Re: [testsuite patch] i386 regression for funcargs.exp [Re: [PATCH 3/3] Remove HP-UX references fom testsuite]

[Pedro Alves <palves at redhat dot com>]

On 01/08/2016 09:15 PM, Simon Marchi wrote:

> I also don't see this on Ubuntu 14.04.3 LTS, gdb master, gcc 4.8.4.
> 

I don't see it either on F20, gcc 4.8.3.  But I do see it with
gcc 6 (20151117, getting old, but what I had handy).

The difference is that with newer GCC there's an extra instruction
after the call which is still assigned to line 583:

$ diff -up /tmp/4.8.3 /tmp/6.0.0 -U 1000
--- /tmp/4.8.3  2016-01-11 12:37:39.611089156 +0000
+++ /tmp/6.0.0  2016-01-11 13:21:00.021127976 +0000
@@ -1,27 +1,30 @@
 583       (*pointer_to_call0a) (c, s, i, l);    /* First step into call0a.  */
    mov    0x804d060,%ebx
    mov    0x804d050,%ecx
    movzwl 0x804d040,%eax
    movswl %ax,%edx
    movzbl 0x804d030,%eax
    movsbl %al,%eax
-   mov    %ebx,0xc(%esp)
-   mov    %ecx,0x8(%esp)
-   mov    %edx,0x4(%esp)
-   mov    %eax,(%esp)
-   mov    0x7c(%esp),%eax
+   push   %ebx
+   push   %ecx
+   push   %edx
+   push   %eax
+   mov    -0x1c(%ebp),%eax
    call   *%eax
+   add    $0x10,%esp

 584       (*pointer_to_call0a) (c, s, i, l);    /* Second step into call0a.  */
    mov    0x804d060,%ebx
    mov    0x804d050,%ecx
    movzwl 0x804d040,%eax
    movswl %ax,%edx
    movzbl 0x804d030,%eax
    movsbl %al,%eax
-   mov    %ebx,0xc(%esp)
-   mov    %ecx,0x8(%esp)
-   mov    %edx,0x4(%esp)
-   mov    %eax,(%esp)
-   mov    0x7c(%esp),%eax
+   push   %ebx
+   push   %ecx
+   push   %edx
+   push   %eax
+   mov    -0x1c(%ebp),%eax
    call   *%eax
+   add    $0x10,%esp
+

I don't know why -m32 changed to push/add instead of mov while 64-bit hasn't.

> But if there is a configuration where this is needed, then yes, it should be re-added.

I agree.

On 01/08/2016 07:30 PM, Jan Kratochvil wrote:
>      gdb_test_multiple "finish" "finish from indirectly called function" {
> +	-re "\\(\\*pointer_to_call0a\\) \\(c, s, i, l\\);.*First.*$gdb_prompt $" {
> +	    # On i386, gdb finishes at one line earlier than on x86_64.
> +	    # Therefore, an extra "step" is necessary to continue the test.

This is most likely needed on non-x86 ports as well.
I think we should say:

 # If the branch instruction is not the last instruction in the function call
 # line, we'll stop at that line, and need an extra "step" to continue the test.

OK with that change.

Thanks,
Pedro Alves

> +	    send_gdb "step\n"
> +	    exp_continue
> +	}
>  	-re ".*\\(\\*pointer_to_call0a\\) \\(c, s, i, l\\);.*Second.*$gdb_prompt $" {
>  	    pass "finish from indirectly called function"
>  	}