[Bug target/46179] New: Codegen/TLS: invalid assembler syntax

[Bug target/46179] New: Codegen/TLS: invalid assembler syntax

[fthain at telegraphics dot com.au]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46179

           Summary: Codegen/TLS: invalid assembler syntax
           Product: gcc
           Version: 4.5.2
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: fthain@telegraphics.com.au
            Target: m68k-linux-gnu


Created attachment 22157
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=22157
pre-processed source from gen_uuid.c from util-linux

gcc version 4.5.2 20101021 (prerelease) (GCC)

$ m68k-linux-gnu-gcc -c gen_uuid.i -O2   
/tmp/cceVtxzz.s: Assembler messages:
/tmp/cceVtxzz.s:709: Error: syntax error -- statement `move.l
%d1,last.5137@TLSLE+4(%a0)' ignored
/tmp/cceVtxzz.s:886: Error: syntax error -- statement `move.l
%d1,last.5137@TLSLE+4(%a0)' ignored

Some background may be found here (relating to debian's gcc-4.4 package which
has TLS backported):
http://lists.debian.org/debian-68k/2010/10/msg00037.html
(In that compiler, the code generated is an offset against TLSLDO, but shows
the same bug.)

The bug was initially filed for binutils (since closed as invalid):
http://sourceware.org/bugzilla/show_bug.cgi?id=12157

Bug may be worked around with -O0.

[Bug target/46179] Codegen/TLS: invalid assembler syntax

[schwab@linux-m68k.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46179

Andreas Schwab <schwab@linux-m68k.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2010.10.26 21:04:21
     Ever Confirmed|0                           |1

--- Comment #1 from Andreas Schwab <schwab@linux-m68k.org> 2010-10-26 21:04:21 UTC ---
Reduced testcase:

struct a { int a, b; };

void
foo (struct a x)
{
  static __thread struct a y;

  y = x;
}

[Bug target/46179] Codegen/TLS: invalid assembler syntax

[mkuvyrkov at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46179

--- Comment #2 from Maxim Kuvyrkov <mkuvyrkov at gcc dot gnu.org> 2010-10-26 22:14:02 UTC ---
Just to check that this is not a lingering issue from backporting, do you have
the following comment in gcc/config/m68k/m68k.c ?  I fixed a very similar issue
before checking the final GCC TLS patch.

  /* Combine and, possibly, other optimizations may do good job
     converting
       (const (unspec [(symbol)]))
     into
       (const (plus (unspec [(symbol)])
                    (const_int N))).
     The problem with this is emitting @TLS or @GOT decorations.
     The decoration is emitted when processing (unspec), so the
     result would be "#symbol@TLSLE+N" instead of "#symbol+N@TLSLE".

     It seems that the easiest solution to this is to convert such
     operands to
       (const (unspec [(plus (symbol)
                             (const_int N))])).
     Note, that the top level of operand remains intact, so we don't have
     to patch up anything outside of the operand.  */

[Bug target/46179] Codegen/TLS: invalid assembler syntax

[fthain at telegraphics dot com.au]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46179

--- Comment #3 from Finn Thain <fthain at telegraphics dot com.au> 2010-10-27 00:47:23 UTC ---
> Just to check that this is not a lingering issue from backporting, do you have
> the following comment in gcc/config/m68k/m68k.c ? 

Yes, I checked that much. After that, I checked that the FINAL_PRESCAN_INSN
macro was being expanded properly:
$ m68k-linux-gnu-nm gcc/final.o | grep m68k
         U m68k_final_prescan_insn
         U m68k_illegitimate_symbolic_constant_p
         U m68k_output_addr_const_extra

At that point, I decided to build the gcc 4.5 snapshot (which needs no
backporting at all), and found the same failure. Then I filed the bug report...

[Bug target/46179] Codegen/TLS: invalid assembler syntax

[schwab@linux-m68k.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46179

--- Comment #4 from Andreas Schwab <schwab@linux-m68k.org> 2010-10-27 08:41:05 UTC ---
The invalid address is generated by output_move_double.

[Bug target/46179] Codegen/TLS: invalid assembler syntax

[fthain at telegraphics dot com.au]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46179

--- Comment #5 from Finn Thain <fthain at telegraphics dot com.au> 2010-10-27 13:28:01 UTC ---
> The invalid address is generated by output_move_double.

I have to say it, the .md files I looked at are hands-down the most complex and
baroque code I've ever seen. So I don't understand them. But, that's where
output_move_double gets called.

And apparently, somehow, ColdFire (and 68000) avoid this issue... What's the
solution? Should I somehow find a test case that breaks on ColdFire so that
Maxim gets to fix this? (Just kidding ;-)

Or maybe have output_move_double call a new routine that does the same
reshuffling that m68k_final_prescan_insn does -- a wrapper around the
*handle_movsi invocation. Or is that going to be too expensive?

[Bug target/46179] Codegen/TLS: invalid assembler syntax

[fthain at telegraphics dot com.au]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46179

--- Comment #6 from Finn Thain <fthain at telegraphics dot com.au> 2010-10-28 15:21:36 UTC ---
Created attachment 22191
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=22191
possible fix

[Bug target/46179] Codegen/TLS: invalid assembler syntax

[fthain at telegraphics dot com.au]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46179

--- Comment #7 from Finn Thain <fthain at telegraphics dot com.au> 2010-10-28 15:34:06 UTC ---
On Sun, 24 Oct 2010, Andreas Schwab wrote:

> This should habe [sic] been fixed up by m68k_final_prescan_insn.

(In reply to comment #4)
> The invalid address is generated by output_move_double.

Hmm. I think Andreas is dropping hints...

Is anyone able to regression test this patch for ColdFire? It seems to fix the
680x0 test cases.

[Bug target/46179] Codegen/TLS: invalid assembler syntax

[tg at mirbsd dot org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46179

--- Comment #8 from Thorsten Glaser <tg at mirbsd dot org> 2010-11-01 16:08:18 UTC ---
I can confirm that attachment 22191 fixes the original problem, namely that
one source file from util-linux, for me. No testsuite run, though ☹

Built on Debian/m68k, natively.

[Bug target/46179] Codegen/TLS: invalid assembler syntax

[mkuvyrkov at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46179

--- Comment #9 from Maxim Kuvyrkov <mkuvyrkov at gcc dot gnu.org> 2010-11-04 13:12:01 UTC ---
(In reply to comment #8)
> I can confirm that attachment 22191 [details] fixes the original problem, namely that
> one source file from util-linux, for me. No testsuite run, though ☹

The patch looks OK to me.  Would you please post it to gcc-patches@?

> 
> Built on Debian/m68k, natively.

Does this mean you bootstrapped the compiler?  I think bootstrapping GCC with
C, C++ and, possibly, Fortran, would provide a good enough test coverage for
m68k.

[Bug target/46179] Codegen/TLS: invalid assembler syntax

[tg at mirbsd dot org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46179

--- Comment #10 from Thorsten Glaser <tg at mirbsd dot org> 2010-11-04 16:27:00 UTC ---
https://www.freewrt.org/~tg/debs68k/dists/sid/main/PS-ok-unrel/gcc-4.4/

I built a Debian source package, which means three-stage bootstrap and then
building the RTL with the just compiled GCC for C, C++, Objective-C, I think
also Objective-C++, and Fortran. (I only used C and C++ compilers later,
though, unless some package snuck in some of the others.)

[Bug target/46179] Codegen/TLS: invalid assembler syntax

[schwab at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46179

--- Comment #11 from Andreas Schwab <schwab at gcc dot gnu.org> 2010-12-12 14:03:59 UTC ---
Author: schwab
Date: Sun Dec 12 14:03:55 2010
New Revision: 167724

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=167724
Log:
Author: Finn Thain <fthain@telegraphics.com.au>

PR target/46179
* gcc/config/m68k/m68k.c (handle_move_double): Insert calls to
m68k_final_prescan_insn to clean up @TLS operand syntax.

Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/config/m68k/m68k.c

[Bug target/46179] Codegen/TLS: invalid assembler syntax

[schwab@linux-m68k.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46179

Andreas Schwab <schwab@linux-m68k.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |RESOLVED
         Resolution|                            |FIXED
   Target Milestone|---                         |4.6.0

--- Comment #12 from Andreas Schwab <schwab@linux-m68k.org> 2010-12-12 14:08:48 UTC ---
Fixed.