<table border="1" cellspacing="0" cellpadding="8">
<tr>
<th>Issue</th>
<td>
<a href=https://github.com/llvm/llvm-project/issues/56540>56540</a>
</td>
</tr>
<tr>
<th>Summary</th>
<td>
[mlir] Problems in slices of Python bindings of BlockArgumentList
</td>
</tr>
<tr>
<th>Labels</th>
<td>
</td>
</tr>
<tr>
<th>Assignees</th>
<td>
</td>
</tr>
<tr>
<th>Reporter</th>
<td>
ingomueller-net
</td>
</tr>
</table>
<pre>
I am running into two problems that I think might both be due to a problem of how slices of the Python bindings of `BlockArgumentList` are handled.
The first problem occurs in this snippet:
```python
print(type(loop))
# <class 'mlir.dialects._ods_common.ForOp'>
print(len(loop.inner_iter_args))
# 1
print([str(a) for a in loop.inner_iter_args])
# ["Value(<block argument> of type 'tensor<1000x2000xf64>' at index: 1)"]
print(len(loop.inner_iter_args.types))
# 2
print(loop.inner_iter_args.types)
# [Type(index), Type(tensor<1000x2000xf64>)]
```
Notice that `loop.inner_iter_args` consists of one "inner" loop argument but not the induction variable of the loop but taking `.types` of that expression suddenly *does* include the induction variable (of type `index`).
The second problem occurs in this snippet (run after the previous one):
```python
print([str(a) for a in wrapped_func.arguments])
# ["Value(<block argument> of type 'tensor<1000x1500xf64>' at index: 0)", "Value(<block argument> of type 'tensor<1500x2000xf64>' at index: 1)", "Value(<block argument> of type 'tensor<1000x2000xf64>' at index: 2)", "Value(<block argument> of type 'memref<?xi64>' at index: 3)"]
print([str(a) for a in wrapped_func.arguments[:-num_results - 1]])
# ["Value(<block argument> of type 'tensor<1000x1500xf64>' at index: 0)", "Value(<block argument> of type 'tensor<1500x2000xf64>' at index: 1)"]
print([str(a) for a in wrapped_func.arguments[:-num_results - 1] + loop.inner_iter_args])
# ["Value(<block argument> of type 'tensor<1000x1500xf64>' at index: 0)", "Value(<block argument> of type 'tensor<1500x2000xf64>' at index: 1)", "Value(<block argument> of type 'tensor<1000x1500xf64>' at index: 0)"]
```
Notice that `loop.inner_iter_args` has type `tensor<1000x2000xf64>` but the last element of the last expression, which is supposed to consist of that value, is of type `tensor<1000x1500xf64>`. Also, if I print the IR, the last value refers to the wrong SSA name.
The behavior can be reproduced on `5da5483ff` (that commit is required; without it, the surrounding code does not run due to other problems.) by applying the following patch:
```patch
diff --git a/mlir/benchmark/python/common.py b/mlir/benchmark/python/common.py
index 3d80892b0fb8..ea9cd0349a2d 100644
--- a/mlir/benchmark/python/common.py
+++ b/mlir/benchmark/python/common.py
@@ -94,10 +94,32 @@ def emit_benchmark_wrapped_main_func(kernel_func, timer_func):
loop = scf.ForOp(zero, n_iterations, one, iter_args)
with ir.InsertionPoint(loop.body):
start = func.CallOp(timer_func, [])
+ print(type(loop))
+ # <class 'mlir.dialects._ods_common.ForOp'>
+ print(len(loop.inner_iter_args))
+ # 1
+ print([str(a) for a in loop.inner_iter_args])
+ # ["Value(<block argument> of type 'tensor<1000x2000xf64>' at index: 1)"]
+ print(len(loop.inner_iter_args.types))
+ # 2
+ print(loop.inner_iter_args.types)
+ # [Type(index), Type(tensor<1000x2000xf64>)]
+ print([str(a) for a in wrapped_func.arguments])
+ # ["Value(<block argument> of type 'tensor<1000x1500xf64>' at index: 0)", "Value(<block argument> of type 'tensor<1500x2000xf64>' at index: 1)", "Value(<block argument> of type 'tensor<1000x2000xf64>' at index: 2)", "Value(<block argument> of type 'memref<?xi64>' at index: 3)"]
+ print([str(a) for a in wrapped_func.arguments[:-num_results - 1]])
+ # ["Value(<block argument> of type 'tensor<1000x1500xf64>' at index: 0)", "Value(<block argument> of type 'tensor<1500x2000xf64>' at index: 1)"]
+ print([
+ str(a) for a in wrapped_func.arguments[:-num_results - 1] +
+ loop.inner_iter_args
+ ])
+ # ["Value(<block argument> of type 'tensor<1000x1500xf64>' at index: 0)", "Value(<block argument> of type 'tensor<1500x2000xf64>' at index: 1)", "Value(<block argument> of type 'tensor<1000x1500xf64>' at index: 0)"]
call = func.CallOp(
kernel_func,
wrapped_func.arguments[:-num_results - 1] + loop.inner_iter_args
)
+ # args = [a for a in wrapped_func.arguments[:-num_results - 1]] + \
+ # [a for a in loop.inner_iter_args]
+ # call = func.CallOp(kernel_func, args)
end = func.CallOp(timer_func, [])
time_taken = arith.SubIOp(end, start)
memref.StoreOp(time_taken, timer_buffer, [loop.induction_variable])
```
The code can then be run with this command: `python build/bin/mlir-mbr --machine $(hostname) --revision $(git rev-parse HEAD) --stop_on_error true --result-stdout mlir/benchmark/python/`. It currently fails with `loc("-":8:10): error: 'func.call' op operand type mismatch: expected operand type 'tensor<1000x2000xf64>', but provided 'tensor<1000x1500xf64>' for operand number 2` because the wrong IR is produced.
The patch also contains a work-around, which is currently commented out. It converts the `BlockArgumentList` into a Python `list`. That makes the benchmarked kernel compile (but execution fails unless you define some environment variables).
</pre>
<img width="1px" height="1px" alt="" src="http://email.email.llvm.org/o/eJzdWVtz4jYY_TXkRYPHMZfAAw-QbKaZ6bQ7m52-MrItx2pkyZXkJPTX93wyJrBAQnaz23YZBrAtfffbEanJV7MbxitmG62lvmNSe8P8o2G1NakSlWO-5J7d4Evqe1bJu9Kz1PiSpYLljWBYzrvFzBSsNI_MKZkJR1e-FOzjypdGs1TqHBzC7d44XiiT3c_tXVMJ7X-VzuMe41awkutciTzqxVe9eN5-fgaZQlrnnzllWWMdxCXBHHNa1rXwvcF8extItu86iNDerC107CUTv6oFvpQxdS-Z0rvdkwxYb3CZKe4c6yUXlZI2yiVXIvMuWprcLTNTVUZH18b-jq0XvcGHXcpK6DXhSGot7FJ6fHB7575kdL67sTdaOG_xg2MVK4yFaaHhQUqjq12RR4tekvzBVUNKQYGU7AuDtgaGjMEd0JmU8kI7Y7HqPI7jp4Q-ivGQFEkuGNwNV4knGBMCksQJcTtRxYh47CmafLH9xY3bSn1uvdQKREQv2frWCzpMN_JuImA7LH4zHvHZBjYeHTQvgjEz2iEuQ8AaTXZLwip8B5dsjMvSxjNtfIh2SNpkXiLgH7iVHMHa5UHYQ0s9v6dUA4-1zmAW1kAe8VRb4Rztd02eC61WYDzPDZlmDuqZanJxjBPMsvHyOG6NBt2T6V42OQH18lfSieihMDBewC6BJ4R7kKZxZBCy8xvy7UhwP1oOVvmyaHQWdRZ91_A-Hx0N73gd3oiqryI_Oi17vpb8K8mZfBX5SlRWFFjVG1w_yYOEB0ey_o0uXIBWXzfVEhHdKCRSHxYBzZ_Kt9_HRhBl8d3r_r9vvG9JjBOkf48uUHK3KajHMxMLQ2WnMs8xpggUVeoMXekP9zalnfR-LGVWMqq1TV0bJ3KapNYtZ9MNHlrLXNK6rcJ-3BBoKWyunAl7CoxtISyDDDef6OZGmkCaoRII1H2a-fDg0Rr0pdvbOdO8Ens9IxUlR_W3LOOaxj8r0D7QgyA7mhB4j3I-Gk4GRUH2oBZNOtC0JD1pYMVfjbQi7w0W7FGiQ8Bi0ndSucZa04QZEXvQ46jnhbZKLWg9amLwRCfqZtOIUixdMaSWWtE-olMYpcwjXdXcZ-XRFhUehnu5LArW799BSiTtNU18-EqFzsqK23v8Xvez5Ho9-tUrlp66suURYpMN8kk8mSZpXKSTKBJ8muXxYDjlSc7gyvFw2C7u9_unS9LVgkX7fqtgvWGMN-tPh3DEObltEX4OErZ-lIuCCbhwuSG37KpZxaUOJQ3OvhdWC7W-gktlhSxqr57nBNa9wjDUG1wxlxXdJD35W9gQuTqkIKfZxtF1GDYQ0Ntj9C45CieGQf1GO2Fp30ezNWmmgDkHpaCX89z6IEqozZdcqSDMtgKXoc7ulN7FNo1XMcXO6m-AGEf4noo49sQ4f5HsV-ORfXV_DDx5s3kOoZU94ZOXiZ-CZQ4Z5FuhzZt8dsqY_a5e-38PF_-Fqft9HHzSEP6Tef51Cx58Tq93mt2P0j9YOw964Wd20A-a-3csn6GxH2jzB1bSa3eaObLo3UDdPvmXfE9bgibgxL-lAgSBeqPLo5zaQNxmcqz3H6NwxOpfDIuHhzp6CZ2_dTbb2U8rl57fCx3IcItBMbpt0ptAB9SJQBgBD-9v63Z0640VG9YtwecpN22Kgs4FgyhrC63P5pbd2dxOPh_Co4SwAu4hdAUg00IsQJ8w24ZjOZoKOURGqG8O2YA7pcpp2Jd6Pfn3q9QCz1Q8K2U4t8Q8PymN84TqqLL1-3SMFw4Z24cEfXCrX3PrBPvlw_yqXea8qZdQQgCaWeYtQBjtpTDCs5wA3EtQI8DRG0BAQDvEolqxgkvlWo0C5ibggIztU9YO5qgE8_O4ndRZ4BlUTS6C9ymUKOMBHEwNeIDICBWikq5aIz3C2BiiCZBur3ilvZPjCL0DVj7IHJtfqzmUDR0DJFYKSJqEIwCR8caJLSx984mAbweT9xB1AKGMA7AT9PdAUw5p9mjsfZ8HMLxzUvBsRgoE_CI9G9-a2OgHIB8XeB_7eyX8ucO7P2TIAe2DiH0mpF4hqlsCG2-CQ5uqxLKW7REz2Uo8iawJh8-tSxutBNDMyjSEFynsnKkE0vdBwhDhLKTLBMr06CyfDfLpYMrPvPRKzJA4IZJQlD52_zyh3Dz_k3TgX6Q9Hc8aq2al97UjqIf4S64R2WWTRpAeF0o9dF99-ORPhAoupXMNCXU9Go-G8Vk5Eyk_53yYjIs4TbPhRTGeZnESj0UxSTjn2ZniqVBu1tadMzlL4iSJLxAkkzgZjqM4O08ngywv8hHPL-IEIFoAKauIGEfG3p3ZWZAhbVD2hsEL7vkhUKG800J09HkDze0MWpuqEUoJ29fCnwWxZ0HmfwDUuVWv">