On 4 April 2017 at 11:06, Heiko Stuebner heiko@sntech.de wrote:
Am Dienstag, 4. April 2017, 12:29:53 CEST schrieb Tom Rini:
On Fri, Mar 31, 2017 at 10:24:07PM -0600, Simon Glass wrote:
On 26 March 2017 at 16:38, Heiko Stuebner heiko@sntech.de wrote:
I've added Tom for comments, executive summary:
- rk3188-tpl is size limited to 1020 bytes
- gcc 6.3 produces a rk3188-tpl of 792 bytes
- gcc 4.9 makes it 1020 bytes
- buildman seems to always use gcc-4.9
- rk3188 board does not compile with buildman
Isn't holding on to a pretty old compiler for everything somewhat strange? ;-)
Well it's not that old. 4.6 would be old. We do need to be careful not to drop old toolchains too aggressively, although for new platforms such as this is doesn't matter. I try to test with older things to avoid problems applying things to mainline (with Tom's automated tests, etc.)
wrt buildman using certain toolchains, it comes down to the order in which it finds matches for a given arch and then it picks (and I don't recall which off the top of my head) the first or last match.
I do agree that gcc-4.9 isn't something we can drop just yet (as for example it's what'll be used in travis-ci today. But it's getting pretty long in the tooth and we will have to at some point say that "platform X requires gcc-6.x or later" as we start running into hard walls that are solved in 6.x.
Finally, I have no objection to adding TPL_USE_ARCH_MEMSET as an option so that in cases like this it can be disabled due to space just as it is on SPL.
The problem wasn't ARCH_MEMSET - which already was way outsize the size constraints, but the general memset also being somewhat big, with its slight speed optimizations.
But thanks to Simon's recent patches [0] we got a really nice size- reduction on the Rock's TPL (from 1020 to 488 bytes) . So with these we're really good with all toolchains now.
Heiko
[0] https://www.mail-archive.com/u-boot@lists.denx.de/msg243443.html
Well, hopefully for the last time, this patch:
Applied to u-boot-rockchip, thanks!