[PATCH u-boot v4 36/36] ARM: enable LTO for some boards

[Tom Rini]

On Mon, May 24, 2021 at 09:19:35PM +0200, Marek Behun wrote:
> On Mon, 24 May 2021 13:44:38 -0400
> Tom Rini <trini at konsulko.com> wrote:
> 
> > On Mon, May 24, 2021 at 01:09:19PM -0400, Tom Rini wrote:
> > > On Mon, May 24, 2021 at 05:58:55PM +0200, Marek Behun wrote:  
> > > > On Mon, 24 May 2021 11:40:53 -0400
> > > > Tom Rini <trini at konsulko.com> wrote:
> > > >   
> > > > > On Fri, May 21, 2021 at 12:56:41PM -0400, Tom Rini wrote:  
> > > > > > On Fri, May 21, 2021 at 06:00:31PM +0200, Marek Behún wrote:    
> > > > > > > On Fri, 21 May 2021 10:11:47 -0400
> > > > > > > Tom Rini <trini at konsulko.com> wrote:
> > > > > > >     
> > > > > > > > On Thu, May 20, 2021 at 01:56:29PM -0500, Adam Ford wrote:    
> > > > > > > > > On Thu, May 20, 2021 at 6:25 AM Marek Behún <marek.behun at nic.cz>
> > > > > > > > > wrote:      
> > > > > > > > > >
> > > > > > > > > > Enable LTO for some boards that were tested by people on U-Boot
> > > > > > > > > > Mailing List.
> > > > > > > > > >
> > > > > > > > > > Signed-off-by: Marek Behún <marek.behun at nic.cz>
> > > > > > > > > > Tested-by: Adam Ford <aford173 at gmail.com>
> > > > > > > > > > Tested-by: Pali Rohár <pali at kernel.org>
> > > > > > > > > > Tested-by: Tim Harvey <tharvey at gateworks.com>      
> > > > > > > > > 
> > > > > > > > > Since the imx8mm beacon boards and the imx8mm venice board both show
> > > > > > > > > promise, does it make sense to 'imply' the LTO for anything enabling
> > > > > > > > > imx8mm?
> > > > > > > > > Same thing for the various omap3 boards, and potentially the renesas
> > > > > > > > > RZ/G2 boards.  I know Tom went through to remove a bunch of boards
> > > > > > > > > that were never converted to DM.  Most of the boards remaining
> > > > > > > > > boards have minimal board files and most of code is common to other
> > > > > > > > > boards in the same platforms.
> > > > > > > > > 
> > > > > > > > > I have an l138_lcdk that I can use to test which I expect to be
> > > > > > > > > similar to the da850evm.      
> > > > > > > > 
> > > > > > > > As much as I am eager to move everything, quickly, over to LTO by
> > > > > > > > default, I think the problems that we've seen thus far show it's best
> > > > > > > > to really make it an explicit enable per board at least for the first
> > > > > > > > release or two.  Once we've hopefully gotten more boards tested and
> > > > > > > > enabled we can see what makes sense for defaults, give a release worth
> > > > > > > > of heads up, and then go.    
> > > > > > > 
> > > > > > > Tom, are there some other issues aside from the one failing CI scenario
> > > > > > > (sandbox_clang)? Would you be willing to merge this if I resolved that
> > > > > > > one fail by disabling LTO for that scenario (until I resolve it)? It
> > > > > > > would help me not having to maintain all 30+ patches...    
> > > > > > 
> > > > > > Yeah, CI needs to keep passing, so if we need to disable
> > > > > > sandbox+clang+lto for now, OK.    
> > > > > 
> > > > > Ah, I see the problem now.  I've worked out a fix after looking at the
> > > > > Linux kernel a bit and I'll post something for us and upstream dtc as
> > > > > well.
> > > > >   
> > > > 
> > > > What do you mean? The problem is in dtc? I see 2 problems:
> > > > - one with DM test
> > > > - one with stack protector test  
> > > 
> > > I don't have a full answer about the stack protector test just yet, but
> > > it almost seems like it's too simple and maybe something is happening
> > > with it being optimized to not a problem?  
> > 
> > Yeah, so clang with LTO optimizes away that memset call, and so the test
> > passes.  I'll do something to make sure the array is used so it won't be
> > optimized away.
> > 
> 
> I am unable to make the compiler to protect the stack of that function
> even with GCC on my local machine. It seems that at least on my gentoo
> with gcc-10.2, when compiling with -ffreestanding, the call to
> __stack_chk_fail is not made at all.
> 
> I even started reading sources of gcc on thursday because of this, but
> it didn't lead anywhere...
> 
> When you compile sandbox_defconfig with gcc, does the test pass on your
> local machine?

It passes here with gcc-7.5 as well as gcc-9.2.0 in CI.  When I moved to
a gcc-11 snapshot in CI, it was also passing there before, so I'm not
sure what's going on in your setup.  But as part of coming up with the
patch I sent to fix the test, I ran it manually and saw the expected
failure message.

-- 
Tom
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 659 bytes
Desc: not available
URL: <https://lists.denx.de/pipermail/u-boot/attachments/20210524/b7c24a85/attachment.sig>