TMS320F280041: Does the compiler optimization level affect the result of arithmetic operation and how to avoid?

Will Gao

Part Number: TMS320F280041

Hi expert,

The customer did the following arithmetic operations with no optimization and with optimization level of 3, respectively.

Unexpectedly, the results of the above two situations are different. If turning on the optimization, the "t_f32Temp552" equals to 0x3F3DE3DF while the "t_f32Temp55max" equals to 0x3F3DE3DE, as you can see from the following figure. However, if turning off the optimization, they both equal to 0x3F3DE3DE.

(with optimization level of 3)

By the way, the version of customer's CCS is v7.3.0, and the version of CGT is showing below:

The customer wants to know how to avoid this problem.

over 4 years ago

0 Ashwini Athalye over 4 years ago

TI__Expert 7715 points

Hi Will Gao,

The two values when optimization are within the margin of error. One reason I can think of for the difference in behavior with optimization turned on and off is that with optimization, the compiler may be replacing the "540.0f / LWIN_VOL_RANGE) by the computed value which is getting loaded into t_f32Temp55max and the compiler may be determining it to be 0x3F3DE3DF. Where as with optimization turned off the hardware is doing both computations for t_f32Temp552 and t_f32Temp55max resulting in the same hex value.

You can confirm this by looking at the generated assembly.

Thanks,

Ashwini

0 Will Gao over 4 years ago in reply to Ashwini Athalye

TI__Intellectual 2335 points

Thank you, Ashwini. I understand the reason behind this difference and has returned this reason to the customer. But they want to further know how to code so that the results are the same under different compilation optimization levels. If possible, please give some advices. Thank you so much!

0 Ashwini Athalye over 4 years ago in reply to Will Gao

TI__Expert 7715 points

Hi Will,

Apologies for the late reply. Please could you post the compiler generated assembly code for the two optimization levels? I will follow up with the compiler team to get inputs on whether the exact same result can be generated with optimization on.

Thanks,

Ashwini

0 Will Gao over 4 years ago in reply to Ashwini Athalye

TI__Intellectual 2335 points

Hi Ashwini, the generated assembly codes for the two optimization levels are shown below. Thanks a lot!

0 Ashwini Athalye over 4 years ago in reply to Will Gao

TI__Expert 7715 points

Hi Will,

Thank you. I will follow up with the compiler team.

Thanks,

Ashwini

0 George Mock over 4 years ago in reply to Ashwini Athalye

TI__Guru**** 251490 points

For the source file with problem division expressions, please follow the directions in the article How to Submit a Compiler Test Case.

Thanks and regards,

-George

0 Will Gao over 4 years ago in reply to George Mock

TI__Intellectual 2335 points

/cfs-file/__key/communityserver-discussions-components-files/171/8686.main.pp.txt

Hi George, the above is the preprocessed source file, please check it. By the way, the customer's compiler version is v16.9.4.

The compiler options are shown in below links, one for no optimization and the other one for optimization level 3:

/cfs-file/__key/communityserver-discussions-components-files/171/with-optimization-level-of-3_5F00_-compiler_2D00_option.txt

/cfs-file/__key/communityserver-discussions-components-files/171/without-optimization-_5F00_-compiler_2D00_option.txt

Thanks for your help.

0 George Mock over 4 years ago in reply to Will Gao

TI__Guru**** 251490 points

Thank you for submitting a test case. I am able to reproduce a very similar result. I filed the issue EXT_EP-10499 to have this investigated. You are welcome to follow it with that link.

Thanks and regards,

-George

0 Will Gao over 4 years ago in reply to George Mock

TI__Intellectual 2335 points

Thanks for your help, George! I will follow it.

0 Will Gao over 4 years ago in reply to George Mock

TI__Intellectual 2335 points

By the way, the customer wants to know if there are some coding patterns or coding standards to avoid those problems. They have concerns about the consistency of optimization so that they want a consistency-guaranteed coding and compiler setting method. If so, please share the method, too.

Thanks.

-Will

0 George Mock over 4 years ago in reply to Will Gao

TI__Guru**** 251490 points

Will Gao said:
the customer wants to know if there are some coding patterns or coding standards to avoid those problems

I'm not aware of a workaround to the problem you reported. After the problem is characterized, a workaround might be developed.

Speaking more generally ... Using optimization should not result in a difference in program behavior. If it does, we view that as a problem that should be addressed.

Thanks and regards,

-George

+1 GregM over 4 years ago in reply to George Mock

TI__Expert 5385 points

Will,

This is not an optimization issue. This is an issue with using --fp_mode as relaxed instead of strict. To avoid above issue you need to use strict fp_mode.

For more details on strict vs relaxed, please see spru514w Compiler User guide section 2.3.3 which lists many of the relaxations that the compiler provides.

This is a trade off with performance vs precision.

Regards,
Greg

0 Will Gao over 4 years ago in reply to GregM

TI__Intellectual 2335 points

Thank you, Greg. That makes sense.

Regards

-Will

C2000™︎ microcontrollers

C2000 microcontrollers forum

TMS320F280041: Does the compiler optimization level affect the result of arithmetic operation and how to avoid?