https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83008
--- Comment #13 from rguenther at suse dot de <rguenther at suse dot de> ---
On Fri, 8 Dec 2017, sergey.shalnov at intel dot com wrote:
> And it uses xmm+ vpbroadcastd to spill tmp[] to stack
> ...
> 1e7: 62 d2 7d 08 7c c9 vpbroadcastd %r9d,%xmm1
> 1ed: c4 c1 79 7e c9 vmovd %xmm1,%r9d
^^^
this is an odd instruction ...
> 1f2: 62 f1 fd 08 7f 8c 24 vmovdqa64 %xmm1,-0x38(%rsp)
> 1f9: c8 ff ff ff
> 1fd: 62 f2 7d 08 7c d7 vpbroadcastd %edi,%xmm2
> 203: c5 f9 7e d7 vmovd %xmm2,%edi
> 207: 62 f1 fd 08 7f 94 24 vmovdqa64 %xmm2,-0x28(%rsp)
> 20e: d8 ff ff ff
> 212: 62 f2 7d 08 7c db vpbroadcastd %ebx,%xmm3
> 218: c5 f9 7e de vmovd %xmm3,%esi
> 21c: 62 f1 fd 08 7f 9c 24 vmovdqa64 %xmm3,-0x18(%rsp)
> 223: e8 ff ff ff
> 227: 01 fe add %edi,%esi
> 229: 45 01 c8 add %r9d,%r8d
> 22c: 41 01 f0 add %esi,%r8d
> 22f: 8b 5c 24 dc mov -0x24(%rsp),%ebx
> 233: 03 5c 24 ec add -0x14(%rsp),%ebx
> 237: 8b 6c 24 bc mov -0x44(%rsp),%ebp
> 23b: 03 6c 24 cc add -0x34(%rsp),%ebp
> ...
>
> I think this is better in case of performance perspective but, as I said
> before, not using vector registers here is the best option if no loops
> vectorized.
As I said we have a basic-block vectorizer. Do you propose to remove it?
What's the rationale for "not using vector registers ... if no loops [are]
vectorized"?
With AVX256/512 an additional cost when using vector registers is
a vzeroupper required at function boundaries. Is this (one of) the
reason?
If it is the case that the instruction sequence
vpbroadcastd %ebx,%xmm3
vmovdqa64 %xmm3,-0x18(%rsp)
is slower than doing
vmovd %ebx,-0x18(%rsp)
vmovd %ebx,-0x22(%rsp)
vmovd %ebx,-0x26(%rsp)
vmovd %ebx,-0x30(%rsp)
then the costing in the backend needs to reflect that.
I see that the vectorization prevents eliding 'tmp' on GIMPLE
but that's a completely separate issue (the value-numbering we
run after vectorization is poor):
vect_cst__108 = {_48, _48, _48, _48};
vect_cst__106 = {_349, _349, _349, _349};
vect_cst__251 = {_97, _97, _97, _97};
MEM[(unsigned int *)&tmp + 16B] = vect_cst__251;
MEM[(unsigned int *)&tmp + 32B] = vect_cst__106;
MEM[(unsigned int *)&tmp + 48B] = vect_cst__108;
_292 = tmp[0][0];
_291 = tmp[1][0];
...
those loads could have been elided but DOM isn't powerful enough to
see that (and never will be).