[PATCH] D107966: [SLP]Do not emit extract elements for insertelements users, replace with shuffles directly.

[Alexey Bataev via Phabricator via llvm-commits]

ABataev added inline comments.


================
Comment at: llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp:8212
+                }
+                Base = cast<InsertElementInst>(Base)->getOperand(0);
+                // After the vectorization the def-use chain has changed, need
----------------
RKSimon wrote:
> I find this control flow very confusing - is the 'cast<InsertElementInst>(Base)' guaranteed to match IEBase? we break after the if() above so we can't get here from there.
We just iterate through insertelements, which are not part of the vectorized buildvector.
For example:
```
%0 = insertelement %..., %a, 0
%1 = insertelement %0, %b, 1
%2 = insertelement %1, %c, 2
```

If %c is vectorized, we start looking through a buildvectror, trying to find the vectorized base.
Start from %2. getTreeEntry(%2) returns nullptr. Go to %1.
getTreeEntry(%1) returns nullptr too (it is not a part of vectorized buildvector). Go to %0.
getTreeEntry(%0) is vectorized and returns E. Iterate through all vectorized insertelements, build a mask.
Put %2 to the list of insertelements, which must be transformed to shuffles.

Later, we do the analysis of all inserts between %1-%2 (including boundaries), If they must be replaced with shuffles - replace them with shuffles, other insertelements remain as is, just change their base properly to the shuffles.



Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D107966/new/

https://reviews.llvm.org/D107966