[Monotone-devel] Re: illegal instruction [was Re: Problem with build of monotone 0.11 on Debian testing]

[graydon hoare]

Robert Helgesson wrote:

> Just checking for the emmintrin.h file doesn't suffice. I have it on my
> system and I have an old Intel Celeron which definately does not have
> support for SSE2.

oh yes, I'm aware of this. emmintrin.h is just the intrinsics 
definitions. those are necessary to support the code generation from 
high level functions at compile time, rather than inline asm. nothing to 
do with whether your chip does sse2 (you could always be cross-compiling 
anyways). I always do detection at runtime anyways, using cpuid.

what's got me confused is that even with my "detect at runtime" code, it 
appears to crash on an athlon. I have checked to make sure the machine 
can execute cpuid on its own, and it can (and successfully reports no 
support for SSE2, bit 26 of edx is zero) so I'm a bit bewildered about 
the fault. I can only assume gcc with -msse2 is spontaneously generating 
sse2 code somewhere *outside* my runtime-switched core, but I didn't 
think it was supposed to do that. it's probably something a bit more 
subtle (read: silly) in my own code.

I'll look at it more tonight. I'm sure there's a way to get it right; I 
only discovered it wasn't working on athlons at the last moment so I 
disabled it by commenting out the -msse2 check in configure.ac for the 
distribution build.

-graydon