On Sun, Jan 05, 2025 at 08:45:44PM +0100, Anders Magnusson wrote: > > Under -std=c11, this rounds the product to float precision first, and > > then computes the sum, as required by the C spec. Under -std=gnu11, > > this may compute everything in double or extended precision, just like > > if it had been `return x*y + c' with no intermediate `float' variable > > assignment. (The finer-grained option here is -fexcess-precision: > > under -std=c11 it's set to `standard'; under -std=gnu11, to `fast'.) > > Hm, interesting optimization.? I understand why they have done this, > but I would not have expected it to be the default behaviour. > Writing code like this is not uncommon to get rounding as expected.
These are the people who came up with -ffast-math. Anyway, I object to C11 without explicitly disabling the memory model stuff, unless we are VERY VERY SURE it isn't going to make half the kernel into UB. Which I don't think we can be. -- David A. Holland dholl...@netbsd.org
