2006-05-11 change to pe-dll.c breaks Windows DLLs

2006-05-11 change to pe-dll.c breaks Windows DLLs

[Christopher Faylor]

The change following the "Huh?" below causes problems for x86 Windows DLLs.

Index: pe-dll.c
===================================================================
RCS file: /cvs/src/src/ld/pe-dll.c,v
retrieving revision 1.83
retrieving revision 1.84
diff -u -r1.83 -r1.84
--- pe-dll.c	31 Jan 2006 22:08:14 -0000	1.83
+++ pe-dll.c	11 May 2006 08:48:58 -0000	1.84
@@ -1131,8 +1137,13 @@
 	  int nsyms, symsize;
 
 	  /* If it's not loaded, we don't need to relocate it this way.  */
+#if 0	  /* Some toolchains can generate .data sections without a SEC_LOAD flag but with relocs.  */
 	  if (!(s->output_section->flags & SEC_LOAD))
 	    continue;
+#endif
+	  /* Huh ?  */
+	  if (strncmp (bfd_get_section_name (abfd, s), ".idata",6) == 0)
+	    continue;
 
 	  /* I don't know why there would be a reloc for these, but I've
 	     seen it happen - DJ  */

I don't understand the rationale for the change.  How can we skip idata$6 sections?

And, as a meta-issue the "#if 0" and /* Huh ?  */ comment don't provide
a lot of information into what is going on here.  Shouldn't there be
some more informative comments there?

This is the ChangeLog from this change:

2006-05-11  Pedro Alves  <pedro_alves...>

	* pe-dll.c (autofilter_symbollist): Add Dllmain,
	DllMainCRTStartup, _DllMainCRTStartup and .text.
	(autofilter_liblist): Add libcegcc.
	(autofilter_symbolprefixlist): Add __imp_ and .idata$.
	(generate_reloc): Do not skip sections without a SEC_LOAD flag,
	they can still contain relocs that need processing.
	Skip the .idata$6 section.
	(jmp_arm_bytes): New array: Contains byte codes for an ARM jump.
	(make_one): Use the new array.
	(make_import_fixup_entry): Use .idata$2 instead of .idata$3.
	* emultempl/pe.em (MajorSubsystemVersion): Set to 3 for armpe.

Pedro can you explain why this is needed?  Should it possibly be
conditionalized only for your ARM target?

cgf

Re: 2006-05-11 change to pe-dll.c breaks Windows DLLs

[Christopher Faylor]

On Thu, Aug 17, 2006 at 01:00:36AM -0400, Christopher Faylor wrote:
>The change following the "Huh?" below causes problems for x86 Windows DLLs.

Btw, the problem that I'm seeing is that Windows DLLs produced by ld can
no longer be relocated -- which sort of makes sense given this change.

cgf

>Index: pe-dll.c
>===================================================================
>RCS file: /cvs/src/src/ld/pe-dll.c,v
>retrieving revision 1.83
>retrieving revision 1.84
>diff -u -r1.83 -r1.84
>--- pe-dll.c	31 Jan 2006 22:08:14 -0000	1.83
>+++ pe-dll.c	11 May 2006 08:48:58 -0000	1.84
>@@ -1131,8 +1137,13 @@
> 	  int nsyms, symsize;
> 
> 	  /* If it's not loaded, we don't need to relocate it this way.  */
>+#if 0	  /* Some toolchains can generate .data sections without a SEC_LOAD flag but with relocs.  */
> 	  if (!(s->output_section->flags & SEC_LOAD))
> 	    continue;
>+#endif
>+	  /* Huh ?  */
>+	  if (strncmp (bfd_get_section_name (abfd, s), ".idata",6) == 0)
>+	    continue;
> 
> 	  /* I don't know why there would be a reloc for these, but I've
> 	     seen it happen - DJ  */
>
>I don't understand the rationale for the change.  How can we skip idata$6 sections?
>
>And, as a meta-issue the "#if 0" and /* Huh ?  */ comment don't provide
>a lot of information into what is going on here.  Shouldn't there be
>some more informative comments there?
>
>This is the ChangeLog from this change:
>
>2006-05-11  Pedro Alves  <pedro_alves...>
>
>	* pe-dll.c (autofilter_symbollist): Add Dllmain,
>	DllMainCRTStartup, _DllMainCRTStartup and .text.
>	(autofilter_liblist): Add libcegcc.
>	(autofilter_symbolprefixlist): Add __imp_ and .idata$.
>	(generate_reloc): Do not skip sections without a SEC_LOAD flag,
>	they can still contain relocs that need processing.
>	Skip the .idata$6 section.
>	(jmp_arm_bytes): New array: Contains byte codes for an ARM jump.
>	(make_one): Use the new array.
>	(make_import_fixup_entry): Use .idata$2 instead of .idata$3.
>	* emultempl/pe.em (MajorSubsystemVersion): Set to 3 for armpe.
>
>Pedro can you explain why this is needed?  Should it possibly be
>conditionalized only for your ARM target?
>
>cgf
>

Re: 2006-05-11 change to pe-dll.c breaks Windows DLLs

[Pedro Alves]

[-- Attachment #1: Type: text/plain, Size: 2698 bytes --]

Christopher Faylor wrote:
> On Thu, Aug 17, 2006 at 01:00:36AM -0400, Christopher Faylor wrote:
>> The change following the "Huh?" below causes problems for x86 Windows DLLs.
> 
> Btw, the problem that I'm seeing is that Windows DLLs produced by ld can
> no longer be relocated -- which sort of makes sense given this change.
> 
> 

Outch!


>>
>> 	  /* If it's not loaded, we don't need to relocate it this way.  */
>> +#if 0	  /* Some toolchains can generate .data sections without a SEC_LOAD flag but with relocs.  */
>> 	  if (!(s->output_section->flags & SEC_LOAD))
>> 	    continue;
>> +#endif
>> +	  /* Huh ?  */
>> +	  if (strncmp (bfd_get_section_name (abfd, s), ".idata",6) == 0)
>> +	    continue;
>>
>> 	  /* I don't know why there would be a reloc for these, but I've
>> 	     seen it happen - DJ  */
>>
>> I don't understand the rationale for the change.  How can we skip idata$6 sections?
>>
>> And, as a meta-issue the "#if 0" and /* Huh ?  */ comment don't provide
>> a lot of information into what is going on here.  Shouldn't there be
>> some more informative comments there?
>>
>> This is the ChangeLog from this change:
>>
>> 2006-05-11  Pedro Alves  <pedro_alves...>
>>
>> 	* pe-dll.c (autofilter_symbollist): Add Dllmain,
>> 	DllMainCRTStartup, _DllMainCRTStartup and .text.
>> 	(autofilter_liblist): Add libcegcc.
>> 	(autofilter_symbolprefixlist): Add __imp_ and .idata$.
>> 	(generate_reloc): Do not skip sections without a SEC_LOAD flag,
>> 	they can still contain relocs that need processing.
>> 	Skip the .idata$6 section.
>> 	(jmp_arm_bytes): New array: Contains byte codes for an ARM jump.
>> 	(make_one): Use the new array.
>> 	(make_import_fixup_entry): Use .idata$2 instead of .idata$3.
>> 	* emultempl/pe.em (MajorSubsystemVersion): Set to 3 for armpe.
>>
>> Pedro can you explain why this is needed?  Should it possibly be
>> conditionalized only for your ARM target?
>>

I can't say I understand it myself. It was part of some pending patches
I had that I found somewhere else when I started hacking binutils (as 
explained on the mail where I posted that patch), for which I don't know
the author (I think it was Mamaich).
That part has been making me nervous too, and I have it removed in my
local tree for a while, in the hope I would see the reason for it.
I don't.
I am *very* sorry that it caused trouble. Proposed patch attached that 
reverts the offending parts.
Please review and commit.

ld/Changelog

  2006-08-17  Pedro Alves  <pedro_alves@portugalmail.pt>

	* pe-dll.c (autofilter_symbolprefixlist): Remove .idata$.
	(generate_reloc): Revert to skipping sections without a
	SEC_LOAD flag, and to not skipping .idata* sections.

Cheers,
Pedro Alves

[-- Attachment #2: pe-dll.c.diff --]
[-- Type: text/plain, Size: 896 bytes --]

--- pe-dll.c.prev	2006-08-17 08:03:18.000000000 +0100
+++ pe-dll.c	2006-08-17 08:12:24.000000000 +0100
@@ -296,7 +296,6 @@ static autofilter_entry_type autofilter_
   { "__imp_", 6 },
   /* Do __imp_ explicitly to save time.  */
   { "__rtti_", 7 },
-  { ".idata$", 7 },
   /* Don't re-export auto-imported symbols.  */
   { "_nm_", 4 },
   { "__builtin_", 10 },
@@ -1155,13 +1154,8 @@ generate_reloc (bfd *abfd, struct bfd_li
 	  int nsyms, symsize;
 
 	  /* If it's not loaded, we don't need to relocate it this way.  */
-#if 0	  /* Some toolchains can generate .data sections without a SEC_LOAD flag but with relocs.  */
 	  if (!(s->output_section->flags & SEC_LOAD))
 	    continue;
-#endif
-	  /* Huh ?  */
-	  if (strncmp (bfd_get_section_name (abfd, s), ".idata",6) == 0)
-	    continue;
 
 	  /* I don't know why there would be a reloc for these, but I've
 	     seen it happen - DJ  */

Re: 2006-05-11 change to pe-dll.c breaks Windows DLLs

[Christopher Faylor]

On Thu, Aug 17, 2006 at 08:19:52AM +0100, Pedro Alves wrote:
>Christopher Faylor wrote:
>>On Thu, Aug 17, 2006 at 01:00:36AM -0400, Christopher Faylor wrote:
>>>The change following the "Huh?" below causes problems for x86 Windows 
>>>DLLs.
>>
>>Btw, the problem that I'm seeing is that Windows DLLs produced by ld can
>>no longer be relocated -- which sort of makes sense given this change.
>>
>>[snip]
>>>Pedro can you explain why this is needed?  Should it possibly be
>>>conditionalized only for your ARM target?
>>>
>
>I can't say I understand it myself. It was part of some pending patches
>I had that I found somewhere else when I started hacking binutils (as 
>explained on the mail where I posted that patch), for which I don't know
>the author (I think it was Mamaich).
>That part has been making me nervous too, and I have it removed in my
>local tree for a while, in the hope I would see the reason for it.
>I don't.
>I am *very* sorry that it caused trouble.

It's really no problem.  I'm afraid that I will have to demand a refund,
however.  :-)

>Proposed patch attached that reverts the offending parts.  Please
>review and commit.
>
>ld/Changelog
>
> 2006-08-17  Pedro Alves  <pedro_alves@portugalmail.pt>
>
>	* pe-dll.c (autofilter_symbolprefixlist): Remove .idata$.
>	(generate_reloc): Revert to skipping sections without a
>	SEC_LOAD flag, and to not skipping .idata* sections.

Reviewed and committed.

Thanks for the quick response.

cgf