[llvm-commits] Patch: avoid legalization assert

[Mon Ping Wang]

Hi Duncan,

Yes, you have it right.  In general, its better to run vector code  
using the SSE instructions and avoid using MMX instructions.  By using  
-disable-mmx, we disable MMX registers and make i64 to no longer be a  
legal type but leave v2i64 as a legal SSE2+ type. Since X86 doesn't  
have an 64 bit divide instruction and we don't have a vector form for  
the library function, we end up scalarizing it and calling library  
routine which introduces those expression nodes with i64 and then hit  
the assert. It is a tricky case and it is one I wonder could happen in  
other cases (probably rarely) where we have library routines that uses  
illegal types that need be promoted.

-- Mon Ping

On Jan 30, 2009, at 7:11 AM, Duncan Sands wrote:

> Hi Mon Ping, if I understand right the problem is that v2i64 is legal,
> but i64 is not, and that when we try to scalarize this
>  0x2b1d9c8: v2i64 = sdiv 0x2b1ddc8, 0x2b1dfc8
> into a pair of libcalls, we end up with expressions in terms of the
> elements (i.e. of i64 type) but those types aren't legal.  The
> point of running llc with -disable-mmx is that there is then no
> sdiv vector instruction I suppose.  Hmmm, tricky :)
>
> Ciao,
>
> Duncan.
> _______________________________________________
> llvm-commits mailing list
> llvm-commits at cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits