[Mlir-commits] [mlir] [mlir][tosa] Fix ability to expand ranks with dynamic shape support (PR #128037)

[llvmlistbot at llvm.org]

================
@@ -88,9 +88,9 @@ computeReshapeOutput(ArrayRef<int64_t> higherRankShape,
     higherRankDim = higherRankShape[i];
     lowerRankDim = lowerRankShape[j];
 
-    if (lowerRankDim == 1 && higherRankDim > 1)
+    if (lowerRankDim == 1 && higherRankDim != 1)
----------------
Jerry-Ge wrote:

what do you specifically mean right aligning? I think we still need some logic to implement different special cases instead of adding 1s to the left of lower rank tensor shape? 
For example: 

low Rank: `[b, c]`
high Rank: `[a, 1, c]`
the output shape would be `[a, b, c]` given the existing logic.

we can't simply adding 1s to the left of `[b, c]` (right aligning?) it would be `[1, b, c]` which is not what we wanted. 



https://github.com/llvm/llvm-project/pull/128037