canonmn_int: round converted focal lengths

[geri1701]

On i686, x87 excess floating-point precision can make products such as 100.0F * 1.4F slightly lower than 140. Converting that value directly to int truncates it to 139, so teleconverter lens entries do not match the focal lengths decoded from metadata.

This rounds the focal lengths after applying the teleconverter factor and before converting them to int. The existing lens tests cover the affected cases.

Verification:

• i686: all 6 CTest suites pass (baseline: 28 lens-test failures)
• x86_64: all 6 CTest suites pass

[kevinbackhouse]

Thanks! This is an improvement.

By the way, do you know whether there's a possibility that the number could be too big, causing the static_cast<int> to overflow? If so, somebody is inevitably going to send us a vulnerability report about it so it might be good to fix it proactively. I recently wrote some code to do a cast with overflow checking in #9433.

[mergify]

Tick the box to add this pull request to the merge queue (same as @mergifyio queue).

• Queue this pull request

[geri1701]

By the way, do you know whether there's a possibility that the number could be too big, causing the static_cast<int> to overflow? If so, somebody is inevitably going to send us a vulnerability report about it so it might be good to fix it proactively. I recently wrote some code to do a cast with overflow checking in #9433.

The values passed to std::lround here come from the built-in canonCsLensType labels rather than metadata. Across the current 533 entries, the largest focal-length × teleconverter value is 2400, so the conversion is safely within the range of int.