On 04/26/2016 08:21 PM, Richard Sandiford wrote:
"H.J. Lu" <hongjiu...@intel.com> writes:
I am working a patch to enable SSE, AVX and AVX512 for memcpy/memset
optimization. x86 backend defines MAX_BITSIZE_MODE_ANY_INT to 128
to keep the OI and XI modes from confusing the compiler into thinking
that these modes could actually be used for computation. But the OI
and XI modes can be used for data movement with vector instructions.
But doesn't this then open the possibility that a memset or memcpy
will be seen as a "normal" integer operation? Routines like
simplify_immed_subreg could in principle create a constant integer
for the stored value, which could then be treated by later passes as
a wide_int, breaking the MAX_BITSIZE_MODE_ANY_INT assumption.
Couldn't we move to allowing vector modes instead? That seems better
than having to pander to the current situation in which every vector
effectively needs an associated integer mode.
I'm actually working on a patch in this area. Haven't gotten to the
point of allowing vector modes yet, but it's something that I've had on
my mind; I think it would be a good idea.
Bernd
