[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

Re: Fix for review



Hi,

On 9/19/25 6:41 PM, tshah wrote:

1. The failure also occurs on arm64, which has always used IEEE . This strongly suggests that the issue is not specific to IBM format. Also most probably  the recent transitions are for long double and here btas is using double.

Mh, so it may be passing on x87 FPUs only, which is brittle even on x86/x86_64 because it depends on using x87 style FPU instructions and keeping intermediates on the FPU stack without spilling them, so different compiler options can influence the result.

In that case, a larger epsilon might be correct, but it could also be that the expected value is wrong (because it was calculated with an x87 style FPU), so the proper fix is to fix the expected value to what an IEEE conformant implementation outputs, and then maybe extend the epsilon for x86.

   Simon


Reply to: