[llvm] [LoopInterchange] Add tests for the vectorization profitability (NFC) (PR #133665)

[Ryotaro Kasuga via llvm-commits]

================
@@ -0,0 +1,108 @@
+; RUN: opt < %s -passes=loop-interchange -cache-line-size=64 \
+; RUN:     -pass-remarks-output=%t -disable-output -loop-interchange-profitabilities=vectorize
+; RUN: FileCheck -input-file %t %s
+
+ at A = dso_local global [256 x [256 x float]] zeroinitializer
+ at B = dso_local global [256 x [256 x float]] zeroinitializer
+ at C = dso_local global [256 x [256 x float]] zeroinitializer
+
+; Check that the below loops are exchanged for vectorization.
+;
+; for (int i = 0; i < 256; i++) {
+;   for (int j = 1; j < 256; j++) {
+;     A[i][j] = A[i][j-1] + B[i][j];
+;     C[i][j] += 1;
----------------
kasuga-fj wrote:

> is the write to array C necessary?

Yes. Due to the existence of the write to `C`, the profitability check makes a wrong decision as "we can vectorize the  j-loop". Please see the description of #133667 for more details if you are interested in.

> and is vectorised with a gather/scatter?

Maybe yes, if the i-loop becomes the innermost one by interchange and it is vectorized. Note that this depends on LoopVectorizer whether the loop is actually vectorized or not.

The current vectorization profitability check in LoopInterchange only checks if a loop can be vectorized or we cannot due to some dependency. To make better decisions, I think we should estimate the vectorization cost.

https://github.com/llvm/llvm-project/pull/133665