#20563: Cannot compile SnapPy for sage because sage's gcc is incompatible with
system's as
-------------------------------------+-------------------------------------
Reporter: mgoerner | Owner:
Type: defect | Status: needs_info
Priority: major | Milestone: sage-7.2
Component: porting | Resolution:
Keywords: | Merged in:
Authors: | Reviewers:
Report Upstream: N/A | Work issues:
Branch: | Commit:
u/mgoerner/binutils2.24 | afbe56b320454d496e82bf62d1a8efda15aededc
Dependencies: | Stopgaps:
-------------------------------------+-------------------------------------
Changes (by leif):
* status: new => needs_info
Comment:
Replying to [comment:7 mgoerner]:
> Replying to [comment:6 dunfield]:
> > My guess is that the trigger is that the code is such that gcc decides
to optimize by issuing certain SSE instructions/operands. Which SSE
instructions/operands it generates depends on what version of "as" was
present when gcc was compiled, not what version is present at run time.
The version of "as" on 10.11 knows more SSE stuff than the one on (say)
10.9, hence the issue.
>
> Nathan and I actually verified it by looking at the results of "gcc
SnapPy.c -S" when run under "sage -sh" which produces the intermittent
assembly code.
>
> My guess is that the configure script of gcc follows a certain pattern
to search for "as" and, if it finds it, probes it for its capabilities. If
gcc is then compiled, the assembly produced by that gcc depends on the
"as" found during configure time.
Yep (from `gcc/configure.ac`):
{{{
#!sh
gcc_GAS_CHECK_FEATURE([interunit movq mnemonic],
gcc_cv_as_ix86_interunit_movq,,,
[.code64
movq %mm0, %rax
movq %rax, %xmm0])
AC_DEFINE_UNQUOTED(HAVE_AS_IX86_INTERUNIT_MOVQ,
[`if test $gcc_cv_as_ix86_interunit_movq = yes; then echo 1; else
echo 0; fi`],
[Define if your assembler supports interunit movq mnemonic.])
}}}
\\
> <SNIP>
>
> In detail:
> The "offending instruction" is a movq instruction where an operand is an
SSE register xmm0, ...
> When using the sage binary distribution, the result of the above gcc
command contains the offending instruction.
Because the assembler (used when building GCC) on the buildbot is less
broken, namely clang's / LLVM's instead of Apple's ancient GAS 1.38 fork.
\\
> When Nathan compiled sage (and thus its gcc) himself on Mac OS 10.9, the
result of the same command does not contain the offending instruction
(though it still contained other instructions with SSE registers).
> "as" on 10.11 accepts the "offending instruction" but on 10.9 it does
not (supporting of the "offending instruction" starting binutils version
2.17 https://gcc.gnu.org/bugzilla/show_bug.cgi?id=43215#c7).
Up to version 4.8 GCC still emitted the traditional, wrong instruction
(`movd`; cf. the problem report above), while it was changed to the
correct one (`movq`) in 4.9.
To stay compatible with "broken" (very old) assemblers, this has again
been changed (also in 4.9) to depend on the assembler found/used ''during
the build of GCC'' (i.e., in its `configure` run, which in the case of
Sage binary dists happens on the buildbot).
The relevant PR is [https://gcc.gnu.org/bugzilla/show_bug.cgi?id=56656],
[https://gcc.gnu.org/ml/gcc-patches/2013-03/msg00789.html corresponding
patch].
Note that in fact the brokenness of bdists on older MacOS X versions
wasn't even triggered by upgrading Sage's GCC, but by upgrading the
buildbot (to a more recent OS version and hence a newer version of
XCode)... XD
----
So it's not really clear (from the current ticket ''description'') whether
this ticket is supposed to deal (only) with the specific `movd`/`movq`
issue which may pop up ''when compiling !SnapPy'' (as the title says), or
the more general problem with old Apple assemblers.
The latter will presumably be solved on #20779, thereby also fixing the
more specific problem here (which could easily be solved in even more ways
though).
--
Ticket URL: <https://trac.sagemath.org/ticket/20563#comment:16>
Sage <http://www.sagemath.org>
Sage: Creating a Viable Open Source Alternative to Magma, Maple, Mathematica,
and MATLAB
--
You received this message because you are subscribed to the Google Groups
"sage-trac" group.
To unsubscribe from this group and stop receiving emails from it, send an email
to [email protected].
To post to this group, send email to [email protected].
Visit this group at https://groups.google.com/group/sage-trac.
For more options, visit https://groups.google.com/d/optout.
