Replace implicit use of fast_memcpy via macro by explicit use to allow

for future optimization.


git-svn-id: svn://svn.mplayerhq.hu/mplayer/trunk@23475 b3059339-0415-0410-9bf9-f77b7e298cf2

58 files changed, 192 insertions, 191 deletions

diff --git a/gui/mplayer/mw.c b/gui/mplayer/mw.c
index 142eca3185..1762eea45d 100644
--- a/gui/mplayer/mw.c
+++ b/gui/mplayer/mw.c
@@ -65,7 +65,7 @@ void mplMainDraw( void )
    btnModify( evSetMoviePosition,guiIntfStruct.Position );
    btnModify( evSetVolume,guiIntfStruct.Volume );
 
-   memcpy( mplDrawBuffer,appMPlayer.main.Bitmap.Image,appMPlayer.main.Bitmap.ImageSize );
+   fast_memcpy( mplDrawBuffer,appMPlayer.main.Bitmap.Image,appMPlayer.main.Bitmap.ImageSize );
    Render( &appMPlayer.mainWindow,appMPlayer.Items,appMPlayer.NumberOfItems,mplDrawBuffer,appMPlayer.main.Bitmap.ImageSize );
    mplMainRender=0;
   }
diff --git a/gui/mplayer/pb.c b/gui/mplayer/pb.c
index 3fad991c28..0a06abf9b0 100644
--- a/gui/mplayer/pb.c
+++ b/gui/mplayer/pb.c
@@ -88,7 +88,7 @@ void mplPBDraw( void )
    
    vo_mouse_autohide=0;
 
-   memcpy( mplPBDrawBuffer,appMPlayer.bar.Bitmap.Image,appMPlayer.bar.Bitmap.ImageSize );
+   fast_memcpy( mplPBDrawBuffer,appMPlayer.bar.Bitmap.Image,appMPlayer.bar.Bitmap.ImageSize );
    Render( &appMPlayer.barWindow,appMPlayer.barItems,appMPlayer.NumberOfBarItems,mplPBDrawBuffer,appMPlayer.bar.Bitmap.ImageSize );
    wsConvert( &appMPlayer.barWindow,mplPBDrawBuffer,appMPlayer.bar.Bitmap.ImageSize );   
   }
diff --git a/libaf/af_format.c b/libaf/af_format.c
index 193a4e7f9b..df95a03aeb 100644
--- a/libaf/af_format.c
+++ b/libaf/af_format.c
@@ -287,7 +287,7 @@ static af_data_t* play(struct af_instance_s* af, af_data_t* data)
       if(c->bps != l->bps)
 	change_bps(c->audio,l->audio,len,c->bps,l->bps);
       else
-	memcpy(l->audio,c->audio,len*c->bps);
+	fast_memcpy(l->audio,c->audio,len*c->bps);
       break;
     }
   }
diff --git a/libao2/ao_dsound.c b/libao2/ao_dsound.c
index 79f70b1b76..5f6bc8daea 100644
--- a/libao2/ao_dsound.c
+++ b/libao2/ao_dsound.c
@@ -199,7 +199,7 @@ static BOOL CALLBACK DirectSoundEnum(LPGUID guid,LPCSTR desc,LPCSTR module,LPVOI
     if(device_num==*device_index){
         mp_msg(MSGT_AO, MSGL_V,"<--");
         if(guid){
-            memcpy(&device,guid,sizeof(GUID));
+            fast_memcpy(&device,guid,sizeof(GUID));
         }
     }
     mp_msg(MSGT_AO, MSGL_V,"\n");
@@ -337,14 +337,14 @@ static int write_buffer(unsigned char *data, int len)
   	    numsamp = dwBytes1 / (ao_data.channels * sampsize);  // number of samples for each channel in this buffer
 
   	    for( i = 0; i < numsamp; i++ ) for( j = 0; j < ao_data.channels; j++ ) {
-  	        memcpy(lpvPtr1+(i*ao_data.channels*sampsize)+(chantable[j]*sampsize),data+(i*ao_data.channels*sampsize)+(j*sampsize),sampsize);
+  	        fast_memcpy(lpvPtr1+(i*ao_data.channels*sampsize)+(chantable[j]*sampsize),data+(i*ao_data.channels*sampsize)+(j*sampsize),sampsize);
   	    }
 
   	    if (NULL != lpvPtr2 )
   	    {
   	        numsamp = dwBytes2 / (ao_data.channels * sampsize);
   	        for( i = 0; i < numsamp; i++ ) for( j = 0; j < ao_data.channels; j++ ) {
-  	            memcpy(lpvPtr2+(i*ao_data.channels*sampsize)+(chantable[j]*sampsize),data+dwBytes1+(i*ao_data.channels*sampsize)+(j*sampsize),sampsize);
+  	            fast_memcpy(lpvPtr2+(i*ao_data.channels*sampsize)+(chantable[j]*sampsize),data+dwBytes1+(i*ao_data.channels*sampsize)+(j*sampsize),sampsize);
   	        }
   	    }
 
@@ -352,8 +352,8 @@ static int write_buffer(unsigned char *data, int len)
   	    if(write_offset>=buffer_size)write_offset=dwBytes2;
   	} else {
   	    // Write to pointers without reordering. 
-	memcpy(lpvPtr1,data,dwBytes1);
-    if (NULL != lpvPtr2 )memcpy(lpvPtr2,data+dwBytes1,dwBytes2);
+	fast_memcpy(lpvPtr1,data,dwBytes1);
+    if (NULL != lpvPtr2 )fast_memcpy(lpvPtr2,data+dwBytes1,dwBytes2);
 	write_offset+=dwBytes1+dwBytes2;
     if(write_offset>=buffer_size)write_offset=dwBytes2;
   	}
diff --git a/libao2/ao_jack.c b/libao2/ao_jack.c
index 556a887a4b..1e9dbff8c8 100644
--- a/libao2/ao_jack.c
+++ b/libao2/ao_jack.c
@@ -111,10 +111,10 @@ static int write_buffer(unsigned char* data, int len) {
   if (len > free) len = free;
   if (first_len > len) first_len = len;
   // till end of buffer
-  memcpy (&buffer[write_pos], data, first_len);
+  fast_memcpy (&buffer[write_pos], data, first_len);
   if (len > first_len) { // we have to wrap around
     // remaining part from beginning of buffer
-    memcpy (buffer, &data[first_len], len - first_len);
+    fast_memcpy (buffer, &data[first_len], len - first_len);
   }
   write_pos = (write_pos + len) % BUFFSIZE;
   return len;
diff --git a/libao2/ao_sdl.c b/libao2/ao_sdl.c
index 6ec56d1665..0da9bb338f 100644
--- a/libao2/ao_sdl.c
+++ b/libao2/ao_sdl.c
@@ -88,10 +88,10 @@ static int write_buffer(unsigned char* data,int len){
   if (len > free) len = free;
   if (first_len > len) first_len = len;
   // till end of buffer
-  memcpy (&buffer[write_pos], data, first_len);
+  fast_memcpy (&buffer[write_pos], data, first_len);
   if (len > first_len) { // we have to wrap around
     // remaining part from beginning of buffer
-    memcpy (buffer, &data[first_len], len - first_len);
+    fast_memcpy (buffer, &data[first_len], len - first_len);
   }
   write_pos = (write_pos + len) % BUFFSIZE;
   return len;
@@ -106,14 +106,14 @@ static int read_buffer(unsigned char* data,int len){
 #ifdef USE_SDL_INTERNAL_MIXER
   SDL_MixAudio (data, &buffer[read_pos], first_len, volume);
 #else
-  memcpy (data, &buffer[read_pos], first_len);
+  fast_memcpy (data, &buffer[read_pos], first_len);
 #endif
   if (len > first_len) { // we have to wrap around
     // remaining part from beginning of buffer
 #ifdef USE_SDL_INTERNAL_MIXER
     SDL_MixAudio (&data[first_len], buffer, len - first_len, volume);
 #else
-    memcpy (&data[first_len], buffer, len - first_len);
+    fast_memcpy (&data[first_len], buffer, len - first_len);
 #endif
   }
   read_pos = (read_pos + len) % BUFFSIZE;
diff --git a/libao2/ao_win32.c b/libao2/ao_win32.c
index 0f39b51767..71146b43a2 100644
--- a/libao2/ao_win32.c
+++ b/libao2/ao_win32.c
@@ -299,7 +299,7 @@ static int write_waveOutBuffer(unsigned char* data,int len){
            waveOutUnprepareHeader(hWaveOut, current, sizeof(WAVEHDR));
 	x=BUFFER_SIZE-buf_write_pos;          
     if(x>len) x=len;                   
-    memcpy(current->lpData+buf_write_pos,data+len2,x); 
+    fast_memcpy(current->lpData+buf_write_pos,data+len2,x); 
     if(buf_write_pos==0)full_buffers++;
     len2+=x; len-=x;                 
 	buffered_bytes+=x; buf_write_pos+=x; 
diff --git a/libmenu/vf_menu.c b/libmenu/vf_menu.c
index 7996c24adc..3b1e1cc2a1 100644
--- a/libmenu/vf_menu.c
+++ b/libmenu/vf_menu.c
@@ -146,8 +146,8 @@ static void get_image(struct vf_instance_s* vf, mp_image_t *mpi){
 
   if(mpi->type == MP_IMGTYPE_TEMP && (!(mpi->flags&MP_IMGFLAG_PRESERVE)) ) {
     dmpi = vf_get_image(vf->next,mpi->imgfmt,mpi->type, mpi->flags, mpi->w, mpi->h);
-    memcpy(mpi->planes,dmpi->planes,MP_MAX_PLANES*sizeof(unsigned char*));
-    memcpy(mpi->stride,dmpi->stride,MP_MAX_PLANES*sizeof(unsigned int));
+    fast_memcpy(mpi->planes,dmpi->planes,MP_MAX_PLANES*sizeof(unsigned char*));
+    fast_memcpy(mpi->stride,dmpi->stride,MP_MAX_PLANES*sizeof(unsigned int));
     mpi->flags|=MP_IMGFLAG_DIRECT;
     mpi->priv=(void*)dmpi;
     return;
diff --git a/libmpcodecs/ad_mpc.c b/libmpcodecs/ad_mpc.c
index e91998df7b..435a6cb0e5 100644
--- a/libmpcodecs/ad_mpc.c
+++ b/libmpcodecs/ad_mpc.c
@@ -54,7 +54,7 @@ static mpc_int32_t cb_read(void *data, void *buf, mpc_int32_t size) {
   if (d->pos < d->header_len) {
     if (s > d->header_len - d->pos)
       s = d->header_len - d->pos;
-    memcpy(p, &d->header[d->pos], s);
+    fast_memcpy(p, &d->header[d->pos], s);
   } else
     s = 0;
   memset(&p[s], 0, size - s);
diff --git a/libmpcodecs/native/nuppelvideo.c b/libmpcodecs/native/nuppelvideo.c
index c7d88eca66..cac116f3eb 100644
--- a/libmpcodecs/native/nuppelvideo.c
+++ b/libmpcodecs/native/nuppelvideo.c
@@ -67,7 +67,7 @@ void decode_nuv( unsigned char *encoded, int encoded_size,
 		switch(encodedh->comptype)
 		{
 		    case '0': /* raw YUV420 */
-			memcpy(decoded, encoded + 12, out_len);
+			fast_memcpy(decoded, encoded + 12, out_len);
 			break;
 		    case '1': /* RTJpeg */
 			RTjpeg_decompressYUV420 ( ( __s8 * ) encoded + 12, decoded );
@@ -102,13 +102,13 @@ void decode_nuv( unsigned char *encoded, int encoded_size,
 			break;
 		    case 'L': /* copy last frame */
 #ifdef KEEP_BUFFER
-			memcpy ( decoded, previous_buffer, width*height*3/2);
+			fast_memcpy ( decoded, previous_buffer, width*height*3/2);
 #endif
 			break;
 		}
 
 #ifdef KEEP_BUFFER
-		memcpy(previous_buffer, decoded, width*height*3/2);
+		fast_memcpy(previous_buffer, decoded, width*height*3/2);
 #endif
 		break;
 	    }
diff --git a/libmpcodecs/vd_mpng.c b/libmpcodecs/vd_mpng.c
index d8725f7c08..f4af973e8d 100644
--- a/libmpcodecs/vd_mpng.c
+++ b/libmpcodecs/vd_mpng.c
@@ -58,7 +58,7 @@ static void pngReadFN( png_structp pngstr,png_bytep buffer,png_size_t size )
 {
  char * p = pngstr->io_ptr;
  if(size>pngLength-pngPointer && pngLength>=pngPointer) size=pngLength-pngPointer;
- memcpy( buffer,(char *)&p[pngPointer],size );
+ fast_memcpy( buffer,(char *)&p[pngPointer],size );
  pngPointer+=size;
 }
 
diff --git a/libmpcodecs/vd_mtga.c b/libmpcodecs/vd_mtga.c
index a328ab037c..f9f2301da2 100644
--- a/libmpcodecs/vd_mtga.c
+++ b/libmpcodecs/vd_mtga.c
@@ -110,20 +110,20 @@ static void decode_rle_tga(TGAInfo *info, unsigned char *data, mp_image_t *mpi)
 	    
 	    if (packet_header & 0x80) /* runlength encoded packet */
 	    {
-		memcpy(final, data, num_bytes);
+		fast_memcpy(final, data, num_bytes);
 		
 		// Note: this will be slow when DR to vram!
 		i=num_bytes;
 		while(2*i<=replen){
-		    memcpy(final+i,final,i);
+		    fast_memcpy(final+i,final,i);
 		    i*=2;
 		}
-		memcpy(final+i,final,replen-i);
+		fast_memcpy(final+i,final,replen-i);
 		data += num_bytes;
 	    }
 	    else /* raw packet */
 	    {
-		memcpy(final, data, replen);
+		fast_memcpy(final, data, replen);
 		data += replen;
 	    }
 	    
@@ -144,7 +144,7 @@ static void decode_uncompressed_tga(TGAInfo *info, unsigned char *data, mp_image
     for (row = info->start_row; (!info->origin && row) || (info->origin && row < info->height); row += info->increment)
     {
 	final = mpi->planes[0] + mpi->stride[0] * row;
-	memcpy(final, data, info->width * num_bytes);
+	fast_memcpy(final, data, info->width * num_bytes);
 	data += info->width * num_bytes;
     }
 
diff --git a/libmpcodecs/vf_bmovl.c b/libmpcodecs/vf_bmovl.c
index 90b6ccd15e..ef7ce19ce9 100644
--- a/libmpcodecs/vf_bmovl.c
+++ b/libmpcodecs/vf_bmovl.c
@@ -388,14 +388,14 @@ put_image(struct vf_instance_s* vf, mp_image_t* mpi, double pts){
 
 	if(vf->priv->opaque) {	// Just copy buffer memory to screen
 		for( ypos=vf->priv->y1 ; ypos < vf->priv->y2 ; ypos++ ) {
-			memcpy( dmpi->planes[0] + (ypos*dmpi->stride[0]) + vf->priv->x1,
+			fast_memcpy( dmpi->planes[0] + (ypos*dmpi->stride[0]) + vf->priv->x1,
 			        vf->priv->bitmap.y + (ypos*vf->priv->w) + vf->priv->x1,
 					vf->priv->x2 - vf->priv->x1 );
 			if(ypos%2) {
-				memcpy( dmpi->planes[1] + ((ypos/2)*dmpi->stride[1]) + (vf->priv->x1/2),
+				fast_memcpy( dmpi->planes[1] + ((ypos/2)*dmpi->stride[1]) + (vf->priv->x1/2),
 				        vf->priv->bitmap.u + (((ypos/2)*(vf->priv->w)/2)) + (vf->priv->x1/2),
 				        (vf->priv->x2 - vf->priv->x1)/2 );
-				memcpy( dmpi->planes[2] + ((ypos/2)*dmpi->stride[2]) + (vf->priv->x1/2),
+				fast_memcpy( dmpi->planes[2] + ((ypos/2)*dmpi->stride[2]) + (vf->priv->x1/2),
 				        vf->priv->bitmap.v + (((ypos/2)*(vf->priv->w)/2)) + (vf->priv->x1/2),
 				        (vf->priv->x2 - vf->priv->x1)/2 );
 			}
diff --git a/libmpcodecs/vf_detc.c b/libmpcodecs/vf_detc.c
index 52932d8d6b..27be5d393c 100644
--- a/libmpcodecs/vf_detc.c
+++ b/libmpcodecs/vf_detc.c
@@ -60,7 +60,7 @@ static inline void *my_memcpy_pic(void * dst, void * src, int bytesPerLine, int
 
 	for(i=0; i<height; i++)
 	{
-		memcpy(dst, src, bytesPerLine);
+		fast_memcpy(dst, src, bytesPerLine);
 		src+= srcStride;
 		dst+= dstStride;
 	}
diff --git a/libmpcodecs/vf_down3dright.c b/libmpcodecs/vf_down3dright.c
index 84e2f693b2..8bccf0cb84 100644
--- a/libmpcodecs/vf_down3dright.c
+++ b/libmpcodecs/vf_down3dright.c
@@ -67,7 +67,7 @@ static void toright(unsigned char *dst[3], unsigned char *src[3],
 					*t++ = *sR++;
 			}
 			if (p->scaleh == 1) {
-				memcpy(to + dst, to, dst);
+				fast_memcpy(to + dst, to, dst);
                                 to += dst;
 			}
 			to += dst;
diff --git a/libmpcodecs/vf_filmdint.c b/libmpcodecs/vf_filmdint.c
index 26896626e2..0fb8ea69d1 100644
--- a/libmpcodecs/vf_filmdint.c
+++ b/libmpcodecs/vf_filmdint.c
@@ -97,7 +97,7 @@ static inline void *my_memcpy_pic(void * dst, void * src, int bytesPerLine, int
 
     for(i=0; i<height; i++)
     {
-	memcpy(dst, src, bytesPerLine);
+	fast_memcpy(dst, src, bytesPerLine);
 	src+= srcStride;
 	dst+= dstStride;
     }
@@ -711,7 +711,7 @@ dint_copy_plane(unsigned char *d, unsigned char *a, unsigned char *b,
     long bos = b - a;
     long cos = c - a;
     if (field) {
-	memcpy(d, b, w);
+	fast_memcpy(d, b, w);
 	h--;
 	d += ds;
 	a += ss;
@@ -720,8 +720,8 @@ dint_copy_plane(unsigned char *d, unsigned char *a, unsigned char *b,
     cos += ss;
     while (h > 2) {
 	if (threshold >= 128) {
-	    memcpy(d, a, w);
-	    memcpy(d+ds, a+bos, w);
+	    fast_memcpy(d, a, w);
+	    fast_memcpy(d+ds, a+bos, w);
 	} else if (mmx2 == 1) {
 	    ret += dint_copy_line_mmx2(d, a, bos, cos, ds, ss, w, threshold);
 	} else
@@ -730,9 +730,9 @@ dint_copy_plane(unsigned char *d, unsigned char *a, unsigned char *b,
 	d += 2*ds;
 	a += 2*ss;
     }
-    memcpy(d, a, w);
+    fast_memcpy(d, a, w);
     if (h == 2)
-	memcpy(d+ds, a+bos, w);
+	fast_memcpy(d+ds, a+bos, w);
     return ret;
 }
 
diff --git a/libmpcodecs/vf_fspp.c b/libmpcodecs/vf_fspp.c
index 541dfb6564..acb39f5c56 100644
--- a/libmpcodecs/vf_fspp.c
+++ b/libmpcodecs/vf_fspp.c
@@ -421,15 +421,15 @@ static void filter(struct vf_priv_s *p, uint8_t *dst, uint8_t *src,
     if (!src || !dst) return; // HACK avoid crash for Y8 colourspace
     for(y=0; y<height; y++){
         int index= 8 + 8*stride + y*stride;
-        memcpy(p->src + index, src + y*src_stride, width);//this line can be avoided by using DR & user fr.buffers
+        fast_memcpy(p->src + index, src + y*src_stride, width);//this line can be avoided by using DR & user fr.buffers
         for(x=0; x<8; x++){ 
             p->src[index         - x - 1]= p->src[index +         x    ];
             p->src[index + width + x    ]= p->src[index + width - x - 1];
         }
     }
     for(y=0; y<8; y++){
-        memcpy(p->src + (      7-y)*stride, p->src + (      y+8)*stride, stride);
-        memcpy(p->src + (height+8+y)*stride, p->src + (height-y+7)*stride, stride);
+        fast_memcpy(p->src + (      7-y)*stride, p->src + (      y+8)*stride, stride);
+        fast_memcpy(p->src + (height+8+y)*stride, p->src + (height-y+7)*stride, stride);
     }
     //FIXME (try edge emu)
 
@@ -456,8 +456,8 @@ static void filter(struct vf_priv_s *p, uint8_t *dst, uint8_t *src,
 		    column_fidct_s((int16_t*)(&p->threshold_mtx[0]), block+x*8, block3+x*8, 8); //yes, this is a HOTSPOT
 		}
 	    row_idct_s(block3+0*8, p->temp + (y&15)*stride+x0+2-(y&1), stride, 2*(BLOCKSZ-1));
-	    memcpy(block, block+(BLOCKSZ-1)*64, 8*8*sizeof(DCTELEM)); //cycling
-	    memcpy(block3, block3+(BLOCKSZ-1)*64, 6*8*sizeof(DCTELEM));  
+	    fast_memcpy(block, block+(BLOCKSZ-1)*64, 8*8*sizeof(DCTELEM)); //cycling
+	    fast_memcpy(block3, block3+(BLOCKSZ-1)*64, 6*8*sizeof(DCTELEM));  
 	}
 	//
 	es=width+8-x0; //  8, ...      
@@ -532,7 +532,7 @@ static int put_image(struct vf_instance_s* vf, mp_image_t *mpi, double pts)
     if(mpi->pict_type != 3 && mpi->qscale && !vf->priv->qp){
 	if(!vf->priv->non_b_qp)
 	    vf->priv->non_b_qp= malloc(mpi->qstride * ((mpi->h + 15) >> 4));
-	memcpy(vf->priv->non_b_qp, mpi->qscale, mpi->qstride * ((mpi->h + 15) >> 4));
+	fast_memcpy(vf->priv->non_b_qp, mpi->qscale, mpi->qstride * ((mpi->h + 15) >> 4));
     }
     if(vf->priv->log2_count || !(mpi->flags&MP_IMGFLAG_DIRECT)){
 	char *qp_tab= vf->priv->non_b_qp;
diff --git a/libmpcodecs/vf_il.c b/libmpcodecs/vf_il.c
index edc847ccff..a5af511c6e 100644
--- a/libmpcodecs/vf_il.c
+++ b/libmpcodecs/vf_il.c
@@ -58,20 +58,20 @@ static void interleave(uint8_t *dst, uint8_t *src, int w, int h, int dstStride,
 	switch(interleave){
 	case -1:
 		for(y=0; y < m; y++){
-			memcpy(dst + dstStride* y     , src + srcStride*(y*2 + a), w);
-			memcpy(dst + dstStride*(y + m), src + srcStride*(y*2 + b), w);
+			fast_memcpy(dst + dstStride* y     , src + srcStride*(y*2 + a), w);
+			fast_memcpy(dst + dstStride*(y + m), src + srcStride*(y*2 + b), w);
 		}
 		break;
 	case 0:
 		for(y=0; y < m; y++){
-			memcpy(dst + dstStride* y*2   , src + srcStride*(y*2 + a), w);
-			memcpy(dst + dstStride*(y*2+1), src + srcStride*(y*2 + b), w);
+			fast_memcpy(dst + dstStride* y*2   , src + srcStride*(y*2 + a), w);
+			fast_memcpy(dst + dstStride*(y*2+1), src + srcStride*(y*2 + b), w);
 		}
 		break;
 	case 1:
 		for(y=0; y < m; y++){
-			memcpy(dst + dstStride*(y*2+a), src + srcStride* y     , w);
-			memcpy(dst + dstStride*(y*2+b), src + srcStride*(y + m), w);
+			fast_memcpy(dst + dstStride*(y*2+a), src + srcStride* y     , w);
+			fast_memcpy(dst + dstStride*(y*2+b), src + srcStride*(y + m), w);
 		}
 		break;
 	}
diff --git a/libmpcodecs/vf_ivtc.c b/libmpcodecs/vf_ivtc.c
index 9c30a02477..4f9ca36d7c 100644
--- a/libmpcodecs/vf_ivtc.c
+++ b/libmpcodecs/vf_ivtc.c
@@ -47,7 +47,7 @@ static inline void *my_memcpy_pic(void * dst, void * src, int bytesPerLine, int
 
 	for(i=0; i<height; i++)
 	{
-		memcpy(dst, src, bytesPerLine);
+		fast_memcpy(dst, src, bytesPerLine);
 		src+= srcStride;
 		dst+= dstStride;
 	}
diff --git a/libmpcodecs/vf_kerndeint.c b/libmpcodecs/vf_kerndeint.c
index f512601fb3..41da44ebfb 100644
--- a/libmpcodecs/vf_kerndeint.c
+++ b/libmpcodecs/vf_kerndeint.c
@@ -128,16 +128,16 @@ static int put_image(struct vf_instance_s* vf, mp_image_t *mpi, double pts){
 		dstp = dstp_saved + (1-order) * dst_pitch;
 
 		for (y=0; y<h; y+=2) {
-			memcpy(dstp, srcp, w);
+			fast_memcpy(dstp, srcp, w);
 			srcp += 2*src_pitch;
 			dstp += 2*dst_pitch;
 		}
 
 		// Copy through the lines that will be missed below.
-		memcpy(dstp_saved + order*dst_pitch, srcp_saved + (1-order)*src_pitch, w);
-		memcpy(dstp_saved + (2+order)*dst_pitch, srcp_saved + (3-order)*src_pitch, w);
-		memcpy(dstp_saved + (h-2+order)*dst_pitch, srcp_saved + (h-1-order)*src_pitch, w);
-		memcpy(dstp_saved + (h-4+order)*dst_pitch, srcp_saved + (h-3-order)*src_pitch, w);
+		fast_memcpy(dstp_saved + order*dst_pitch, srcp_saved + (1-order)*src_pitch, w);
+		fast_memcpy(dstp_saved + (2+order)*dst_pitch, srcp_saved + (3-order)*src_pitch, w);
+		fast_memcpy(dstp_saved + (h-2+order)*dst_pitch, srcp_saved + (h-1-order)*src_pitch, w);
+		fast_memcpy(dstp_saved + (h-4+order)*dst_pitch, srcp_saved + (h-3-order)*src_pitch, w);
 		/* For the other field choose adaptively between using the previous field
 		   or the interpolant from the current field. */
 
@@ -270,7 +270,7 @@ static int put_image(struct vf_instance_s* vf, mp_image_t *mpi, double pts){
 		srcp = mpi->planes[z];
 		dstp = pmpi->planes[z];
 		for (y=0; y<h; y++) {
-			memcpy(dstp, srcp, w);
+			fast_memcpy(dstp, srcp, w);
 			srcp += src_pitch;
 			dstp += psrc_pitch;
 		}
diff --git a/libmpcodecs/vf_noise.c b/libmpcodecs/vf_noise.c
index d430ef3281..360cd4dfec 100644
--- a/libmpcodecs/vf_noise.c
+++ b/libmpcodecs/vf_noise.c
@@ -278,12 +278,12 @@ static void noise(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int
 	{
 		if(src==dst) return;
 
-		if(dstStride==srcStride) memcpy(dst, src, srcStride*height);
+		if(dstStride==srcStride) fast_memcpy(dst, src, srcStride*height);
 		else
 		{
 			for(y=0; y<height; y++)
 			{
-				memcpy(dst, src, width);
+				fast_memcpy(dst, src, width);
 				dst+= dstStride;
 				src+= srcStride;
 			}
diff --git a/libmpcodecs/vf_phase.c b/libmpcodecs/vf_phase.c
index 42957cd77e..45c76a1466 100644
--- a/libmpcodecs/vf_phase.c
+++ b/libmpcodecs/vf_phase.c
@@ -45,8 +45,8 @@ static void do_plane(unsigned char *to, unsigned char *from,
 
    for(end=to+h*ts, buf=*bufp, top=1; to<end; from+=fs, to+=ts, buf+=w, top^=1)
       {
-      memcpy(to, mode==(top?BOTTOM_FIRST:TOP_FIRST)?buf:from, w);
-      memcpy(buf, from, w);
+      fast_memcpy(to, mode==(top?BOTTOM_FIRST:TOP_FIRST)?buf:from, w);
+      fast_memcpy(buf, from, w);</