Re: Bountysource campaign for gcc-vax

["Maciej W. Rozycki" <macro%linux-mips.org@localhost>]

On Mon, 30 Nov 2020, Maciej W. Rozycki wrote:

> > Any update on when these will get merged? Looks like Jeff is fine with most
> > of the patches.
> 
>  Well, 01/31 has turned out the most troublesome, required review effort 
> from several people, and taken an extra verification cycle, i.e. ~2 days, 
> but has been approved by Ulrich as of last Fri, and acknowledged by Jeff 
> too.  All the remaining patches (except for 29/31, a PDP-11 fix, obviously 
> not strictly required for our purpose, approved separately by Paul Koning) 
> have been previously reviewed and approved by Jeff.

 As it turned out I previously slipped over the review of 08/31 and missed 
concerns Jeff had as to how the change may make jump optimisation happen 
where it is not supposed to.  I came across the message while making sure 
I haven't missed anything and the resulting investigation cost me extra 
few days and another two regression verification cycles.

 In the end Jeff's concerns were not unfounded, though my original change 
was actually sound and I have only updated a comment there according to my 
findings.  In the course of the investigation however I discovered and 
fixed no less than five middle end bugs where jumps were handled in a 
dangerous manner, the fixes for all of which have been accepted for GCC 
11, and a somewhat related buglet the fix for which has been accepted for 
GCC 12 only on the side of caution.

>  I'm going to verify and submit an extra change that I'd like to go in 
> ahead of the actual MODE_CC conversion (i.e. 30/31), which is to fix the 
> LTO issue observed mentioned here and along with the upstream submission, 
> as I think it is important and might be worth backporting.

 The fix has been posted:

<https://gcc.gnu.org/pipermail/gcc-patches/2020-November/560640.html>

and accepted, commit e552abe2ba27 ("PR target/95294: VAX: Convert backend 
to MODE_CC representation").

>  This will remove the sole regression, which I found a sore point in the 
> whole effort.  Not a perfect fix, as a better one I have made turned out 
> to trip on a linker bug in BFD, but certainly a working one, and good 
> enough for now.  The better fix will have to wait for GCC 12 along with 
> any other improvements.

 The BFD fix has also been posted:

<https://sourceware.org/pipermail/binutils/2020-December/114389.html>

and accepted, commit 0fcf331bb133 ("VAX/BFD: Do not warn about GOT addend 
mismatches if no GOT entry is made").

>  So I think the current ETC is this coming Wed.  Do you have any specific 
> reason to ask?

 I pushed all the changes over the weekend, with the MODE_CC conversion 
specifically:

<https://gcc.gnu.org/pipermail/gcc-cvs/2020-December/338657.html>

and all the patches, the total count of which ultimately worked out at 37, 
included in the same push.  I'll yet see if there are any loose ends to 
tie, but in any case the objective has been reached.

 Questions, comments?

  Maciej