Compiler Bug in 7.4.14

Ralf Goebel

I think I found a compiler bug in CG Tools v7.4.14 when using -O2 on a C6678 CPU. It can be reproduced with the following test code:

void test()
{
	unsigned char Buffer[64];
	int i;

	for (i=0; i<64; i++)
	{
		int pos = i%8;
		if (pos < 3)
			Buffer[3*(i/8)+pos] = i;
	}

	while (1)
		printf("%u\n", Buffer[0]);
}

When calling this function, the for() loop spins forever and the printf() line is never reached. The problem goes away after remove the while(1) line.

Thanks,
Ralf

over 9 years ago

0 George Mock over 9 years ago

TI__Guru**** 239845 points

Please show the exact compiler build options you use.

Thanks and regards,

-George

0 Ralf Goebel over 9 years ago in reply to George Mock

Genius 3715 points

Hi,

these are the compiler options without the include paths:

-mv6600 --abi=eabi -O2 -g --quiet --display_error_number --diag_warning=225 --gen_func_subsections=on --no_bad_aliases --mem_model:data=far --preproc_with_compile --preproc_dependency="test.pp"

Ralf

0 Archaeologist over 9 years ago in reply to Ralf Goebel

TI__Guru* 84225 points

Just to be sure, did you include stdio.h before calling printf?

0 Ralf Goebel over 9 years ago in reply to Archaeologist

Genius 3715 points

Yes, I did include stdio.h. The printf is reached as soon as I remove only the while(1) line.

But the printf() seems not important, I can also put different function call there. Example:

while (1)
memset(Buffer, 0, sizeof(Buffer));

Again, memset() is never reached. The DSP keeps spinning in the for() loop after halting the CPU.

Thanks,
Ralf

0 Archaeologist over 9 years ago in reply to Ralf Goebel

TI__Guru* 84225 points

I can reproduce the problem, but I haven't quite characterized it yet, so I don't know if I can find a workaround. I've submitted SDSCM00052214 to track the issue.

0 pf over 9 years ago in reply to Archaeologist

TI__Expert 4930 points

A workaround is to add "#pragma UNROLL(1)" ahead of the for-loop.

0 pf over 9 years ago

TI__Expert 4930 points

A lovely and bizarre little compiler bug. To make it happen, one needs a loop whose body is a one-armed IF (ie, no "else"), and that that loop be suitable for unrolling and present in a function that ends in an infinite loop. The presence of the infinite loop messes up one of the compiler analyses used in the unrolling.

We've identified a couple of possible solutions and are discussing which one is best. It is likely that one will appear in 7.4.15.

In the meantime, use "#pragma UNROLL(1)" to disable the unrolling, or move the infinite loop out of the function containing the troubled loop; either will avoid the problem.

Code Composer Studio™︎

Code Composer Studio forum

Compiler Bug in 7.4.14