Nxp Semiconductors UM10237 ユーザーズマニュアル

UM10237_2

© NXP B.V. 2008. All rights reserved.

 

The Memory Accelerator Module operates in combination with the flash controller and is 
available in parts LPC2458/68/78.

The MAM block in the LPC2400 maximizes the performance of the ARM processor when 
it is running code in Flash memory using a single Flash bank.

Simply put, the Memory Accelerator Module (MAM) attempts to have the next ARM 
instruction that will be needed in its latches in time to prevent CPU fetch stalls. The 
LPC2400 uses one bank of Flash memory, compared to the two banks used on 
predecessor devices. It includes three 128 bit buffers called the Prefetch buffer, the 
Branch Trail Buffer and the data buffer. When an Instruction Fetch is not satisfied by either 
the Prefetch or Branch Trail buffer, nor has a prefetch been initiated for that line, the ARM 
is stalled while a fetch is initiated for the 128 bit line. If a prefetch has been initiated but not 
yet completed, the ARM is stalled for a shorter time. Unless aborted by a data access, a 
prefetch is initiated as soon as the Flash has completed the previous access. The 
prefetched line is latched by the Flash module, but the MAM does not capture the line in 
its prefetch buffer until the ARM core presents the address from which the prefetch has 
been made. If the core presents a different address from the one from which the prefetch 
has been made, the prefetched line is discarded.

The prefetch and Branch Trail buffers each include four 32 bit ARM instructions or eight 
16 bit Thumb instructions. During sequential code execution, typically the prefetch buffer 
contains the current instruction and the entire Flash line that contains it.

The MAM uses the LPROT[0] line to differentiate between instruction and data accesses. 
Code and data accesses use separate 128 bit buffers. 3 of every 4 sequential 32 bit code 
or data accesses "hit" in the buffer without requiring a Flash access (7 of 8 sequential 
16 bit accesses, 15 of every 16 sequential byte accesses). The fourth (eighth, 16th) 
sequential data access must access Flash, aborting any prefetch in progress. When a 
Flash data access is concluded, any prefetch that had been in progress is re-initiated.

Timing of Flash read operations is programmable and is described later in this section.

In this manner, there is no code fetch penalty for sequential instruction execution when the 
CPU clock period is greater than or equal to one fourth of the Flash access time. The 
average amount of time spent doing program branches is relatively small (less than 25%) 
and may be minimized in ARM (rather than Thumb) code through the use of the 
conditional execution feature present in all ARM instructions. This conditional execution 
may often be used to avoid small forward branches that would otherwise be necessary.