[PATCH] D111460: [X86][LoopVectorize] "Fix" `X86TTIImpl::getAddressComputationCost()`

Mon Oct 18 02:46:45 PDT 2021

lebedev.ri added a comment.

In D111460#3069578 <https://reviews.llvm.org/D111460#3069578>, @pengfei wrote:

> @lebedev.ri , I saw you mentioned this patch on D111942 <https://reviews.llvm.org/D111942>. Could you explain what's the relationship between this one and the interleaving cost patches?

It's mostly tangential, the longer this patch drags out, the more tests may need to be updated, which isn't fun.

Naive interleaving sequence is: wide load + extracts + inserts,
while scalarized gather sequence is: narrow loads + inserts,

Right now "gather" sequence is modelled as having artificially-high cost
(what this patch changes), effectively prohibiting using it for vectorization. 
But afterwards, naturally, the cost of scalarized gather sequence lowers,
becoming more attractive for vectorization, and sometimes even more than 
for the naive interleaving sequence, as you can see in `interleaved-load-i16-stride-5.ll`.

But in general, we can assume that sequence of shuffles may be better than 
a sequence of extracts+inserts, so we may be better off with interleaving
rather than gather. But we can't really improve the cost modelling for the
naive interleaving sequence (because we can't really properly cost model shuffles),
so what we have to do is hardcode the costs for the interestting interleaving tuples
in the cost tables, like i have done in those 90+ patches.

================
Comment at: llvm/lib/Transforms/Vectorize/LoopVectorize.cpp:7061
   Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
+  // FIXME: PtrTy should not be a vector, it's a hack.

----------------
pengfei wrote:
> It would be more productive if it includes the reason why we can’t change it right away.
> Lack of such explanation led to the original version of the patch that repeats D93129.
I'm open to suggestions as to how this should be amended.

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D111460/new/

https://reviews.llvm.org/D111460