On 2015-05-25 14:57, Arun Raghavan wrote: > On 25 May 2015 at 12:49, hwang4 <hui.wang at canonical.com> wrote: >> >> >> On 2015å¹´05æ??25æ?¥ 14:40, David Henningsson wrote: >>> >>> >>> >>> On 2015-05-25 06:49, Hui Wang wrote: >>>> >>>> On 32bits OS, this test case fails. The reason is when rewinding to >>>> the middle of a block, some of float parameters in the saved_state >>>> are stored in the memory from FPU registers, and those parameters will >>>> be used for next time to process data with lfe. Here if FPU register >>>> is over 32bits, the storing from FPU register to memory will introduce >>>> some variation, and this small variation will introduce small >>>> variation to the rewinding result. >>> >>> >>> Very interesting finding. I didn't know that storing things back and forth >>> to memory could change the computation result. >>> >>> And the fact that it only happens on 32-bit platforms and only with >>> optimisations makes it even stranger. Makes me wonder if this is actually an >>> gcc optimisation bug. >>> >> Probably. > > Did I misunderstand something here? Fixing a problem with the compiler > in our test seems to be incorrect. My assumption is that as long as we have -ffast-math on, then these are things we have to live with (and thus calling it a "gcc optimisation bug" is inaccurate, sorry about that). Being pragmatic: since the difference is one bit only, no one will hear any difference. But if you feel like digging into it further, I'm curious to know the result. :-) -- David Henningsson, Canonical Ltd. https://launchpad.net/~diwic
