Re: [HACKERS] LLVM miscompiles numeric.c access to short numeric var headers

Fwiw it looks like the LLVM folk think this is an asan bug. https://llvm.org/bugs/show_bug.cgi?id=25550 It looks like the fix is that the compiler should avoid this optimization if the code is being compiled with instrumentation. This worries me a bit but I think our code is safe as the Datum

[HACKERS] LLVM miscompiles numeric.c access to short numeric var headers

I've been using LLVM's sanitizers and asan turned up a new bit of compiler behaviour that I hadn't had before. I don't see it in 3.7 or before, only in their HEAD so I don't know if it's a bug or intentional. In numeric.c we have the short numeric headers that have one uint16 (in addition to the

Re: [HACKERS] LLVM miscompiles numeric.c access to short numeric var headers

Greg Stark writes: > In numeric.c we have the short numeric headers that have one uint16 > (in addition to the varlena header) followed by digits. When compiling > with -O2 on x86-64 LLVM now seems to use a 4-byte access. Either that's a reportable compiler bug, or someplace

Re: [HACKERS] LLVM miscompiles numeric.c access to short numeric var headers

Greg Stark writes: > On Thu, Nov 12, 2015 at 2:39 PM, Tom Lane wrote: >> Either that's a reportable compiler bug, or someplace nearby we've >> casted the pointer to something that would require a 4-byte struct. >> I'm not sure which code you're looking at

Re: [HACKERS] LLVM miscompiles numeric.c access to short numeric var headers

On Thu, Nov 12, 2015 at 2:39 PM, Tom Lane wrote: > Either that's a reportable compiler bug, or someplace nearby we've > casted the pointer to something that would require a 4-byte struct. > I'm not sure which code you're looking at exactly, but maybe we're > using "union

Re: [HACKERS] LLVM miscompiles numeric.c access to short numeric var headers

On Thu, Nov 12, 2015 at 3:51 PM, Tom Lane wrote: > Meh. The palloc to create an aligned array of digits would eat up > any possible performance win --- it'd be just about as expensive > as the existing unpack operation. I suppose we would only need to palloc the digits if we

Re: [HACKERS] LLVM miscompiles numeric.c access to short numeric var headers

Greg Stark writes: > On Thu, Nov 12, 2015 at 3:51 PM, Tom Lane wrote: >> I think we could fix the immediate issue by redeclaring numeric >> headers as arrays of (u)int16 rather than structs. I'm not >> very excited about the packed-header case. > That would