Re: [PATCH] PR51280 LTO/trans-mem ICE with TM builtins

[Richard Guenther <rguenther at suse dot de>]

On Mon, 19 Dec 2011, Patrick Marlier wrote:

> On 12/16/2011 03:54 AM, Richard Guenther wrote:
> > On Thu, 15 Dec 2011, Patrick Marlier wrote:
> > 
> > > In PR51280, LTO does ICE because the object file uses TM builtin but the
> > > TM is
> > > not enabled.
> > > In the patch, it displays a error message if the builtin is not defined
> > > and
> > > due to TM.
> > > I moved is_tm_builtin() from calls.c to trans-mem.c. I splitted it into 2
> > > functions is_tm_builtin/is_tm_builtin_code. In is_tm_builtin_code, I added
> > > some missing builtins (TM_START, TM_GETTMCLONE_SAFE, TM_MALLOC, TM_CALLOC,
> > > TM_FREE). Finally, I declared them into tree.h to be usable in calls.c and
> > > tree-streamer-in.c.
> > > 
> > > Bootstrapped and LTO/TM regtested on Linux/i686.
> > > (If ok, please commit it. Thanks.)
> > 
> > No - why should this matter?  All of TM should be lowered to a point
> > where only target specific code should be needed.
> > 
> > Richard.
> 
> Thanks Richard.
> 
> In lto file, there is GIMPLE_TRANSACTION statement and a builtin call
> (__builtin_ITM_commitTransaction) to delimit the end of the transaction
> region. The transaction is not yet instrumented. So all of TM are not lowered.
> 
> I guess this could be also added even if we should always break at the missing
> _ITM_commitTransaction builtin declaration.
> 
> Index: gimple-streamer-in.c
> ===================================================================
> --- gimple-streamer-in.c        (revision 182487)
> +++ gimple-streamer-in.c        (working copy)
> @@ -234,6 +234,9 @@ input_gimple_stmt (struct lto_input_block *ib, str
>        break;
> 
>      case GIMPLE_TRANSACTION:
> +      if (!flag_tm)
> +        error_at (gimple_location (stmt),
> +                 "use of transactional memory without support enabled");
>        gimple_transaction_set_label (stmt, stream_read_tree (ib, data_in));
>        break;
> 
> 
> It seems a bit out of my scope of my GCC knowledge so I guess I will let GCC
> guys solve this in a proper way.

I'd say we should stream the new IL elements and enable TM at link-time
once we encounter such IL element (similar to how we enable exceptions
when one TU contains EH regions).

Richard.