TVMScript also provide `doc.NodeVisitor` and `doc.NodeTransformer` for AST
level mutations. In our private fork we collect all lhs variables with multiple
`Assign` or `AugAssign` bindings. They are implicitly transformed to scalar
buffer load & store during parsing to structure same as @spectrometerHBH's
answer.
```python
from tvm.script import tir as T
@T.prim_func
def function(X: T.Buffer([16], "int32")):
v = 0
X[0] = v
for i in range(10):
v = i
X[i + 1] = v
print(function)
# default tir parser output
# different assignment to `v` treat as new let bindings
@T.prim_func
def function(X: T.Buffer((16,), "int32")):
v: T.int32 = 0
X[0] = v
for i in range(10):
v_1: T.int32 = i
X[i + 1] = v_1
# scalar transformed output
@T.prim_func
def function(X: T.Buffer((16,), "int32")):
# with T.block("root"):
v = T.alloc_buffer((1,), "int32")
v[0] = 0
X[0] = v[0]
for i in range(10):
v[0] = i
X[i + 1] = v[0]
```
In llvm-based backends, we could also introduce new storage scope or someother
annotations to ensure the scalar buffer allocation map to alloca and finally
transform to SSA value via `mem2reg`.
One important note is though this behavior is more "pythonic", it break the
concise scoping behaviour for default IR parsing. And it may be hard to keep
round trip property. So from my understanding it could be useful when we use
TVMScript as a programming language, but may not be proper or default
behaviour for TVMScript's origin serving purpose.
---
[Visit
Topic](https://discuss.tvm.apache.org/t/how-to-design-a-local-variable/18490/3)
to respond.
You are receiving this because you enabled mailing list mode.
To unsubscribe from these emails, [click
here](https://discuss.tvm.apache.org/email/unsubscribe/45c1cfcbc90a272d36cfca1a92438166918a16175a51f99e467659abbda7d9aa).
