| Issue |
179247
|
| Summary |
[mlir] `MLIRContext::getLoadedDialect` asserts when using TD schedule
|
| Labels |
mlir
|
| Assignees |
|
| Reporter |
banach-space
|
### OVERVIEW
I hit the following assert: https://github.com/llvm/llvm-project/blob/02804501af9c764126964a965c02dbee84690227/mlir/lib/IR/MLIRContext.cpp?plain=1#L420-L422
while running [ConvertToLLVM](https://github.com/llvm/llvm-project/blob/02804501af9c764126964a965c02dbee84690227/mlir/lib/Conversion/ConvertToLLVM/ConvertToLLVMPass.cpp) and [ConvertToEmitC](https://github.com/llvm/llvm-project/blob/02804501af9c764126964a965c02dbee84690227/mlir/lib/Conversion/ConvertToEmitC/ConvertToEmitCPass.cpp), **via** Transform Dialect's [transform.apply_registered_pass](https://mlir.llvm.org/docs/Dialects/Transform/#transformapply_registered_pass-transformapplyregisteredpassop).
**NOTE:** No assert is hit when _invoking the passes directly_ via `mlir-opt`, i.e.
```bash
mlir-opt -convert-to-{emitc|llvm} file.mlir
```
This implies that there is an issue with how dialects are registered when using Transform Dialect.
**Expected behavior:** `transform.apply_registered_pass` should behave equivalently to invoking the same pass via `mlir-opt`.
### HOW TO REPRODUCE
Take one of the examples below and run:
```bash
mlir-opt -transform-interpeter file.mlir
```
While I am able to reproduce this reliably on my M3 Pro using Apple clang version 16.0.0, it does not reproduce in Compiler Explorer:
* https://godbolt.org/z/1ox78rzsr
It does, however, reproduce in LLVM's pre-commit CI, as documented here:
* https://github.com/llvm/llvm-project/pull/177339#issuecomment-3786626607
Note that in pre-commit CI the assertions is hit for **X86**, but not for **AArch64**.
LLVM version locally: 59f815b047c8
### REPRO EXAMPLES
**EXAMPLE 1 (ConvertToLLVM)**
```mlir
module {
func.func @foo(%arg0: vector<f32>) -> vector<i32> {
%0 = vector.bitcast %arg0 : vector<f32> to vector<i32>
return %0 : vector<i32>
}
}
module attributes {transform.with_named_sequence} {
transform.named_sequence @__transform_main(%module:
!transform.any_op{transform.readonly}) {
%func_h = transform.structured.match ops{["func.func"]} in %module
: (!transform.any_op) -> !transform.any_op
%module_h = transform.get_parent_op %func_h {isolated_from_above} : (!transform.any_op) -> !transform.any_op
// NOTE THE PASS BELOW!
transform.apply_registered_pass "convert-to-llvm" to %module_h
: (!transform.any_op) -> !transform.op<"builtin.module">
transform.yield
}
}
```
**EXAMPLE 2 (ConvertToEmitC)**
```mlir
module {
func.func private @foo(%arg0: memref<2xf32>, %arg1: memref<2xf32>, %arg2: memref<2xf32>) {
%c0 = arith.constant 0 : index
%c2 = arith.constant 2 : index
%c1 = arith.constant 1 : index
scf.for %arg3 = %c0 to %c2 step %c1 {
%0 = memref.load %arg0[%arg3] : memref<2xf32>
%1 = memref.load %arg1[%arg3] : memref<2xf32>
%2 = arith.addf %0, %1 : f32
memref.store %2, %arg2[%arg3] : memref<2xf32>
}
return
}
}
module attributes {transform.with_named_sequence} {
transform.named_sequence @__transform_main(%module:
!transform.any_op{transform.readonly}) {
%func_h = transform.structured.match ops{["func.func"]} in %module
: (!transform.any_op) -> !transform.any_op
%module_h = transform.get_parent_op %func_h {isolated_from_above} : (!transform.any_op) -> !transform.any_op
// NOTE THE PASS BELOW!
transform.apply_registered_pass "convert-to-emitc" to %module_h
: (!transform.any_op) -> !transform.op<"builtin.module">
transform.yield
}
}
```
### NEXT STEPS
I am not that familiar with the TD infra and would appreciate some pointers. @ftynse , any suggestions?
![]()
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
