[PATCH] Fix dwarf2out ICE (PR debug/65771)

[PATCH] Fix dwarf2out ICE (PR debug/65771)

[Jakub Jelinek]

Hi!

As mentioned in the PR, on the following testcase we ICE, because for
  # DEBUG D#2 => b
  # DEBUG D#1 => a[D#2].t
  # DEBUG c => D#1
during expansion we get the a[D#2].t added as MEM_EXPR of a MEM, and because
we can't mem_loc_descriptor that MEM (I'll post separately a trunk only
patch that fixes that in this case, but generally not all MEMs can be
represented in debug info), we try harder and try to use MEM_EXPR in
loc_list_from_tree, but that one ICEs on DEBUG_EXPR_DECL.  There is nothing
we can do for those at this point, debug_exprs are only useful to
var-tracking, so returning NULL is the only thing we can do for those.

Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk and 5.1?

2015-04-16  Jakub Jelinek  <jakub@redhat.com>

	PR debug/65771
	* dwarf2out.c (loc_list_from_tree): Return NULL
	for DEBUG_EXPR_DECL.

	* gcc.dg/debug/pr65771.c: New test.

--- gcc/dwarf2out.c.jj	2015-04-16 16:51:52.000000000 +0200
+++ gcc/dwarf2out.c	2015-04-16 16:57:28.866134980 +0200
@@ -14642,6 +14642,7 @@ loc_list_from_tree (tree loc, int want_a
 
     case TARGET_MEM_REF:
     case SSA_NAME:
+    case DEBUG_EXPR_DECL:
       return NULL;
 
     case COMPOUND_EXPR:
--- gcc/testsuite/gcc.dg/debug/pr65771.c.jj	2015-04-16 17:00:23.811328842 +0200
+++ gcc/testsuite/gcc.dg/debug/pr65771.c	2015-04-16 17:00:13.000000000 +0200
@@ -0,0 +1,15 @@
+/* PR debug/65771 */
+/* { dg-do link } */
+/* { dg-require-effective-target tls } */
+
+struct S { int s; int t; };
+__thread struct S a[10];
+int b;
+
+int
+main ()
+{
+  int c = a[b].t;
+  (void) c;
+  return 0;
+}

	Jakub

[PATCH] Improve debug info generation for TLS + const (PR debug/65771)

[Jakub Jelinek]

On Thu, Apr 16, 2015 at 11:11:06PM +0200, Jakub Jelinek wrote:
> during expansion we get the a[D#2].t added as MEM_EXPR of a MEM, and because
> we can't mem_loc_descriptor that MEM (I'll post separately a trunk only
> patch that fixes that in this case, but generally not all MEMs can be
> represented in debug info), we try harder and try to use MEM_EXPR in

Here is the patch for it, bootstrapped/regtested on x86_64-linux and
i686-linux, ok for trunk?
We handle TLS SYMBOL_REFs in mem_loc_descriptor, but in the testcase it
is surrounded by (const (plus (symbol_ref) (const_int))).

2015-04-16  Jakub Jelinek  <jakub@redhat.com>

	PR debug/65771
	* dwarf2out.c (mem_loc_descriptor): For CONST, fallback to
	trying mem_loc_descriptor on XEXP (rtl, 0).

--- gcc/dwarf2out.c.jj	2015-02-26 16:37:09.000000000 +0100
+++ gcc/dwarf2out.c	2015-04-16 16:45:44.584429608 +0200
@@ -12799,7 +12799,12 @@ mem_loc_descriptor (rtx rtl, machine_mod
 	}
 
       if (!const_ok_for_output (rtl))
-	break;
+	{
+	  if (GET_CODE (rtl) == CONST)
+	    mem_loc_result = mem_loc_descriptor (XEXP (rtl, 0), mode, mem_mode,
+						 initialized);
+	  break;
+	}
 
     symref:
       mem_loc_result = new_addr_loc_descr (rtl, dtprel_false);


	Jakub

Re: [PATCH] Fix dwarf2out ICE (PR debug/65771)

[Richard Biener]

On Thu, Apr 16, 2015 at 11:11 PM, Jakub Jelinek <jakub@redhat.com> wrote:
> Hi!
>
> As mentioned in the PR, on the following testcase we ICE, because for
>   # DEBUG D#2 => b
>   # DEBUG D#1 => a[D#2].t
>   # DEBUG c => D#1
> during expansion we get the a[D#2].t added as MEM_EXPR of a MEM, and because
> we can't mem_loc_descriptor that MEM (I'll post separately a trunk only
> patch that fixes that in this case, but generally not all MEMs can be
> represented in debug info), we try harder and try to use MEM_EXPR in
> loc_list_from_tree, but that one ICEs on DEBUG_EXPR_DECL.  There is nothing
> we can do for those at this point, debug_exprs are only useful to
> var-tracking, so returning NULL is the only thing we can do for those.
>
> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk and 5.1?

Ok.

Thanks,
Richard.

> 2015-04-16  Jakub Jelinek  <jakub@redhat.com>
>
>         PR debug/65771
>         * dwarf2out.c (loc_list_from_tree): Return NULL
>         for DEBUG_EXPR_DECL.
>
>         * gcc.dg/debug/pr65771.c: New test.
>
> --- gcc/dwarf2out.c.jj  2015-04-16 16:51:52.000000000 +0200
> +++ gcc/dwarf2out.c     2015-04-16 16:57:28.866134980 +0200
> @@ -14642,6 +14642,7 @@ loc_list_from_tree (tree loc, int want_a
>
>      case TARGET_MEM_REF:
>      case SSA_NAME:
> +    case DEBUG_EXPR_DECL:
>        return NULL;
>
>      case COMPOUND_EXPR:
> --- gcc/testsuite/gcc.dg/debug/pr65771.c.jj     2015-04-16 17:00:23.811328842 +0200
> +++ gcc/testsuite/gcc.dg/debug/pr65771.c        2015-04-16 17:00:13.000000000 +0200
> @@ -0,0 +1,15 @@
> +/* PR debug/65771 */
> +/* { dg-do link } */
> +/* { dg-require-effective-target tls } */
> +
> +struct S { int s; int t; };
> +__thread struct S a[10];
> +int b;
> +
> +int
> +main ()
> +{
> +  int c = a[b].t;
> +  (void) c;
> +  return 0;
> +}
>
>         Jakub

Re: [PATCH] Improve debug info generation for TLS + const (PR debug/65771)

[Jason Merrill]

OK.

Jason