The link you posted refers only to lr(R13), sp(R14), pc(R5) for legacy code support of code for old versions of ARM ADS and does not apply to general-purpose registers.
In ARM's compiler (also used in Keil's MDK-ARM):
The inline assembler provides no direct access to the physical registers
of an ARM processor. If an ARM register name is used as an operand in
an inline assembler instruction it becomes a reference to a variable
of the same name, and not the physical ARM register.
(Ref: Inline assembler and register access)
Inline assembler in ARM's compiler is subject to optimisation like the C or C++ code it is in-lined within, as such the compiler may generate code that differs from that you have written in any case. If you want assembler code to be generated exactly as you have written you must use embedded assembler rather the inline assembler