255;
-            }
             if (flags & (SDL_COPY_BLEND|SDL_COPY_ADD)) {
                 /* This goes away if we ever use premultiplied alpha */
                 if (srcA < 255) {
@@ -469,7 +449,7 @@ static void SDL_Blit_RGB888_BGR888_Blend(SDL_BlitInfo *info)
 {
     const int flags = info->flags;
     Uint32 srcpixel;
-    Uint32 srcR, srcG, srcB, srcA;
+    Uint32 srcR, srcG, srcB;
     Uint32 dstpixel;
     Uint32 dstR, dstG, dstB;
 
@@ -479,22 +459,14 @@ static void SDL_Blit_RGB888_BGR888_Blend(SDL_BlitInfo *info)
         int n = info->dst_w;
         while (n--) {
             srcpixel = *src;
-            srcR = (Uint8)(srcpixel >> 16); srcG = (Uint8)(srcpixel >> 8); srcB = (Uint8)srcpixel; srcA = 0xFF;
+            srcR = (Uint8)(srcpixel >> 16); srcG = (Uint8)(srcpixel >> 8); srcB = (Uint8)srcpixel;
             dstpixel = *dst;
             dstB = (Uint8)(dstpixel >> 16); dstG = (Uint8)(dstpixel >> 8); dstR = (Uint8)dstpixel;
-            if (flags & (SDL_COPY_BLEND|SDL_COPY_ADD)) {
-                /* This goes away if we ever use premultiplied alpha */
-                if (srcA < 255) {
-                    srcR = (srcR * srcA) / 255;
-                    srcG = (srcG * srcA) / 255;
-                    srcB = (srcB * srcA) / 255;
-                }
-            }
             switch (flags & (SDL_COPY_BLEND|SDL_COPY_ADD|SDL_COPY_MOD)) {
             case SDL_COPY_BLEND:
-                dstR = srcR + ((255 - srcA) * dstR) / 255;
-                dstG = srcG + ((255 - srcA) * dstG) / 255;
-                dstB = srcB + ((255 - srcA) * dstB) / 255;
+                dstR = srcR;
+                dstG = srcG;
+                dstB = srcB;
                 break;
             case SDL_COPY_ADD:
                 dstR = srcR + dstR; if (dstR > 255) dstR = 255;
@@ -521,7 +493,7 @@ static void SDL_Blit_RGB888_BGR888_Blend_Scale(SDL_BlitInfo *info)
 {
     const int flags = info->flags;
     Uint32 srcpixel;
-    Uint32 srcR, srcG, srcB, srcA;
+    Uint32 srcR, srcG, srcB;
     Uint32 dstpixel;
     Uint32 dstR, dstG, dstB;
     int srcy, srcx;
@@ -552,22 +524,14 @@ static void SDL_Blit_RGB888_BGR888_Blend_Scale(SDL_BlitInfo *info)
                 src = (Uint32 *)(info->src + (srcy * info->src_pitch) + (srcx * 4));
             }
             srcpixel = *src;
-            srcR = (Uint8)(srcpixel >> 16); srcG = (Uint8)(srcpixel >> 8); srcB = (Uint8)srcpixel; srcA = 0xFF;
+            srcR = (Uint8)(srcpixel >> 16); srcG = (Uint8)(srcpixel >> 8); srcB = (Uint8)srcpixel;
             dstpixel = *dst;
             dstB = (Uint8)(dstpixel >> 16); dstG = (Uint8)(dstpixel >> 8); dstR = (Uint8)dstpixel;
-            if (flags & (SDL_COPY_BLEND|SDL_COPY_ADD)) {
-                /* This goes away if we ever use premultiplied alpha */
-                if (srcA < 255) {
-                    srcR = (srcR * srcA) / 255;
-                    srcG = (srcG * srcA) / 255;
-                    srcB = (srcB * srcA) / 255;
-                }
-            }
             switch (flags & (SDL_COPY_BLEND|SDL_COPY_ADD|SDL_COPY_MOD)) {
             case SDL_COPY_BLEND:
-                dstR = srcR + ((255 - srcA) * dstR) / 255;
-                dstG = srcG + ((255 - srcA) * dstG) / 255;
-                dstB = srcB + ((255 - srcA) * dstB) / 255;
+                dstR = srcR;
+                dstG = srcG;
+                dstB = srcB;
                 break;
             case SDL_COPY_ADD:
                 dstR = srcR + dstR; if (dstR > 255) dstR = 255;
@@ -681,7 +645,8 @@ static void SDL_Blit_RGB888_BGR888_Modulate_Blend(SDL_BlitInfo *info)
     const Uint32 modulateB = info->b;
     const Uint32 modulateA = info->a;
     Uint32 srcpixel;
-    Uint32 srcR, srcG, srcB, srcA;
+    const Uint32 srcA = (flags & SDL_COPY_MODULATE_ALPHA) ? modulateA : 0xFF;
+    Uint32 srcR, srcG, srcB;
     Uint32 dstpixel;
     Uint32 dstR, dstG, dstB;
 
@@ -691,7 +656,7 @@ static void SDL_Blit_RGB888_BGR888_Modulate_Blend(SDL_BlitInfo *info)
         int n = info->dst_w;
         while (n--) {
             srcpixel = *src;
-            srcR = (Uint8)(srcpixel >> 16); srcG = (Uint8)(srcpixel >> 8); srcB = (Uint8)srcpixel; srcA = 0xFF;
+            srcR = (Uint8)(srcpixel >> 16); srcG = (Uint8)(srcpixel >> 8); srcB = (Uint8)srcpixel;
             dstpixel = *dst;
             dstB = (Uint8)(dstpixel >> 16); dstG = (Uint8)(dstpixel >> 8); dstR = (Uint8)dstpixel;
             if (flags & SDL_COPY_MODULATE_COLOR) {
@@ -699,9 +664,6 @@ static void SDL_Blit_RGB888_BGR888_Modulate_Blend(SDL_BlitInfo *info)
                 srcG = (srcG * modulateG) / 255;
                 srcB = (srcB * modulateB) / 255;
             }
-            if (flags & SDL_COPY_MODULATE_ALPHA) {
-                srcA = (srcA * modulateA) / 255;
-            }
             if (flags & (SDL_COPY_BLEND|SDL_COPY_ADD)) {
                 /* This goes away if we ever use premultiplied alpha */
                 if (srcA < 255) {
@@ -745,7 +707,8 @@ static void SDL_Blit_RGB888_BGR888_Modulate_Blend_Scale(SDL_BlitInfo *info)
     const Uint32 modulateB = info->b;
     const Uint32 modulateA = info->a;
     Uint32 srcpixel;
-    Uint32 srcR, srcG, srcB, srcA;
+    const Uint32 srcA = (flags & SDL_COPY_MODULATE_ALPHA) ? modulateA : 0xFF;
+    Uint32 srcR, srcG, srcB;
     Uint32 dstpixel;
     Uint32 dstR, dstG, dstB;
     int srcy, srcx;
@@ -776,7 +739,7 @@ static void SDL_Blit_RGB888_BGR888_Modulate_Blend_Scale(SDL_BlitInfo *info)
                 src = (Uint32 *)(info->src + (srcy * info->src_pitch) + (srcx * 4));
             }
             srcpixel = *src;
-            srcR = (Uint8)(srcpixel >> 16); srcG = (Uint8)(srcpixel >> 8); srcB = (Uint8)srcpixel; srcA = 0xFF;
+            srcR = (Uint8)(srcpixel >> 16); srcG = (Uint8)(srcpixel >> 8); srcB = (Uint8)srcpixel;
             dstpixel = *dst;
             dstB = (Uint8)(dstpixel >> 16); dstG = (Uint8)(dstpixel >> 8); dstR = (Uint8)dstpixel;
             if (flags & SDL_COPY_MODULATE_COLOR) {
@@ -784,9 +747,6 @@ static void SDL_Blit_RGB888_BGR888_Modulate_Blend_Scale(SDL_BlitInfo *info)
                 srcG = (srcG * modulateG) / 255;
                 srcB = (srcB * modulateB) / 255;
             }
-            if (flags & SDL_COPY_MODULATE_ALPHA) {
-                srcA = (srcA * modulateA) / 255;
-            }
             if (flags & (SDL_COPY_BLEND|SDL_COPY_ADD)) {
                 /* This goes away if we ever use premultiplied alpha */
                 if (srcA < 255) {
@@ -825,7 +785,8 @@ static void SDL_Blit_RGB888_BGR888_Modulate_Blend_Scale(SDL_BlitInfo *info)
 static void SDL_Blit_RGB888_ARGB8888_Scale(SDL_BlitInfo *info)
 {
     Uint32 pixel;
-    Uint32 R, G, B, A;
+    const Uint32 A = 0xFF;
+    Uint32 R, G, B;
     int srcy, srcx;
     int posy, posx;
     int incy, incx;
@@ -854,7 +815,7 @@ static void SDL_Blit_RGB888_ARGB8888_Scale(SDL_BlitInfo *info)
                 src = (Uint32 *)(info->src + (srcy * info->src_pitch) + (srcx * 4));
             }
             pixel = *src;
-            R = (Uint8)(pixel >> 16); G = (Uint8)(pixel >> 8); B = (Uint8)pixel; A = 0xFF;
+            R = (Uint8)(pixel >> 16); G = (Uint8)(pixel >> 8); B = (Uint8)pixel;
             pixel = ((Uint32)A << 24) | ((Uint32)R << 16) | ((Uint32)G << 8) | B;
             *dst = pixel;
             posx += incx;
@@ -869,7 +830,7 @@ static void SDL_Blit_RGB888_ARGB8888_Blend(SDL_BlitInfo *info)
 {
     const int flags = info->flags;
     Uint32 srcpixel;
-    Uint32 srcR, srcG, srcB, srcA;
+    Uint32 srcR, srcG, srcB;
     Uint32 dstpixel;
     Uint32 dstR, dstG, dstB, dstA;
 
@@ -879,23 +840,15 @@ static void SDL_Blit_RGB888_ARGB8888_Blend(SDL_BlitInfo *info)
         int n = info->dst_w;
         while (n--) {
             srcpixel = *src;
-            srcR = (Uint8)(srcpixel >> 16); srcG = (Uint8)(srcpixel >> 8); srcB = (Uint8)srcpixel; srcA = 0xFF;
+            srcR = (Uint8)(srcpixel >> 16); srcG = (Uint8)(srcpixel >> 8); srcB = (Uint8)srcpixel;
             dstpixel = *dst;
             dstR = (Uint8)(dstpixel >> 16); dstG = (Uint8)(dstpixel >> 8); dstB = (Uint8)dstpixel; dstA = (Uint8)(dstpixel >> 24);
-            if (flags & (SDL_COPY_BLEND|SDL_COPY_ADD)) {
-                /* This goes away if we ever use premultiplied alpha */
-                if (srcA < 255) {
-                    srcR = (srcR * srcA) / 255;
-                    srcG = (srcG * srcA) / 255;
-                    srcB = (srcB * srcA) / 255;
-                }
-            }
             switch (flags & (SDL_COPY_BLEND|SDL_COPY_ADD|SDL_COPY_MOD)) {
             case SDL_COPY_BLEND:
-                dstR = srcR + ((255 - srcA) * dstR) / 255;
-                dstG = srcG + ((255 - srcA) * dstG) / 255;
-                dstB = srcB + ((255 - srcA) * dstB) / 255;
-                dstA = srcA + ((255 - srcA) * dstA) / 255;
+                dstR = srcR;
+                dstG = srcG;
+                dstB = srcB;
+                dstA = 0xFF;
                 break;
             case SDL_COPY_ADD:
                 dstR = srcR + dstR; if (dstR > 255) dstR = 255;
@@ -922,7 +875,7 @@ static void SDL_Blit_RGB888_ARGB8888_Blend_Scale(SDL_BlitInfo *info)
 {
     const int flags = info->flags;
     Uint32 srcpixel;
-    Uint32 srcR, srcG, srcB, srcA;
+    Uint32 srcR, srcG, srcB;
     Uint32 dstpixel;
     Uint32 dstR, dstG, dstB, dstA;
     int srcy, srcx;
@@ -953,23 +906,15 @@ static void SDL_Blit_RGB888_ARGB8888_Blend_Scale(SDL_BlitInfo *info)
                 src = (Uint32 *)(info->src + (srcy * info->src_pitch) + (srcx * 4));
             }
             srcpixel = *src;
-            srcR = (Uint8)(srcpixel >> 16); srcG = (Uint8)(srcpixel >> 8); srcB = (Uint8)srcpixel; srcA = 0xFF;
+            srcR = (Uint8)(srcpixel >> 16); srcG = (Uint8)(srcpixel >> 8); srcB = (Uint8)srcpixel;
             dstpixel = *dst;
             dstR = (Uint8)(dstpixel >> 16); dstG = (Uint8)(dstpixel >> 8); dstB = (Uint8)dstpixel; dstA = (Uint8)(dstpixel >> 24);
-            if (flags & (SDL_COPY_BLEND|SDL_COPY_ADD)) {
-                /* This goes away if we ever use premultiplied alpha */
-                if (srcA < 255) {
-                    srcR = (srcR * srcA) / 255;
-                    srcG = (srcG * srcA) / 255;
-                    srcB = (srcB * srcA) / 255;
-                }
-            }
             switch (flags & (SDL_COPY_BLEND|SDL_COPY_ADD|SDL_COPY_MOD)) {
             case SDL_COPY_BLEND:
-                dstR = srcR + ((255 - srcA) * dstR) / 255;
-                dstG = srcG + ((255 - srcA) * dstG) / 255;
-                dstB = srcB + ((255 - srcA) * dstB) / 255;
-                dstA = srcA + ((255 - srcA) * dstA) / 255;
+                dstR = srcR;
+                dstG = srcG;
+                dstB = srcB;
+                dstA = 0xFF;
                 break;
             case SDL_COPY_ADD:
                 dstR = srcR + dstR; if (dstR > 255) dstR = 255;
@@ -1000,7 +945,8 @@ static void SDL_Blit_RGB888_ARGB8888_Modulate(SDL_BlitInfo *info)
     const Uint32 modulateB = info->b;
     const Uint32 modulateA = info->a;
     Uint32 pixel;
-    Uint32 R, G, B, A;
+    const Uint32 A = (flags & SDL_COPY_MODULATE_ALPHA) ? modulateA : 0xFF;
+    Uint32 R, G, B;
 
     while (info->dst_h--) {
         Uint32 *src = (Uint32 *)info->src;
@@ -1008,15 +954,12 @@ static void SDL_Blit_RGB888_ARGB8888_Modulate(SDL_BlitInfo *info)
         int n = info->dst_w;
         while (n--) {
             pixel = *src;
-            R = (Uint8)(pixel >> 16); G = (Uint8)(pixel >> 8); B = (Uint8)pixel; A = 0xFF;
+            R = (Uint8)(pixel >> 16); G = (Uint8)(pixel >> 8); B = (Uint8)pixel;
             if (flags & SDL_COPY_MODULATE_COLOR) {
                 R = (R * modulateR) / 255;
                 G = (G * modulateG) / 255;
                 B = (B * modulateB) / 255;
             }
-            if (flags & SDL_COPY_MODULATE_ALPHA) {
-                A = (A * modulateA) / 255;
-            }
             pixel = ((Uint32)A << 24) | ((Uint32)R << 16) | ((Uint32)G << 8) | B;
             *dst = pixel;
             ++src;
@@ -1035,7 +978,8 @@ static void SDL_Blit_RGB888_ARGB8888_Modulate_Scale(SDL_BlitInfo *info)
     const Uint32 modulateB = info->b;
     const Uint32 modulateA = info->a;
     Uint32 pixel;
-    Uint32 R, G, B, A;
+    const Uint32 A = (flags & SDL_COPY_MODULATE_ALPHA) ? modulateA : 0xFF;
+    Uint32 R, G, B;
     int srcy, srcx;
     int posy, posx;
     int incy, incx;
@@ -1064,15 +1008,12 @@ static void SDL_Blit_RGB888_ARGB8888_Modulate_Scale(SDL_BlitInfo *info)
                 src = (Uint32 *)(info->src + (srcy * info->src_pitch) + (srcx * 4));
             }
             pixel = *src;
-            R = (Uint8)(pixel >> 16); G = (Uint8)(pixel >> 8); B = (Uint8)pixel; A = 0xFF;
+            R = (Uint8)(pixel >> 16); G = (Uint8)(pixel >> 8); B = (Uint8)pixel;
             if (flags & SDL_COPY_MODULATE_COLOR) {
                 R = (R * modulateR) / 255;
                 G = (G * modulateG) / 255;
                 B = (B * modulateB) / 255;
             }
-            if (flags & SDL_COPY_MODULATE_ALPHA) {
-                A = (A * modulateA) / 255;
-            }
             pixel = ((Uint32)A << 24) | ((Uint32)R << 16) | ((Uint32)G << 8) | B;
             *dst = pixel;
             posx += incx;
@@ -1091,7 +1032,8 @@ static void SDL_Blit_RGB888_ARGB8888_Modulate_Blend(SDL_BlitInfo *info)
     const Uint32 modulateB = info->b;
     const Uint32 modulateA = info->a;
     Uint32 srcpixel;
-    Uint32 srcR, srcG, srcB, srcA;
+    const Uint32 srcA = (flags & SDL_COPY_MODULATE_ALPHA) ? modulateA : 0xFF;
+    Uint32 srcR, srcG, srcB;
     Uint32 dstpixel;
     Uint32 dstR, dstG, dstB, dstA;
 
@@ -1101,7 +1043,7 @@ static void SDL_Blit_RGB888_ARGB8888_Modulate_Blend(SDL_BlitInfo *info)
         int n = info->dst_w;
         while (n--) {
             srcpixel = *src;
-            srcR = (Uint8)(srcpixel >> 16); srcG = (Uint8)(srcpixel >> 8); srcB = (Uint8)srcpixel; srcA = 0xFF;
+            srcR = (Uint8)(srcpixel >> 16); srcG = (Uint8)(srcpixel >> 8); srcB = (Uint8)srcpixel;
             dstpixel = *dst;
             dstR = (Uint8)(dstpixel >> 16); dstG = (Uint8)(dstpixel >> 8); dstB = (Uint8)dstpixel; dstA = (Uint8)(dstpixel >> 24);
             if (flags & SDL_COPY_MODULATE_COLOR) {
@@ -1109,9 +1051,6 @@ static void SDL_Blit_RGB888_ARGB8888_Modulate_Blend(SDL_BlitInfo *info)
                 srcG = (srcG * modulateG) / 255;
                 srcB = (srcB * modulateB) / 255;
             }
-            if (flags & SDL_COPY_MODULATE_ALPHA) {
-                srcA = (srcA * modulateA) / 255;
-            }
             if (flags & (SDL_COPY_BLEND|SDL_COPY_ADD)) {
                 /* This goes away if we ever use premultiplied alpha */
                 if (srcA < 255) {
@@ -1156,7 +1095,8 @@ static void SDL_Blit_RGB888_ARGB8888_Modulate_Blend_Scale(SDL_BlitInfo *info)
     const Uint32 modulateB = info->b;
     const Uint32 modulateA = info->a;
     Uint32 srcpixel;
-    Uint32 srcR, srcG, srcB, srcA;
+    const Uint32 srcA = (flags & SDL_COPY_MODULATE_ALPHA) ? modulateA : 0xFF;
+    Uint32 srcR, srcG, srcB;
     Uint32 dstpixel;
     Uint32 dstR, dstG, dstB, dstA;
     int srcy, srcx;
@@ -1187,7 +1127,7 @@ static void SDL_Blit_RGB888_ARGB8888_Modulate_Blend_Scale(SDL_BlitInfo *info)
                 src = (Uint32 *)(info->src + (srcy * info->src_pitch) + (srcx * 4));
             }
             srcpixel = *src;
-            srcR = (Uint8)(srcpixel >> 16); srcG = (Uint8)(srcpixel >> 8); srcB = (Uint8)srcpixel; srcA = 0xFF;
+            srcR = (Uint8)(srcpixel >> 16); srcG = (Uint8)(srcpixel >> 8); srcB = (Uint8)srcpixel;
             dstpixel = *dst;
             dstR = (Uint8)(dstpixel >> 16); dstG = (Uint8)(dstpixel >> 8); dstB = (Uint8)dstpixel; dstA = (Uint8)(dstpixel >> 24);
             if (flags & SDL_COPY_MODULATE_COLOR) {
@@ -1195,9 +1135,6 @@ static void SDL_Blit_RGB888_ARGB8888_Modulate_Blend_Scale(SDL_BlitInfo *info)
                 srcG = (srcG * modulateG) / 255;
                 srcB = (srcB * modulateB) / 255;
             }
-            if (flags & SDL_COPY_MODULATE_ALPHA) {
-                srcA = (srcA * modulateA) / 255;
-            }
             if (flags & (SDL_COPY_BLEND|SDL_COPY_ADD)) {
                 /* This goes away if we ever use premultiplied alpha */
                 if (srcA < 255) {
@@ -1281,7 +1218,7 @@ static void SDL_Blit_BGR888_RGB888_Blend(SDL_BlitInfo *info)
 {
     const int flags = info->flags;
     Uint32 srcpixel;
-    Uint32 srcR, srcG, srcB, srcA;
+    Uint32 srcR, srcG, srcB;
     Uint32 dstpixel;
     Uint32 dstR, dstG, dstB;
 
@@ -1291,22 +1228,14 @@ static void SDL_Blit_BGR888_RGB888_Blend(SDL_BlitInfo *info)
         int n = info->dst_w;
         while (n--) {
             srcpixel = *src;
-            srcB = (Uint8)(srcpixel >> 16); srcG = (Uint8)(srcpixel >> 8); srcR = (Uint8)srcpixel; srcA = 0xFF;
+            srcB = (Uint8)(srcpixel >> 16); srcG = (Uint8)(srcpixel >> 8); srcR = (Uint8)srcpixel;
             dstpixel = *dst;
             dstR = (Uint8)(dstpixel >> 16); dstG = (Uint8)(dstpixel >> 8); dstB = (Uint8)dstpixel;
-            if (flags & (SDL_COPY_BLEND|SDL_COPY_ADD)) {
-                /* This goes away if we ever use premultiplied alpha */
-                if (srcA < 255) {
-                    srcR = (srcR * srcA) / 255;
-                    srcG = (srcG * srcA) / 255;
-                    srcB = (srcB * srcA) / 255;
-                }
-            }
             switch (flags & (SDL_COPY_BLEND|SDL_COPY_ADD|SDL_COPY_MOD)) {
             case SDL_COPY_BLEND:
-                dstR = srcR + ((255 - srcA) * dstR) / 255;
-                dstG = srcG + ((255 - srcA) * dstG) / 255;
-                dstB = srcB + ((255 - srcA) * dstB) / 255;
+                dstR = srcR;
+                dstG = srcG;
+                dstB = srcB;
                 break;
             case SDL_COPY_ADD:
                 dstR = srcR + dstR; if (dstR > 255) dstR = 255;
@@ -1333,7 +1262,7 @@ static void SDL_Blit_BGR888_RGB888_Blend_Scale(SDL_BlitInfo *info)
 {
     const int flags = info->flags;
     Uint32 srcpixel;
-    Uint32 srcR, srcG, srcB, srcA;
+    Uint32 srcR, srcG, srcB;
     Uint32 dstpixel;
     Uint32 dstR, dstG, dstB;
     int srcy, srcx;
@@ -1364,22 +1293,14 @@ static void SDL_Blit_BGR888_RGB888_Blend_Scale(SDL_BlitInfo *info)
                 src = (Uint32 *)(info->src + (srcy * info->src_pitch) + (srcx * 4));
             }
             srcpixel = *src;
-            srcB = (Uint8)(srcpixel >> 16); srcG = (Uint8)(srcpixel >> 8); srcR = (Uint8)srcpixel; srcA = 0xFF;
+            srcB = (Uint8)(srcpixel >> 16); srcG = (Uint8)(srcpixel >> 8); srcR = (Uint8)srcpixel;
             dstpixel = *dst;
             dstR = (Uint8)(dstpixel >> 16); dstG = (Uint8)(dstpixel >> 8); dstB = (Uint8)dstpixel;
-            if (flags & (SDL_COPY_BLEND|SDL_COPY_ADD)) {
-                /* This goes away if we ever use premultiplied alpha */
-                if (srcA < 255) {
-                    srcR = (srcR * srcA) / 255;
-                    srcG = (srcG * srcA) / 255;
-                    srcB = (srcB * srcA) / 255;
-                }
-            }
             switch (flags & (SDL_COPY_BLEND|SDL_COPY_ADD|SDL_COPY_MOD)) {
             case SDL_COPY_BLEND:
-                dstR = srcR + ((255 - srcA) * dstR) / 255;
-                dstG = srcG + ((255 - srcA) * dstG) / 255;
-                dstB = srcB + ((255 - srcA) * dstB) / 255;
+                dstR = srcR;
+                dstG = srcG;
+                dstB = srcB;
                 break;
             case SDL_COPY_ADD:
                 dstR = srcR + dstR; if (dstR > 255) dstR = 255;
@@ -1493,7 +1414,8 @@ static void SDL_Blit_BGR888_RGB888_Modulate_Blend(SDL_BlitInfo *info)
     const Uint32 modulateB = info->b;
     const Uint32 modulateA = info->a;
     Uint32 srcpixel;
-    Uint32 srcR, srcG, srcB, srcA;
+    const Uint32 srcA = (flags & SDL_COPY_MODULATE_ALPHA) ? modulateA : 0xFF;
+    Uint32 srcR, srcG, srcB;
     Uint32 dstpixel;
     Uint32 dstR, dstG, dstB;
 
@@ -1503,7 +1425,7 @@ static void SDL_Blit_BGR888_RGB888_Modulate_Blend(SDL_BlitInfo *info)
         int n = info->dst_w;
         while (n--) {
             srcpixel = *src;
-            srcB = (Uint8)(srcpixel >> 16); srcG = (Uint8)(srcpixel >> 8); srcR = (Uint8)srcpixel; srcA = 0xFF;
+            srcB = (Uint8)(srcpixel >> 16); srcG = (Uint8)(srcpixel >> 8); srcR = (Uint8)srcpixel;
             dstpixel = *dst;
             dstR = (Uint8)(dstpixel >> 16); dstG = (Uint8)(dstpixel >> 8); dstB = (Uint8)dstpixel;
             if (flags & SDL_COPY_MODULATE_COLOR) {
@@ -1511,9 +1433,6 @@ static void SDL_Blit_BGR888_RGB888_Modulate_Blend(SDL_BlitInfo *info)
                 srcG = (srcG * modulateG) / 255;
                 srcB = (srcB * modulateB) / 255;
             }
-            if (flags & SDL_COPY_MODULATE_ALPHA) {
-                srcA = (srcA * modulateA) / 255;
-            }
             if (flags & (SDL_COPY_BLEND|SDL_COPY_ADD)) {
                 /* This goes away if we ever use premultiplied alpha */
                 if (srcA < 255) {
@@ -1557,7 +1476,8 @@ static void SDL_Blit_BGR888_RGB888_Modulate_Blend_Scale(SDL_BlitInfo *info)
     const Uint32 modulateB = info->b;
     const Uint32 modulateA = info->a;
     Uint32 srcpixel;
-    Uint32 srcR, srcG, srcB, srcA;
+    const Uint32 srcA = (flags & SDL_COPY_MODULATE_ALPHA) ? modulateA : 0xFF;
+    Uint32 srcR, srcG, srcB;
     Uint32 dstpixel;
     Uint32 dstR, dstG, dstB;
     int srcy, srcx;
@@ -1588,7 +1508,7 @@ static void SDL_Blit_BGR888_RGB888_Modulate_Blend_Scale(SDL_BlitInfo *info)
                 src = (Uint32 *)(info->src + (srcy * info->src_pitch) + (srcx * 4));
             }
             srcpixel = *src;
-            srcB = (Uint8)(srcpixel >> 16); srcG = (Uint8)(srcpixel >> 8); srcR = (Uint8)srcpixel; srcA = 0xFF;
+            srcB = (Uint8)(srcpixel >> 16); srcG = (Uint8)(srcpixel >> 8); srcR = (Uint8)srcpixel;
             dstpixel = *dst;
             dstR = (Uint8)(dstpixel >> 16); dstG = (Uint8)(dstpixel >> 8); dstB = (Uint8)dstpixel;
             if (flags & SDL_COPY_MODULATE_COLOR) {
@@ -1596,9 +1516,6 @@ static void SDL_Blit_BGR888_RGB888_Modulate_Blend_Scale(SDL_BlitInfo *info)
                 srcG = (srcG * modulateG) / 255;
                 srcB = (srcB * modulateB) / 255;
             }
-            if (flags & SDL_COPY_MODULATE_ALPHA) {
-                srcA = (srcA * modulateA) / 255;
-            }
             if (flags & (SDL_COPY_BLEND|SDL_COPY_ADD)) {
                 /* This goes away if we ever use premultiplied alpha */
                 if (srcA < 255) {
@@ -1676,7 +1593,7 @@ static void SDL_Blit_BGR888_BGR888_Blend(SDL_BlitInfo *info)
 {
     const int flags = info->flags;
     Uint32 srcpixel;
-    Uint32 srcR, srcG, srcB, srcA;
+    Uint32 srcR, srcG, srcB;
     Uint32 dstpixel;
     Uint32 dstR, dstG, dstB;
 
@@ -1686,22 +1603,14 @@ static void SDL_Blit_BGR888_BGR888_Blend(SDL_BlitInfo *info)
         int n = info->dst_w;
         while (n--) {
             srcpixel = *src;
-            srcB = (Uint8)(srcpixel >> 16); srcG = (Uint8)(srcpixel >> 8); srcR = (Uint8)srcpixel; srcA = 0xFF;
+            srcB = (Uint8)(srcpixel >> 16); srcG = (Uint8)(srcpixel >> 8); srcR = (Uint8)srcpixel;
             dstpixel = *dst;
             dstB = (Uint8)(dstpixel >> 16); dstG = (Uint8)(dstpixel >> 8); dstR = (Uint8)dstpixel;
-            if (flags & (SDL_COPY_BLEND|SDL_COPY_ADD)) {
-                /* This goes away if we ever use premultiplied alpha */
-                if (srcA < 255) {
-                    srcR = (srcR * srcA) / 255;
-                    srcG = (srcG * srcA) / 255;
-                    srcB = (srcB * srcA) / 255;
-                }
-            }
             switch (flags & (SDL_COPY_BLEND|SDL_COPY_ADD|SDL_COPY_MOD)) {
             case SDL_COPY_BLEND:
-                dstR = srcR + ((255 - srcA) * dstR) / 255;
-                dstG = srcG + ((255 - srcA) * dstG) / 255;
-                dstB = srcB + ((255 - srcA) * dstB) / 255;
+                dstR = srcR;
+                dstG = srcG;
+                dstB = srcB;
                 break;
             case SDL_COPY_ADD:
                 dstR = srcR + dstR; if (dstR > 255) dstR = 255;
@@ -1728,7 +1637,7 @@ static void SDL_Blit_BGR888_BGR888_Blend_Scale(SDL_BlitInfo *info)
 {
     const int flags = info->flags;
     Uint32 srcpixel;
-    Uint32 srcR, srcG, srcB, srcA;
+    Uint32 srcR, srcG, srcB;
     Uint32 dstpixel;
     Uint32 dstR, dstG, dstB;
     int srcy, srcx;
@@ -1759,22 +1668,14 @@ static void SDL_Blit_BGR888_BGR888_Blend_Scale(SDL_BlitInfo *info)
                 src = (Uint32 *)(info->src + (srcy * info->src_pitch) + (srcx * 4));
             }
             srcpixel = *src;
-            srcB = (Uint8)(srcpixel >> 16); srcG = (Uint8)(srcpixel >> 8); srcR = (Uint8)srcpixel; srcA = 0xFF;
+            srcB = (Uint8)(srcpixel >> 16); srcG = (Uint8)(srcpixel >> 8); srcR = (Uint8)srcpixel;
             dstpixel = *dst;
             dstB = (Uint8)(dstpixel >> 16); dstG = (Uint8)(dstpixel >> 8); dstR = (Uint8)dstpixel;
-            if (flags & (SDL_COPY_BLEND|SDL_COPY_ADD)) {
-                /* This goes away if we ever use premultiplied alpha */
-                if (srcA < 255) {
-                    srcR = (srcR * srcA) / 255;
-                    srcG = (srcG * srcA) / 255;
-                    srcB = (srcB * srcA) / 255;
-                }
-            }
             switch (flags & (SDL_COPY_BLEND|SDL_COPY_ADD|SDL_COPY_MOD)) {
             case SDL_COPY_BLEND:
-                dstR = srcR + ((255 - srcA) * dstR) / 255;
-                dstG = srcG + ((255 - srcA) * dstG) / 255;
-                dstB = srcB + ((255 - srcA) * dstB) / 255;
+                dstR = srcR;
+                dstG = srcG;
+                dstB = srcB;
                 break;
             case SDL_COPY_ADD:
                 dstR = srcR + dstR; if (dstR > 255) dstR = 255;
@@ -1888,7 +1789,8 @@ static void SDL_Blit_BGR888_BGR888_Modulate_Blend(SDL_BlitInfo *info)
     const Uint32 modulateB = info->b;
     const Uint32 modulateA = info->a;
     Uint32 srcpixel;
-    Uint32 srcR, srcG, srcB, srcA;
+    const Uint32 srcA = (flags & SDL_COPY_MODULATE_ALPHA) ? modulateA : 0xFF;
+    Uint32 srcR, srcG, srcB;
     Uint32 dstpixel;
     Uint32 dstR, dstG, dstB;
 
@@ -1898,7 +1800,7 @@ static void SDL_Blit_BGR888_BGR888_Modulate_Blend(SDL_BlitInfo *info)
         int n = info->dst_w;
         while (n--) {
             srcpixel = *src;
-            srcB = (Uint8)(srcpixel >> 16); srcG = (Uint8)(srcpixel >> 8); srcR = (Uint8)srcpixel; srcA = 0xFF;
+            srcB = (Uint8)(srcpixel >> 16); srcG = (Uint8)(srcpixel >> 8); srcR = (Uint8)srcpixel;
             dstpixel = *dst;
             dstB = (Uint8)(dstpixel >> 16); dstG = (Uint8)(dstpixel >> 8); dstR = (Uint8)dstpixel;
             if (flags & SDL_COPY_MODULATE_COLOR) {
@@ -1906,9 +1808,6 @@ static void SDL_Blit_BGR888_BGR888_Modulate_Blend(SDL_BlitInfo *info)
                 srcG = (srcG * modulateG) / 255;
                 srcB = (srcB * modulateB) / 255;
             }
-            if (flags & SDL_COPY_MODULATE_ALPHA) {
-                srcA = (srcA * modulateA) / 255;
-            }
             if (flags & (SDL_COPY_BLEND|SDL_COPY_ADD)) {
                 /* This goes away if we ever use premultiplied alpha */
                 if (srcA < 255) {
@@ -1952,7 +1851,8 @@ static void SDL_Blit_BGR888_BGR888_Modulate_Blend_Scale(SDL_BlitInfo *info)
     const Uint32 modulateB = info->b;
     const Uint32 modulateA = info->a;
     Uint32 srcpixel;
-    Uint32 srcR, srcG, srcB, srcA;
+    const Uint32 srcA = (flags & SDL_COPY_MODULATE_ALPHA) ? modulateA : 0xFF;
+    Uint32 srcR, srcG, srcB;
     Uint32 dstpixel;
     Uint32 dstR, dstG, dstB;
     int srcy, srcx;
@@ -1983,7 +1883,7 @@ static void SDL_Blit_BGR888_BGR888_Modulate_Blend_Scale(SDL_BlitInfo *info)
                 src = (Uint32 *)(info->src + (srcy * info->src_pitch) + (srcx * 4));
             }
             srcpixel = *src;
-            srcB = (Uint8)(srcpixel >> 16); srcG = (Uint8)(srcpixel >> 8); srcR = (Uint8)srcpixel; srcA = 0xFF;
+            srcB = (Uint8)(srcpixel >> 16); srcG = (Uint8)(srcpixel >> 8); srcR = (Uint8)srcpixel;
             dstpixel = *dst;
             dstB = (Uint8)(dstpixel >> 16); dstG = (Uint8)(dstpixel >> 8); dstR = (Uint8)dstpixel;
             if (flags & SDL_COPY_MODULATE_COLOR) {
@@ -1991,9 +1891,6 @@ static void SDL_Blit_BGR888_BGR888_Modulate_Blend_Scale(SDL_BlitInfo *info)
                 srcG = (srcG * modulateG) / 255;
                 srcB = (srcB * modulateB) / 255;
             }
-            if (flags & SDL_COPY_MODULATE_ALPHA) {
-                srcA = (srcA * modulateA) / 255;
-            }
             if (flags & (SDL_COPY_BLEND|SDL_COPY_ADD)) {
                 /* This goes away if we ever use premultiplied alpha */
                 if (srcA < 255) {
@@ -2032,7 +1929,8 @@ static void SDL_Blit_BGR888_BGR888_Modulate_Blend_Scale(SDL_BlitInfo *info)
 static void SDL_Blit_BGR888_ARGB8888_Scale(SDL_BlitInfo *info)
 {
     Uint32 pixel;
-    Uint32 R, G, B, A;
+    const Uint32 A = 0xFF;
+    Uint32 R, G, B;
     int srcy, srcx;
     int posy, posx;
     int incy, incx;
@@ -2061,7 +1959,7 @@ static void SDL_Blit_BGR888_ARGB8888_Scale(SDL_BlitInfo *info)
                 src = (Uint32 *)(info->src + (srcy * info->src_pitch) + (srcx * 4));
             }
             pixel = *src;
-            B = (Uint8)(pixel >> 16); G = (Uint8)(pixel >> 8); R = (Uint8)pixel; A = 0xFF;
+            B = (Uint8)(pixel >> 16); G = (Uint8)(pixel >> 8); R = (Uint8)pixel;
             pixel = ((Uint32)A << 24) | ((Uint32)R << 16) | ((Uint32)G << 8) | B;
             *dst = pixel;
             posx += incx;
@@ -2076,7 +1974,7 @@ static void SDL_Blit_BGR888_ARGB8888_Blend(SDL_BlitInfo *info)
 {
     const int flags = info->flags;
     Uint32 srcpixel;
-    Uint32 srcR, srcG, srcB, srcA;
+    Uint32 srcR, srcG, srcB;
     Uint32 dstpixel;
     Uint32 dstR, dstG, dstB, dstA;
 
@@ -2086,23 +1984,15 @@ static void SDL_Blit_BGR888_ARGB8888_Blend(SDL_BlitInfo *info)
         int n = info->dst_w;
         while (n--) {
             srcpixel = *src;
-            srcB = (Uint8)(srcpixel >> 16); srcG = (Uint8)(srcpixel >> 8); srcR = (Uint8)srcpixel; srcA = 0xFF;
+            srcB = (Uint8)(srcpixel >> 16); srcG = (Uint8)(srcpixel >> 8); srcR = (Uint8)srcpixel;
             dstpixel = *dst;
             dstR = (Uint8)(dstpixel >> 16); dstG = (Uint8)(dstpixel >> 8); dstB = (Uint8)dstpixel; dstA = (Uint8)(dstpixel >> 24);
-            if (flags & (SDL_COPY_BLEND|SDL_COPY_ADD)) {
-                /* This goes away if we ever use premultiplied alpha */
-                if (srcA < 255) {
-                    srcR = (srcR * srcA) / 255;
-                    srcG = (srcG * srcA) / 255;
-                    srcB = (srcB * srcA) / 255;
-                }
-            }
             switch (flags & (SDL_COPY_BLEND|SDL_COPY_ADD|SDL_COPY_MOD)) {
             case SDL_COPY_BLEND:
-                dstR = srcR + ((255 - srcA) * dstR) / 255;
-                dstG = srcG + ((255 - srcA) * dstG) / 255;
-                dstB = srcB + ((255 - srcA) * dstB) / 255;
-                dstA = srcA + ((255 - srcA) * dstA) / 255;
+                dstR = srcR;
+                dstG = srcG;
+                dstB = srcB;
+                dstA = 0xFF;
                 break;
             case SDL_COPY_ADD:
                 dstR = srcR + dstR; if (dstR > 255) dstR = 255;
@@ -2129,7 +2019,7 @@ static void SDL_Blit_BGR888_ARGB8888_Blend_Scale(SDL_BlitInfo *info)
 {
     const int flags = info->flags;
     Uint32 srcpixel;
-    Uint32 srcR, srcG, srcB, srcA;
+    Uint32 srcR, srcG, srcB;
     Uint32 dstpixel;
     Uint32 dstR, dstG, dstB, dstA;
     int srcy, srcx;
@@ -2160,23 +2050,15 @@ static void SDL_Blit_BGR888_ARGB8888_Blend_Scale(SDL_BlitInfo *info)
                 src = (Uint32 *)(info->src + (srcy * info->src_pitch) + (srcx * 4));
             }
             srcpixel = *src;
-            srcB = (Uint8)(srcpixel >> 16); srcG = (Uint8)(srcpixel >> 8); srcR = (Uint8)srcpixel; srcA = 0xFF;
+            srcB = (Uint8)(srcpixel >> 16); srcG = (Uint8)(srcpixel >> 8); srcR = (Uint8)srcpixel;
             dstpixel = *dst;
             dstR = (Uint8)(dstpixel >> 16); dstG = (Uint8)(dstpixel >> 8); dstB = (Uint8)dstpixel; dstA = (Uint8)(dstpixel >> 24);
-            if (flags & (SDL_COPY_BLEND|SDL_COPY_ADD)) {
-                /* This goes away if we ever use premultiplied alpha */
-                if (srcA < 255) {
-                    srcR = (srcR * srcA) / 255;
-                    srcG = (srcG * srcA) / 255;
-                    srcB = (srcB * srcA) / 255;
-                }
-            }
             switch (flags & (SDL_COPY_BLEND|SDL_COPY_ADD|SDL_COPY_MOD)) {
             case SDL_COPY_BLEND:
-                dstR = srcR + ((255 - srcA) * dstR) / 255;
-                dstG = srcG + ((255 - srcA) * dstG) / 255;
-                dstB = srcB + ((255 - srcA) * dstB) / 255;
-                dstA = srcA + ((255 - srcA) * dstA) / 255;
+                dstR = srcR;
+                dstG = srcG;
+                dstB = srcB;
+                dstA = 0xFF;
                 break;
             case SDL_COPY_ADD:
                 dstR = srcR + dstR; if (dstR > 255) dstR = 255;
@@ -2207,7 +2089,8 @@ static void SDL_Blit_BGR888_ARGB8888_Modulate(SDL_BlitInfo *info)
     const Uint32 modulateB = info->b;
     const Uint32 modulateA = info->a;
     Uint32 pixel;
-    Uint32 R, G, B, A;
+    const Uint32 A = (flags & SDL_COPY_MODULATE_ALPHA) ? modulateA : 0xFF;
+    Uint32 R, G, B;
 
     while (info->dst_h--) {
         Uint32 *src = (Uint32 *)info->src;
@@ -2215,15 +2098,12 @@ static void SDL_Blit_BGR888_ARGB8888_Modulate(SDL_BlitInfo *info)
         int n = info->dst_w;
         while (n--) {
             pixel = *src;
-            B = (Uint8)(pixel >> 16); G = (Uint8)(pixel >> 8); R = (Uint8)pixel; A = 0xFF;
+            B = (Uint8)(pixel >> 16); G = (Uint8)(pixel >> 8); R = (Uint8)pixel;
             if (flags & SDL_COPY_MODULATE_COLOR) {
                 R = (R * modulateR) / 255;
                 G = (G * modulateG) / 255;
                 B = (B * modulateB) / 255;
             }
-            if (flags & SDL_COPY_MODULATE_ALPHA) {
-                A = (A * modulateA) / 255;
-            }
             pixel = ((Uint32)A << 24) | ((Uint32)R << 16) | ((Uint32)G << 8) | B;
             *dst = pixel;
             ++src;
@@ -2242,7 +2122,8 @@ static void SDL_Blit_BGR888_ARGB8888_Modulate_Scale(SDL_BlitInfo *info)
     const Uint32 modulateB = info->b;
     const Uint32 modulateA = info->a;
     Uint32 pixel;
-    Uint32 R, G, B, A;
+    const Uint32 A = (flags & SDL_COPY_MODULATE_ALPHA) ? modulateA : 0xFF;
+    Uint32 R, G, B;
     int srcy, srcx;
     int posy, posx;
     int incy, incx;
@@ -2271,15 +2152,12 @@ static void SDL_Blit_BGR888_ARGB8888_Modulate_Scale(SDL_BlitInfo *info)
                 src = (Uint32 *)(info->src + (srcy * info->src_pitch) + (srcx * 4));
             }
             pixel = *src;
-            B = (Uint8)(pixel >> 16); G = (Uint8)(pixel >> 8); R = (Uint8)pixel; A = 0xFF;
+            B = (Uint8)(pixel >> 16); G = (Uint8)(pixel >> 8); R = (Uint8)pixel;
             if (flags & SDL_COPY_MODULATE_COLOR) {
                 R = (R * modulateR) / 255;
                 G = (G * modulateG) / 255;
                 B = (B * modulateB) / 255;
             }
-            if (flags & SDL_COPY_MODULATE_ALPHA) {
-                A = (A * modulateA) / 255;
-            }
             pixel = ((Uint32)A << 24) | ((Uint32)R << 16) | ((Uint32)G << 8) | B;
             *dst = pixel;
             posx += incx;
@@ -2298,7 +2176,8 @@ static void SDL_Blit_BGR888_ARGB8888_Modulate_Blend(SDL_BlitInfo *info)
     const Uint32 modulateB = info->b;
     const Uint32 modulateA = info->a;
     Uint32 srcpixel;
-    Uint32 srcR, srcG, srcB, srcA;
+    const Uint32 srcA = (flags & SDL_COPY_MODULATE_ALPHA) ? modulateA : 0xFF;
+    Uint32 srcR, srcG, srcB;
     Uint32 dstpixel;
     Uint32 dstR, dstG, dstB, dstA;
 
@@ -2308,7 +2187,7 @@ static void SDL_Blit_BGR888_ARGB8888_Modulate_Blend(SDL_BlitInfo *info)
         int n = info->dst_w;
         while (n--) {
             srcpixel = *src;
-            srcB = (Uint8)(srcpixel >> 16); srcG = (Uint8)(srcpixel >> 8); srcR = (Uint8)srcpixel; srcA = 0xFF;
+            srcB = (Uint8)(srcpixel >> 16); srcG = (Uint8)(srcpixel >> 8); srcR = (Uint8)srcpixel;
             dstpixel = *dst;
             dstR = (Uint8)(dstpixel >> 16); dstG = (Uint8)(dstpixel >> 8); dstB = (Uint8)dstpixel; dstA = (Uint8)(dstpixel >> 24);
             if (flags & SDL_COPY_MODULATE_COLOR) {
@@ -2316,9 +2195,6 @@ static void SDL_Blit_BGR888_ARGB8888_Modulate_Blend(SDL_BlitInfo *info)
                 srcG = (srcG * modulateG) / 255;
                 srcB = (srcB * modulateB) / 255;
             }
-            if (flags & SDL_COPY_MODULATE_ALPHA) {
-                srcA = (srcA * modulateA) / 255;
-            }
             if (flags & (SDL_COPY_BLEND|SDL_COPY_ADD)) {
                 /* This goes away if we ever use premultiplied alpha */
                 if (srcA < 255) {
@@ -2363,7 +2239,8 @@ static void SDL_Blit_BGR888_ARGB8888_Modulate_Blend_Scale(SDL_BlitInfo *info)
     const Uint32 modulateB = info->b;
     const Uint32 modulateA = info->a;
     Uint32 srcpixel;
-    Uint32 srcR, srcG, srcB, srcA;
+    const Uint32 srcA = (flags & SDL_COPY_MODULATE_ALPHA) ? modulateA : 0xFF;
+    Uint32 srcR, srcG, srcB;
     Uint32 dstpixel;
     Uint32 dstR, dstG, dstB, dstA;
     int srcy, srcx;
@@ -2394,7 +2271,7 @@ static void SDL_Blit_BGR888_ARGB8888_Modulate_Blend_Scale(SDL_BlitInfo *info)
                 src = (Uint32 *)(info->src + (srcy * info->src_pitch) + (srcx * 4));
             }
             srcpixel = *src;
-            srcB = (Uint8)(srcpixel >> 16); srcG = (Uint8)(srcpixel >> 8); srcR = (Uint8)srcpixel; srcA = 0xFF;
+            srcB = (Uint8)(srcpixel >> 16); srcG = (Uint8)(srcpixel >> 8); srcR = (Uint8)srcpixel;
             dstpixel = *dst;
             dstR = (Uint8)(dstpixel >> 16); dstG = (Uint8)(dstpixel >> 8); dstB = (Uint8)dstpixel; dstA = (Uint8)(dstpixel >> 24);
             if (flags & SDL_COPY_MODULATE_COLOR) {
@@ -2402,9 +2279,6 @@ static void SDL_Blit_BGR888_ARGB8888_Modulate_Blend_Scale(SDL_BlitInfo *info)
                 srcG = (srcG * modulateG) / 255;
                 srcB = (srcB * modulateB) / 255;
             }
-            if (flags & SDL_COPY_MODULATE_ALPHA) {
-                srcA = (srcA * modulateA) / 255;
-            }
             if (flags & (SDL_COPY_BLEND|SDL_COPY_ADD)) {
                 /* This goes away if we ever use premultiplied alpha */
                 if (srcA < 255) {
diff --git a/src/video/sdlgenblit.pl b/src/video/sdlgenblit.pl
index 9481e6d95..800bf9884 100755
--- a/src/video/sdlgenblit.pl
+++ b/src/video/sdlgenblit.pl
@@ -68,8 +68,8 @@ my %get_rgba_string_ignore_alpha = (
 );
 
 my %get_rgba_string = (
-    "RGB888" => $get_rgba_string_ignore_alpha{"RGB888"} . " _A = 0xFF;",
-    "BGR888" => $get_rgba_string_ignore_alpha{"BGR888"} . " _A = 0xFF;",
+    "RGB888" => $get_rgba_string_ignore_alpha{"RGB888"},
+    "BGR888" => $get_rgba_string_ignore_alpha{"BGR888"},
     "ARGB8888" => $get_rgba_string_ignore_alpha{"ARGB8888"} . " _A = (Uint8)(_pixel >> 24);",
     "RGBA8888" => $get_rgba_string_ignore_alpha{"RGBA8888"} . " _A = (Uint8)_pixel;",
     "ABGR8888" => $get_rgba_string_ignore_alpha{"ABGR8888"} . " _A = (Uint8)(_pixel >> 24);",
@@ -212,6 +212,8 @@ sub output_copycore
     my $dst = shift;
     my $modulate = shift;
     my $blend = shift;
+    my $is_modulateA_done = shift;
+    my $A_is_const_FF = shift;
     my $s = "";
     my $d = "";
 
@@ -243,7 +245,7 @@ __EOF__
                 ${s}B = (${s}B * modulateB) / 255;
             }
 __EOF__
-        if (not $ignore_dst_alpha) {
+        if (!$ignore_dst_alpha && !$is_modulateA_done) {
             print FILE <<__EOF__;
             if (flags & SDL_COPY_MODULATE_ALPHA) {
                 ${s}A = (${s}A * modulateA) / 255;
@@ -252,7 +254,8 @@ __EOF__
         }
     }
     if ( $blend ) {
-        print FILE <<__EOF__;
+        if (!$A_is_const_FF) {
+            print FILE <<__EOF__;
             if (flags & (SDL_COPY_BLEND|SDL_COPY_ADD)) {
                 /* This goes away if we ever use premultiplied alpha */
                 if (${s}A < 255) {
@@ -261,17 +264,35 @@ __EOF__
                     ${s}B = (${s}B * ${s}A) / 255;
                 }
             }
+__EOF__
+        }
+        print FILE <<__EOF__;
             switch (flags & (SDL_COPY_BLEND|SDL_COPY_ADD|SDL_COPY_MOD)) {
             case SDL_COPY_BLEND:
+__EOF__
+        if ($A_is_const_FF) {
+            print FILE <<__EOF__;
+                ${d}R = ${s}R;
+                ${d}G = ${s}G;
+                ${d}B = ${s}B;
+__EOF__
+        } else {
+            print FILE <<__EOF__;
                 ${d}R = ${s}R + ((255 - ${s}A) * ${d}R) / 255;
                 ${d}G = ${s}G + ((255 - ${s}A) * ${d}G) / 255;
                 ${d}B = ${s}B + ((255 - ${s}A) * ${d}B) / 255;
 __EOF__
-
+        }
         if ( $dst_has_alpha ) {
-            print FILE <<__EOF__;
+            if ($A_is_const_FF) {
+                print FILE <<__EOF__;
+                ${d}A = 0xFF;
+__EOF__
+            } else {
+                print FILE <<__EOF__;
                 ${d}A = ${s}A + ((255 - ${s}A) * ${d}A) / 255;
 __EOF__
+            }
         }
 
         print FILE <<__EOF__;
@@ -306,6 +327,11 @@ sub output_copyfunc
 
     my $dst_has_alpha = ($dst =~ /A/) ? 1 : 0;
     my $ignore_dst_alpha = !$dst_has_alpha && !$blend;
+    
+    my $src_has_alpha = ($src =~ /A/) ? 1 : 0;
+
+    my $is_modulateA_done = 0;
+    my $A_is_const_FF = 0;
 
     output_copyfuncname("static void", $src, $dst, $modulate, $blend, $scale, 1, "\n");
     print FILE <<__EOF__;
@@ -331,7 +357,25 @@ __EOF__
     if ( $blend ) {
         print FILE <<__EOF__;
     Uint32 srcpixel;
+__EOF__
+        if (!$ignore_dst_alpha && !$src_has_alpha) {
+            if ($modulate){
+                $is_modulateA_done = 1;
+                print FILE <<__EOF__;
+    const Uint32 srcA = (flags & SDL_COPY_MODULATE_ALPHA) ? modulateA : 0xFF;
+__EOF__
+            } else {
+                $A_is_const_FF = 1;
+            }
+            print FILE <<__EOF__;
+    Uint32 srcR, srcG, srcB;
+__EOF__
+        } else {
+            print FILE <<__EOF__;
     Uint32 srcR, srcG, srcB, srcA;
+__EOF__
+        }
+        print FILE <<__EOF__;
     Uint32 dstpixel;
 __EOF__
         if ($dst_has_alpha) {
@@ -347,7 +391,22 @@ __EOF__
         print FILE <<__EOF__;
     Uint32 pixel;
 __EOF__
-        if (!$ignore_dst_alpha) {
+        if (!$ignore_dst_alpha && !$src_has_alpha) {
+            if ($modulate){
+                $is_modulateA_done = 1;
+                print FILE <<__EOF__;
+    const Uint32 A = (flags & SDL_COPY_MODULATE_ALPHA) ? modulateA : 0xFF;
+__EOF__
+            } else {
+                $A_is_const_FF = 1;
+                print FILE <<__EOF__;
+    const Uint32 A = 0xFF;
+__EOF__
+            }
+            print FILE <<__EOF__;
+    Uint32 R, G, B;
+__EOF__
+        } elsif (!$ignore_dst_alpha) {
             print FILE <<__EOF__;
     Uint32 R, G, B, A;
 __EOF__
@@ -392,7 +451,7 @@ __EOF__
         print FILE <<__EOF__;
             }
 __EOF__
-        output_copycore($src, $dst, $modulate, $blend);
+        output_copycore($src, $dst, $modulate, $blend, $is_modulateA_done, $A_is_const_FF);
         print FILE <<__EOF__;
             posx += incx;
             ++dst;
@@ -410,7 +469,7 @@ __EOF__
         int n = info->dst_w;
         while (n--) {
 __EOF__
-        output_copycore($src, $dst, $modulate, $blend);
+        output_copycore($src, $dst, $modulate, $blend, $is_modulateA_done, $A_is_const_FF);
         print FILE <<__EOF__;
             ++src;
             ++dst;