Re: [PATCH 0/4 v3] Fix PR/63679 when --param sra-max-scalarization-size specified

[Bill Schmidt <wschmidt at linux dot vnet dot ibm dot com>]

On Tue, 2015-12-22 at 17:36 +0000, Alan Lawrence wrote:
> On 22/12/15 16:05, Bill Schmidt wrote:
> > On Tue, 2015-12-22 at 15:54 +0000, Alan Lawrence wrote:
> >> On 21/12/15 13:13, Alan Lawrence wrote:
> >>> This is a respin of previous patch series:
> >>> https://gcc.gnu.org/ml/gcc-patches/2015-10/msg03271.html
> >>> Minus three of the smaller patches having already been committed; with the
> >>> updated version of the main patch to SRA; and the patches to DOM reworked
> >>> to avoid constructing gimple stmt's.
> >>>
> >>> IMHO this feels quite invasive for stage 3, however, I note the PR/63679
> >>> (ssa-dom-cse-2.c being XFAILed on a bunch of platforms, and currently still
> >>> FAILing on ARM) is a regression from gcc 4.9.1. So I'd ask maintainer's thoughts
> >>> as to whether we should take this patch set for gcc 6.
> >>>
> >>>
> >>
> >> I've now tested the series on powerpc64-none-linux-gnu (gcc110) and
> >> powerpc64le-none-linux-gnu (gcc112).
> >>
> >> On powerpc64le, this causes the same guality failures as on AArch64:
> >> gcc.dg/guality/pr54970.c   -O1  line 15 a[0] == 1
> >> gcc.dg/guality/pr54970.c   -O1  line 20 a[0] == 1
> >> gcc.dg/guality/pr54970.c   -O1  line 25 a[0] == 1
> >> and the same with other optimization flags. (I note that there are quite a lot
> >> of guality failures on both powerpc and also aarch64, which are generally not
> >> included in the posts on the gcc-testresults mailing list).
> >
> > That's interesting.  I never see these for powerpc64le on my internal
> > build system.  I wonder what the difference is?
> 
> I've ssh'd into gcc112.fsffrance.org, a power8 powerpc64le system;
> and configure --with-cpu=power8 --disable-libsanitizer --with-long-double-128 
> --enable-languages=c,c++,lto.
> 
> Hmmm...ISTR one variable that can make a big difference is the version (or 
> absence!) of gdb...gcc112 has gdb "Fedora 7.8.2-38.fc21", copyright 2014, and 
> GDB 7.8 looks to have been released at the end of 2014.

I no longer use --disable-libsanitizer, but that doesn't seem relevant.
I also use --disable-multilib, though I think thats the default for
powerpc64le now; not sure.

The GDB on my internal system is actually even older (7.7.1, from
earlier in 2014), so I don't think that's necessarily to blame.  Still
seems mysterious.

That said, for better or worse we've had a history of ignoring the
guality failures for big-endian because we've seen so many of them, and
they don't seem to reflect a real problem with the debugger.
Understanding why they break hasn't managed to become a priority yet.  I
personally won't be upset if we have a few more than we did before, as
it probably doesn't indicate anything relevant about your patch.

Bill

> 
> >> The same pr54970 tests still seem to pass on powerpc64 big-endian even with the
> >> patches.
> 
> Ach, not quite. In fact those three are failing on powerpc64 bigendian even 
> *without* the patches (at all optimization levels besides -O0 where they pass). 
> This is on gcc110, configure --enable-languages=c,c++,lto --disable-libsanitizer 
> --with-cpu=power7 --with-long-double-128. GDB is Fedora 7.7.1-21.fc20, a bit 
> older, and I tested both unix/-m32 and unix/-m64 (following Bill Seurer's posts 
> on gcc-testresults).
> 
> I'll email a list of the failures I'm seeing offlist (summary: 186 on gcc112,
> 148 on gcc110 with -m32, 395 on gcc112 with -m64); however, I suspect gdb 
> version is the difference we are looking for.
> 
> Which *may* mean that with a more up-to-date GDB, it's possible those failures 
> may not be introduced on ppc64le. (Or similarly AArch64.) Hmmm....
> 
> --Alan
>