Cortex A: fix extra memory read and non-word sizes

Cortex A: fix extra memory read and non-word sizes

Without this patch, to perform a memory read, OpenOCD first issues an
LDC instruction into DBGITR in Stall mode (thus executing the
instruction), then switches to Fast mode and reads from DBGDTRTX once
for each word to transfer.

At the very end of the transfer, the final Fast mode read of DBGDTRTX
has, as always, the side effect of re-issuing the LDC instruction. This
causes two problems:

(1) If the word immediately beyond the end of the requested region is
inaccessible, this spurious LDC will cause a fault. On a fast CPU, the
LDC will finish executing by the time the poll of DSCR takes place,
failing the entire memory read. On a slow CPU, the LDC might finish
executing later, leaving an unexpected and confusing sticky fault lying
around for the next operation to see.

(2) If the LDC succeeds, it will leave the loaded word in DBGDTRTX, thus
setting DBGDSCR.TXFULL=1. The cortex_a_read_apb_ab_memory routine
completes without consuming that last word, thus confusing the next
routine that tries to use DBGDTRTX (this may not have any visible effect
on some implementations, because writing to DBGDTRTXint when TXFULL=1 is
defined as Unpredictable, but I believe it caused a visible problem for
me).

With this patch, the bulk mem_ap_sel_read_buf_noincr is modified to omit
the last word of the block. The second-to-last read of DBGDTRTX by that
function will cause the issue of the LDC for the last word. After
switching back to Normal mode and waiting for that instruction to
finish, do a final read of DBGDTRTX to extract the last word into the
buffer, leaving TXFULL=0.

Without this patch, memory accesses are always expanded such that they
are aligned to the access size. With this patch, accesses are issued
exactly as ordered by the caller. The caller is expected to handle
fragments at the beginning and end of the transfer if the address is
unaligned and an unaligned access is not desired.

Without this patch, the DFAR and DFSR registers, which report the
location and status of data faults, are ignored while performing memory
accesses, which could cause problems debugging an OS page fault handler.
With this patch, DFAR and DFSR are preserved across memory accesses, and
DFSR is decoded in the event of a synchronous fault to provide the
caller with more information about the reason for failure.

Thanks to Boris Brezillon for the original patch whose ideas led to the
non-word access mechanism implemented here and to various code reviewers
for their comments.

Change-Id: I11ae7104fbe69a522efadefc705c9a217a7eef41
Signed-off-by: Christopher Head <chead@zaber.com>
Reviewed-on: http://openocd.zylin.com/2381
Tested-by: jenkins
Reviewed-by: Olivier Schonken <olivier.schonken@gmail.com>
Reviewed-by: Andreas Fritiofson <andreas.fritiofson@gmail.com>