public inbox for fortran@gcc.gnu.org
 help / color / mirror / Atom feed
* Re: [gomp4] declare directive [3/5]
       [not found] ` <5575AEEB.6020902@codesourcery.com>
@ 2015-06-17  9:59   ` Thomas Schwinge
  2015-06-18 16:02     ` James Norris
  2015-10-30 13:27   ` Thomas Schwinge
  1 sibling, 1 reply; 3+ messages in thread
From: Thomas Schwinge @ 2015-06-17  9:59 UTC (permalink / raw)
  To: James Norris; +Cc: Tobias Burnus, gcc-patches, fortran

[-- Attachment #1: Type: text/plain, Size: 3749 bytes --]

Hi Jim!

I had mentioned that the Fortran front end changes cause regressions in a
few libgomp execution tests, if configured for Intel MIC (emulation)
offloading.  I have now located *where* this is coming from, but would
you please work on figuring out *why*?

Fortunately, you'll be able to work on the problem even without Intel MIC
(emulation) offloading configured: to reproduce, just look at the
difference in -fdump-tree-original without and with your patch applied.
You'll notice that clauses are getting "lost" from OpenMP target update
directives; for example, for
libgomp/testsuite/libgomp.fortran/declare-target-1.f90 I see:

    --- GOOD/declare-target-2.f90.003t.original     2015-06-16 18:16:07.472763339 +0200
    +++ ./declare-target-2.f90.003t.original        2015-06-16 19:28:22.706845250 +0200
    @@ -3,14 +3,14 @@
       extern integer(kind=4) var_x;
     
       var_x = 10;
    -  #pragma omp target update to(var_x)
    +  #pragma omp target update
       #pragma omp target
         {
           {
             var_x = var_x * 2;
           }
         }
    -  #pragma omp target update from(var_x)
    +  #pragma omp target update
       if (var_x != 20)
         {
           _gfortran_abort ();

(This is the only test case that I looked at, so far.)

I tracked this down to:

On Mon, 8 Jun 2015 10:04:11 -0500, James Norris <jnorris@codesourcery.com> wrote:
> --- a/gcc/fortran/trans-decl.c
> +++ b/gcc/fortran/trans-decl.c

> +static void
> +find_module_oacc_declare_clauses (gfc_symbol *sym)
> +{
> +  if (sym->attr.use_assoc)
> +    {
> +      gfc_omp_map_op map_op;
> +
> +      sym->attr.referenced = sym->attr.oacc_declare_create
> +			     | sym->attr.oacc_declare_copyin
> +			     | sym->attr.oacc_declare_deviceptr
> +			     | sym->attr.oacc_declare_device_resident;
> +
> +      if (sym->attr.oacc_declare_create)
> +	map_op = OMP_MAP_FORCE_ALLOC;
> +
> +      if (sym->attr.oacc_declare_copyin)
> +	map_op = OMP_MAP_FORCE_TO;
> +
> +      if (sym->attr.oacc_declare_deviceptr)
> +	map_op = OMP_MAP_FORCE_DEVICEPTR;
> +
> +      if (sym->attr.oacc_declare_device_resident)
> +	map_op = OMP_MAP_DEVICE_RESIDENT;
> +
> +      if (sym->attr.referenced)
> +	add_clause (sym, map_op);
> +    }
> +}

... this function apparently doing "something inappropriate".  It gets
(unconditionally) called from:

> +finish_oacc_declare (gfc_namespace *ns, enum sym_flavor flavor)
>  {
> [...]
> +  gfc_traverse_ns (ns, find_module_oacc_declare_clauses);

... here, which in turn gets (unconditionally) called from:

> @@ -5946,8 +6237,7 @@ gfc_generate_function_code (gfc_namespace * ns)
>      add_argument_checking (&body, sym);
>  
>    /* Generate !$ACC DECLARE directive. */
> -  if (ns->oacc_declare)
> -    insert_oacc_declare (ns);
> +  finish_oacc_declare (ns, sym->attr.flavor);
>  
>    tmp = gfc_trans_code (ns->code);
>    gfc_add_expr_to_block (&body, tmp);

... here, and:

> --- a/gcc/fortran/trans-stmt.c
> +++ b/gcc/fortran/trans-stmt.c
> @@ -1588,8 +1588,7 @@ gfc_trans_block_construct (gfc_code* code)
>    code->exit_label = exit_label;
>  
>    /* Generate !$ACC DECLARE directive. */
> -  if (ns->oacc_declare)
> -    insert_oacc_declare (ns);
> +  finish_oacc_declare (ns, FL_UNKNOWN);
>  
>    gfc_add_expr_to_block (&body, gfc_trans_code (ns->code));
>    gfc_add_expr_to_block (&body, build1_v (LABEL_EXPR, exit_label));

... here.

Is that sufficient information for you to reproduce the problem?

As soon as you have a patch to bring back the lost clauses in the
-fdump-tree-original, I'll be happy to test it in my Intel MIC (emulated)
offloading build.


Grüße,
 Thomas

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 472 bytes --]

^ permalink raw reply	[flat|nested] 3+ messages in thread

* Re: [gomp4] declare directive [3/5]
  2015-06-17  9:59   ` [gomp4] declare directive [3/5] Thomas Schwinge
@ 2015-06-18 16:02     ` James Norris
  0 siblings, 0 replies; 3+ messages in thread
From: James Norris @ 2015-06-18 16:02 UTC (permalink / raw)
  To: Thomas Schwinge; +Cc: Tobias Burnus, gcc-patches, fortran

[-- Attachment #1: Type: text/plain, Size: 1170 bytes --]

Hi Thomas!

On 06/17/2015 04:59 AM, Thomas Schwinge wrote:
> Hi Jim!
>
> I had mentioned that the Fortran front end changes cause regressions in a
> few libgomp execution tests, if configured for Intel MIC (emulation)
> offloading.  I have now located *where* this is coming from, but would
> you please work on figuring out *why*?
>

There are actually two bugs in find_module_oacc_declare_clauses which
are causing the issues you are seeing.

With the first bug, if none of the 'oacc_declare_*' bits were asserted,
then the referenced field within the attribute structure was set to
zero. If the referenced field was already set to one prior to
find_module_oacc_declare_clauses being called, then the field gets
incorrectly set to zero, if none of the 'oacc_declare_*' bits were
asserted.

With the second bug, if the referenced field within the attribute
structure is already set to one prior to
find_module_oacc_declare_clauses being called, then the subroutine
add_clause was called. The subroutine add_clause should only be
called if one of the 'oacc_declare_*' bits are asserted.

The attached patch resolves the above issues.

Committed to gomp-4_0-branch

Jim


[-- Attachment #2: declare.patch --]
[-- Type: text/x-patch, Size: 1013 bytes --]

diff --git a/gcc/fortran/trans-decl.c b/gcc/fortran/trans-decl.c
index 77fdc8b..7387a80 100644
--- a/gcc/fortran/trans-decl.c
+++ b/gcc/fortran/trans-decl.c
@@ -5848,11 +5848,6 @@ find_module_oacc_declare_clauses (gfc_symbol *sym)
     {
       gfc_omp_map_op map_op;
 
-      sym->attr.referenced = sym->attr.oacc_declare_create
-			     | sym->attr.oacc_declare_copyin
-			     | sym->attr.oacc_declare_deviceptr
-			     | sym->attr.oacc_declare_device_resident;
-
       if (sym->attr.oacc_declare_create)
 	map_op = OMP_MAP_FORCE_ALLOC;
 
@@ -5865,8 +5860,14 @@ find_module_oacc_declare_clauses (gfc_symbol *sym)
       if (sym->attr.oacc_declare_device_resident)
 	map_op = OMP_MAP_DEVICE_RESIDENT;
 
-      if (sym->attr.referenced)
-	add_clause (sym, map_op);
+      if (sym->attr.oacc_declare_create
+	  || sym->attr.oacc_declare_copyin
+	  || sym->attr.oacc_declare_deviceptr
+	  || sym->attr.oacc_declare_device_resident)
+	{
+	  sym->attr.referenced = 1;
+	  add_clause (sym, map_op);
+	}
     }
 }
 

^ permalink raw reply	[flat|nested] 3+ messages in thread

* Re: [gomp4] declare directive [3/5]
       [not found] ` <5575AEEB.6020902@codesourcery.com>
  2015-06-17  9:59   ` [gomp4] declare directive [3/5] Thomas Schwinge
@ 2015-10-30 13:27   ` Thomas Schwinge
  1 sibling, 0 replies; 3+ messages in thread
From: Thomas Schwinge @ 2015-10-30 13:27 UTC (permalink / raw)
  To: James Norris, gcc-patches, fortran; +Cc: Tobias Burnus

[-- Attachment #1: Type: text/plain, Size: 2245 bytes --]

Hi!

On Mon, 8 Jun 2015 10:04:11 -0500, James Norris <jnorris@codesourcery.com> wrote:
> --- a/gcc/fortran/gfortran.h
> +++ b/gcc/fortran/gfortran.h
> @@ -1174,6 +1183,7 @@ enum
>    OMP_LIST_FROM,
>    OMP_LIST_REDUCTION,
>    OMP_LIST_DEVICE_RESIDENT,
> +  OMP_LIST_LINK,
>    OMP_LIST_USE_DEVICE,
>    OMP_LIST_CACHE,
>    OMP_LIST_NUM

I noticed (my means of hitting a segmentation fault) that this was
missing an update to the clause_names in
gcc/fortran/openmp.c:resolve_omp_clauses.  (Yes, I agree that is a
strange, non-obvious dependency that this function needs to be updated
for OMP_LIST_* changes...)  Fixed on gomp-4_0-branch in r229576:

commit a5246d7b6c91e0800eeb6355bf5e4c63d27aafb2
Author: tschwinge <tschwinge@138bc75d-0d04-0410-961f-82ee72b054a4>
Date:   Fri Oct 30 13:24:35 2015 +0000

    Fix OMP_LIST_LINK handling
    
    	gcc/fortran/
    	* openmp.c (resolve_omp_clauses): Add "LINK" to clause_names.
    
    git-svn-id: svn+ssh://gcc.gnu.org/svn/gcc/branches/gomp-4_0-branch@229576 138bc75d-0d04-0410-961f-82ee72b054a4
---
 gcc/fortran/ChangeLog.gomp | 4 ++++
 gcc/fortran/openmp.c       | 2 +-
 2 files changed, 5 insertions(+), 1 deletion(-)

diff --git gcc/fortran/ChangeLog.gomp gcc/fortran/ChangeLog.gomp
index 7fe3eac..592dd8d 100644
--- gcc/fortran/ChangeLog.gomp
+++ gcc/fortran/ChangeLog.gomp
@@ -1,3 +1,7 @@
+2015-10-30  Thomas Schwinge  <thomas@codesourcery.com>
+
+	* openmp.c (resolve_omp_clauses): Add "LINK" to clause_names.
+
 2015-10-29  Thomas Schwinge  <thomas@codesourcery.com>
 
 	* openmp.c (gfc_match_omp_map_clause): Remove allow_sections
diff --git gcc/fortran/openmp.c gcc/fortran/openmp.c
index a2c5105..32779f7 100644
--- gcc/fortran/openmp.c
+++ gcc/fortran/openmp.c
@@ -3197,7 +3197,7 @@ resolve_omp_clauses (gfc_code *code, gfc_omp_clauses *omp_clauses,
   static const char *clause_names[]
     = { "PRIVATE", "FIRSTPRIVATE", "LASTPRIVATE", "COPYPRIVATE", "SHARED",
 	"COPYIN", "UNIFORM", "ALIGNED", "LINEAR", "DEPEND", "MAP",
-	"TO", "FROM", "REDUCTION", "DEVICE_RESIDENT", "USE_DEVICE",
+	"TO", "FROM", "REDUCTION", "DEVICE_RESIDENT", "LINK", "USE_DEVICE",
 	"CACHE" };
 
   if (omp_clauses == NULL)


Grüße
 Thomas

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 472 bytes --]

^ permalink raw reply	[flat|nested] 3+ messages in thread

end of thread, other threads:[~2015-10-30 13:27 UTC | newest]

Thread overview: 3+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
     [not found] <5575ADCB.6030107@codesourcery.com>
     [not found] ` <5575AEEB.6020902@codesourcery.com>
2015-06-17  9:59   ` [gomp4] declare directive [3/5] Thomas Schwinge
2015-06-18 16:02     ` James Norris
2015-10-30 13:27   ` Thomas Schwinge

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).