[mlir][tosa] Work around GCC bug in tosa-to-tensor #91521
Conversation
|
@llvm/pr-subscribers-mlir Author: Spenser Bauman (sabauma) ChangesGCC 12 and 13 generate incorrect code for a pattern in the tosa-to-tensor pass responsible for lowering tosa.reshape. This results in the tosa.reshape lowering producing IR which fails to verify. I've narrowed down the set of cmake flags needed to reproduce the issue to this: This is the failing test case: This should lower to a tensor.expand_shape operation like so: Under GCC 12/13 with the above cmake configuration, the tensor.expand_shape looks like this The key difference is the computed output type of The problematic code is calculating the intermediate shape of the generated tensor.expand_shape operation in the The failure point seems to be the update of the resultIsStatic variable in the lambda body. The assignment of false is not propagated to the use in the if-statement, resulting in the branch being taken when it should not. I've found several modification to the code that gets around the bug. The version I settled on is one which makes the logic a little more obvious. Full diff: https://github.com/llvm/llvm-project/pull/91521.diff 1 Files Affected:
diff --git a/mlir/lib/Conversion/TosaToTensor/TosaToTensor.cpp b/mlir/lib/Conversion/TosaToTensor/TosaToTensor.cpp
index cd6da35582469..488c24e90f5a0 100644
--- a/mlir/lib/Conversion/TosaToTensor/TosaToTensor.cpp
+++ b/mlir/lib/Conversion/TosaToTensor/TosaToTensor.cpp
@@ -84,6 +84,7 @@ TensorType inferReshapeExpandedType(TensorType inputType,
return totalSize / totalSizeNoPlaceholder;
});
+
// A syntactic restriction in 'tensor.expand_shape' forbids a dynamically
// shaped input from being reshaped into a statically shaped result. We may
// simply turn the first result dimension dynamic to address this.
|
|
✅ With the latest revision this PR passed the C/C++ code formatter. |
|
Is there a GCC bug report about this? If not please file one and double check to make sure this is a GCC bug or just a working around something else. |
I'm working on a GCC bug report at the moment. Currently, I'm blocked on account creation on their bug tracker. I'll update this report when I have an issue filed. Based on GCC's IR dumps, I do think this is a bug in GCC, though I'm not that knowledgeable in their IR. There is a rather obvious removal of the relevant branch after the full redundancy elimination pass. |
GCC 12 and 13 generate incorrect code for a pattern in the
tosa-to-tensor pass responsible for lowering tosa.reshape.
This results in the tosa.reshape lowering producing IR which fails to
verify. I've narrowed down the set of cmake flags needed to reproduce
the issue to this:
cmake -G Ninja ../llvm \
-DLLVM_ENABLE_PROJECTS="mlir" \
-DLLVM_TARGETS_TO_BUILD=host \
-DLLVM_ENABLE_PROJECTS=mlir \
-DCMAKE_BUILD_TYPE="Release" \
-DCMAKE_CXX_FLAGS_RELEASE="-O2" \
-DCMAKE_CXX_FLAGS="-O2" \
-DCMAKE_CXX_COMPILER=g++ \
-DCMAKE_C_COMPILER=gcc
This is the failing test case:
func.func @fails_in_gcc_12(%arg0: tensor<?xf32>) -> tensor<1x1x1x?xf32> {
%0 = tosa.reshape %arg0 {new_shape = array<i64: 1, 1, 1, -1>} : (tensor<?xf32>) -> tensor<1x1x1x?xf32>
return %0 : tensor<1x1x1x?xf32>
}
This should correctly lower to a single tensor.expand_shape operation
like so:
func.func @foo(%arg0: tensor<?xf32>) -> tensor<1x1x1x?xf32> {
%c0 = arith.constant 0 : index
%dim = tensor.dim %arg0, %c0 : tensor<?xf32>
%c1 = arith.constant 1 : index
%expanded = tensor.expand_shape %arg0 [[0, 1, 2, 3]] output_shape [1, 1, 1, %dim] : tensor<?xf32> into tensor<1x1x1x?xf32>
return %expanded : tensor<1x1x1x?xf32>
}
Under GCC 12/13 with the above cmake configuration, the tensor.expand_shape
looks like this
%2 = "tensor.expand_shape"(%arg0) <{reassociation = [[0, 1, 2, 3]], static_output_shape = array<i64>}> : (tensor<?xf32>) -> tensor<?x1x1x?xf32>
This expand_shape fails to verify with this error message:
error: 'tensor.expand_shape' op expected number of static shape dims to be equal to the output rank (4) but found 0 inputs instead
The problematic code is calculating the intermediate shape of the
generated tensor.expand_shape operation in the
expand_shape/collapse_shape sequence that implements tosa.reshape.
// Compute result shape
bool resultIsStatic = true;
auto resultShape = llvm::map_to_vector(newShape, [&](int64_t size) {
// Omitted
// If we do not know the total size of the tensor, keep this dimension
// dynamic in the result shape.
if (!inputIsStatic) {
resultIsStatic = false;
return ShapedType::kDynamic;
}
});
if (resultIsStatic) {
// do something
return;
}
// do something else
return;
The failure point seems to be the update of the resultIsStatic variable
in the lambda body. The assignment of false is not propagated to the use
in the if-statement, resulting in the branch being taken when it should
not.
I've found several modification to the code that gets around the bug.
The version I settled on is one which makes the logic a little more
obvious.
|
Just to close the loop on this, I've filed the GCC bug report here: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115033 |
|
Interesting GCC bug report, and great double contribution, Spenser. Hopefully GCC folks can take it from there. Makes me wonder about the potentially critical ramifications of this bug in lambda variable capture. |
Thanks again for filing the GCC bug report. It took a while but I was able to reduce it enough to get a decent runtime testcase out of the issue. It is lambda related only in the sense copying the closure around and it seems like GCC loses track of the aliasing set of the closure though it could be any struct and it does not have to be a closure as shown by my last testcase. |
GCC 12 and 13 generate incorrect code for a pattern in the tosa-to-tensor pass responsible for lowering tosa.reshape. This results in the tosa.reshape lowering producing IR which fails to verify. I've narrowed down the set of cmake flags needed to reproduce the issue to this:
This is the failing test case:
This should lower to a tensor.expand_shape operation like so:
Under GCC 12/13 with the above cmake configuration, the tensor.expand_shape looks like this
The key difference is the computed output type of
tensor<?x1x1x?xf32>rather than the expectedtensor<1x1x1x?xf32>. This expand_shape fails to verify with this error message:The problematic code is calculating the intermediate shape of the generated tensor.expand_shape operation in the
expand_shape/collapse_shape sequence that implements tosa.reshape.
The failure point seems to be the update of the resultIsStatic variable in the lambda body. The assignment of false is not propagated to the use in the if-statement, resulting in the branch being taken when it should not.
I've found several modification to the code that gets around the bug. The version I settled on is one which makes the logic a little more obvious.