Re: [PATCH] Modify altivec-regs.exp testcase for AIX

[Aditya Kamath1 <Aditya.Kamath1@ibm.com>]

[-- Attachment #1: Type: text/plain, Size: 2510 bytes --]

Hi Ulrich and community,

>I think it would be preferable to instead extend the
>test case (when compiled on AIX only) by adding some
>other instruction early in main, but before that
>assignment to x, that touches a vector register,
>and then perform the GDB register tests after that
>new instruction and before the assignment to x.

So I do not have the knowledge to do that. I had seen the align-c test case which uses a tcl script and can do the same. However in .exp file how can we do it? Do you have an example which I can look into and learn how to does it, anything different also is fine. Kindly let me know. It will be of help to me, to contribute the same.

Have a nice day ahead.

Thanks and regards,
Aditya.

From: Ulrich Weigand <Ulrich.Weigand@de.ibm.com>
Date: Friday, 24 February 2023 at 9:07 PM
To: gdb-patches@sourceware.org <gdb-patches@sourceware.org>, Aditya Kamath1 <Aditya.Kamath1@ibm.com>, simark@simark.ca <simark@simark.ca>
Cc: Sangamesh Mallayya <sangamesh.swamy@in.ibm.com>
Subject: Re: [PATCH] Modify altivec-regs.exp testcase for AIX
Aditya Kamath1 <Aditya.Kamath1@ibm.com> wrote:

>Kindly review the patch in the previous email which changed
>the altivec-regs.exp so that AIX folks can also use this while
>in development. We are currently working on this feature.

This still needs at the very least more comments.  Note that
the existing comments in the test case specifically say that
this test *wants* to verify that VRs can be written before
they were ever used - which works on Linux but apparently
doesn't work on AIX.

Your patch changes this, but doesn't change those comments,
which makes the resulting test confusing to read.

Also, as a functional concern: you now execute the GDB
commands to modify vector register *after* the line
   x = ((vector unsigned int) vec_splat_u8 (-2));
was executed.

If the compiler chose to hold the value of x in a
register at this point, those GDB commands would
clobber the value of x, causing subsequent execution
to fail.

While this might happen to work with the executable
produced with the current compiler and/or settings,
this is not guaranteed.

I think it would be preferable to instead extend the
test case (when compiled on AIX only) by adding some
other instruction early in main, but before that
assignment to x, that touches a vector register,
and then perform the GDB register tests after that
new instruction and before the assignment to x.

Bye,
Ulrich