[llvm-branch-commits] [llvm] [VPlan] Implement VPlan-based unit-strideness speculation (PR #182595)

[Andrei Elovikov via llvm-branch-commits]

================
@@ -1,57 +1,53 @@
 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --check-globals none --version 6
 
 ; RUN: opt < %s -p loop-vectorize -force-vector-width=4 -S \
-; RUN:     -enable-mem-access-versioning=false  2>&1 | FileCheck %s --check-prefix COMPARE-NO-MV
+; RUN:     -enable-mem-access-versioning=false -enable-vplan-based-stride-mv 2>&1 | FileCheck %s --check-prefix COMPARE-NO-MV
 
 ; RUN: opt < %s -p loop-vectorize -force-vector-width=4 -S \
-; RUN:     -enable-mem-access-versioning=true  2>&1 | FileCheck %s --check-prefix COMPARE-LAA-MV
+; RUN:     -enable-mem-access-versioning=false -enable-vplan-based-stride-mv 2>&1 | FileCheck %s --check-prefix COMPARE-LAA-MV
----------------
eas wrote:

I was sure I replied to your original question but somehow I didn't. To answer that first 

> How come these RUN lines are the same? 

RUN lines are identical and one of them will be dropped in a subsequent PR after this lands. The reason I have two is that they generate diffs against two different baselines to give more context to the reviewers.

> Am I right in that the VPlan flag is not meant as an experimental flag, but rather to prevent multiversioning anyway when -enable-mem-access-versioning=false?

In my mind it is meant as an experimental flag until LAA is improved not to bailout on RT-strided dependence checks.

> the vplan based unit-strideness speculation doesn't kick in unless -enable-mem-access-versioning=false is also false.

Not sure how important that question is given my answer about "experimental" nature above, but I think the state is as follows:
- In current version using `RecipeBuilder.isConsecutiveWithoutVPlanBasedStrideSpeculation` would ignore unit-strided accesses speculated in LAI
- I'm not sure if `widenConsecutiveMemOps` from #211315 would widen those or not, but we shouldn't be running in a mode when both LAA and VPlan stride speculation are enabled simultaneously anyway.

https://github.com/llvm/llvm-project/pull/182595