Wrong optimization in icx when using arch-consistency

[Frank_R_1]

Hi

We found a problem in Intel oneAPI 2025.3.1 which was not present in 2024.2.1

Please have a look at:

2024.2.1 (here everything is bit identical)

https://godbolt.org/z/3E6YerxWW

2025.3.1 (output for Re = 1.0000e+07 is different)

https://godbolt.org/z/67rjM1jWr

It boils down to usage of 

__bwr_log@PLT vs  log@PLT

 left debug, right release

The problem is the same on both platforms Windows and Linux.

We heavily rely on bit identical results between relase and debug builds, which was given over the last years. (On an older parallel studio version I reported a similar bug which was then fixed later on)

Best regards

Frank

[Frank_R_1]

FYI:

We found out that with 2025.1.0 the problem emerges, before it was ok. You can verify this also with the godbolt links above.

[Frank_R_1]

FYI:

When using -fp-model=strict instead of -fp-model=precise it delivers bit identical results. We didn't find information about it in the release notes about that change.

Best regards

Frank

[Sravani_K_Intel]

HI @Frank_R_1 , I just tested this with our latest development branch compiler and can no longer reproduce the issue. Please retest your code when 2026.0 Compiler is available in a few weeks.