[gentoo-dev] Should this be considered a gcc bug?

[gentoo-dev] Should this be considered a gcc bug?

[grozin]

Hello *,

There was a bug #526194 - dev-lisp/sbcl does not respect CFLAGS. It was 
"fixed" by Mark Wright <gienah@gentoo.org> on Jan 31 - Feb 1. However, 
after this fix the upstream CFLAGS were appended to the user-supplyed 
${CFLAGS}. And the upstream CFLAGS contain -O3. So, is a user has, e.g., 
-O2 in his/her ${CFLAGS}, it was silently replaced by -O3. For some time, 
nobody noticed this: gcc-4.8 happily compiled the C stuff in sbcl with 
-O3.

However, after the upgrade to gcc-4.9 problems began (bug #544070). On 
amd64, gcc is still happy co compile sbcl with -O3. However, on x86 this 
leads to a crash of a freshly compiled sbcl runtime. Namely, the 
combinations

-O2 -march=<something>
-O3

behave correctly, and produce a working sbcl; but

-O3 -march=<something>

lead to the crush. I have changed the above "fix" in sbcl-1.2.10 in such a 
way that now it appends only -g -Wall -Wsign-compare to ${CFLAGS}, but 
not -O3. This resolves the bug #544070, unless a user has -O3 
-march=<something> in his/her ${CFLAGS}.

Shouldn't gcc-4.9 on x86 produce with -O3 something functionally 
equivalent to the -O2 case, only more optimized? Should this be considered 
a gcc-4.9 bug?

Andrey

Re: [gentoo-dev] Should this be considered a gcc bug?

[Andrew Savchenko]

[-- Attachment #1: Type: text/plain, Size: 1740 bytes --]

Hi,

On Tue, 21 Apr 2015 09:57:16 +0600 (NOVT) grozin@gentoo.org wrote:
> Hello *,
> 
> There was a bug #526194 - dev-lisp/sbcl does not respect CFLAGS. It was 
> "fixed" by Mark Wright <gienah@gentoo.org> on Jan 31 - Feb 1. However, 
> after this fix the upstream CFLAGS were appended to the user-supplyed 
> ${CFLAGS}. And the upstream CFLAGS contain -O3. So, is a user has, e.g., 
> -O2 in his/her ${CFLAGS}, it was silently replaced by -O3. For some time, 
> nobody noticed this: gcc-4.8 happily compiled the C stuff in sbcl with 
> -O3.
> 
> However, after the upgrade to gcc-4.9 problems began (bug #544070). On 
> amd64, gcc is still happy co compile sbcl with -O3. However, on x86 this 
> leads to a crash of a freshly compiled sbcl runtime. Namely, the 
> combinations
> 
> -O2 -march=<something>
> -O3
> 
> behave correctly, and produce a working sbcl; but
> 
> -O3 -march=<something>
> 
> lead to the crush. I have changed the above "fix" in sbcl-1.2.10 in such a 
> way that now it appends only -g -Wall -Wsign-compare to ${CFLAGS}, but 
> not -O3. This resolves the bug #544070, unless a user has -O3 
> -march=<something> in his/her ${CFLAGS}.
> 
> Shouldn't gcc-4.9 on x86 produce with -O3 something functionally 
> equivalent to the -O2 case, only more optimized? Should this be considered 
> a gcc-4.9 bug?

Please look at gcc-4.9 manual for the list of -O3 expansion and
find what flag exactly causes this issue. There may be two reasons:
gcc bug and sbcl bug. While you have correctly pointed out that
this may be a problem in gcc, another possibility is that extra
optimization triggers some problem in the code itself, which causes
a segfault.

Best regards,
Andrew Savchenko

[-- Attachment #2: Type: application/pgp-signature, Size: 836 bytes --]

Re: [gentoo-dev] Should this be considered a gcc bug?

[Jeroen Roovers]

On Tue, 21 Apr 2015 09:57:16 +0600 (NOVT)
grozin@gentoo.org wrote:

> However, after this fix the upstream CFLAGS were appended to the
> user-supplyed ${CFLAGS}. And the upstream CFLAGS contain -O3.

There is your problem. We filter out those compiler flags that affect
the output, so -O3 should go (and e.g. -Wall could stay).


     jer

Re: [gentoo-dev] Should this be considered a gcc bug?

[William Hubbs]

[-- Attachment #1: Type: text/plain, Size: 663 bytes --]

On Tue, Apr 21, 2015 at 09:57:16AM +0600, grozin@gentoo.org wrote:
> Hello *,
> 
> There was a bug #526194 - dev-lisp/sbcl does not respect CFLAGS. It was 
> "fixed" by Mark Wright <gienah@gentoo.org> on Jan 31 - Feb 1. However, 
> after this fix the upstream CFLAGS were appended to the user-supplyed 
> ${CFLAGS}. And the upstream CFLAGS contain -O3. So, is a user has, e.g., 
> -O2 in his/her ${CFLAGS}, it was silently replaced by -O3. For some time, 
> nobody noticed this: gcc-4.8 happily compiled the C stuff in sbcl with 
> -O3.

The best fix here would be to ask upstream to stop putting -O3 in their
default CFLAGS.

Thanks,

William


[-- Attachment #2: Digital signature --]
[-- Type: application/pgp-signature, Size: 181 bytes --]