[PATCH] Try harder to emit DW_AT_location for TLS vars

[PATCH] Try harder to emit DW_AT_location for TLS vars

[Jakub Jelinek]

Hi!

Now that resolve_addr in dwarf2out nukes at the end expressions
referencing SYMBOL_REFs not emitted to assembly we can do better for
debuginfo of TLS VAR_DECLs that don't have DECL_RTL set yet
(e.g. if a Fortran TLS common is referenced just in an OpenMP parallel
region, in the parent fn it might not have DECL_RTL set yet).

Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?

2009-11-09  Jakub Jelinek  <jakub@redhat.com>

	* dwarf2out.c (loc_list_from_tree): For TLS vars try harder to get
	rtl for DW_OP_GNU_push_tls_address addr.

--- gcc/dwarf2out.c.jj	2009-11-04 08:16:04.000000000 +0100
+++ gcc/dwarf2out.c	2009-11-06 16:11:47.000000000 +0100
@@ -14281,7 +14281,26 @@ loc_list_from_tree (tree loc, int want_a
 
 	  rtl = rtl_for_decl_location (loc);
 	  if (rtl == NULL_RTX)
-	    return 0;
+	    {
+	      if (DECL_EXTERNAL (loc))
+		return 0;
+
+	      gcc_assert (TREE_STATIC (loc)
+			  && DECL_NAME (loc)
+			  && !DECL_HARD_REGISTER (loc)
+			  && DECL_MODE (loc) != VOIDmode);
+
+	      /* Try harder to get a rtl.  If this symbol
+		 ends up not being emitted in the current CU,
+		 resolve_addr will remove the expression
+		 referencing it.  */
+	      rtl = DECL_RTL (loc);
+	      SET_DECL_RTL (loc, NULL);
+	      if (!MEM_P (rtl)
+		  || GET_CODE (XEXP (rtl, 0)) != SYMBOL_REF
+		  || SYMBOL_REF_DECL (XEXP (rtl, 0)) != loc)
+		return 0;
+	    }
 
 	  if (!MEM_P (rtl))
 	    return 0;

	Jakub

Re: [PATCH] Try harder to emit DW_AT_location for TLS vars

[Ian Lance Taylor]

Jakub Jelinek <jakub@redhat.com> writes:

> Now that resolve_addr in dwarf2out nukes at the end expressions
> referencing SYMBOL_REFs not emitted to assembly we can do better for
> debuginfo of TLS VAR_DECLs that don't have DECL_RTL set yet
> (e.g. if a Fortran TLS common is referenced just in an OpenMP parallel
> region, in the parent fn it might not have DECL_RTL set yet).

What is special about TLS variables that makes them subject to this
problem?  Why not do something like this in rtl_for_decl_location?

Calling DECL_RTL and then SET_DECL_RTL is rather awkward and requires
a comment.

Ian


> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
>
> 2009-11-09  Jakub Jelinek  <jakub@redhat.com>
>
> 	* dwarf2out.c (loc_list_from_tree): For TLS vars try harder to get
> 	rtl for DW_OP_GNU_push_tls_address addr.
>
> --- gcc/dwarf2out.c.jj	2009-11-04 08:16:04.000000000 +0100
> +++ gcc/dwarf2out.c	2009-11-06 16:11:47.000000000 +0100
> @@ -14281,7 +14281,26 @@ loc_list_from_tree (tree loc, int want_a
>  
>  	  rtl = rtl_for_decl_location (loc);
>  	  if (rtl == NULL_RTX)
> -	    return 0;
> +	    {
> +	      if (DECL_EXTERNAL (loc))
> +		return 0;
> +
> +	      gcc_assert (TREE_STATIC (loc)
> +			  && DECL_NAME (loc)
> +			  && !DECL_HARD_REGISTER (loc)
> +			  && DECL_MODE (loc) != VOIDmode);
> +
> +	      /* Try harder to get a rtl.  If this symbol
> +		 ends up not being emitted in the current CU,
> +		 resolve_addr will remove the expression
> +		 referencing it.  */
> +	      rtl = DECL_RTL (loc);
> +	      SET_DECL_RTL (loc, NULL);
> +	      if (!MEM_P (rtl)
> +		  || GET_CODE (XEXP (rtl, 0)) != SYMBOL_REF
> +		  || SYMBOL_REF_DECL (XEXP (rtl, 0)) != loc)
> +		return 0;
> +	    }
>  
>  	  if (!MEM_P (rtl))
>  	    return 0;
>
> 	Jakub

[PATCH] Try harder to emit DW_AT_location (take 2)

[Jakub Jelinek]

On Tue, Nov 10, 2009 at 07:20:07AM -0800, Ian Lance Taylor wrote:
> What is special about TLS variables that makes them subject to this
> problem?  Why not do something like this in rtl_for_decl_location?

Nothing, just I thought at least in some cases rtl_for_decl_location
would try to do that unnecessarily (e.g. when just checking if the RTL
is CONST_STRING or CONST).

Anyway, if you prefer it that way, here is a patch that has been also
bootstrapped/regtested on x86_64-linux and i686-linux.

2009-11-11  Jakub Jelinek  <jakub@redhat.com>

	* dwarf2out.c (loc_list_from_tree): Don't call rtl_for_decl_location
	unnecessarily.
	(rtl_for_decl_location): Try harder to get a rtl for TREE_STATIC vars.

--- gcc/dwarf2out.c.jj	2009-11-09 20:39:21.000000000 +0100
+++ gcc/dwarf2out.c	2009-11-11 15:13:34.000000000 +0100
@@ -14311,13 +14311,17 @@ loc_list_from_tree (tree loc, int want_a
     case RESULT_DECL:
     case FUNCTION_DECL:
       {
-	rtx rtl = rtl_for_decl_location (loc);
+	rtx rtl;
 	var_loc_list *loc_list = lookup_decl_loc (loc);
 
 	if (loc_list && loc_list->first
 	    && (list_ret = dw_loc_list (loc_list, loc, want_address)))
-	  have_address = want_address != 0;
-	else if (rtl == NULL_RTX)
+	  {
+	    have_address = want_address != 0;
+	    break;
+	  }
+	rtl = rtl_for_decl_location (loc);
+	if (rtl == NULL_RTX)
 	  {
 	    expansion_failed (loc, NULL_RTX, "DECL has no RTL");
 	    return 0;
@@ -15607,6 +15611,27 @@ rtl_for_decl_location (tree decl)
   if (rtl)
     rtl = avoid_constant_pool_reference (rtl);
 
+  /* Try harder to get a rtl.  If this symbol ends up not being emitted
+     in the current CU, resolve_addr will remove the expression referencing
+     it.  */
+  if (rtl == NULL_RTX
+      && TREE_CODE (decl) == VAR_DECL
+      && !DECL_EXTERNAL (decl)
+      && TREE_STATIC (decl)
+      && DECL_NAME (decl)
+      && !DECL_HARD_REGISTER (decl)
+      && DECL_MODE (decl) != VOIDmode)
+    {
+      rtl = DECL_RTL (decl);
+      /* Reset DECL_RTL back, as various parts of the compiler expects
+	 DECL_RTL set meaning it is actually going to be output.  */
+      SET_DECL_RTL (decl, NULL);
+      if (!MEM_P (rtl)
+	  || GET_CODE (XEXP (rtl, 0)) != SYMBOL_REF
+	  || SYMBOL_REF_DECL (XEXP (rtl, 0)) != decl)
+	rtl = NULL_RTX;
+    }
+
   return rtl;
 }
 


	Jakub

Re: [PATCH] Try harder to emit DW_AT_location (take 2)

[Richard Henderson]

On 11/11/2009 08:01 AM, Jakub Jelinek wrote:
> 	* dwarf2out.c (loc_list_from_tree): Don't call rtl_for_decl_location
> 	unnecessarily.
> 	(rtl_for_decl_location): Try harder to get a rtl for TREE_STATIC vars.

Ok.


r~