Optimization cause exception on C6657 EVM board.

Denny%20Yang99373

Hi all,

I used -O3 to optimized attached project, it will cause a exception as below:

External exception happened. MEXPFLAG[3]=0x4000000.
Event 122: DMC_CMPA CPU memory protection fault for L1D (and other memory read finally goes through the L1D controller)
memory protection exception caused by local access at 0x10
Supervisor Read violation
NRP=0xc00684c, NTSR=0x1000e, IRP=0x0, ITSR=0x0, TSCH= 0x2, TSCL= 0x41f23e57
B3=0xa, A4=0x0, B4= 0x0, B14= 0x8307f0, B15= 0x827e90
Exception happened at a place can not safely return!

And I found if change out[k] = (int32)(count>>16); to out[k] = (int32)(count/0xffff); everything is fine. In function inline int32 cal_difcurerror(int32 *difcur,int16 *cosbuf).

And if I change cgt from v7.14 to v8.1, everything is fine.

This code run on C6657 EVM, could you please to check why this happened?

thanks!

Robust.zipBR,

Denny

over 9 years ago

Denny%20Yang99373 over 9 years ago

TI__Expert 3805 points

If I change
for(k=0;k<HALF_SMPL_STEP;k++)
{
count=0;
for(l=0;l<SMPL_STEP;l++)
{
count += tempbuf[k+l]*cosbuf[l];
}
out[k] = (int32)(count>>16);
}
to
for(k=0;k<HALF_SMPL_STEP;k++)
{
count=0;
for(l=0;l<SMPL_STEP;l++)
{
count += tempbuf[k+l]*cosbuf[l];
}
_mfence();
out[k] = (int32)(count>>16);
}
Also everything work fine. It may have some issue about pipeline with 64bit data >>16.
thanks!

Denny%20Yang99373 over 9 years ago in reply to Denny%20Yang99373

TI__Expert 3805 points

Also post the trace file as well.Robust_trace.zip

Denny%20Yang99373 over 9 years ago in reply to Denny%20Yang99373

TI__Expert 3805 points

This is the register snapshot for help.

Archaeologist over 9 years ago

TI__Guru* 84285 points

Does the exception happen consistently? Does it always happen around the same cycle count (TSCH,TSCL)?

lding over 9 years ago in reply to Archaeologist

TI__Guru* 95265 points

Sorry, not able to build your project. Missing some header file in the package you uploaded.

"../src/Robust_INTC.c", line 31: fatal error #5: could not open source file "KeyStone_common.h"
1 fatal error detected in the compilation of "../src/Robust_INTC.c".

Regards, Eric

Denny%20Yang99373 over 9 years ago in reply to lding

TI__Expert 3805 points

Sorry for miss a folder.common.zip

Denny%20Yang99373 over 9 years ago in reply to Archaeologist

TI__Expert 3805 points

Hi Archaeologist,

Yes, the exception happened consistently. The cycle may be a little difference of tscl.
thanks!

BR,
Denny

lding over 9 years ago in reply to Denny%20Yang99373

TI__Guru* 95265 points

Denny,

Yes, the problem can be reproduced and noticed the difference between CGT 7.4.14 and 8.1.0.. Also, the _mfence() made the difference. I think this is more of tool/optimization issue.

Archaeologist, are you in CCS or CGT team? Can you follow this up with Denny?

Regards, Eric

Archaeologist over 9 years ago in reply to lding

TI__Guru* 84285 points

Yes, I'm on the compiler team. We need a bit more before we can start looking at this as a compiler issue. We need to be able to reproduce the issue, but I don' t have an appropriate EVM. Next, we need to know exactly what sequence of instructions triggers the problem. These are things the compiler team is not equipped to do. Once we know the incorrect sequence, the compiler team can investigate why the compiler generates it.

AartiG over 9 years ago in reply to lding

TI__Guru**** 173310 points

Denny, Eric,

I was also able to reproduce the problem on a C6657 EVM and noted that the code changes mentioned in earlier posts, as well as usage of CGT 8.1.0 works around the issue. I can see the sequence of instructions after which the exception is generated, but I don't have the expertise to tell exactly what or why it is triggered.

The compiler team may need more details and assistance when they start looking into it, but I have filed a bug report to track its progress. The tracking # is CODEGEN-1272.

Denny%20Yang99373 over 9 years ago in reply to AartiG

TI__Expert 3805 points

Hi AartiG,

Could you please help to update this issue's status? Customer is busy on this issue. thanks!

BR,
Denny

mjoyner over 9 years ago in reply to Denny%20Yang99373

TI__Intellectual 2010 points

This issue has not been assigned to an engineer yet. It should get assigned to an engineer soon to analyze.

Denny%20Yang99373 over 9 years ago in reply to mjoyner

TI__Expert 3805 points

Hi mjoyner,

Thanks for your update, could you please help to speed up the process flow? thanks!

BR,
Denny

Code Composer Studio™︎

Code Composer Studio forum

Optimization cause exception on C6657 EVM board.