[rtl, patch] combine concat+shuffle

[Marc Glisse <marc dot glisse at inria dot fr>]

Hello,

this patch combines for vectors a concat and a shuffle. An example on x86 
would be:

__m128d f(double d){
  __m128d x=_mm_setr_pd(-d,d);
  return _mm_shuffle_pd(x,x,1);
}

which was compiled as:

	vmovsd	.LC0(%rip), %xmm1
	vxorpd	%xmm0, %xmm1, %xmm1
	vunpcklpd	%xmm0, %xmm1, %xmm0
	vshufpd	$1, %xmm0, %xmm0, %xmm0

and with the patch:

	vmovsd	.LC0(%rip), %xmm1
	vxorpd	%xmm0, %xmm1, %xmm1
	vunpcklpd	%xmm1, %xmm0, %xmm0

This happens a lot in my code, for interval arithmetics, where I have a 
number d, build an interval (-d,d) from it, then subtract that interval 
from an other one, and subtraction is implemented as shufpd+addpd.

The patch is quite specialized, but I guessed I could start there, and it 
can always be generalized later.

For the testsuite, since the patch is not in a particular target, it would 
be better to have a generic test (in gcc.dg?), but I don't really know how 
to write a generic one, so would a test in gcc.target/i386 that scans 
the asm for shuf or perm be ok?

Ah, and if I use __builtin_shuffle instead of _mm_shuffle_pd, the patch 
works without -mavx, but -mavx uses vpermilpd (ie a vec_select:V2DF 
(reg:V2DF) ...) instead of a vshufpd, so I'll probably want to handle that 
too later. I thought about doing a general transformation from 
vec_select(vec_concat(x,x),*) to vec_select(x,*) (reducing the indexes in 
* so they fit), but that seemed way too dangerous.

--
Marc Glisse

Attachment: P2
Description: Text document