[llvm] r268036 - [mips][ias] Make section sizes a multiple of the alignment.

[Daniel Sanders via llvm-commits]

> -----Original Message-----
> From: Rafael Espíndola [mailto:rafael.espindola at gmail.com]
> Sent: 03 May 2016 21:11
> To: Daniel Sanders
> Cc: llvm-commits
> Subject: Re: [llvm] r268036 - [mips][ias] Make section sizes a multiple of the
> alignment.
> 
> On 2 May 2016 at 10:54, Daniel Sanders <Daniel.Sanders at imgtec.com>
> wrote:
> > Hi,
> >
> > It's part of my strategy for enabling IAS by default. I'm currently using
> objdump and diff to verify that IAS and GAS produce the same objects
> (within reason, I'm ignoring debug sections for example) and section size
> rounding had been the source of a lot of noise in the diffs. Making the
> section sizes match was a key part of finding the double expansion bugs
> (delay slots being filled twice for -save-temps, and similar), and the
> initialization order bug (.cprestore not working correctly without -save-
> temps). Without this change, those bugs would have been very difficult to
> spot amongst the noise.
> >
> > Once we're fully on IAS then the slight size saving will be more valuable
> than the ability to verify IAS and it will make sense to either revert this
> change or disable it by default. There are some big problems to fix before
> everything is on IAS though. The main one is that N32 is badly broken in IAS
> which also prevents N64 from being enabled by default (because the code
> can't distinguish them yet). It's probably going to take a couple months to fix
> that assuming all goes well.
> 
> I really worry about cargo cults like this. I say so because I am
> guilty of having done it in the past. For example, we used to create
> the same dummy sections as gas, and it gets hard to move out of it.
> 
> Can you move this behind a command line option that is disabled by
> default and use that in your testing? That way it is way less likely
> that we will develop a dependency on it.
> 
> Cheers,
> Rafael

Makes sense to me. I've made this change in r268496.

Thanks