update kernel to 5.10.198

hc

2024-01-03 2f7c68cb55ecb7331f2381deb497c27155f32faf

 kernel/drivers/video/fbdev/core/sysimgblt.c
..
..
@@ -188,23 +188,29 @@
188
188
 {
189
189
 	u32 fgx = fgcolor, bgx = bgcolor, bpp = p->var.bits_per_pixel;
190
190
 	u32 ppw = 32/bpp, spitch = (image->width + 7)/8;
191
-	u32 bit_mask, end_mask, eorx, shift;
192
-	const char *s = image->data, *src;
191
+	u32 bit_mask, eorx, shift;
192
+	const u8 *s = image->data, *src;
193
193
 	u32 *dst;
194
-	const u32 *tab = NULL;
194
+	const u32 *tab;
195
+	size_t tablen;
196
+	u32 colortab[16];
195
197
 	int i, j, k;
196
198
 
197
199
 	switch (bpp) {
198
200
 	case 8:
199
201
 		tab = fb_be_math(p) ? cfb_tab8_be : cfb_tab8_le;
202
+		tablen = 16;
200
203
 		break;
201
204
 	case 16:
202
205
 		tab = fb_be_math(p) ? cfb_tab16_be : cfb_tab16_le;
206
+		tablen = 4;
203
207
 		break;
204
208
 	case 32:
205
-	default:
206
209
 		tab = cfb_tab32;
210
+		tablen = 2;
207
211
 		break;
212
+	default:
213
+		return;
208
214
 	}
209
215
 
210
216
 	for (i = ppw-1; i--; ) {
..
..
@@ -218,20 +224,62 @@
218
224
 	eorx = fgx ^ bgx;
219
225
 	k = image->width/ppw;
220
226
 
227
+	for (i = 0; i < tablen; ++i)
228
+		colortab[i] = (tab[i] & eorx) ^ bgx;
229
+
221
230
 	for (i = image->height; i--; ) {
222
231
 		dst = dst1;
223
232
 		shift = 8;
224
233
 		src = s;
225
234
 
226
-		for (j = k; j--; ) {
235
+		/*
236
+		 * Manually unroll the per-line copying loop for better
237
+		 * performance. This works until we processed the last
238
+		 * completely filled source byte (inclusive).
239
+		 */
240
+		switch (ppw) {
241
+		case 4: /* 8 bpp */
242
+			for (j = k; j >= 2; j -= 2, ++src) {
243
+				*dst++ = colortab[(*src >> 4) & bit_mask];
244
+				*dst++ = colortab[(*src >> 0) & bit_mask];
245
+			}
246
+			break;
247
+		case 2: /* 16 bpp */
248
+			for (j = k; j >= 4; j -= 4, ++src) {
249
+				*dst++ = colortab[(*src >> 6) & bit_mask];
250
+				*dst++ = colortab[(*src >> 4) & bit_mask];
251
+				*dst++ = colortab[(*src >> 2) & bit_mask];
252
+				*dst++ = colortab[(*src >> 0) & bit_mask];
253
+			}
254
+			break;
255
+		case 1: /* 32 bpp */
256
+			for (j = k; j >= 8; j -= 8, ++src) {
257
+				*dst++ = colortab[(*src >> 7) & bit_mask];
258
+				*dst++ = colortab[(*src >> 6) & bit_mask];
259
+				*dst++ = colortab[(*src >> 5) & bit_mask];
260
+				*dst++ = colortab[(*src >> 4) & bit_mask];
261
+				*dst++ = colortab[(*src >> 3) & bit_mask];
262
+				*dst++ = colortab[(*src >> 2) & bit_mask];
263
+				*dst++ = colortab[(*src >> 1) & bit_mask];
264
+				*dst++ = colortab[(*src >> 0) & bit_mask];
265
+			}
266
+			break;
267
+		}
268
+
269
+		/*
270
+		 * For image widths that are not a multiple of 8, there
271
+		 * are trailing pixels left on the current line. Print
272
+		 * them as well.
273
+		 */
274
+		for (; j--; ) {
227
275
 			shift -= ppw;
228
-			end_mask = tab[(*src >> shift) & bit_mask];
229
-			*dst++ = (end_mask & eorx) ^ bgx;
276
+			*dst++ = colortab[(*src >> shift) & bit_mask];
230
277
 			if (!shift) {
231
278
 				shift = 8;
232
-				src++;
279
+				++src;
233
280
 			}
234
281
 		}
282
+
235
283
 		dst1 += p->fix.line_length;
236
284
 		s += spitch;
237
285
 	}