[PATCH][AArch64] Improve legitimize_address

[Wilco Dijkstra]

Improve aarch64_legitimize_address - avoid splitting the offset if it is
supported.  When we do split, take the mode size into account.  BLKmode
falls into the unaligned case but should be treated like LDP/STP.
This improves codesize slightly due to fewer base address calculations:

int f(int *p) { return p[5000] + p[7000]; }

Now generates:

f:
        add     x0, x0, 16384
        ldr     w1, [x0, 3616]
        ldr     w0, [x0, 11616]
        add     w0, w1, w0
        ret

instead of:

f:
        add     x1, x0, 16384
        add     x0, x0, 24576
        ldr     w1, [x1, 3616]
        ldr     w0, [x0, 3424]
        add     w0, w1, w0
        ret

OK for trunk?

ChangeLog:
2016-09-06  Wilco Dijkstra  <wdijk...@arm.com>

    gcc/
        * config/aarch64/aarch64.c (aarch64_legitimize_address):
        Avoid use of base_offset if offset already in range.
--

diff --git a/gcc/config/aarch64/aarch64.c b/gcc/config/aarch64/aarch64.c
index 
27bbdbad8cddc576f9ed4fd0670116bd6d318412..119ff0aecb0c9f88899fa141b2c7f9158281f9c3
 100644
--- a/gcc/config/aarch64/aarch64.c
+++ b/gcc/config/aarch64/aarch64.c
@@ -5058,9 +5058,19 @@ aarch64_legitimize_address (rtx x, rtx /* orig_x  */, 
machine_mode mode)
       /* For offsets aren't a multiple of the access size, the limit is
         -256...255.  */
       else if (offset & (GET_MODE_SIZE (mode) - 1))
-       base_offset = (offset + 0x100) & ~0x1ff;
+       {
+         base_offset = (offset + 0x100) & ~0x1ff;
+
+         /* BLKmode typically uses LDP of X-registers.  */
+         if (mode == BLKmode)
+           base_offset = (offset + 512) & ~0x3ff;
+       }
+      /* Small negative offsets are supported.  */
+      else if (IN_RANGE (offset, -256, 0))
+       base_offset = 0;
+      /* Use 12-bit offset by access size.  */
       else
-       base_offset = offset & ~0xfff;
+       base_offset = offset & (~0xfff * GET_MODE_SIZE (mode));
 
       if (base_offset != 0)
        {