DependenceAnalysis is not access size aware.

[etherzhhb]

Bugzilla Link	15811
Version	trunk
OS	All
Attachments	The patch to add the fail case.
CC	@etherzhhb

Extended Description

Consider the following function:

define i64 @​alias(i32* nocapture %A) nounwind uwtable {
entry:
%arrayidx = getelementptr inbounds i32* %A, i64 1
store i32 2, i32* %arrayidx, align 4
%0 = bitcast i32* %A to i64*
%1 = load i64* %0, align 8
ret i64 %1
}

Ignoring size of the memory location accessed in
store i32 2, i32* %arrayidx, align 4
and
%1 = load i64* %0, align 8,

The addresses pointed by %arrayidx and %0 doesn't alias each other, and hence DependenceAnalysis claims there is not dependencies from the store to the load, and gives the following output:
da analyze - none!
da analyze - none!
da analyze - none!

However, the above result is incorrect (at least for x86), there should be a dependency from from the store to the load, because the memory location written by the store overlap with the memory location read by the load.

[sebpop]

The current implementation of the dependence analysis test does not handle access functions reading/writing a different number of bytes. There are ways to extend DA to check for all the covered bytes by each R/W, however that would require more code for little benefit. This is really a corner case that does not occur often in codes.

The fix is to disable such R/W memory accesses as "aliasing". I will post a patch like this:

--- a/llvm/lib/Analysis/DependenceAnalysis.cpp
+++ b/llvm/lib/Analysis/DependenceAnalysis.cpp
@@ -722,9 +722,14 @@ static AliasResult underlyingObjectsAlias(AAResults *AA, LoopInfo *LI,
   const Value *AObj = getUnderlyingObject(LocA.Ptr);
   const Value *BObj = getUnderlyingObject(LocB.Ptr);

-  // If the underlying objects are the same, they must alias.
-  if (AObj == BObj)
+  if (AObj == BObj) {
+    // The dependence test gets confused if the size of the memory accesses differ.
+    if (LocA.Size != LocB.Size)
+      return AliasResult::MayAlias;
+
+    // If the underlying objects are the same, they must alias.
     return AliasResult::MustAlias;
+  }

[sebpop]

Fixed with #123436