[patch] Fix EH breakage in LTO mode

[patch] Fix EH breakage in LTO mode

[Eric Botcazou]

[-- Attachment #1: Type: text/plain, Size: 1893 bytes --]

Hi,

the patch

2017-07-06  Jan Hubicka  <hubicka@ucw.cz>

	* lto-wrapper.c (merge_and_complain): Do not merge
	fexceptions, fnon_call_exceptions, ftrapv, ffp_contract_, fmath_errno,
	fsigned_zeros, ftrapping_math, fwrapv.
	(append_compiler_options): Do not track these options.
	(append_linker_options): Likewie

has broken exception handling in LTO mode for targets which don't force either 
-funwind-tables or -fasynchronous-unwind-tables (i.e not x86) and do not 
default to -fdwarf2-cfi-asm (i.e. not Linux), for example SPARC/Solaris.

The problem is that the eh_frame section is no longer generated because it is 
emitted from compile_file:

    #if defined DWARF2_DEBUGGING_INFO || defined DWARF2_UNWIND_INFO
      if (dwarf2out_do_frame ())
	dwarf2out_frame_finish ();
    #endif

and this requires that either flag_unwind_tables or flag_exceptions be set.

But both -funwind-tables and -fexceptions are optimization options defaulting 
to 0 and they are reset to 0 through optimization_default_node after the last 
function is processed by the cgraph machinery.

Note that flag_exceptions is generally set to 1 in lto_init_eh very early, but 
the above mechanism overrides it.  However, flag_exceptions may *never* be set 
to 1 for a specific partition if it contains only EH-neutral functions; that's 
why I think that there is no other way than restoring the build-wide handling 
of -fexceptions before Jan's patch (hopefully it's the only affected switch).

Tested (lightly for now) on x86-64/Linux and SPARC/Solaris, OK for mainline?


2017-11-09  Eric Botcazou  <ebotcazou@adacore.com>

	* lto-opts.c (lto_write_options): Do not save -fnon-call-exceptions,
	-ffp-contract, -fmath-errno, -fsigned-zeros, -ftrapping-math, -ftrapv
	and -fwrapv.
	* lto-wrapper.c (merge_and_complain): Merge again -fexceptions.
	(append_compiler_options): Pass again -fexceptions.

-- 
Eric Botcazou

[-- Attachment #2: p.diff --]
[-- Type: text/x-patch, Size: 4181 bytes --]

Index: lto-opts.c
===================================================================
--- lto-opts.c	(revision 254560)
+++ lto-opts.c	(working copy)
@@ -82,61 +82,10 @@ lto_write_options (void)
       && global_options.x_flag_exceptions)
     append_to_collect_gcc_options (&temporary_obstack, &first_p,
 				   "-fexceptions");
-  /* -fnon-call-exceptions changes the generation of exception
-      regions.  It is enabled implicitly by the Go frontend.  */
-  if (!global_options_set.x_flag_non_call_exceptions
-      && global_options.x_flag_non_call_exceptions)
-    append_to_collect_gcc_options (&temporary_obstack, &first_p,
-				   "-fnon-call-exceptions");
-  /* The default -ffp-contract changes depending on the language
-     standard.  Pass thru conservative standard settings.  */
-  if (!global_options_set.x_flag_fp_contract_mode)
-    switch (global_options.x_flag_fp_contract_mode)
-      {
-      case FP_CONTRACT_OFF:
-	append_to_collect_gcc_options (&temporary_obstack, &first_p,
-				       "-ffp-contract=off");
-	break;
-      case FP_CONTRACT_ON:
-	append_to_collect_gcc_options (&temporary_obstack, &first_p,
-				       "-ffp-contract=on");
-	break;
-      case FP_CONTRACT_FAST:
-	/* Nothing.  That merges conservatively and is the default for LTO.  */
-	break;
-      default:
-	gcc_unreachable ();
-      }
-  /* The default -fmath-errno, -fsigned-zeros and -ftrapping-math change
-     depending on the language (they can be disabled by the Ada front-end).
-     Pass thru conservative standard settings.  */
-  if (!global_options_set.x_flag_errno_math)
-    append_to_collect_gcc_options (&temporary_obstack, &first_p,
-				   global_options.x_flag_errno_math
-				   ? "-fmath-errno"
-				   : "-fno-math-errno");
-  if (!global_options_set.x_flag_signed_zeros)
-    append_to_collect_gcc_options (&temporary_obstack, &first_p,
-				   global_options.x_flag_signed_zeros
-				   ? "-fsigned-zeros"
-				   : "-fno-signed-zeros");
-  if (!global_options_set.x_flag_trapping_math)
-    append_to_collect_gcc_options (&temporary_obstack, &first_p,
-				   global_options.x_flag_trapping_math
-				   ? "-ftrapping-math"
-				   : "-fno-trapping-math");
-  /* We need to merge -f[no-]strict-overflow, -f[no-]wrapv and -f[no-]trapv
-     conservatively, so stream out their defaults.  */
-  if (!global_options_set.x_flag_wrapv
-      && global_options.x_flag_wrapv)
-    append_to_collect_gcc_options (&temporary_obstack, &first_p, "-fwrapv");
-  if (!global_options_set.x_flag_trapv
-      && !global_options.x_flag_trapv)
-    append_to_collect_gcc_options (&temporary_obstack, &first_p, "-fno-trapv");
-
   if (!global_options_set.x_flag_openmp
       && !global_options.x_flag_openmp)
-    append_to_collect_gcc_options (&temporary_obstack, &first_p, "-fno-openmp");
+    append_to_collect_gcc_options (&temporary_obstack, &first_p,
+				   "-fno-openmp");
   if (!global_options_set.x_flag_openacc
       && !global_options.x_flag_openacc)
     append_to_collect_gcc_options (&temporary_obstack, &first_p,
Index: lto-wrapper.c
===================================================================
--- lto-wrapper.c	(revision 254560)
+++ lto-wrapper.c	(working copy)
@@ -280,6 +280,7 @@ merge_and_complain (struct cl_decoded_op
 	  pie_option = foption;
 	  break;
 
+	case OPT_fexceptions:
 	case OPT_fopenmp:
 	case OPT_fopenacc:
 	case OPT_fcilkplus:
@@ -290,10 +291,11 @@ merge_and_complain (struct cl_decoded_op
 	      break;
 	  if (j == *decoded_options_count)
 	    append_option (decoded_options, decoded_options_count, foption);
-	  /* -fopenmp > -fno-openmp,
+	  /* -fexceptions -> -fno-exceptions
+	     -fopenmp > -fno-openmp,
 	     -fopenacc > -fno-openacc,
 	     -fcilkplus > -fno-cilkplus,
-	     -fcheck_pointer_bounds > -fcheck_pointer_bounds  */
+	     -fcheck-pointer-bounds > -fno-check-pointer-bounds  */
 	  else if (foption->value > (*decoded_options)[j].value)
 	    (*decoded_options)[j] = *foption;
 	  break;
@@ -546,6 +548,7 @@ append_compiler_options (obstack *argv_o
 	case OPT_fpie:
 	case OPT_fcommon:
 	case OPT_fgnu_tm:
+	case OPT_fexceptions:
 	case OPT_fopenmp:
 	case OPT_fopenacc:
 	case OPT_fopenacc_dim_:

Re: [patch] Fix EH breakage in LTO mode

[Richard Biener]

On November 9, 2017 5:11:39 PM GMT+01:00, Eric Botcazou <ebotcazou@adacore.com> wrote:
>Hi,
>
>the patch
>
>2017-07-06  Jan Hubicka  <hubicka@ucw.cz>
>
>	* lto-wrapper.c (merge_and_complain): Do not merge
>	fexceptions, fnon_call_exceptions, ftrapv, ffp_contract_, fmath_errno,
>	fsigned_zeros, ftrapping_math, fwrapv.
>	(append_compiler_options): Do not track these options.
>	(append_linker_options): Likewie
>
>has broken exception handling in LTO mode for targets which don't force
>either 
>-funwind-tables or -fasynchronous-unwind-tables (i.e not x86) and do
>not 
>default to -fdwarf2-cfi-asm (i.e. not Linux), for example
>SPARC/Solaris.
>
>The problem is that the eh_frame section is no longer generated because
>it is 
>emitted from compile_file:
>
>    #if defined DWARF2_DEBUGGING_INFO || defined DWARF2_UNWIND_INFO
>      if (dwarf2out_do_frame ())
>	dwarf2out_frame_finish ();
>    #endif
>
>and this requires that either flag_unwind_tables or flag_exceptions be
>set.
>
>But both -funwind-tables and -fexceptions are optimization options
>defaulting 
>to 0 and they are reset to 0 through optimization_default_node after
>the last 
>function is processed by the cgraph machinery.
>
>Note that flag_exceptions is generally set to 1 in lto_init_eh very
>early, but 
>the above mechanism overrides it.  However, flag_exceptions may *never*
>be set 
>to 1 for a specific partition if it contains only EH-neutral functions;
>that's 
>why I think that there is no other way than restoring the build-wide
>handling 
>of -fexceptions before Jan's patch (hopefully it's the only affected
>switch).
>
>Tested (lightly for now) on x86-64/Linux and SPARC/Solaris, OK for
>mainline?

OK. 

Richard. 

>
>2017-11-09  Eric Botcazou  <ebotcazou@adacore.com>
>
>	* lto-opts.c (lto_write_options): Do not save -fnon-call-exceptions,
>	-ffp-contract, -fmath-errno, -fsigned-zeros, -ftrapping-math, -ftrapv
>	and -fwrapv.
>	* lto-wrapper.c (merge_and_complain): Merge again -fexceptions.
>	(append_compiler_options): Pass again -fexceptions.

Re: [patch] Fix EH breakage in LTO mode

[Jan Hubicka]

> Hi,
> 
> the patch
> 
> 2017-07-06  Jan Hubicka  <hubicka@ucw.cz>
> 
> 	* lto-wrapper.c (merge_and_complain): Do not merge
> 	fexceptions, fnon_call_exceptions, ftrapv, ffp_contract_, fmath_errno,
> 	fsigned_zeros, ftrapping_math, fwrapv.
> 	(append_compiler_options): Do not track these options.
> 	(append_linker_options): Likewie
> 
> has broken exception handling in LTO mode for targets which don't force either 
> -funwind-tables or -fasynchronous-unwind-tables (i.e not x86) and do not 
> default to -fdwarf2-cfi-asm (i.e. not Linux), for example SPARC/Solaris.
> 
> The problem is that the eh_frame section is no longer generated because it is 
> emitted from compile_file:
> 
>     #if defined DWARF2_DEBUGGING_INFO || defined DWARF2_UNWIND_INFO
>       if (dwarf2out_do_frame ())
> 	dwarf2out_frame_finish ();
>     #endif
> 
> and this requires that either flag_unwind_tables or flag_exceptions be set.
> 
> But both -funwind-tables and -fexceptions are optimization options defaulting 
> to 0 and they are reset to 0 through optimization_default_node after the last 
> function is processed by the cgraph machinery.
> 
> Note that flag_exceptions is generally set to 1 in lto_init_eh very early, but 
> the above mechanism overrides it.  However, flag_exceptions may *never* be set 
> to 1 for a specific partition if it contains only EH-neutral functions; that's 
> why I think that there is no other way than restoring the build-wide handling 
> of -fexceptions before Jan's patch (hopefully it's the only affected switch).

I am not sure I follow your argumentation here.  flag_exceptions is optimization
which means that one should not test it globally. For that reason I would just
change dwarf2out to simply trigger dwarf2out_frame_finish if any of the compiled
functions needed it (had flag_exceptions set).

My readon of lto-wrapper is that it will set flag_exceptions for whole compilation
unit when one of compilation units were copmiled with flag_exceptions.

How that helps the neutral functions?
Honza

Re: [patch] Fix EH breakage in LTO mode

[Eric Botcazou]

> I am not sure I follow your argumentation here.  flag_exceptions is
> optimization which means that one should not test it globally. For that
> reason I would just change dwarf2out to simply trigger
> dwarf2out_frame_finish if any of the compiled functions needed it (had
> flag_exceptions set).

That's not sufficient for a partition with only EH-neutral functions.

> My readon of lto-wrapper is that it will set flag_exceptions for whole
> compilation unit when one of compilation units were copmiled with
> flag_exceptions.
> 
> How that helps the neutral functions?

flag_exceptions is never set in a partition with only EH-neutral functions.

-- 
Eric Botcazou

Re: [patch] Fix EH breakage in LTO mode

[Rainer Orth]

Hi Eric,

> the patch
>
> 2017-07-06  Jan Hubicka  <hubicka@ucw.cz>
>
> 	* lto-wrapper.c (merge_and_complain): Do not merge
> 	fexceptions, fnon_call_exceptions, ftrapv, ffp_contract_, fmath_errno,
> 	fsigned_zeros, ftrapping_math, fwrapv.
> 	(append_compiler_options): Do not track these options.
> 	(append_linker_options): Likewie
>
> has broken exception handling in LTO mode for targets which don't force either 
> -funwind-tables or -fasynchronous-unwind-tables (i.e not x86) and do not 
> default to -fdwarf2-cfi-asm (i.e. not Linux), for example SPARC/Solaris.

this is PR lto/81351.

	Rainer

-- 
-----------------------------------------------------------------------------
Rainer Orth, Center for Biotechnology, Bielefeld University

Re: [patch] Fix EH breakage in LTO mode

[Jan Hubicka]

> > I am not sure I follow your argumentation here.  flag_exceptions is
> > optimization which means that one should not test it globally. For that
> > reason I would just change dwarf2out to simply trigger
> > dwarf2out_frame_finish if any of the compiled functions needed it (had
> > flag_exceptions set).
> 
> That's not sufficient for a partition with only EH-neutral functions.
> 
> > My readon of lto-wrapper is that it will set flag_exceptions for whole
> > compilation unit when one of compilation units were copmiled with
> > flag_exceptions.
> > 
> > How that helps the neutral functions?
> 
> flag_exceptions is never set in a partition with only EH-neutral functions.

Hmm, I am still not following. By EH-neutral you mean function compiled
with -fno-exceptions so it has no unwind info at all or one that has unwind
info but does not do anything with EH?

Honza
> 
> -- 
> Eric Botcazou

Re: [patch] Fix EH breakage in LTO mode

[Richard Biener]

On November 9, 2017 7:03:18 PM GMT+01:00, Jan Hubicka <hubicka@ucw.cz> wrote:
>> > I am not sure I follow your argumentation here.  flag_exceptions is
>> > optimization which means that one should not test it globally. For
>that
>> > reason I would just change dwarf2out to simply trigger
>> > dwarf2out_frame_finish if any of the compiled functions needed it
>(had
>> > flag_exceptions set).
>> 
>> That's not sufficient for a partition with only EH-neutral functions.
>> 
>> > My readon of lto-wrapper is that it will set flag_exceptions for
>whole
>> > compilation unit when one of compilation units were copmiled with
>> > flag_exceptions.
>> > 
>> > How that helps the neutral functions?
>> 
>> flag_exceptions is never set in a partition with only EH-neutral
>functions.
>
>Hmm, I am still not following. By EH-neutral you mean function compiled
>with -fno-exceptions so it has no unwind info at all or one that has
>unwind
>info but does not do anything with EH?

The latter. We choose a 'neutral' EH personality for those and thus if all behave that way we fail to init EH. 

I suppose we could record a TU wide setting of options attached to the TRANSLATION_UNIT_DECL and read global flags from that? 

Richard. 

>Honza
>> 
>> -- 
>> Eric Botcazou

Re: [patch] Fix EH breakage in LTO mode

[Eric Botcazou]

> this is PR lto/81351.

Totally missed it...  Do not hesitate to CC me for SPARC regressions.

-- 
Eric Botcazou

Re: [patch] Fix EH breakage in LTO mode

[Eric Botcazou]

> Hmm, I am still not following. By EH-neutral you mean function compiled
> with -fno-exceptions so it has no unwind info at all or one that has unwind
> info but does not do anything with EH?

OK, I see what you mean now.  So, even if lto_init_eh is never invoked and 
thus never sets flag_exceptions for the partition,  you would record whether 
the flag was saved in the optimization node of at least one function, right?

But this would require to compile C code with -fexceptions if it's included in 
an EH-enabled LTO link, whereas you didn't need do it previously (of course 
this is all masked on x86 because -funwind-tables is always enabled).  On the 
other hand, this is required for a normal link too.

I guess that would be more consistent with what happens with -fdwarf2-cfi-asm.

-- 
Eric Botcazou