sh: Optimized cache handling for SH-4/SH-4A caches.

This reworks some of the SH-4 cache handling code to more easily
accomodate newer-style caches (particularly for the > direct-mapped
case), as well as optimizing some of the old code.

Signed-off-by: Richard Curnow <richard.curnow@st.com>
Signed-off-by: Paul Mundt <lethal@linux-sh.org>

1 files changed, 430 insertions, 87 deletions

diff --git a/arch/sh/mm/cache-sh4.c b/arch/sh/mm/cache-sh4.c
index 846b63d6f5e8..c036c2b4ac2b 100644
--- a/arch/sh/mm/cache-sh4.c
+++ b/arch/sh/mm/cache-sh4.c
@@ -25,28 +25,95 @@
 #include <asm/mmu_context.h>
 #include <asm/cacheflush.h>
 
-extern void __flush_cache_4096(unsigned long addr, unsigned long phys,
+static void __flush_dcache_segment_1way(unsigned long start,
+                                        unsigned long extent);
+static void __flush_dcache_segment_2way(unsigned long start,
+                                        unsigned long extent);
+static void __flush_dcache_segment_4way(unsigned long start,
+                                        unsigned long extent);
+
+static void __flush_cache_4096(unsigned long addr, unsigned long phys,
                                unsigned long exec_offset);
-extern void __flush_cache_4096_all(unsigned long start);
-static void __flush_cache_4096_all_ex(unsigned long start);
-extern void __flush_dcache_all(void);
-static void __flush_dcache_all_ex(void);
+
+/*
+ * This is initialised here to ensure that it is not placed in the BSS.  If
+ * that were to happen, note that cache_init gets called before the BSS is
+ * cleared, so this would get nulled out which would be hopeless.
+ */
+static void (*__flush_dcache_segment_fn)(unsigned long, unsigned long) =
+        (void (*)(unsigned long, unsigned long))0xdeadbeef;
+
+static void compute_alias(struct cache_info *c)
+{
+        c->alias_mask = ((c->sets - 1) << c->entry_shift) & ~(PAGE_SIZE - 1);
+        c->n_aliases = (c->alias_mask >> PAGE_SHIFT) + 1;
+}
+
+static void __init emit_cache_params(void)
+{
+        printk("PVR=%08x CVR=%08x PRR=%08x\n",
+                ctrl_inl(CCN_PVR),
+                ctrl_inl(CCN_CVR),
+                ctrl_inl(CCN_PRR));
+        printk("I-cache : n_ways=%d n_sets=%d way_incr=%d\n",
+                cpu_data->icache.ways,
+                cpu_data->icache.sets,
+                cpu_data->icache.way_incr);
+        printk("I-cache : entry_mask=0x%08x alias_mask=0x%08x n_aliases=%d\n",
+                cpu_data->icache.entry_mask,
+                cpu_data->icache.alias_mask,
+                cpu_data->icache.n_aliases);
+        printk("D-cache : n_ways=%d n_sets=%d way_incr=%d\n",
+                cpu_data->dcache.ways,
+                cpu_data->dcache.sets,
+                cpu_data->dcache.way_incr);
+        printk("D-cache : entry_mask=0x%08x alias_mask=0x%08x n_aliases=%d\n",
+                cpu_data->dcache.entry_mask,
+                cpu_data->dcache.alias_mask,
+                cpu_data->dcache.n_aliases);
+
+        if (!__flush_dcache_segment_fn)
+                panic("unknown number of cache ways\n");
+}
 
 /*
  * SH-4 has virtually indexed and physically tagged cache.
  */
 
-struct semaphore p3map_sem[4];
+/* Worst case assumed to be 64k cache, direct-mapped i.e. 4 synonym bits. */
+#define MAX_P3_SEMAPHORES 16
+
+struct semaphore p3map_sem[MAX_P3_SEMAPHORES];
 
 void __init p3_cache_init(void)
 {
-        if (remap_area_pages(P3SEG, 0, PAGE_SIZE*4, _PAGE_CACHABLE))
+        int i;
+
+        compute_alias(&cpu_data->icache);
+        compute_alias(&cpu_data->dcache);
+
+        switch (cpu_data->dcache.ways) {
+        case 1:
+                __flush_dcache_segment_fn = __flush_dcache_segment_1way;
+                break;
+        case 2:
+                __flush_dcache_segment_fn = __flush_dcache_segment_2way;
+                break;
+        case 4:
+                __flush_dcache_segment_fn = __flush_dcache_segment_4way;
+                break;
+        default:
+                __flush_dcache_segment_fn = NULL;
+                break;
+        }
+
+        emit_cache_params();
+
+        if (remap_area_pages(P3SEG, 0, PAGE_SIZE * 4, _PAGE_CACHABLE))
                 panic("%s failed.", __FUNCTION__);
 
-        sema_init (&p3map_sem[0], 1);
-        sema_init (&p3map_sem[1], 1);
-        sema_init (&p3map_sem[2], 1);
-        sema_init (&p3map_sem[3], 1);
+        for (i = 0; i < cpu_data->dcache.n_aliases; i++)
+                sema_init(&p3map_sem[i], 1);
 }
 
 /*
@@ -91,7 +158,6 @@ void __flush_purge_region(void *start, int size)
         }
 }
 
-
 /*
  * No write back please
  */
@@ -110,46 +176,6 @@ void __flush_invalidate_region(void *start, int size)
         }
 }
 
-static void __flush_dcache_all_ex(void)
-{
-        unsigned long addr, end_addr, entry_offset;
-
-        end_addr = CACHE_OC_ADDRESS_ARRAY +
-                (cpu_data->dcache.sets << cpu_data->dcache.entry_shift) *
-                 cpu_data->dcache.ways;
-
-        entry_offset = 1 << cpu_data->dcache.entry_shift;
-        for (addr = CACHE_OC_ADDRESS_ARRAY;
-             addr < end_addr;
-             addr += entry_offset) {
-                ctrl_outl(0, addr);
-        }
-}
-
-static void __flush_cache_4096_all_ex(unsigned long start)
-{
-        unsigned long addr, entry_offset;
-        int i;
-
-        entry_offset = 1 << cpu_data->dcache.entry_shift;
-        for (i = 0; i < cpu_data->dcache.ways;
-             i++, start += cpu_data->dcache.way_incr) {
-                for (addr = CACHE_OC_ADDRESS_ARRAY + start;
-                     addr < CACHE_OC_ADDRESS_ARRAY + 4096 + start;
-                     addr += entry_offset) {
-                        ctrl_outl(0, addr);
-                }
-        }
-}
-
-void flush_cache_4096_all(unsigned long start)
-{
-        if (cpu_data->dcache.ways == 1)
-                __flush_cache_4096_all(start);
-        else
-                __flush_cache_4096_all_ex(start);
-}
-
 /*
  * Write back the range of D-cache, and purge the I-cache.
  *
@@ -180,9 +206,11 @@ void flush_cache_sigtramp(unsigned long addr)
 
         local_irq_save(flags);
         jump_to_P2();
+
         for (i = 0; i < cpu_data->icache.ways;
              i++, index += cpu_data->icache.way_incr)
                 ctrl_outl(0, index);    /* Clear out Valid-bit */
+
         back_to_P1();
         wmb();
         local_irq_restore(flags);
@@ -194,8 +222,8 @@ static inline void flush_cache_4096(unsigned long start,
         unsigned long flags;
 
         /*
-         * SH7751, SH7751R, and ST40 have no restriction to handle cache.
-         * (While SH7750 must do that at P2 area.)
+         * All types of SH-4 require PC to be in P2 to operate on the I-cache.
+         * Some types of SH-4 require PC to be in P2 to operate on the D-cache.
          */
         if ((cpu_data->flags & CPU_HAS_P2_FLUSH_BUG)
            || start < CACHE_OC_ADDRESS_ARRAY) {
@@ -217,12 +245,13 @@ void flush_dcache_page(struct page *page)
 {
         if (test_bit(PG_mapped, &page->flags)) {
                 unsigned long phys = PHYSADDR(page_address(page));
+                unsigned long addr = CACHE_OC_ADDRESS_ARRAY;
+                int i, n;
 
                 /* Loop all the D-cache */
-                flush_cache_4096(CACHE_OC_ADDRESS_ARRAY,          phys);
-                flush_cache_4096(CACHE_OC_ADDRESS_ARRAY | 0x1000, phys);
-                flush_cache_4096(CACHE_OC_ADDRESS_ARRAY | 0x2000, phys);
-                flush_cache_4096(CACHE_OC_ADDRESS_ARRAY | 0x3000, phys);
+                n = cpu_data->dcache.n_aliases;
+                for (i = 0; i < n; i++, addr += PAGE_SIZE)
+                        flush_cache_4096(addr, phys);
         }
 
         wmb();
@@ -246,10 +275,7 @@ static inline void flush_icache_all(void)
 
 void flush_dcache_all(void)
 {
-        if (cpu_data->dcache.ways == 1)
-                __flush_dcache_all();
-        else
-                __flush_dcache_all_ex();
+        (*__flush_dcache_segment_fn)(0UL, cpu_data->dcache.way_size);
         wmb();
 }
 
@@ -261,6 +287,16 @@ void flush_cache_all(void)
 
 void flush_cache_mm(struct mm_struct *mm)
 {
+        /*
+         * Note : (RPC) since the caches are physically tagged, the only point
+         * of flush_cache_mm for SH-4 is to get rid of aliases from the
+         * D-cache.  The assumption elsewhere, e.g. flush_cache_range, is that
+         * lines can stay resident so long as the virtual address they were
+         * accessed with (hence cache set) is in accord with the physical
+         * address (i.e. tag).  It's no different here.  So I reckon we don't
+         * need to flush the I-cache, since aliases don't matter for that.  We
+         * should try that.
+         */
         flush_cache_all();
 }
 
@@ -273,24 +309,36 @@ void flush_cache_mm(struct mm_struct *mm)
 void flush_cache_page(struct vm_area_struct *vma, unsigned long address, unsigned long pfn)
 {
         unsigned long phys = pfn << PAGE_SHIFT;
+        unsigned int alias_mask;
+
+        alias_mask = cpu_data->dcache.alias_mask;
 
         /* We only need to flush D-cache when we have alias */
-        if ((address^phys) & CACHE_ALIAS) {
+        if ((address^phys) & alias_mask) {
                 /* Loop 4K of the D-cache */
                 flush_cache_4096(
-                        CACHE_OC_ADDRESS_ARRAY | (address & CACHE_ALIAS),
+                        CACHE_OC_ADDRESS_ARRAY | (address & alias_mask),
                         phys);
                 /* Loop another 4K of the D-cache */
                 flush_cache_4096(
-                        CACHE_OC_ADDRESS_ARRAY | (phys & CACHE_ALIAS),
+                        CACHE_OC_ADDRESS_ARRAY | (phys & alias_mask),
                         phys);
         }
 
-        if (vma->vm_flags & VM_EXEC)
-                /* Loop 4K (half) of the I-cache */
+        alias_mask = cpu_data->icache.alias_mask;
+        if (vma->vm_flags & VM_EXEC) {
+                /*
+                 * Evict entries from the portion of the cache from which code
+                 * may have been executed at this address (virtual).  There's
+                 * no need to evict from the portion corresponding to the
+                 * physical address as for the D-cache, because we know the
+                 * kernel has never executed the code through its identity
+                 * translation.
+                 */
                 flush_cache_4096(
-                        CACHE_IC_ADDRESS_ARRAY | (address & 0x1000),
+                        CACHE_IC_ADDRESS_ARRAY | (address & alias_mask),
                         phys);
+        }
 }
 
 /*
@@ -305,14 +353,28 @@ void flush_cache_page(struct vm_area_struct *vma, unsigned long address, unsigne
 void flush_cache_range(struct vm_area_struct *vma, unsigned long start,
                        unsigned long end)
 {
-        unsigned long p = start & PAGE_MASK;
+        unsigned long d = 0, p = start & PAGE_MASK;
+        unsigned long alias_mask = cpu_data->dcache.alias_mask;
+        unsigned long n_aliases = cpu_data->dcache.n_aliases;
+        unsigned long select_bit;
+        unsigned long all_aliases_mask;
+        unsigned long addr_offset;
+        unsigned long phys;
         pgd_t *dir;
         pmd_t *pmd;
         pud_t *pud;
         pte_t *pte;
         pte_t entry;
-        unsigned long phys;
-        unsigned long d = 0;
+        int i;
+
+        /*
+         * If cache is only 4k-per-way, there are never any 'aliases'.  Since
+         * the cache is physically tagged, the data can just be left in there.
+         */
+        if (n_aliases == 0)
+                return;
+
+        all_aliases_mask = (1 << n_aliases) - 1;
 
         /*
          * Don't bother with the lookup and alias check if we have a
@@ -335,39 +397,52 @@ void flush_cache_range(struct vm_area_struct *vma, unsigned long start,
 
         do {
                 if (pmd_none(*pmd) || pmd_bad(*pmd)) {
-                        p &= ~((1 << PMD_SHIFT) -1);
+                        p &= ~((1 << PMD_SHIFT) - 1);
                         p += (1 << PMD_SHIFT);
                         pmd++;
+
                         continue;
                 }
+
                 pte = pte_offset_kernel(pmd, p);
+
                 do {
                         entry = *pte;
+
                         if ((pte_val(entry) & _PAGE_PRESENT)) {
-                                phys = pte_val(entry)&PTE_PHYS_MASK;
-                                if ((p^phys) & CACHE_ALIAS) {
-                                        d |= 1 << ((p & CACHE_ALIAS)>>12);
-                                        d |= 1 << ((phys & CACHE_ALIAS)>>12);
-                                        if (d == 0x0f)
+                                phys = pte_val(entry) & PTE_PHYS_MASK;
+
+                                if ((p ^ phys) & alias_mask) {
+                                        d |= 1 << ((p & alias_mask) >> PAGE_SHIFT);
+                                        d |= 1 << ((phys & alias_mask) >> PAGE_SHIFT);
+
+                                        if (d == all_aliases_mask)
                                                 goto loop_exit;
                                 }
                         }
+
                         pte++;
                         p += PAGE_SIZE;
                 } while (p < end && ((unsigned long)pte & ~PAGE_MASK));
                 pmd++;
         } while (p < end);
- loop_exit:
-        if (d & 1)
-                flush_cache_4096_all(0);
-        if (d & 2)
-                flush_cache_4096_all(0x1000);
-        if (d & 4)
-                flush_cache_4096_all(0x2000);
-        if (d & 8)
-                flush_cache_4096_all(0x3000);
-        if (vma->vm_flags & VM_EXEC)
+
+loop_exit:
+        for (i = 0, select_bit = 0x1, addr_offset = 0x0; i < n_aliases;
+             i++, select_bit <<= 1, addr_offset += PAGE_SIZE)
+                if (d & select_bit) {
+                        (*__flush_dcache_segment_fn)(addr_offset, PAGE_SIZE);
+                        wmb();
+                }
+
+        if (vma->vm_flags & VM_EXEC) {
+                /*
+                 * TODO: Is this required???  Need to look at how I-cache
+                 * coherency is assured when new programs are loaded to see if
+                 * this matters.
+                 */
                 flush_icache_all();
+        }
 }
 
 /*
@@ -384,3 +459,271 @@ void flush_icache_user_range(struct vm_area_struct *vma,
         mb();
 }
 
+/**
+ * __flush_cache_4096
+ *
+ * @addr:  address in memory mapped cache array
+ * @phys:  P1 address to flush (has to match tags if addr has 'A' bit
+ *         set i.e. associative write)
+ * @exec_offset: set to 0x20000000 if flush has to be executed from P2
+ *               region else 0x0
+ *
+ * The offset into the cache array implied by 'addr' selects the
+ * 'colour' of the virtual address range that will be flushed.  The
+ * operation (purge/write-back) is selected by the lower 2 bits of
+ * 'phys'.
+ */
+static void __flush_cache_4096(unsigned long addr, unsigned long phys,
+                               unsigned long exec_offset)
+{
+        int way_count;
+        unsigned long base_addr = addr;
+        struct cache_info *dcache;
+        unsigned long way_incr;
+        unsigned long a, ea, p;
+        unsigned long temp_pc;
+
+        dcache = &cpu_data->dcache;
+        /* Write this way for better assembly. */
+        way_count = dcache->ways;
+        way_incr = dcache->way_incr;
+
+        /*
+         * Apply exec_offset (i.e. branch to P2 if required.).
+         *
+         * FIXME:
+         *
+         *      If I write "=r" for the (temp_pc), it puts this in r6 hence
+         *      trashing exec_offset before it's been added on - why?  Hence
+         *      "=&r" as a 'workaround'
+         */
+        asm volatile("mov.l 1f, %0\n\t"
+                     "add   %1, %0\n\t"
+                     "jmp   @%0\n\t"
+                     "nop\n\t"
+                     ".balign 4\n\t"
+                     "1:  .long 2f\n\t"
+                     "2:\n" : "=&r" (temp_pc) : "r" (exec_offset));
+
+        /*
+         * We know there will be >=1 iteration, so write as do-while to avoid
+         * pointless nead-of-loop check for 0 iterations.
+         */
+        do {
+                ea = base_addr + PAGE_SIZE;
+                a = base_addr;
+                p = phys;
+
+                do {
+                        *(volatile unsigned long *)a = p;
+                        /*
+                         * Next line: intentionally not p+32, saves an add, p
+                         * will do since only the cache tag bits need to
+                         * match.
+                         */
+                        *(volatile unsigned long *)(a+32) = p;
+                        a += 64;
+                        p += 64;
+                } while (a < ea);
+
+                base_addr += way_incr;
+        } while (--way_count != 0);
+}
+
+/*
+ * Break the 1, 2 and 4 way variants of this out into separate functions to
+ * avoid nearly all the overhead of having the conditional stuff in the function
+ * bodies (+ the 1 and 2 way cases avoid saving any registers too).
+ */
+static void __flush_dcache_segment_1way(unsigned long start,
+                                        unsigned long extent_per_way)
+{
+        unsigned long orig_sr, sr_with_bl;
+        unsigned long base_addr;
+        unsigned long way_incr, linesz, way_size;
+        struct cache_info *dcache;
+        register unsigned long a0, a0e;
+
+        asm volatile("stc sr, %0" : "=r" (orig_sr));
+        sr_with_bl = orig_sr | (1<<28);
+        base_addr = ((unsigned long)&empty_zero_page[0]);
+
+        /*
+         * The previous code aligned base_addr to 16k, i.e. the way_size of all
+         * existing SH-4 D-caches.  Whilst I don't see a need to have this
+         * aligned to any better than the cache line size (which it will be
+         * anyway by construction), let's align it to at least the way_size of
+         * any existing or conceivable SH-4 D-cache.  -- RPC
+         */
+        base_addr = ((base_addr >> 16) << 16);
+        base_addr |= start;
+
+        dcache = &cpu_data->dcache;
+        linesz = dcache->linesz;
+        way_incr = dcache->way_incr;
+        way_size = dcache->way_size;
+
+        a0 = base_addr;
+        a0e = base_addr + extent_per_way;
+        do {
+                asm volatile("ldc %0, sr" : : "r" (sr_with_bl));
+                asm volatile("movca.l r0, @%0\n\t"
+                             "ocbi @%0" : : "r" (a0));
+                a0 += linesz;
+                asm volatile("movca.l r0, @%0\n\t"
+                             "ocbi @%0" : : "r" (a0));
+                a0 += linesz;
+                asm volatile("movca.l r0, @%0\n\t"
+                             "ocbi @%0" : : "r" (a0));
+                a0 += linesz;
+                asm volatile("movca.l r0, @%0\n\t"
+                             "ocbi @%0" : : "r" (a0));
+                asm volatile("ldc %0, sr" : : "r" (orig_sr));
+                a0 += linesz;
+        } while (a0 < a0e);
+}
+
+static void __flush_dcache_segment_2way(unsigned long start,
+                                        unsigned long extent_per_way)
+{
+        unsigned long orig_sr, sr_with_bl;
+        unsigned long base_addr;
+        unsigned long way_incr, linesz, way_size;
+        struct cache_info *dcache;
+        register unsigned long a0, a1, a0e;
+
+        asm volatile("stc sr, %0" : "=r" (orig_sr));
+        sr_with_bl = orig_sr | (1<<28);
+        base_addr = ((unsigned long)&empty_zero_page[0]);
+
+        /* See comment under 1-way above */
+        base_addr = ((base_addr >> 16) << 16);
+        base_addr |= start;
+
+        dcache = &cpu_data->dcache;
+        linesz = dcache->linesz;
+        way_incr = dcache->way_incr;
+        way_size = dcache->way_size;
+
+        a0 = base_addr;
+        a1 = a0 + way_incr;
+        a0e = base_addr + extent_per_way;
+        do {
+                asm volatile("ldc %0, sr" : : "r" (sr_with_bl));
+                asm volatile("movca.l r0, @%0\n\t"
+                             "movca.l r0, @%1\n\t"
+                             "ocbi @%0\n\t"
+                             "ocbi @%1" : :
+                             "r" (a0), "r" (a1));
+                a0 += linesz;
+                a1 += linesz;
+                asm volatile("movca.l r0, @%0\n\t"
+                             "movca.l r0, @%1\n\t"
+                             "ocbi @%0\n\t"
+                             "ocbi @%1" : :
+                             "r" (a0), "r" (a1));
+                a0 += linesz;
+                a1 += linesz;
+                asm volatile("movca.l r0, @%0\n\t"
+                             "movca.l r0, @%1\n\t"
+                             "ocbi @%0\n\t"
+                             "ocbi @%1" : :
+                             "r" (a0), "r" (a1));
+                a0 += linesz;
+                a1 += linesz;
+                asm volatile("movca.l r0, @%0\n\t"
+                             "movca.l r0, @%1\n\t"
+                             "ocbi @%0\n\t"
+                             "ocbi @%1" : :
+                             "r" (a0), "r" (a1));
+                asm volatile("ldc %0, sr" : : "r" (orig_sr));
+                a0 += linesz;
+                a1 += linesz;
+        } while (a0 < a0e);
+}
+
+static void __flush_dcache_segment_4way(unsigned long start,
+                                        unsigned long extent_per_way)
+{
+        unsigned long orig_sr, sr_with_bl;
+        unsigned long base_addr;
+        unsigned long way_incr, linesz, way_size;
+        struct cache_info *dcache;
+        register unsigned long a0, a1, a2, a3, a0e;
+
+        asm volatile("stc sr, %0" : "=r" (orig_sr));
+        sr_with_bl = orig_sr | (1<<28);
+        base_addr = ((unsigned long)&empty_zero_page[0]);
+
+        /* See comment under 1-way above */
+        base_addr = ((base_addr >> 16) << 16);
+        base_addr |= start;
+
+        dcache = &cpu_data->dcache;
+        linesz = dcache->linesz;
+        way_incr = dcache->way_incr;
+        way_size = dcache->way_size;
+
+        a0 = base_addr;
+        a1 = a0 + way_incr;
+        a2 = a1 + way_incr;
+        a3 = a2 + way_incr;
+        a0e = base_addr + extent_per_way;
+        do {
+                asm volatile("ldc %0, sr" : : "r" (sr_with_bl));
+                asm volatile("movca.l r0, @%0\n\t"
+                             "movca.l r0, @%1\n\t"
+                             "movca.l r0, @%2\n\t"
+                             "movca.l r0, @%3\n\t"
+                             "ocbi @%0\n\t"
+                             "ocbi @%1\n\t"
+                             "ocbi @%2\n\t"
+                             "ocbi @%3\n\t" : :
+                             "r" (a0), "r" (a1), "r" (a2), "r" (a3));
+                a0 += linesz;
+                a1 += linesz;
+                a2 += linesz;
+                a3 += linesz;
+                asm volatile("movca.l r0, @%0\n\t"
+                             "movca.l r0, @%1\n\t"
+                             "movca.l r0, @%2\n\t"
+                             "movca.l r0, @%3\n\t"
+                             "ocbi @%0\n\t"
+                             "ocbi @%1\n\t"
+                             "ocbi @%2\n\t"
+                             "ocbi @%3\n\t" : :
+                             "r" (a0), "r" (a1), "r" (a2), "r" (a3));
+                a0 += linesz;
+                a1 += linesz;
+                a2 += linesz;
+                a3 += linesz;
+                asm volatile("movca.l r0, @%0\n\t"
+                             "movca.l r0, @%1\n\t"
+                             "movca.l r0, @%2\n\t"
+                             "movca.l r0, @%3\n\t"
+                             "ocbi @%0\n\t"
+                             "ocbi @%1\n\t"
+                             "ocbi @%2\n\t"
+                             "ocbi @%3\n\t" : :
+                             "r" (a0), "r" (a1), "r" (a2), "r" (a3));
+                a0 += linesz;
+                a1 += linesz;
+                a2 += linesz;
+                a3 += linesz;
+                asm volatile("movca.l r0, @%0\n\t"
+                             "movca.l r0, @%1\n\t"
+                             "movca.l r0, @%2\n\t"
+                             "movca.l r0, @%3\n\t"
+                             "ocbi @%0\n\t"
+                             "ocbi @%1\n\t"
+                             "ocbi @%2\n\t"
+                             "ocbi @%3\n\t" : :
+                             "r" (a0), "r" (a1), "r" (a2), "r" (a3));
+                asm volatile("ldc %0, sr" : : "r" (orig_sr));
+                a0 += linesz;
+                a1 += linesz;
+                a2 += linesz;
+                a3 += linesz;
+        } while (a0 < a0e);
+}
+