Move the gp register, to allow the compiler to leverage gp to access …

[cnlohr]

…the first part of ram. This produces smaller executables in general. I don't know why the original startup code pointed at 0x20000800. You can reach all of the first part of ram by setting it to 0x200003fc

[TommyMurphyTM1234]

I don't know why the original startup code pointed at 0x20000800.

I don't know if this is discussion about $gp might help at all?

• https://groups.google.com/a/groups.riscv.org/g/sw-dev/c/60IdaZj27dY?pli=1

[monte-monte]

Yeah, it seems they just copepasted it from the chips with bigger ram. But if gp is used as described here https://gnu-mcu-eclipse.github.io/arch/riscv/programmer/#the-gp-global-pointer-register why does this change produces effect? If all 2K fit in negative region of 12bit offset?

[cnlohr]

It looks like it might be primarily a GCC bug. It SHOULD be able to leverage 0x20000800 But, in practice, even with GCC v14.2.0-3 it still doesn't know how to use it appropriately.

[TommyMurphyTM1234]

It looks like it might be primarily a GCC bug. It SHOULD be able to leverage 0x20000800 But, in practice, even with GCC v14.2.0-3 it still doesn't know how to use it appropriately.

Do you happen to have a reproducible test case? If this is a bug in GCC then it should ideally be logged upstream. Although isn't the $gp primarily used for linker relaxation so any issue in this context may actually be with the Binutils linker rather than GCC per se?

[monte-monte]

@TommyMurphyTM1234 I compiled a project with different values for gp. Only when offset is a big negative number gcc seems to ignore the existence of gp. If you set gp = 0 then it works (the positive portion of 12bit)
I was building a rv003usb bootloader (for reference).

[cnlohr]

Before submitting this, I also want to try adding -ffunction-sections - there seems to be some situations where GCC will not strip unused functions.

Also, as we discussed on the Discord. We cannot find a reason why GCC can't use the full +/- 2k Range, but, it does seem to effectively use the +/- 1k range.

[cnlohr]

We already use -ffunction-sections it turned out, and I was just misusing the cflags in my test. All is well. Still need to make this change so I am submitting it.