FW: Minutes of Meeting (29-Aug-2012): Status and Queries in icyflex2 Work Package 1 and 2

[Deepti Sharma]

-----Original Message-----
From: Deepti Sharma [mailto:deepti.sharma@rf-silicon.com] 
Sent: 29 August 2012 17:36
To: 'MORGAN Marc-Nicolas'
Cc: 'rahul.sud@silicon-capital.com'; 'Gopal Dharmaraj';
'satish.bhonsle@rf-silicon.com'; 'rakesh.adhikari@rf-silicon.com';
'damodar.sonone@rf-silicon.com'; 'Hari Shankar Singh'; 'VOLET Patrick';
'NAGEL Jean-Luc'
Subject: Minutes of Meeting (29-Aug-2012): Status and Queries in icyflex2
Work Package 1 and 2

Hello Marc,

Please find below the Minutes of today's meeting.

Regards,
Deepti


Topic: Status and Queries icfx2 WP1 and WP2
Date : 29-Aug-12 (Wednesday)
Time : 14:00 Hrs - 14:35 Hrs IST

Members Present:
-----------------
[CSEM]
1. Marc MORGAN
2. Patrick VOLET
3. Jean-Luc NAGEL

[RF-Silicon]
1. Deepti Sharma
2. Damodar Sonone
3. Gaganpreet
4. Shweta Gupta
5. Amit Ranjan

Summary
--------
1. The status and queries regarding WP1 and WP2 were mailed to CSEM.

Mail Reference: Agenda document for call on 29-Aug-2012
Dated         : 28-Aug-2012
Document      :
Icyflex2-GCC-Compiler-WorkPackage-StatusAndQueries_29Aug2012.pdf

2. The status of WP1 was discussed as follows:

Activities Completed
---------------------
a. RF-Silicon has resolved the problem in WP1 sent by CSEM. The test suite
contains 2 unexpected (execution) errors for the WP1 solution.

b. The latest release was done to CSEM on 28-Aug-2012. 

c. Marc said that they are happy with the fix done by RFS.

Bottlenecks
------------
a. The test suite contains 2 unexpected (execution) errors for the WP1
solution.

b. Deepti asked if CSEM has graphical simulator to step through the
instructions.

c. Marc replied that they do not have a graphical simulator. We can use the
icfx2-run utility to see the cycles taken by the program.

d. Deepti asked how we can dump the content of registers at each
instruction. Marc said that we can try the command line options
print-after-cycles or verbose option to see the registers dump.

e. Deepti asked if CSEM has got a chance to have a look at the execution
errors. Jean said that after a quick look, it was observed that the program
aborts, but the exact cause is yet to be debugged.

f. Marc suggested RFS to check with icfx2-run utility. If RFS is getting
stuck, or is unable to debug further, we should inform them so that CSEM
might have a look at the issue to help us moving further.

g. Marc said that if RFS is not able to understand the icfx2-run utility, he
will provide some documentation on usage of the tool.


Additional Issues Resolved
----------------------------
a. RFS had observed additional patterns having same problem as WP1. Most of
the patterns have been fixed by RFS, except issue 2. 

movi r2, #.LC1 
movb (px7,0), r2

b. Deepti told that this issue is related with removal of PSImode. So, she
suggested if we can consider this issue to be fixed along with WP2.

c. Jean said that it might not be an issue at all, since the first
instruction moves an address to r2 and second instruction reads it. It is
not a load instruction, and might not create any stalls. However, RFS should
verify this to confirm that the latency of the first instruction is zero.

d. The verbose option of the scheduler may be used to check the stalls for
these instructions.


Use of -frename-registers
---------------------------
a. Marc confirmed that the test case that he had sent yesterday, needs to
use frename-registers, as suggested by RFS.

b. Deepti asked whether they want RFS to work on to enable this option by
default for icfx2?

c. Jean replied that it can be done, if it is not very complex. Also, gcc
test suite should be run again to verify that all tests are passing.

d. Also, RFS should suggest if there are additional options that should be
enabled by default for icfx2.

e. Deepti asked if CSEM would like some command line option like -Oicfx2, to
enable all options required by icfx2. 

f. Jean replied that they don’t want the user to specify additional option.
So, it must be enabled in the code itself. 
In addition, a command line option like -fnorename-registers can be
provided, to disable default renaming of registers.

g. Deepti confirmed that RFS can take up this task.


3. The status of WP2 was discussed as follows:

Activities Completed
---------------------
a. Deepti mentioned that RFS has completed setup for WP2, and is able to
achieve the same test result as sent by CSEM. She further mentioned that 16
unexpected errors of segmentation fault are resolved by increasing the stack
size and currently we are left with 60 unexpected errors.

b. RFS can now start the work on WP2.

c. Jean mentioned that they had deactivated some command line options (dce,
if-conversion2), to suppress some errors in the gcc test suite. If they
enable these options, the test suite results in more errors. RFS should
check on these options.

d. Out of the 60 errors obtained for gcc 4.6.3 (without applying the no-psi
patch), some are due to known issue of gcc, for which patch is available.
CSEM will share the link to the patch. On applying the patch, the errors
reduce further.

e. Deepti told that RFS team will investigate on this and resolve the
errors. 



> -----Original Message-----
> From: Deepti Sharma [mailto:deepti.sharma@rf-silicon.com]
> Sent: 29 August 2012 09:45
> To: 'MORGAN Marc-Nicolas'
> Cc: rahul.sud@silicon-capital.com; 'Gopal Dharmaraj';
> satish.bhonsle@rf-silicon.com; rakesh.adhikari@rf-silicon.com;
> damodar.sonone@rf-silicon.com; 'Hari Shankar Singh'; 'VOLET Patrick';
> 'NAGEL Jean-Luc'
> Subject: RE: Agenda document for call on 29-Aug-2012
> 
> Hello Marc,
> 
> We thank you for your feedback. We look forward for the discussion.
> 
> We have checked the attached test case, and we understand that you have
> not used command line option "frename-registers". Please note that gcc
> by default allocates registers optimally, and to reduce false
> dependencies, we need to use -frename-registers option. We have
> explained this in the Problem Analysis report for WP1.
> 
> We are obtaining different assembly with "-O2 -frename-registers". The
> file is attached for your reference. Please confirm.
> 
> Regards,
> Deepti
> 
> 
> 
> 
> > -----Original Message-----
> > From: MORGAN Marc-Nicolas [mailto:marc.morgan@csem.ch]
> > Sent: 28 August 2012 20:55
> > To: Deepti Sharma
> > Cc: rahul.sud@silicon-capital.com; 'Gopal Dharmaraj';
> > satish.bhonsle@rf-silicon.com; rakesh.adhikari@rf-silicon.com;
> > damodar.sonone@rf-silicon.com; 'Hari Shankar Singh'; VOLET Patrick;
> > NAGEL Jean-Luc; MORGAN Marc-Nicolas
> > Subject: RE: Agenda document for call on 29-Aug-2012
> >
> > Hi Deepti,
> >
> > Thanks for the agenda and the questions. We had a short meeting here
> > among ourselves to prepare the discussion for tomorrow.
> >
> > Regarding WP1, we are quite impressed with the result you have been
> > able to achieve. We understand the work is not completely finalized
> > since there are still 2 unexpected errors to sort out.
> >
> > The question which came up is whether or not your solution is
> generic.
> > It seems like the solution works well for memory to memory copies. It
> > is not completely generic in that if a C code does something similar
> > to a memory to memory copy, the solution no longer applies. I'm
> > attaching a sample C code and the corresponding disassembled code
> > which shows a counter example. This obviously does not diminish in
> any
> > way your solution which I'm sure we will be using.
> >
> > We were wondering whether you had considered a solution based on
> > modifying the scheduler to distinguish the latencies of dependencies
> > of the individual resources of the instruction. It seems that gcc
> > supports hooks to do this. The team here seems to think this would be
> > a path worth exploring. Let's discuss it tomorrow as well.
> >
> > Best regards,
> > Marc
> >
> > -----Original Message-----
> > From: Deepti Sharma [mailto:deepti.sharma@rf-silicon.com]
> > Sent: mardi 28 août 2012 13:45
> > To: MORGAN Marc-Nicolas
> > Cc: rahul.sud@silicon-capital.com; 'Gopal Dharmaraj';
> > satish.bhonsle@rf-silicon.com; rakesh.adhikari@rf-silicon.com;
> > damodar.sonone@rf-silicon.com; 'Hari Shankar Singh'; VOLET Patrick;
> > NAGEL Jean-Luc
> > Subject: Agenda document for call on 29-Aug-2012
> >
> > Hello Marc,
> >
> > Please find attached the agenda document for tomorrow's call. The
> > document also contains our queries that we would like to discuss
> > during the call.
> >
> > Regards,
> > Deepti