public inbox for binutils@sourceware.org
 help / color / mirror / Atom feed
* Fix some of the ARM GAS failures
@ 2005-05-24 21:24 Zack Weinberg
  2005-05-25 13:42 ` Richard Earnshaw
  0 siblings, 1 reply; 3+ messages in thread
From: Zack Weinberg @ 2005-05-24 21:24 UTC (permalink / raw)
  To: Nick Clifton, binutils

This patch should improve the ARM-GAS situation somewhat.  md_apply_fix3
needs to truncate the relocation to 32 signed bits manually; if the host
is 64-bit, fx_addnumber and fx_offset are 64-bit, and may generate
spurious overflows.  Also, we need a double cast on both sides of the
hack to pass an integer around in tc_fix_data.

zw
	* config/tc-arm.c (fix_new_arm): Cast thumb_mode to size_t
	before casting it to PTR.
	(md_apply_fix3): Truncate value, *valP, fixP->fx_addnumber,
	and fixP->fx_offset to 32 signed bits.

===================================================================
Index: config/tc-arm.c
--- config/tc-arm.c     18 May 2005 05:40:07 -0000      1.204
+++ config/tc-arm.c     24 May 2005 20:08:45 -0000
@@ -7296,8 +7296,9 @@ fix_new_arm (fragS *         frag,
     }

   /* Mark whether the fix is to a THUMB instruction, or an ARM
-     instruction.  */
-  new_fix->tc_fix_data = (PTR) thumb_mode;
+     instruction.  The double cast prevents warnings about
+     creating a pointer from an integer of different size.  */
+  new_fix->tc_fix_data = (PTR) (size_t) thumb_mode;
 }

 static void
@@ -10098,9 +10099,21 @@ md_apply_fix3 (fixS *  fixP,
        value += md_pcrel_from (fixP);
     }

-  /* Remember value for emit_reloc.  */
+  /* On a 64-bit host, silently truncate 'value' to 32 bits for
+     consistency with the behavior on 32-bit hosts.  Remember value
+     for emit_reloc.  */
+  value &= 0xffffffff;
+  value ^= 0x80000000;
+  value -= 0x80000000;
+
+  *valP = value;
   fixP->fx_addnumber = value;

+  /* Same treatment for fixP->fx_offset.  */
+  fixP->fx_offset &= 0xffffffff;
+  fixP->fx_offset ^= 0x80000000;
+  fixP->fx_offset -= 0x80000000;
+
   switch (fixP->fx_r_type)
     {
     case BFD_RELOC_NONE:


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

* Re: Fix some of the ARM GAS failures
  2005-05-24 21:24 Fix some of the ARM GAS failures Zack Weinberg
@ 2005-05-25 13:42 ` Richard Earnshaw
  2005-05-27 11:42   ` Zack Weinberg
  0 siblings, 1 reply; 3+ messages in thread
From: Richard Earnshaw @ 2005-05-25 13:42 UTC (permalink / raw)
  To: Zack Weinberg; +Cc: Nick Clifton, binutils

On Tue, 2005-05-24 at 21:17, Zack Weinberg wrote:
> This patch should improve the ARM-GAS situation somewhat.  md_apply_fix3
> needs to truncate the relocation to 32 signed bits manually; if the host
> is 64-bit, fx_addnumber and fx_offset are 64-bit, and may generate
> spurious overflows.  Also, we need a double cast on both sides of the
> hack to pass an integer around in tc_fix_data.
> 

AFAICT tc_fix_data is a private type to the ARM-specific parts of the
assembler.  Since storing the thumbness is the only use, and the type of
tc_fix_data is defined in tc-arm.h (TC_FIX_TYPE), it would make much
more sense to just change TC_FIX_TYPE to 'int' and get rid of all those
nasty casts.

Otherwise OK.

R.
> zw
> 	* config/tc-arm.c (fix_new_arm): Cast thumb_mode to size_t
> 	before casting it to PTR.
> 	(md_apply_fix3): Truncate value, *valP, fixP->fx_addnumber,
> 	and fixP->fx_offset to 32 signed bits.

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

* Re: Fix some of the ARM GAS failures
  2005-05-25 13:42 ` Richard Earnshaw
@ 2005-05-27 11:42   ` Zack Weinberg
  0 siblings, 0 replies; 3+ messages in thread
From: Zack Weinberg @ 2005-05-27 11:42 UTC (permalink / raw)
  To: Richard Earnshaw; +Cc: Nick Clifton, binutils

Richard Earnshaw <rearnsha@gcc.gnu.org> writes:

> On Tue, 2005-05-24 at 21:17, Zack Weinberg wrote:
>> This patch should improve the ARM-GAS situation somewhat.  md_apply_fix3
>> needs to truncate the relocation to 32 signed bits manually; if the host
>> is 64-bit, fx_addnumber and fx_offset are 64-bit, and may generate
>> spurious overflows.  Also, we need a double cast on both sides of the
>> hack to pass an integer around in tc_fix_data.
>> 
>
> AFAICT tc_fix_data is a private type to the ARM-specific parts of the
> assembler.  Since storing the thumbness is the only use, and the type of
> tc_fix_data is defined in tc-arm.h (TC_FIX_TYPE), it would make much
> more sense to just change TC_FIX_TYPE to 'int' and get rid of all those
> nasty casts.
>
> Otherwise OK.

Gosh, I had no idea the target could define the type of tc_fix_data.
Thanks!

Here's the patch I'll check in.

zw
        * config/tc-arm.h (TC_FIX_TYPE): Change to int.
        (TC_INIT_FIX_DATA): Initialize to 0, not NULL.
        * config/tc-arm.c (fix_new_arm): Remove now-unnecessary cast.
        (md_apply_fix3): Delete fix_is_thumb variable; refer to
        fixP->tc_fix_data directly in the sole place it was used.
        Explicitly truncate value, *valP, fixP->fx_addnumber, and
        fixP->fx_offset to 32 bits, for consistent behavior between 32-
        and 64-bit hosts.

===================================================================
Index: config/tc-arm.h
--- config/tc-arm.h	18 May 2005 05:40:08 -0000	1.28
+++ config/tc-arm.h	27 May 2005 06:43:18 -0000
@@ -96,8 +96,8 @@ struct fix;
    deliberately not been updated to mark assembler created stabs
    symbols as Thumb.  */
 
-#define TC_FIX_TYPE PTR
-#define TC_INIT_FIX_DATA(FIX) ((FIX)->tc_fix_data = NULL)
+#define TC_FIX_TYPE int
+#define TC_INIT_FIX_DATA(FIX) ((FIX)->tc_fix_data = 0)
 
 /* We need to keep some local information on symbols.  */
 
===================================================================
Index: config/tc-arm.c
--- config/tc-arm.c	18 May 2005 05:40:07 -0000	1.204
+++ config/tc-arm.c	27 May 2005 06:46:10 -0000
@@ -7297,7 +7297,7 @@ fix_new_arm (fragS *	   frag,
 
   /* Mark whether the fix is to a THUMB instruction, or an ARM
      instruction.  */
-  new_fix->tc_fix_data = (PTR) thumb_mode;
+  new_fix->tc_fix_data = thumb_mode;
 }
 
 static void
@@ -10077,9 +10077,6 @@ md_apply_fix3 (fixS *	fixP,
   unsigned long	 temp;
   int		 sign;
   char *	 buf = fixP->fx_where + fixP->fx_frag->fr_literal;
-  /* The double cast here prevents warnings about converting a pointer
-     to an integer of different size.  We know the value is 0, 1, or 2.	 */
-  int		 fix_is_thumb = (int) (size_t) fixP->tc_fix_data;
 
   assert (fixP->fx_r_type <= BFD_RELOC_UNUSED);
 
@@ -10098,9 +10095,21 @@ md_apply_fix3 (fixS *	fixP,
 	value += md_pcrel_from (fixP);
     }
 
-  /* Remember value for emit_reloc.  */
+  /* On a 64-bit host, silently truncate 'value' to 32 bits for
+     consistency with the behavior on 32-bit hosts.  Remember value
+     for emit_reloc.  */
+  value &= 0xffffffff;
+  value ^= 0x80000000;
+  value -= 0x80000000; 
+
+  *valP = value;
   fixP->fx_addnumber = value;
 
+  /* Same treatment for fixP->fx_offset.  */
+  fixP->fx_offset &= 0xffffffff;
+  fixP->fx_offset ^= 0x80000000;
+  fixP->fx_offset -= 0x80000000;
+
   switch (fixP->fx_r_type)
     {
     case BFD_RELOC_NONE:
@@ -10436,7 +10445,7 @@ md_apply_fix3 (fixS *	fixP,
       break;
 
     case BFD_RELOC_ARM_SWI:
-      if (fix_is_thumb)
+      if (fixP->tc_fix_data != 0)
 	{
 	  if (((unsigned long) value) > 0xff)
 	    as_bad_where (fixP->fx_file, fixP->fx_line,

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

end of thread, other threads:[~2005-05-27  6:52 UTC | newest]

Thread overview: 3+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2005-05-24 21:24 Fix some of the ARM GAS failures Zack Weinberg
2005-05-25 13:42 ` Richard Earnshaw
2005-05-27 11:42   ` Zack Weinberg

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).