Hey all,
As Allen Martin noted, on tegra platforms a bug is exposed when using certain toolchains that currently calls to lowlevel_init must be calls to another assembly function as the stack is not explicitly setup / saved and thus register corruption can occur. Over in TI-land Aneesh V wrote a lowlevel_init that sets up the stack and calls s_init which is a C function that does what lowlevel_init does on other platforms. The code is generic to at least v7, so this series moves it to arch/arm/cpu/armv7. The initial version of the code made use of LOW_LEVEL_SRAM_STACK as the stack to be setup. This define is essentially another name for CONFIG_SYS_INIT_SP_ADDR without taking GENERATED_GBL_DATA_SIZE into consideration. So we switch to that instead. Finally, Wolfgang question me as to if the stack really only needed the alignment that we had been giving it. To be safe, I grabbed the alignmnet that we do when setting up the initial stack for real and placed it into lowlevel_init as well.
Tested on AM335x GP EVM and OMAP4 Pandaboard ES2.