[openib-general] Re: [PATCH] Fix ib_pack/unpack for 64 bits

[Grant Grundler]

On Thu, Jun 23, 2005 at 07:45:55PM +0300, Michael S. Tsirkin wrote:
> It seems that compiler generates a single instruction for this shift.

If so, then gcc or the code gen backend should generate a warning
about overflowing the encoding.  This sounds like a toolchain bug.

> And CPU designers apparently have cut a corner by simply using only low 6 bits
> in the operand of a shift, so that 1ull<<64 is same as 1ull<<(64&0x37), or 1.

It might be the instruction encoding is the limiting factor and not
the CPU implementation. I don't have my dead-tree IA32 arch books
to look up the encoding.

Ie gcc might need to learn a new instruction for 64-bit platforms.
I've run into this on the parisc support in the gnu tool chain.

grant