[PATCH] D20789: Consecutive memory access in Loop Vectorizer - fixed and simplified

[Adam Nemet via llvm-commits]

anemet added inline comments.

================
Comment at: ../lib/Analysis/LoopAccessAnalysis.cpp:913-917
@@ -861,1 +912,7 @@
 
+  if (!AR)
+    // One more attempt to get stride from GEP with index calculated in runtime
+    if (auto *Gep = dyn_cast<GetElementPtrInst>(Ptr))
+      if (int Stride = getConsecutiveAccessFromGEPIndex(PSE, Gep, Lp, StridesMap))
+        return Stride;
+
----------------
delena wrote:
> anemet wrote:
> > Sorry for jumping in late but I was on vacation.  Elena, does your testcase actually exercise this part of the code or it's handled as a regular AddRec above?  I am assuming not since the Ptr is not a GEP.
> > 
> > Do we actually have coverage exercising this part?
> > 
> The test that exercises this code is in version-mem-access.ll. This is the existing test that fails without this extension. 
> My test case does not have GEP (but it has AddRec ) and this case wan't handled properly.
OK and what is the SCEV that fails to become an AddRec after replaceSymbolicStrideSCEV?

I am wondering if there is a fundamental reason we need getConsecutiveAccessFromGEPIndex.  My guess is that it's this in getConsecutiveAccessFromGEPIndex:

  // Because of the multiplication by a stride we can have a s/zext cast.
  // We are going to replace this stride by 1 so the cast is safe to ignore.

which does not seem fundamental.


Repository:
  rL LLVM

http://reviews.llvm.org/D20789