On Thu, Dec 19, 2013 at 10:16:16AM +0100, Stefan Esser wrote: > Am 30.11.2013 14:56, schrieb Konstantin Belousov: > > I propose to unconditionally add the switch -fno-strict-overflow > > to the kernel compilation. See the patch at the end of message for > > exact change proposed. > > > > What does it do. It disallows useless and counter-intuitive > > behaviour of the compiler(s) for the signed overflow. Basically, > > the issue is that the C standard left signed overflow as undefined > > to allow for different hardware implementation of signess to be > > used for signed arithmetic. De-facto, all architectures where > > FreeBSD works or have a chance to be ported, use two-complement > > signed integer representation, and developers intuition is right > > about it. > > > > The compiler authors take the undefined part there as a blanket to > > perform optimizations which are assuming that signed overflow > > cannot happen. The problem with that approach is that typical > > checks for bounds are exactly the place where the overflow can > > happen. Instead of making some artificial example, I would just > > point to my own r258088 and r258397. > > > > What makes the things much worse is that the behaviour is highly > > depended on the optimization level of the exact version of > > compiler. > > > > What other projects did in this regard. They turned the same knob > > unconditionally. I can point at least to Linux kernel and > > Postgresql. Python uses -fwrapv, which is equivalent to the > > -fno-strict-overflow on the two-complement machines. Linux used > > -fwrapv before switched to -fno-strict-overflow. > > Hi Konstantin, > > you may put back -fno-strict-overflow after I found and fixed the > problem uncovered by enabling it in -CURRENT (SVN rev. 259609). > > The problem was an overflow in the conversion of timeout values to > sbintine, which lead to negative values being detected with > -fno-strict-overflow, while the compiler performed the signedness > test before the multiplication, without that option. > > I found that timeout values of more than 1000 years were requested > by some programs, which are now capped at 68 years (the maximum that > can be represented by sbintime, 2^31 seconds). > > So, -fno-strict-overflow has already proved itself to be useful > in uncovering a bug that would have been hard to find, otherwise. Feel free to restore the commit, I have no plans to do this.
This archive was generated by hypermail 2.4.0 : Wed May 19 2021 - 11:40:45 UTC