ICE for fortran reasoned by wrong sizetype size for x86_64 mingw32 target

ICE for fortran reasoned by wrong sizetype size for x86_64 mingw32 target

[Kai Tietz]

[-- Attachment #1: Type: text/plain, Size: 1397 bytes --]

Hello,

I reported last week an ICE for target x86_64-pc-mingw32 target while 
building libgfortran. There is a bug in gfc_init_decl_processing in file 
f95-lang.c. Within this function sizetype is getting initialized as a 
"long_unsigned_type_node" - which is for most targets valid - but the 
target x86_64-pc-mingw32 has a size_t of type 
"long_long_unsigned_type_node". The following patch solves this problem:

ChangeLog: /gcc/fortran

2007-07-23  Kai Tietz  <kai.tietz@onevision.com>
        * f95-lang.c: (gfc_init_decl_processing): Use 
long_long_unsigned_type_node
        for target x86_64 mingw32 as sizetype.
 
Cheers,
 i.A. Kai Tietz



|  (\_/)  This is Bunny. Copy and paste Bunny
| (='.'=) into your signature to help him gain
| (")_(") world domination.

------------------------------------------------------------------------------------------
  OneVision Software Entwicklungs GmbH & Co. KG
  Dr.-Leo-Ritter-Straße 9 - 93049 Regensburg
  Tel: +49.(0)941.78004.0 - Fax: +49.(0)941.78004.489 - www.OneVision.com
  Commerzbank Regensburg - BLZ 750 400 62 - Konto 6011050
  Handelsregister: HRA 6744, Amtsgericht Regensburg
  Komplementärin: OneVision Software Entwicklungs Verwaltungs GmbH
  Dr.-Leo-Ritter-Straße 9 – 93049 Regensburg
  Handelsregister: HRB 8932, Amtsgericht Regensburg - Geschäftsführer: 
Ulrike Döhler, Manuela Kluger

[-- Attachment #2: w64fortran.txt --]
[-- Type: text/plain, Size: 683 bytes --]

Index: gcc/gcc/fortran/f95-lang.c
===================================================================
--- gcc.orig/gcc/fortran/f95-lang.c
+++ gcc/gcc/fortran/f95-lang.c
@@ -616,6 +616,11 @@ gfc_init_decl_processing (void)
      only use it for actual characters, not for INTEGER(1). Also, we
      want double_type_node to actually have double precision.  */
   build_common_tree_nodes (false, false);
+#ifdef TARGET_64BIT_MS_ABI
+  if (TARGET_64BIT_MS_ABI)
+    set_sizetype (long_long_unsigned_type_node);
+  else
+#endif
   set_sizetype (long_unsigned_type_node);
   build_common_tree_nodes_2 (0);
   void_list_node = build_tree_list (NULL_TREE, void_type_node);
=

Re: ICE for fortran reasoned by wrong sizetype size for x86_64 mingw32 target

[Paolo Bonzini]

> 2007-07-23  Kai Tietz  <kai.tietz@onevision.com>
>         * f95-lang.c: (gfc_init_decl_processing): Use 
> long_long_unsigned_type_node
>         for target x86_64 mingw32 as sizetype.

You should not introduce dependence on the target defines in the 
front-ends.  I suggest that you try instead

   set_sizetype (gfc_type_for_mode (Pmode, 1));

Paolo

Re: ICE for fortran reasoned by wrong sizetype size for x86_64 mingw32 target

[Kai Tietz]

[-- Attachment #1: Type: text/plain, Size: 636 bytes --]

Hi Paolo,

> 
> > 2007-07-23  Kai Tietz
> >         * f95-lang.c: (gfc_init_decl_processing): Use 
> > long_long_unsigned_type_node
> >         for target x86_64 mingw32 as sizetype.
> 
> You should not introduce dependence on the target defines in the 
> front-ends.  I suggest that you try instead
> 
>    set_sizetype (gfc_type_for_mode (Pmode, 1));

This would not work, because the type fields used by gfc_type_for_mode are 
not initialized at this point. They are initialized by gfc_init_types () 6 
lines down. I would suggest to do this by using the TREE_MODE of both 
types to make the correct choice.

Cheers,
i.A. Kai Tietz


[-- Attachment #2: w64fortran2.txt --]
[-- Type: text/plain, Size: 1029 bytes --]

Index: gcc/gcc/fortran/f95-lang.c
===================================================================
--- gcc.orig/gcc/fortran/f95-lang.c
+++ gcc/gcc/fortran/f95-lang.c
@@ -616,7 +616,15 @@ gfc_init_decl_processing (void)
      only use it for actual characters, not for INTEGER(1). Also, we
      want double_type_node to actually have double precision.  */
   build_common_tree_nodes (false, false);
-  set_sizetype (long_unsigned_type_node);
+  /* x86_64 minw32 has a sizetype of "unsigned long long", most other hosts
+     have a sizetype of "unsigned long". Therefore choose the correct size
+     in mostly target independent way.  */
+  if (TYPE_MODE (long_unsigned_type_node) == Pmode)
+    set_sizetype (long_unsigned_type_node);
+  else if (TYPE_MODE (long_long_unsigned_type_node) == Pmode)
+    set_sizetype (long_long_unsigned_type_node);
+  else
+    set_sizetype (long_unsigned_type_node);
   build_common_tree_nodes_2 (0);
   void_list_node = build_tree_list (NULL_TREE, void_type_node);
 
=

Re: ICE for fortran reasoned by wrong sizetype size for x86_64 mingw32 target

[Paolo Bonzini]

>> You should not introduce dependence on the target defines in the 
>> front-ends.  I suggest that you try instead
>>
>>    set_sizetype (gfc_type_for_mode (Pmode, 1));
> 
> This would not work, because the type fields used by gfc_type_for_mode are 
> not initialized at this point. They are initialized by gfc_init_types () 6 
> lines down. I would suggest to do this by using the TREE_MODE of both 
> types to make the correct choice.

While undoubtedly better, this may still be unsatisfactory because some 
ports may use 16-bit pointers too (I'm not sure if the Fortran run-time 
library runs there).

The Fortran maintainers could accept this patch and I've no intention to 
object, but you may also try moving gfc_init_types up, or set_sizetype 
down, so that you can use gfc_type_for_mode.

Paolo

Re: ICE for fortran reasoned by wrong sizetype size for x86_64 mingw32 target

[Kai Tietz]

Hi Paul,

Paolo Bonzini wrote on 24.07.2007 09:07:36:
> >> You should not introduce dependence on the target defines in the 
> >> front-ends.  I suggest that you try instead
> >>
> >>    set_sizetype (gfc_type_for_mode (Pmode, 1));
> > 
> > This would not work, because the type fields used by gfc_type_for_mode 
are 
> > not initialized at this point. They are initialized by gfc_init_types 
() 6 
> > lines down. I would suggest to do this by using the TREE_MODE of both 
> > types to make the correct choice.
> 
> While undoubtedly better, this may still be unsatisfactory because some 
> ports may use 16-bit pointers too (I'm not sure if the Fortran run-time 
> library runs there).
Hmm, I am uncertain to move the gfc_type_for_mode for getting the proper 
sizetype, because AFAICS it would return the nearest type, which means it 
returns an 'unsigned int' for 32-bit instead of the 'unsigned long' type 
node. Therefore I think my updated patch makes thinks more correct.
Do you know, if 16-bit targets are currently supported by gfortan ?

> The Fortran maintainers could accept this patch and I've no intention to 

> object, but you may also try moving gfc_init_types up, or set_sizetype 
> down, so that you can use gfc_type_for_mode.
Maybe you are the right person to comment on this ?

Cheers,
 i.A. Kai Tietz

|  (\_/)  This is Bunny. Copy and paste Bunny
| (='.'=) into your signature to help him gain
| (")_(") world domination.

------------------------------------------------------------------------------------------
  OneVision Software Entwicklungs GmbH & Co. KG
  Dr.-Leo-Ritter-Straße 9 - 93049 Regensburg
  Tel: +49.(0)941.78004.0 - Fax: +49.(0)941.78004.489 - www.OneVision.com
  Commerzbank Regensburg - BLZ 750 400 62 - Konto 6011050
  Handelsregister: HRA 6744, Amtsgericht Regensburg
  Komplementärin: OneVision Software Entwicklungs Verwaltungs GmbH
  Dr.-Leo-Ritter-Straße 9 – 93049 Regensburg
  Handelsregister: HRB 8932, Amtsgericht Regensburg - Geschäftsführer: 
Ulrike Döhler, Manuela Kluger

Re: ICE for fortran reasoned by wrong sizetype size for x86_64 mingw32 target

[Paul Thomas]

Kai,

It might be better to ask Paul Brook, since he wrote this part of 
gfortran. Alternatively, go to the list; I honestly have no useful 
advice to offer you.

Cheers

Paul
> Hi Paul,
>
> Paolo Bonzini wrote on 24.07.2007 09:07:36:
>   
>>>> You should not introduce dependence on the target defines in the 
>>>> front-ends.  I suggest that you try instead
>>>>
>>>>    set_sizetype (gfc_type_for_mode (Pmode, 1));
>>>>         
>>> This would not work, because the type fields used by gfc_type_for_mode 
>>>       
> are 
>   
>>> not initialized at this point. They are initialized by gfc_init_types 
>>>       
> () 6 
>   
>>> lines down. I would suggest to do this by using the TREE_MODE of both 
>>> types to make the correct choice.
>>>       
>> While undoubtedly better, this may still be unsatisfactory because some 
>> ports may use 16-bit pointers too (I'm not sure if the Fortran run-time 
>> library runs there).
>>     
> Hmm, I am uncertain to move the gfc_type_for_mode for getting the proper 
> sizetype, because AFAICS it would return the nearest type, which means it 
> returns an 'unsigned int' for 32-bit instead of the 'unsigned long' type 
> node. Therefore I think my updated patch makes thinks more correct.
> Do you know, if 16-bit targets are currently supported by gfortan ?
>
>   
>> The Fortran maintainers could accept this patch and I've no intention to 
>>     
>
>   
>> object, but you may also try moving gfc_init_types up, or set_sizetype 
>> down, so that you can use gfc_type_for_mode.
>>     
> Maybe you are the right person to comment on this ?
>
> Cheers,
>  i.A. Kai Tietz
>
> |  (\_/)  This is Bunny. Copy and paste Bunny
> | (='.'=) into your signature to help him gain
> | (")_(") world domination.
>
> ------------------------------------------------------------------------------------------
>   OneVision Software Entwicklungs GmbH & Co. KG
>   Dr.-Leo-Ritter-Straße 9 - 93049 Regensburg
>   Tel: +49.(0)941.78004.0 - Fax: +49.(0)941.78004.489 - www.OneVision.com
>   Commerzbank Regensburg - BLZ 750 400 62 - Konto 6011050
>   Handelsregister: HRA 6744, Amtsgericht Regensburg
>   Komplementärin: OneVision Software Entwicklungs Verwaltungs GmbH
>   Dr.-Leo-Ritter-Straße 9 – 93049 Regensburg
>   Handelsregister: HRB 8932, Amtsgericht Regensburg - Geschäftsführer: 
> Ulrike Döhler, Manuela Kluger
>
>   

Re: ICE for fortran reasoned by wrong sizetype size for x86_64 mingw32 target

[FX Coudert]

Hi Kai,

> I got from Paul Thomas the hint that you are the author of the f95- 
> lang.c
> file of gcc. So may you are the right person to comment or approve  
> this
> patch.I found the problem in gfc_init_decl_processing choosing the  
> wrong
> 'sizetype' type for some targets. The code assumes, that the  
> 'sizetype' is
> a 'unsigned long', which is not true for the x86_64-pc-mingw32 target.

Don't you have to change gcc/c-common.c (c_common_nodes_and_builtins)  
also, then?

> Additionally there was coming up some question about the support of  
> the
> fortan compiler for 16-bit targets (but this is an off topic issue  
> IMHO).

It depends: do you expect your patch to change the situation in any  
way for a 16-bit target? (I'm sorry that I don't know enough of these  
targets to actually answer that question myself)

> 2007-07-23  Kai Tietz  <kai.tietz@onevision.com>
>         * f95-lang.c: (gfc_init_decl_processing): Choose sizetype  
> by using
> Pmode.

Although I'm not Paul, this looks OK to me, but it's still safer to  
ping Paul if you don't hear from him (holidays?).

FX

Re: ICE for fortran reasoned by wrong sizetype size for x86_64 mingw32 target

[Kai Tietz]

[-- Attachment #1: Type: text/plain, Size: 1668 bytes --]

Hi Paul,

I got from Paul Thomas the hint that you are the author of the f95-lang.c 
file of gcc. So may you are the right person to comment or approve this 
patch.I found the problem in gfc_init_decl_processing choosing the wrong 
'sizetype' type for some targets. The code assumes, that the 'sizetype' is 
a 'unsigned long', which is not true for the x86_64-pc-mingw32 target. 
This target defines the 'long' type 32-bit and a pointer is 64-bit (a 
'long long'). This results in an ICE for the fortan compiler for the 
x86_64-pc-mingw32 target. The attached patch fixes this problem.
Additionally there was coming up some question about the support of the 
fortan compiler for 16-bit targets (but this is an off topic issue IMHO).

2007-07-23  Kai Tietz  <kai.tietz@onevision.com>
        * f95-lang.c: (gfc_init_decl_processing): Choose sizetype by using 
Pmode.




Thanks in advance and cheers,
 i.A. Kai Tietz

|  (\_/)  This is Bunny. Copy and paste Bunny
| (='.'=) into your signature to help him gain
| (")_(") world domination.

------------------------------------------------------------------------------------------
  OneVision Software Entwicklungs GmbH & Co. KG
  Dr.-Leo-Ritter-Straße 9 - 93049 Regensburg
  Tel: +49.(0)941.78004.0 - Fax: +49.(0)941.78004.489 - www.OneVision.com
  Commerzbank Regensburg - BLZ 750 400 62 - Konto 6011050
  Handelsregister: HRA 6744, Amtsgericht Regensburg
  Komplementärin: OneVision Software Entwicklungs Verwaltungs GmbH
  Dr.-Leo-Ritter-Straße 9 – 93049 Regensburg
  Handelsregister: HRB 8932, Amtsgericht Regensburg - Geschäftsführer: 
Ulrike Döhler, Manuela Kluger


[-- Attachment #2: w64fortran2.txt --]
[-- Type: text/plain, Size: 1029 bytes --]

Index: gcc/gcc/fortran/f95-lang.c
===================================================================
--- gcc.orig/gcc/fortran/f95-lang.c
+++ gcc/gcc/fortran/f95-lang.c
@@ -616,7 +616,15 @@ gfc_init_decl_processing (void)
      only use it for actual characters, not for INTEGER(1). Also, we
      want double_type_node to actually have double precision.  */
   build_common_tree_nodes (false, false);
-  set_sizetype (long_unsigned_type_node);
+  /* x86_64 minw32 has a sizetype of "unsigned long long", most other hosts
+     have a sizetype of "unsigned long". Therefore choose the correct size
+     in mostly target independent way.  */
+  if (TYPE_MODE (long_unsigned_type_node) == Pmode)
+    set_sizetype (long_unsigned_type_node);
+  else if (TYPE_MODE (long_long_unsigned_type_node) == Pmode)
+    set_sizetype (long_long_unsigned_type_node);
+  else
+    set_sizetype (long_unsigned_type_node);
   build_common_tree_nodes_2 (0);
   void_list_node = build_tree_list (NULL_TREE, void_type_node);
 
=