x1000: optimize crt0.S, improve correctness

Replace inline section copy/fill loops with subroutines, which
reduces code size a bit and and handle zero size copies properly.

Remove the cache initialization loop as well. There's no actual
reason for this because the SPL initializes the caches and just
dropping the cache can even be harmful (in this case it wasn't,
because the SPL flushes the whole cache right before calling in).

Change-Id: I7cddc9ed6d060b1f1bdd75544297883d014cad2d

firmware/target/mips/ingenic_x1000/crt0.S

@@ -49,75 +49,38 @@ _header:
     .ascii  "ENDH"          /* end of header structure */
 
 _realstart:
-    /* Cache init */
-    li      v0, 0x80000000
-    ori     v1, v0, 0x4000
-    mtc0    zero, C0_TAGLO
-    mtc0    zero, C0_TAGHI
-_cache_loop:
-    cache   ICIndexStTag, 0(v0)
-    cache   DCIndexStTag, 0(v0)
-    addiu   v0, v0, 32
-    bne     v0, v1, _cache_loop
-    nop
-
-    /* Invalidate BTB */
-    mfc0    v0, C0_Config, 7
-    nop
-    ori     v0, v0, 2
-    mtc0    v0, C0_Config, 7
-    nop
-
     /* Copy IRAM from BSS to low memory. */
-    la      t0, _iramcopy
-    la      t1, _iramstart
-    la      t2, _iramend
-_iram_loop:
-    lw      t3, 0(t0)
-    addiu   t1, 4
-    addiu   t0, 4
-    bne     t1, t2, _iram_loop
-    sw      t3, -4(t1)
+    la      a0, _iramcopy
+    la      a1, _iramstart
+    la      a2, _iramend
+    bal     _copy
+    nop
 
-#if 0
     /* Copy TCSM from BSS */
-    la      t0, _tcsmcopy
-    la      t1, _tcsmstart
-    la      t2, _tcsmend
-_tcsm_loop:
-    lw      t3, 0(t0)
-    addiu   t0, 4
-    sw      t3, 0(t1)
-    bne     t1, t2, _tcsm_loop
-    addiu   t1, 4
-#endif
+    la      a0, _tcsmcopy
+    la      a1, _tcsmstart
+    la      a2, _tcsmend
+    bal     _copy
+    nop
 
     /* Clear the BSS segment (needed to zero-initialize C static values) */
-    la      t0, _bssbegin
-    la      t1, _bssend
-    beq     t0, t1, _bss_done
-_bss_loop:
-    addiu   t0, 4
-    bne     t0, t1, _bss_loop
-    sw      zero, -4(t0)
-_bss_done:
+    la      a0, _bssbegin
+    la      a1, _bssend
+    bal     _clear
+    move    a2, $0
 
     /* Set stack pointer and clear the stack */
     la      sp, stackend
-    la      t0, stackbegin
-    li      t1, 0xDEADBEEF
-_stack_loop:
-    addiu   t0, 4
-    bne     t0, sp, _stack_loop
-    sw      t1, -4(t0)
+    la      a0, stackbegin
+    li      a2, 0xDEADBEEF
+    bal     _clear
+    move    a1, sp
 
     /* Clear the IRQ stack */
     la      k0, _irqstackend
-    la      t0, _irqstackbegin
-_irqstack_loop:
-    addiu   t0, 4
-    bne     t0, k0, _irqstack_loop
-    sw      t1, -4(t0)
+    la      a0, _irqstackbegin
+    bal     _clear
+    move    a1, k0
 
     /* Jump to C code */
     jal     system_early_init
@@ -125,6 +88,28 @@ _irqstack_loop:
     j       main
     nop
 
+    /* copy(void* src, void* dst, void* dst_end) */
+_copy:
+    beq     a1, a2, 1f
+    addiu   a1, 4
+    lw      t0, 0(a0)
+    addiu   a0, 4
+    b       _copy
+    sw      t0, -4(a1)
+1:
+    jr      ra
+    nop
+
+    /* clear(void* dst, void* dst_end, int value) */
+_clear:
+    beq     a0, a1, 1f
+    addiu   a0, 4
+    b       _clear
+    sw      a2, -4(a0)
+1:
+    jr      ra
+    nop
+
     /* Exception entry points */
     .section .vectors.1, "ax", %progbits
     j       tlb_refill_handler