cache: make sceSifWriteBackDCache share func impl

sceSifWriteBackDCache used Hit writeback invalidate ops, this makes it a
good match to use SyncDCache now that it also does.

This misses some loop unrolling that was there for large invalidations,
not sure that particularly matters.

Author: Ziemas

ee/kernel/Makefile

@@ -82,6 +82,7 @@ GLUE_OBJS += iEnableIntc.o iDisableIntc.o iEnableDmac.o iDisableDmac.o
 
 ### Cache
 CACHE_OBJS = SyncDCache.o iSyncDCache.o InvalidDCache.o iInvalidDCache.o
+CACHE_OBJS += sceSifWriteBackDCache.o
 
 ### SIO objects
 
@@ -220,8 +221,8 @@ KERNEL_OBJS = ResetEE.o SetGsCrt.o $(EXEC_SYSCALLS) \
 	SetPgifHandler.o SetVSyncFlag.o SetSyscall.o sceSifDmaStat.o isceSifDmaStat.o \
 	sceSifSetDma.o isceSifSetDma.o sceSifSetDChain.o isceSifSetDChain.o sceSifSetReg.o \
 	sceSifGetReg.o $(EXECOSD_SYSCALL) Deci2Call.o PSMode.o MachineType.o GetMemorySize.o _GetGsDxDyOffset.o \
-	_InitTLB.o SetMemoryMode.o GetMemoryMode.o sceSifWriteBackDCache.o \
-	__errno.o errno.o strncpy.o strlen.o memcpy.o memset.o __syscall.o GPfuncs.o _print.o
+	_InitTLB.o SetMemoryMode.o GetMemoryMode.o __errno.o errno.o strncpy.o strlen.o memcpy.o \
+	memset.o __syscall.o GPfuncs.o _print.o
 
 EE_OBJS = $(KERNEL_OBJS) $(SIFCMD_OBJS) $(SIFRPC_OBJS) $(FILEIO_OBJS) \
 	$(LOADFILE_OBJS) $(IOPHEAP_OBJS) $(IOPCONTROL_OBJS) $(ROM0_OBJS) $(CONFIG_OBJS) \

ee/kernel/src/cache.c

@@ -42,6 +42,13 @@ static inline void _InvalidDCache(u32 start, u32 end)
     }
 }
 
+#ifdef F_sceSifWriteBackDCache
+void sceSifWriteBackDCache(void *ptr, int size)
+{
+    _SyncDCache((u32)ptr & LINE_MASK, ((u32)(ptr) + size - 1) & LINE_MASK);
+}
+#endif
+
 #ifdef F_SyncDCache
 void SyncDCache(void *start, void *end)
 {

ee/kernel/src/kernel.S

@@ -695,68 +695,6 @@ GetGP:
 .end GetGP
 #endif
 
-#ifdef F_sceSifWriteBackDCache
-
-	.globl	sceSifWriteBackDCache
-	.ent	sceSifWriteBackDCache
-	.set	push
-	.set	noreorder
-
-sceSifWriteBackDCache:	/* DHWBIN: Data cache Hit WriteBack INvalidate.  */
-
-	lui	$25, 0xffff
-	ori	$25, $25, 0xffc0
-	blez	$5, last
-	addu	$10, $4, $5
-	and	$8, $4, $25
-	addiu	$10, $10, -1
-	and	$9, $10, $25
-	subu	$10, $9, $8
-	srl	$11, $10, 0x6
-	addiu	$11, $11, 1
-	andi	$9, $11, 0x7
-	beqz	$9, eight
-	srl	$10, $11, 0x3
-loop1:
-	sync
-	cache	0x18, 0($8)
-	sync
-	addiu	$9, $9, -1
-	nop
-	bgtz	$9, loop1
-	addiu	$8, $8, 64
-
-eight:
-	beqz	$10, last
-loop8:
-	addiu	$10, $10, -1
-	sync
-	cache	0x18, 0($8)
-	sync
-	cache	0x18, 64($8)
-	sync
-	cache	0x18, 128($8)
-	sync
-	cache	0x18, 192($8)
-	sync
-	cache	0x18, 256($8)
-	sync
-	cache	0x18, 320($8)
-	sync
-	cache	0x18, 384($8)
-	sync
-	cache	0x18, 448($8)
-	sync
-	bgtz	$10, loop8
-	addiu	$8, $8, 512
-last:
-	jr	$31
-	nop
-
-	.set	pop
-	.end	sceSifWriteBackDCache
-#endif
-
 #ifdef F___errno
 /* This is needed in case we are linked against libm (the math library) but
    not libc.  */