Lpc2000

Maybe someone can help me to clarify the question with GPIO timing.

My intention is to move some data from flash to port0 on LPC2138 as fast as 
possible.

I tried few ways using Keil compiler and MCB2130 board.

1. LDR   R0, =flash_array       ;set flash array pointer
    LDR   R1,=0xE0028000    ;port0 PIN address (to set 0s and 1s 
simultaneously)

   LDR    R2,[R0],#4              ;load flash word to R2
   STR    R2,[R1]                   ;send R2 to port0

   LDR    R2,[R0],#4              ;next word
   STR    R2,[R1]                   ;
....... more words sent

according to the Keil simulator (and ARM7-TDMI-S timing description from the 
manual)
loading from flash to register takes 2 cycles, storing at the port address - 
3 cycles

Scoping port shows 10 cycles.

2. Doing the same loading data from RAM array gives the same results - 10 
cycles (should be 5 according to simulator and manual).

3. Preloading data to registers and moving to port:

    LDR   R0, =flash_array       ;set flash array pointer
    LDR   R1,=0xE0028000    ;port0 PIN address

    LDR   R2,[R0],#4
    LDR   R3,[R0],#4
    LDR   R4,[R0],#4
    .....

    STR   R2,[R1]
    STR   R3,[R1]
    STR   R4,[R1]
   ...
according to simulator and manual - 2 cycles.
scope - 7 cycles.
=================================
As far as I understand, storing data at the port address adds 5 misterious 
cycles with any method, loading data timing is accurate.
I tried different settings, it seems that only one affecting timing is VPB 
clock setting - above timing is for PCLK = CCLK.

P.S. I am using in-line assembly (one for all array move), because I can't 
get assembly function to work with Keil, but I don't think it matters.

Any idea where do these cycles come from?

Thank you,
Gennady