Re: inlining + optimization = nuisance bugs

> [I've never understood why people are so eager to have fast wrong
> answers. -John]


It's not a matter of fast, wrong answers. It's a matter of
reproducible results. If the compiler is silently using 80-bit
temporaries some of the time, and I rearrange my code slightly, or the
compiler changes how it rearranges expressions slightly, I'll
needlessly get a different answer.


This same issue is causing vendors to insert switches to disable
generation of multiply/add instructions, or to change their chips so
that multiply/add rounds exactly like a multiply followed by an add
(MIPS R8000 vs. R1000). Yes, they're losing a tiny bit of precision.
But different answers should be an option, not forced on the user.


Given that Java has these fp rules so that answers will be consistent
across platforms, it would be stupid to force inconsistent answers. If
a compiler writer wants to make additional precision consistently
available, I'm all for it. If a compiler writer wants to make it
impossible for me to validate any code to the last bit, I'll have them
shot.


-- greg
[I think we agree here. -John]
--

Return to the comp.compilers page.
Search the comp.compilers archives again.