Patchwork [PR46655] Stop emitting invalid numbers to '.line' directive on AIX

login
register
mail settings
Submitter Michael Haubenwallner
Date Nov. 30, 2010, 4:05 p.m.
Message ID <4CF520C7.1000900@salomon.at>
Download mbox | patch
Permalink /patch/73619/
State New
Headers show

Comments

Michael Haubenwallner - Nov. 30, 2010, 4:05 p.m.
Hello!

As of http://gcc.gnu.org/PR46655, recent AIX-as does bounds checking
on '.line' numbers to be >0, as well as <64k in 32bit mode.

Additionally, this patch emits the '.line' numbers as 'unsigned' ("%u"),
to match XCOFF declaration in /usr/include/linenum.h, which is either
'unsigned short' in 32bit mode or 'unsigned int' in 64bit mode.

For gcc/ChangeLog:

2010-11-30  Olivier Hainque  <hainque@adacore.com>
            Michael Haubenwallner <michael.haubenwallner@salomon.at>

	PR target/46655
	* xcoffout.c (ASM_OUTPUT_LINE): Output line as unsigned,
	when line > 0, and line <= USHRT_MAX in 32bit mode.

Thank you!
Michael Haubenwallner
2010-11-30  Olivier Hainque  <hainque@adacore.com>
            Michael Haubenwallner <michael.haubenwallner@salomon.at>

	PR target/46655
	* xcoffout.c (ASM_OUTPUT_LINE): Output line as unsigned,
	when line > 0, and line <= USHRT_MAX in 32bit mode.
Eric Botcazou - Jan. 17, 2011, 11:27 a.m.
> As of http://gcc.gnu.org/PR46655, recent AIX-as does bounds checking
> on '.line' numbers to be >0, as well as <64k in 32bit mode.
>
> Additionally, this patch emits the '.line' numbers as 'unsigned' ("%u"),
> to match XCOFF declaration in /usr/include/linenum.h, which is either
> 'unsigned short' in 32bit mode or 'unsigned int' in 64bit mode.
>
> For gcc/ChangeLog:
>
> 2010-11-30  Olivier Hainque  <hainque@adacore.com>
>             Michael Haubenwallner <michael.haubenwallner@salomon.at>
>
> 	PR target/46655
> 	* xcoffout.c (ASM_OUTPUT_LINE): Output line as unsigned,
> 	when line > 0, and line <= USHRT_MAX in 32bit mode.

Can anyone approve this?  This fixes an annoying problem on AIX.
Richard Guenther - Jan. 17, 2011, 11:51 a.m.
On Mon, Jan 17, 2011 at 12:27 PM, Eric Botcazou <ebotcazou@adacore.com> wrote:
>> As of http://gcc.gnu.org/PR46655, recent AIX-as does bounds checking
>> on '.line' numbers to be >0, as well as <64k in 32bit mode.
>>
>> Additionally, this patch emits the '.line' numbers as 'unsigned' ("%u"),
>> to match XCOFF declaration in /usr/include/linenum.h, which is either
>> 'unsigned short' in 32bit mode or 'unsigned int' in 64bit mode.
>>
>> For gcc/ChangeLog:
>>
>> 2010-11-30  Olivier Hainque  <hainque@adacore.com>
>>             Michael Haubenwallner <michael.haubenwallner@salomon.at>
>>
>>       PR target/46655
>>       * xcoffout.c (ASM_OUTPUT_LINE): Output line as unsigned,
>>       when line > 0, and line <= USHRT_MAX in 32bit mode.
>
> Can anyone approve this?  This fixes an annoying problem on AIX.

Ok.

Thanks,
Richard.

> --
> Eric Botcazou
>

Patch

--- gcc/xcoffout.c.orig	2010-11-30 14:46:44 +0100
+++ gcc/xcoffout.c	2010-11-30 15:04:21 +0100
@@ -80,8 +80,18 @@ 
 #define ASM_OUTPUT_LINE(FILE,LINENUM)					   \
   do									   \
     {									   \
+      /* Make sure we're in a function and prevent output of .line 0, as   \
+	 line # 0 is meant for symbol addresses in Xcoff.  We could get    \
+	 this for insns which inherited BUILTINS_LOCATION or the function  \
+	 decl location somehow, typically created post-gimplification for  \
+	 statements inserted by some optimizers.			   \
+	 Additionally, line numbers are 'unsigned short' in 32bit mode. */ \
       if (xcoff_begin_function_line >= 0)				   \
-	fprintf (FILE, "\t.line\t%d\n", ABS_OR_RELATIVE_LINENO (LINENUM)); \
+	{								   \
+	  unsigned int lno = ABS_OR_RELATIVE_LINENO (LINENUM);		   \
+	  if (lno > 0 && (TARGET_64BIT || lno <= USHRT_MAX))		   \
+	    fprintf (FILE, "\t.line\t%u\n", lno);			   \
+	}								   \
     }									   \
   while (0)