some wierd thing about wrapping tensor.

[PressFforlove]

Hi there~ I found a bug(i think) in the performance of wrapping “arith.constant dense<0.0> : tensor<512x512xf32>” in a function like
func.func @create_tensor0() -> tensor<512x512xf32> {
%cst = arith.constant dense<0.0> : tensor<512x512xf32>
return %cst : tensor<512x512xf32>
}`
, and to call it as %output in linalg.matmul's output arg.
Warp or not give me the different result, which confuse me a lot.
See detail in discourse page please.
I tried in main branch，still act like the same.
In main branch, the compile cmd shoud delete the "allow-return-allocs" part.