Am Di., 12. Mai 2020 um 13:53 Uhr schrieb Damjan Marion <dmar...@me.com>:
> > > > On 12 May 2020, at 11:41, Andreas Schultz < > andreas.schu...@travelping.com> wrote: > > > > Hi, > > > > There are few places in VPP (most notable when trying to convert a > netmask into a length) that can pass 0 (zero) into min_log2 and expect to > get a meaningful result. > > > > Obviously log2(0) is undefined. It turns out that this also applies to > the return of min_log2. Under the hood the function uses __builtin_clzl, > and the return of that function is also undefined for input 0. > > > > __builtin_clzl could be replaced with __builtin_ia32_lzcnt_u64 on > supported CPUs to avoid the undefined behaviour. That would still not fix > the problem that passing 0 into a log function is broken by design. > > Compiler uses LZCNT for __builtin_clzl when target -march supports it. > Today we use -march=corei7 as baseline and that one doesn’t know about > LZCNT, so BSR is emitted instead, which indeed have undefined behaviour. In > other words replacing __builtin_clzl with __builtin_ia32_lzcnt_u64 will not > work, but simply changing baseline -march= to something newer will address > that transparently... > Even using -march doesn't change the fact that __builtin_clzl is documented to have undefined behavior for zero inputs. Relying on undocumented side effects is IMHO not a good idea. Not to mention that clang and gcc might behave differently. Andreas
-=-=-=-=-=-=-=-=-=-=-=- Links: You receive all messages sent to this group. View/Reply Online (#16335): https://lists.fd.io/g/vpp-dev/message/16335 Mute This Topic: https://lists.fd.io/mt/74155196/21656 Group Owner: vpp-dev+ow...@lists.fd.io Unsubscribe: https://lists.fd.io/g/vpp-dev/unsub [arch...@mail-archive.com] -=-=-=-=-=-=-=-=-=-=-=-
