[Bug target/66312] New: [SH] Regression: Bootstrap failure gcc/d/ctfeexpr.dmd.o differs with gcc-4.8/4.9

[Bug target/66312] New: [SH] Regression: Bootstrap failure gcc/d/ctfeexpr.dmd.o differs with gcc-4.8/4.9

[glaubitz at physik dot fu-berlin.de]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66312

            Bug ID: 66312
           Summary: [SH] Regression: Bootstrap failure
                    gcc/d/ctfeexpr.dmd.o differs with gcc-4.8/4.9
           Product: gcc
           Version: 4.9.3
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
          Assignee: unassigned at gcc dot gnu.org
          Reporter: glaubitz at physik dot fu-berlin.de
                CC: kkojima at gcc dot gnu.org, olegendo at gcc dot gnu.org
  Target Milestone: ---
            Target: sh*-*-*

Hello!

As previously stated in [1], filing a new bug report for the gcc-4.8/4.9
bootstrap failure which was introduced recently. As Kaz already pointed out,
the failing reason is most certainly:

> gcc/d/ctfeexpr.dmd.o differs

Build logs:

gcc-4.8:
http://buildd.debian-ports.org/status/fetch.php?pkg=gcc-4.8&arch=sh4&ver=4.8.4-2&stamp=1432515052

gcc-4.9:
http://buildd.debian-ports.org/status/fetch.php?pkg=gcc-4.9&arch=sh4&ver=4.9.2-18&stamp=1432566734

Unfortunately, Matthias Klose doesn't split the patches in his SVN snapshots,
so I'm not sure how to properly bisect the problem. Any suggestions welcome.

The regression must have come through any of these changes [2].

> [1] https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65979#c35
> [2] https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65979#c8

[Bug target/66312] [SH] Regression: Bootstrap failure gcc/d/ctfeexpr.dmd.o differs with gcc-4.8/4.9

[olegendo at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66312

--- Comment #1 from Oleg Endo <olegendo at gcc dot gnu.org> ---



As Kaz mentioned in PR 65979 #c8, first revert all the SH specific patches. 
The SVN revisions are r221686, r221305, r221166, r220957, r220917, r219258. 
Thus you can pull the patches from the GCC SVN and reverse-apply them to your
source tree.

If bootstrap is OK without all the patches, but otherwise the latest 4.8/4.9
source, then obviously the problem must be one of those patches and it can be
narrowed down.  If it doesn't bootstrap even without all those patches, it
means the bugs have been around for a while.

[Bug target/66312] [SH] Regression: Bootstrap failure gcc/d/ctfeexpr.dmd.o differs with gcc-4.8/4.9

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66312

--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
D is not part of FSF GCC so I am inclined to close as invalid.

[Bug target/66312] [SH] Regression: Bootstrap failure gcc/d/ctfeexpr.dmd.o differs with gcc-4.8/4.9

[glaubitz at physik dot fu-berlin.de]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66312

--- Comment #4 from John Paul Adrian Glaubitz <glaubitz at physik dot fu-berlin.de> ---
(In reply to Oleg Endo from comment #1)
> As Kaz mentioned in PR 65979 #c8, first revert all the SH specific patches. 
> The SVN revisions are r221686, r221305, r221166, r220957, r220917, r219258. 
> Thus you can pull the patches from the GCC SVN and reverse-apply them to
> your source tree.

That'll be something for the weekend then. I'm still a noob when it comes to
messing with the gcc source code, so the learning curve will be a bit steep.

But I really want to solve this issue :).

Adrian

[Bug target/66312] [SH] Regression: Bootstrap failure gcc/d/ctfeexpr.dmd.o differs with gcc-4.8/4.9

[glaubitz at physik dot fu-berlin.de]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66312

--- Comment #5 from John Paul Adrian Glaubitz <glaubitz at physik dot fu-berlin.de> ---
Hello!

Just as a heads up: This particular problem did not occur with the snapshot as
of 2014-12-20 (r218987) and we actually always built gdc in Debian. So it's
definitely a regression as a result of the various changes to the SH target
since r218987 as we discussed earlier.

This problem affects gcc-4.8 and 4.9 and can be avoided by disabling the gdc
package. We have currently disabled gdc in gcc-4.9 and 5 which results in
gcc-4.9 being built (gcc-5 is still not being built for known reasons). But it
_seems_ that this particular bug shows itself now when trying to build gcc-4.9
as the compiler reproducibly segfaults now [1].

I still have to confirm this by trying another buildd, but I have the suspicion
that this segfault is directly related to the aforementioned regression. If
this is actually a bug, building gcc-4.9 on tirpitz should fail within
approximately an hour. I will report back then!

Adrian

> [1] http://buildd.debian-ports.org/status/fetch.php?pkg=gcc-4.9&arch=sh4&ver=4.9.2-21&stamp=1434293660

[Bug target/66312] [SH] Regression: Bootstrap failure gcc/d/ctfeexpr.dmd.o differs with gcc-4.8/4.9

[olegendo at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66312

--- Comment #6 from Oleg Endo <olegendo at gcc dot gnu.org> ---
(In reply to John Paul Adrian Glaubitz from comment #5)

> Just as a heads up: This particular problem did not occur with the snapshot
> as of 2014-12-20 (r218987) and we actually always built gdc in Debian. So
> it's definitely a regression as a result of the various changes to the SH
> target since r218987 as we discussed earlier.

Please try to find out which revision/patch caused the regression as mentioned
above.  That would be really helpful.

[Bug target/66312] [SH] Regression: Bootstrap failure gcc/d/ctfeexpr.dmd.o differs with gcc-4.8/4.9

[glaubitz at physik dot fu-berlin.de]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66312

--- Comment #7 from John Paul Adrian Glaubitz <glaubitz at physik dot fu-berlin.de> ---
(In reply to Oleg Endo from comment #6)
> Please try to find out which revision/patch caused the regression as
> mentioned above.  That would be really helpful.

I am currently waiting for the build queue to empty completely (about two more
days), then I'll have a look.

Oh, and the segfault seems to be hardware-related, it doesn't occur on my own
sh4 board.

Adrian

[Bug target/66312] [SH] Regression: Bootstrap failure gcc/d/ctfeexpr.dmd.o differs with gcc-4.8/4.9

[olegendo at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66312

--- Comment #9 from Oleg Endo <olegendo at gcc dot gnu.org> ---
(In reply to John Paul Adrian Glaubitz from comment #8)
> Would it be helpful to have a diff of the disassembly attached to this bug
> report as in PR target/67002?

Yes, maybe.  Please attach.

[Bug target/66312] [SH] Regression: Bootstrap failure gcc/d/ctfeexpr.dmd.o differs with gcc-4.8/4.9

[glaubitz at physik dot fu-berlin.de]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66312

--- Comment #10 from John Paul Adrian Glaubitz <glaubitz at physik dot fu-berlin.de> ---
(In reply to Oleg Endo from comment #9)
> Yes, maybe.  Please attach.

Ok, please give me a few days :).

[Bug target/66312] [SH] Regression: Bootstrap failure gcc/d/ctfeexpr.dmd.o differs with gcc-4.8/4.9

[glaubitz at physik dot fu-berlin.de]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66312

--- Comment #11 from John Paul Adrian Glaubitz <glaubitz at physik dot fu-berlin.de> ---
Created attachment 36223
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=36223&action=edit
Stage 2 compiled version of ctfeexpr.dmd.o (unstripped)

[Bug target/66312] [SH] Regression: Bootstrap failure gcc/d/ctfeexpr.dmd.o differs with gcc-4.8/4.9

[glaubitz at physik dot fu-berlin.de]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66312

--- Comment #12 from John Paul Adrian Glaubitz <glaubitz at physik dot fu-berlin.de> ---
Created attachment 36224
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=36224&action=edit
Stage 3 compiled version of ctfeexpr.dmd.o (unstripped)

[Bug target/66312] [SH] Regression: Bootstrap failure gcc/d/ctfeexpr.dmd.o differs with gcc-4.8/4.9

[glaubitz at physik dot fu-berlin.de]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66312

--- Comment #13 from John Paul Adrian Glaubitz <glaubitz at physik dot fu-berlin.de> ---
Created attachment 36225
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=36225&action=edit
Diff of the disassembly of the stripped versions of cfteexpr.dmd.o for stage 2
and 3

[Bug target/66312] [SH] Regression: Bootstrap failure gcc/d/ctfeexpr.dmd.o differs with gcc-4.8/4.9

[glaubitz at physik dot fu-berlin.de]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66312

--- Comment #14 from John Paul Adrian Glaubitz <glaubitz at physik dot fu-berlin.de> ---
The issue still reproduces with gcc_4.9_4.9.3-3 in Debian which corresponds to
SVN revision r226107 of the gcc-4.9 branch:

Comparing stages 2 and 3
warning: gcc/cc1plus-checksum.o differs
warning: gcc/cc1obj-checksum.o differs
warning: gcc/cc1objplus-checksum.o differs
warning: gcc/cc1-checksum.o differs
Bootstrap comparison failure!
gcc/d/ctfeexpr.dmd.o differs
Makefile:19827: recipe for target 'compare' failed

I have attached the stage 2 and 3 versions of gcc/d/ctfeexpr.dmd.o respectively
and also disassembled both after stripping them to be able to generate a diff
for the disassembled code. Unlike PR/target 67002, the diff seems rather large
this time.

Let me know if you need anything else, I have kept the full build of the
compiler after it failed the compare target.

Adrian

[Bug target/66312] [SH] Regression: Bootstrap failure gcc/d/ctfeexpr.dmd.o differs with gcc-4.8/4.9

[glaubitz at physik dot fu-berlin.de]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66312

--- Comment #15 from John Paul Adrian Glaubitz <glaubitz at physik dot fu-berlin.de> ---
Oh, and according to the Debian changelog, it must be a regression that was
introduced somewhere between r218987 and r222750 of the gcc-4.9 branch.

Currently, the buildds did not build any snapshot version between those two
revisions unfortunately, but I can just do more test builds to narrow down the
revision range where we need to look for the change that introduced the
regression.

Adrian

[Bug target/66312] [SH] Regression: Bootstrap failure gcc/d/ctfeexpr.dmd.o differs with gcc-4.8/4.9

[olegendo at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66312

--- Comment #17 from Oleg Endo <olegendo at gcc dot gnu.org> ---
Adrian, the recent GCC 5 build/bootstrap on SH included D, didn't it?

[Bug target/66312] [SH] Regression: Bootstrap failure gcc/d/ctfeexpr.dmd.o differs with gcc-4.8/4.9

[glaubitz at physik dot fu-berlin.de]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66312

--- Comment #18 from John Paul Adrian Glaubitz <glaubitz at physik dot fu-berlin.de> ---
(In reply to Oleg Endo from comment #17)
> Adrian, the recent GCC 5 build/bootstrap on SH included D, didn't it?

If you mean that gdc is built with gcc-5 on sh4, then no. The gdc packages are
missing for gcc-5 on sh4 [1].

Adrian

> [1] ftp://ftp.debian-ports.org/debian/pool-sh4/main/g/gcc-5/

[Bug target/66312] [SH] Regression: Bootstrap failure gcc/d/ctfeexpr.dmd.o differs with gcc-4.8/4.9

[olegendo at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66312

--- Comment #19 from Oleg Endo <olegendo at gcc dot gnu.org> ---
(In reply to John Paul Adrian Glaubitz from comment #18)
> (In reply to Oleg Endo from comment #17)
> > Adrian, the recent GCC 5 build/bootstrap on SH included D, didn't it?
> 
> If you mean that gdc is built with gcc-5 on sh4, then no. The gdc packages
> are missing for gcc-5 on sh4 [1].

Yes, that's what I meant.  Could you try building gdc with gcc 5?

[Bug target/66312] [SH] Regression: Bootstrap failure gcc/d/ctfeexpr.dmd.o differs with gcc-4.8/4.9

[olegendo at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66312

--- Comment #20 from Oleg Endo <olegendo at gcc dot gnu.org> ---
(In reply to John Paul Adrian Glaubitz from comment #12)
> Created attachment 36224 [details]
> Stage 3 compiled version of ctfeexpr.dmd.o (unstripped)

(In reply to John Paul Adrian Glaubitz from comment #11)
> Created attachment 36223 [details]
> Stage 2 compiled version of ctfeexpr.dmd.o (unstripped)

I could spot only one difference:

stage2:
    cmp/pz   r1       // T = r1 >= 0

stage3:
    not      r1,r0
    shll     r0       // T = ~r0 >> 31

Both versions are functionally equivalent.
The increase from 1 insn to 2 insns then seems to cause other differences,
which are mainly only offsets.  Sometimes register allocation could also be
affected.

The question is why this is happening...

[Bug target/66312] [SH] Regression: Bootstrap failure gcc/d/ctfeexpr.dmd.o differs with gcc-4.8/4.9

[glaubitz at physik dot fu-berlin.de]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66312

--- Comment #21 from John Paul Adrian Glaubitz <glaubitz at physik dot fu-berlin.de> ---
(In reply to Oleg Endo from comment #19)
> Yes, that's what I meant.  Could you try building gdc with gcc 5?

Sure. But I will have to wait a little. We are currently missing some buildds
that went offline which are all in Japan and I am waiting for Nobuhiro to
restore their operation.

I don't want to take away buildd resources at the moment for this reason. But I
will follow up as soon as possible.

Adrian

[Bug target/66312] [SH] Regression: Bootstrap failure gcc/d/ctfeexpr.dmd.o differs with gcc-4.8/4.9

[olegendo at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66312

--- Comment #22 from Oleg Endo <olegendo at gcc dot gnu.org> ---
Adrian, any updates on this issue?