libc: faster memset

An optimization suggested by TSWilliamson, which pushes not only RAM,
but also on-chip memory and the CPU pipeline to their limits.

TODO

@@ -1,8 +1,3 @@
-For the 2.1.0 release:
-* bopti:     remove the deprecated image_t definition
-* project:   remove the compat branch
-* project:   remove the gray aliases
-
 Extensions on existing code:
 * tmu:       make interrupt handlers more elegant
 * bopti:     try to display fullscreen images with TLB access + DMA on fxcg50

include/gint/display-fx.h

@@ -83,10 +83,6 @@ typedef struct
 
 } GPACKED(4) bopti_image_t;
 
-/* Old alias to image_t, now deprecated because of libimg */
-typedef bopti_image_t image_t __attribute__((deprecated(
-	"image_t has been renamed to bopti_image_t")));
-
 #endif /* FX9860G */
 
 #endif /* GINT_DISPLAY_FX */

include/gint/gray.h

@@ -144,34 +144,4 @@ void dgray_getdelays(uint32_t *light, uint32_t *dark);
    These pointers can be used for custom rendering functions. */
 void dgray_getvram(uint32_t **light, uint32_t **dark);
 
-//---
-//	Aliases for older programs
-//---
-
-/* Functions for the gray engine used to be called g*() to mirror the d*()
-   functions of the display module. They are now bundled together, so these
-   aliases are for older programs that still use the g*() naming scheme */
-
-#ifdef GINT_GRAY_ALIASES
-
-#define gclear dclear
-#define grect drect
-#define gpixel dpixel
-#define gline dline
-#define ghline dhline
-#define gvline dvline
-#define gimage dimage
-#define gsubimage dsubimage
-#define gtext dtext
-#define gtext_opt dtext_opt
-#define gprint dprint
-#define gprint_opt dprint_opt
-
-#define gupdate dupdate
-#define gray_delays dgray_setdelays
-#define gray_config dgray_getdelays
-#define gvram dgray_getvram
-
-#endif /* GINT_GRAY_ALIASES */
-
 #endif /* GINT_GRAY */

src/std/memcpy.s

@@ -24,7 +24,7 @@ _memcpy_align_dst:
 
 	/* If source is 4-aligned, use mov.l */
 	tst	r2, r5
-	bt/s	.aligned4
+	bt/s	.aligned4_32
 	mov	#4, r2
 
 	/* If unaligned but SH4, use movua.l */
@@ -42,13 +42,40 @@ _memcpy_align_dst:
 	bra	_naive_memcpy
 	nop
 
-.aligned4:
+.aligned4_32:
+	mov	#36, r2
+
+	/* Copy 32 bytes at a time until at most 32 bytes are left */
+	mov.l	@r5+, r0
+	mov.l	@r5+, r1
+	mov.l	@r5+, r7
+	mov.l	r0, @r4
+	mov.l	r1, @(4,r4)
+	mov.l	r7, @(8,r4)
+	mov.l	@r5+, r0
+	mov.l	@r5+, r1
+	mov.l	@r5+, r7
+	mov.l	r0, @(12,r4)
+	mov.l	r1, @(16,r4)
+	mov.l	r7, @(20,r4)
+	mov.l	@r5+, r0
+	mov.l	@r5+, r1
+	add	#-32, r6
+	mov.l	r0, @(24,r4)
+	mov.l	r1, @(28,r4)
+	cmp/ge	r6, r2
+	bf/s	.aligned4_32
+	add	#32, r4
+
+.aligned4_4:
+	mov	#4, r2
+
 	/* Copy 4 bytes at a time until at most 4 bytes are left */
 	mov.l	@r5+, r0
 	mov.l	r0, @r4
 	add	#-4, r6
 	cmp/ge	r6, r2
-	bf/s	.aligned4
+	bf/s	.aligned4_4
 	add	#4, r4
 
 	bra	_naive_memcpy

src/std/memset.s

@@ -31,12 +31,28 @@ _memset_align:
 	bf/s	_memset_align
 	dt	r6
 
+	mov	#40, r2
+
+.aligned4_32:
+	add	#-32, r4
+	add	#-32, r6
+	mov.l	r0, @(28,r4)
+	mov.l	r0, @(24,r4)
+	mov.l	r0, @(20,r4)
+	mov.l	r0, @(16,r4)
+	mov.l	r0, @(12,r4)
+	mov.l	r0, @(8,r4)
+	mov.l	r0, @(4,r4)
+	cmp/ge	r6, r2
+	bf/s	.aligned4_32
+	mov.l	r0, @r4
+
 	mov	#8, r2
 
-.aligned4:
+.aligned4_4:
 	mov.l	r0, @-r4
 	cmp/ge	r6, r2
-	bf/s	.aligned4
+	bf/s	.aligned4_4
 	add	#-4, r6
 
 _naive_memset: