Unexpected DW_AT_type for Ada strings on s390x?

Unexpected DW_AT_type for Ada strings on s390x?

[Andreas Arnez]

Upstream gnat seems to create different debug data for string parameters
on s390x than on x86, x86_64 or s390.  On s390x the DW_AT_type of a
string parameter is a constant /reference/ to the two-element structure
"string___XUP".  On the other platforms it is the constant structure,
not a reference.

Due to this indirection the second "bt" in "lang_switch.exp" generates
unexpected output; gdb now shows the "msg" parameter with an address
like this:

     msg=@0x3ffffffdc48: {...}

The test case fails on s390x because of that.

Is the DW_AT_type output intentional?  Why does it depend on the
architecture?  Note that for C or C++, gcc correctly declares structure
parameters as /structures/ (not references) in their DW_AT_type.

If the output is intentional, I suggest to relax the pattern in the test
case accordingly.  In any case it would be nice to differentiate the two
"backtrace" test points, e.g. like this:


--- a/gdb/testsuite/gdb.ada/lang_switch.exp
+++ b/gdb/testsuite/gdb.ada/lang_switch.exp
@@ -43,5 +43,5 @@ gdb_test "bt" \
 gdb_test "set lang c"
 gdb_test "bt" \
          ".*#1.*lang_switch\\.ada_procedure\\s*\\(msg={.*\\).*" \
-         "backtrace"
+	 "backtrace (C)"

Re: Unexpected DW_AT_type for Ada strings on s390x?

[Joel Brobecker]

[-- Attachment #1: Type: text/plain, Size: 1474 bytes --]

[Cc'ing Eric, as he is the compiler expert - GDB just represents what
the compiler tells us]

> Upstream gnat seems to create different debug data for string parameters
> on s390x than on x86, x86_64 or s390.  On s390x the DW_AT_type of a
> string parameter is a constant /reference/ to the two-element structure
> "string___XUP".  On the other platforms it is the constant structure,
> not a reference.

My first guess is that the variantion in type definition is caused
by differences in ABI. I don't know s390 platforms, so I can't
confirm offhand. But it seems like a plausible way to pass the
argument.

If Eric confirms my guess, the attached patch should fix the problem.

[the rest of your email is kept below - as a reference for Eric, in case
he needs it]

> Is the DW_AT_type output intentional?  Why does it depend on the
> architecture?  Note that for C or C++, gcc correctly declares structure
> parameters as /structures/ (not references) in their DW_AT_type.

> 
> If the output is intentional, I suggest to relax the pattern in the test
> case accordingly.  In any case it would be nice to differentiate the two
> "backtrace" test points, e.g. like this:
> 
> 
> --- a/gdb/testsuite/gdb.ada/lang_switch.exp
> +++ b/gdb/testsuite/gdb.ada/lang_switch.exp
> @@ -43,5 +43,5 @@ gdb_test "bt" \
>  gdb_test "set lang c"
>  gdb_test "bt" \
>           ".*#1.*lang_switch\\.ada_procedure\\s*\\(msg={.*\\).*" \
> -         "backtrace"
> +	 "backtrace (C)"

-- 
Joel

[-- Attachment #2: 0001-gdb.ada-lang_switch-Allow-msg-parameter-to-be-a-refe.patch --]
[-- Type: text/x-diff, Size: 987 bytes --]

From cdb4ea7d663ca7eab453f31a45f6fae068753803 Mon Sep 17 00:00:00 2001
From: Joel Brobecker <brobecker@adacore.com>
Date: Tue, 4 Dec 2012 16:04:26 +0400
Subject: [PATCH] gdb.ada/lang_switch: Allow msg parameter to be a reference.

gdb/testsuite/ChangeLog:

        * gdb.ada/lang_switch.exp: The "msg" parameter may be a reference.
---
 gdb/testsuite/gdb.ada/lang_switch.exp |    5 ++---
 1 file changed, 2 insertions(+), 3 deletions(-)

diff --git a/gdb/testsuite/gdb.ada/lang_switch.exp b/gdb/testsuite/gdb.ada/lang_switch.exp
index 6ee4c23..a217d2e 100644
--- a/gdb/testsuite/gdb.ada/lang_switch.exp
+++ b/gdb/testsuite/gdb.ada/lang_switch.exp
@@ -42,6 +42,5 @@ gdb_test "bt" \
 # if the current language is not "auto".
 gdb_test "set lang c"
 gdb_test "bt" \
-         ".*#1.*lang_switch\\.ada_procedure\\s*\\(msg={.*\\).*" \
-         "backtrace"
-
+         ".*#1.*lang_switch\\.ada_procedure\\s*\\(msg=(@$hex: +)?{.*\\).*" \
+         "backtrace with lang set to C"
-- 
1.7.10.4

Re: Unexpected DW_AT_type for Ada strings on s390x?

[Andreas Arnez]

Joel Brobecker <brobecker@adacore.com> writes:
> [Cc'ing Eric, as he is the compiler expert - GDB just represents what
> the compiler tells us]

Sure.  Should we also CC the gcc mailing list?

> My first guess is that the variantion in type definition is caused
> by differences in ABI. I don't know s390 platforms, so I can't
> confirm offhand. But it seems like a plausible way to pass the
> argument.

Right, the two-element structure is passed by reference on s390x.  But
that also applies to the same structure in C/C++.  Nevertheless, with C
the argument's DW_AT_type reflects the real type and gcc adjusts the
DW_AT_location instead.

> If Eric confirms my guess, the attached patch should fix the problem.

Thanks.  Eric, do you confirm that it's intentional to adjust the
DW_AT_type based on ABI?

Joel, in the meantime I suggest to commit the second line of your patch.
It's independent of the outcome, and we might otherwise lose track of
it.

Re: Unexpected DW_AT_type for Ada strings on s390x?

[Eric Botcazou]

> Right, the two-element structure is passed by reference on s390x.  But
> that also applies to the same structure in C/C++.  Nevertheless, with C
> the argument's DW_AT_type reflects the real type and gcc adjusts the
> DW_AT_location instead.

It's passed by invisible reference in C and explicit reference in Ada.

> Thanks.  Eric, do you confirm that it's intentional to adjust the
> DW_AT_type based on ABI?

The ABIs are indeed different so the debug info will be different.

-- 
Eric Botcazou

Re: Unexpected DW_AT_type for Ada strings on s390x?

[Joel Brobecker]

> Joel, in the meantime I suggest to commit the second line of your patch.
> It's independent of the outcome, and we might otherwise lose track of
> it.

I'm just waiting for you to confirm that the patch fixes the FAIL,
and I will commit the whole patch. (I have been very busy these days)

-- 
Joel

Re: Unexpected DW_AT_type for Ada strings on s390x?

[Andreas Arnez]

Joel Brobecker <brobecker@adacore.com> writes:

> I'm just waiting for you to confirm that the patch fixes the FAIL, and
> I will commit the whole patch.

OK, confirmed.  The patch fixes the FAIL on s390x.

> (I have been very busy these days)

I can imagine.  Thanks for your help!

Re: Unexpected DW_AT_type for Ada strings on s390x?

[Joel Brobecker]

> > I'm just waiting for you to confirm that the patch fixes the FAIL, and
> > I will commit the whole patch.
> 
> OK, confirmed.  The patch fixes the FAIL on s390x.

Great, thanks for confirming. The patch is now in. See:
http://www.sourceware.org/ml/gdb-patches/2012-12/msg00127.html

-- 
Joel