gdb.cp/templates.exp failures

[Ulrich Weigand]

FAIL: gdb.cp/templates.exp: ptype T5<int>
FAIL: gdb.cp/templates.exp: ptype t5i

Further analysis required.

[Yao Qi (yao-codesourcery)]

[CodeSourcery #1713]

[Ulrich Weigand (uweigand)]

This looks like a GCC issue. GDB wrongly shows a return type on class constructors:
    T5<int> * T5(int);
    T5<int> * T5(const T5<int> &);
which actually comes straight from DWARF debug information:

 <2><266>: Abbrev Number: 20 (DW_TAG_subprogram)
    <267> DW_AT_external : 1
    <268> DW_AT_name : T5
    <26b> DW_AT_decl_file : 1
    <26c> DW_AT_decl_line : 469
    <26e> DW_AT_type : <0x30d>
    <272> DW_AT_declaration : 1
    <273> DW_AT_sibling : <0x283>
 <3><277>: Abbrev Number: 9 (DW_TAG_formal_parameter)
    <278> DW_AT_type : <0x30d>
    <27c> DW_AT_artificial : 1
 <3><27d>: Abbrev Number: 10 (DW_TAG_formal_parameter)
    <27e> DW_AT_type : <0x29>

Note where the return type (at <26e>) is equal to the type of the "this" parameter (at <278>).

I don't quite understand why the same compiler version on Intel doesn't appear to have this problem; I'd have expected debug info generation problems to be platform-independent. More analysis on the GCC side is required ...

[Ulrich Weigand (uweigand)]

Further investigation shows that this is deliberate, and complies with the AAPCS ABI. To quote the "C++ ABI for the ARM Architecture":

GC++ABI §3.1.5 Constructor return values
This ABI requires C1 and C2 constructors to return this (instead of being void functions) so that a C3 constructor can tail call the C1 constructor and the C1 constructor can tail call C2.

GDB has a bug here anyway as constructors really ought to be shown without any return type:
http://sourceware.org/bugzilla/show_bug.cgi?id=8216
but the test case currently accepts a return type of "void" as KFAIL.

I think we should add a second KFAIL case accepting a pointer to the class type as return type and/or fix the underlying GDB problem.

[Ulrich Weigand (uweigand)]

>I think we should add a second KFAIL case accepting a pointer
>to the class type as return type and/or fix the underlying GDB
>problem.

I've implemented the first solution (updating the KFAIL detection) for now. This gets ARM to the same state as other platforms: the output is wrong, but detected as known failure by the testsuite.