[clang] Patch series to reapply #118734 and substantially improve it (PR #120534)

[via cfe-commits]

dyung wrote:

> > > > > > I'll try it and let you know. Give me about an hour or so.
> > > > > 
> > > > > 
> > > > > Awesome! But no huge rush, mostly just hoping this happens to dodge whatever has been tripping up things here.
> > > > 
> > > > 
> > > > Sorry for the delay, but the failures still seem to be present. :( (The tests are still running, but the amount of output being generated is slowing things down).
> > > 
> > > 
> > > Strange... Is it the _same_ output??
> > > Can you upload the output somewhere?
> > 
> > 
> > Just to add some more details now that I've slept a bit...
> > Previously there were errors in AArch64 and RISCV -- it'll be really useful to know if those are the only errors with this patch, are there new ones, and especially if the RISCV errors go away that might be helpful. If the errors are _only_ in AArch64, then I have ideas to further improve this.
> > Also, last time that @zmodem helped look at the `.obj` file for this, I think we ended up looking at the wrong one -- I don't think it is the `clang/lib/Basic/Builtins.obj` that would show signs of the issue, but `clang/lib/Basic/Targets/AArch64.obj` (and potentially the `RISCV.obj`). Either those `.obj` files should be clearly wrong (IE, some of the arrays truncated / zeroed /etc), or this is a mis-_link_ not a mis-_compile_...
> > But my goal is to somewhat experimentally find the threshold below which things start working. With this patch series there is an architecture to mechanically cut things down more and more if needed, but need some idea of the target I should aim for.
> 
> Sure. I deleted the object and log files, but can easily run it again. IIRC the tests that I saw failing in the console were LoongArch, but I think many/most targets were affected.
> 
> I'll take the buildbot offline, run it again and then archive the logs and object files somewhere.

Here is the logfile from the run. This is doing the exact same run as the buildbot (I ran the same commands as the buildbot would run). The compressed logfile is around 15MB, the uncompressed is around 300MB.

https://www.dropbox.com/scl/fi/urcanfwpf6da1dz2i5t85/log_test.zip?rlkey=5ludb4vpoe32mfp7hlnelrr47&dl=0

It's going to take a while to extract the build directory as it is 8GB in size, but if you have specific files you want to see let me know and I can pick them out.

Note that our bot doesn't build ARM/AArch64, so it would make sense that we don't see failures there (although I can't remember if we did previously...)

https://github.com/llvm/llvm-project/pull/120534