[Patch, Fortran] PR87556 – for FORM TEAM also use argse.pre/post

[Patch, Fortran] PR87556 – for FORM TEAM also use argse.pre/post

[Tobias Burnus]

[-- Attachment #1: Type: text/plain, Size: 787 bytes --]

as the subject states, FORM TEAM was only using the resulting tree 
expression, ignoring code which was generated before (or afterward).

I am not sure how to best convert it to a test-suite test case. For

    form team (team(this_image()), my_team2)

the old dump was:

     integer(kind=4) D.3829;
…
     _gfortran_caf_form_team (team (&D.3829), &my_team2, 0);

the new one is:

   {
     integer(kind=4) D.3822;

     D.3822 = _gfortran_caf_this_image (0);
     _gfortran_caf_form_team (team (&D.3822), &my_team2, 0);
   }

[Does it make sense to check for 5 "this_image (0)" calls? or for 4 
"D.\[0-9\]+ = _gfortran_caf_this_image (0);" calls?]


Build and on-going regtesting on x86-64-gnu-linux.

OK for the trunk?

Tobias


[-- Attachment #2: pr87556.diff --]
[-- Type: text/x-patch, Size: 3501 bytes --]

2018-10-15  Tobias Burnus  <burnus@net-b.de>

	PR fortran/87556
	* trans-stmt.c (form_team, change_team, sync_team):
	Don't ignore argse.pre/argse.post.

diff --git a/gcc/fortran/trans-stmt.c b/gcc/fortran/trans-stmt.c
index 6256e3fa805..130e67ba1e4 100644
--- a/gcc/fortran/trans-stmt.c
+++ b/gcc/fortran/trans-stmt.c
@@ -707,19 +707,30 @@ gfc_trans_form_team (gfc_code *code)
 {
   if (flag_coarray == GFC_FCOARRAY_LIB)
     {
-      gfc_se argse;
-      tree team_id,team_type;
-      gfc_init_se (&argse, NULL);
-      gfc_conv_expr_val (&argse, code->expr1);
-      team_id = fold_convert (integer_type_node, argse.expr);
-      gfc_init_se (&argse, NULL);
-      gfc_conv_expr_val (&argse, code->expr2);
-      team_type = gfc_build_addr_expr (ppvoid_type_node, argse.expr);
+      gfc_se se;
+      gfc_se argse1, argse2;
+      tree team_id, team_type, tmp;
 
-      return build_call_expr_loc (input_location,
-				  gfor_fndecl_caf_form_team, 3,
-				  team_id, team_type,
-				  build_int_cst (integer_type_node, 0));
+      gfc_init_se (&se, NULL);
+      gfc_init_se (&argse1, NULL);
+      gfc_init_se (&argse2, NULL);
+      gfc_start_block (&se.pre);
+
+      gfc_conv_expr_val (&argse1, code->expr1);
+      gfc_conv_expr_val (&argse2, code->expr2);
+      team_id = fold_convert (integer_type_node, argse1.expr);
+      team_type = gfc_build_addr_expr (ppvoid_type_node, argse2.expr);
+
+      gfc_add_block_to_block (&se.pre, &argse1.pre);
+      gfc_add_block_to_block (&se.pre, &argse2.pre);
+      tmp = build_call_expr_loc (input_location,
+				 gfor_fndecl_caf_form_team, 3,
+				 team_id, team_type,
+				 build_int_cst (integer_type_node, 0));
+      gfc_add_expr_to_block (&se.pre, tmp);
+      gfc_add_block_to_block (&se.pre, &argse1.post);
+      gfc_add_block_to_block (&se.pre, &argse2.post);
+      return gfc_finish_block (&se.pre);
     }
   else
     {
@@ -738,15 +749,18 @@ gfc_trans_change_team (gfc_code *code)
   if (flag_coarray == GFC_FCOARRAY_LIB)
     {
       gfc_se argse;
-      tree team_type;
+      tree team_type, tmp;
 
       gfc_init_se (&argse, NULL);
       gfc_conv_expr_val (&argse, code->expr1);
       team_type = gfc_build_addr_expr (ppvoid_type_node, argse.expr);
 
-      return build_call_expr_loc (input_location,
-				  gfor_fndecl_caf_change_team, 2, team_type,
-				  build_int_cst (integer_type_node, 0));
+      tmp = build_call_expr_loc (input_location,
+				 gfor_fndecl_caf_change_team, 2, team_type,
+				 build_int_cst (integer_type_node, 0));
+      gfc_add_expr_to_block (&argse.pre, tmp);
+      gfc_add_block_to_block (&argse.pre, &argse.post);
+      return gfc_finish_block (&argse.pre);
     }
   else
     {
@@ -785,16 +799,19 @@ gfc_trans_sync_team (gfc_code *code)
   if (flag_coarray == GFC_FCOARRAY_LIB)
     {
       gfc_se argse;
-      tree team_type;
+      tree team_type, tmp;
 
       gfc_init_se (&argse, NULL);
       gfc_conv_expr_val (&argse, code->expr1);
       team_type = gfc_build_addr_expr (ppvoid_type_node, argse.expr);
 
-      return build_call_expr_loc (input_location,
-				  gfor_fndecl_caf_sync_team, 2,
-				  team_type,
-				  build_int_cst (integer_type_node, 0));
+      tmp = build_call_expr_loc (input_location,
+				 gfor_fndecl_caf_sync_team, 2,
+				 team_type,
+				 build_int_cst (integer_type_node, 0));
+      gfc_add_expr_to_block (&argse.pre, tmp);
+      gfc_add_block_to_block (&argse.pre, &argse.post);
+      return gfc_finish_block (&argse.pre);
     }
   else
     {

Re: [Patch, Fortran] PR87556 – for FORM TEAM also use argse.pre/post

[Paul Richard Thomas]

Hi Tobias,

Following our exchange off-list, I rather agree with you that a
testcase is pointless. Besides which, I do not see this regressing :-)

OK for trunk and, if you are feeling strong, 8-branch.

Thanks and welcome back.

Paul
On Mon, 15 Oct 2018 at 20:46, Tobias Burnus <burnus@net-b.de> wrote:
>
> as the subject states, FORM TEAM was only using the resulting tree
> expression, ignoring code which was generated before (or afterward).
>
> I am not sure how to best convert it to a test-suite test case. For
>
>     form team (team(this_image()), my_team2)
>
> the old dump was:
>
>      integer(kind=4) D.3829;
> …
>      _gfortran_caf_form_team (team (&D.3829), &my_team2, 0);
>
> the new one is:
>
>    {
>      integer(kind=4) D.3822;
>
>      D.3822 = _gfortran_caf_this_image (0);
>      _gfortran_caf_form_team (team (&D.3822), &my_team2, 0);
>    }
>
> [Does it make sense to check for 5 "this_image (0)" calls? or for 4
> "D.\[0-9\]+ = _gfortran_caf_this_image (0);" calls?]
>
>
> Build and on-going regtesting on x86-64-gnu-linux.
>
> OK for the trunk?
>
> Tobias
>


-- 
"If you can't explain it simply, you don't understand it well enough"
- Albert Einstein