Segfault in some matmul cases about iree-llvm-sandbox HOT 8 CLOSED

Hi @hanhanW ,
Yes, after rebasing I am having the same issues. I am trying to investigate a bit more, but if you have any hint, please let me know.

Thanks,
Giuseppe

from iree-llvm-sandbox.

Actually, what I found is that some memref.alloca() instructions creep in during LowerVector transformation. If you set ulimit -s unlimited it should be fine, but I am wondering if this was intentional.

from iree-llvm-sandbox.

And the winner is: https://reviews.llvm.org/D119743 :)

By adding AutomaticAllocationScope to scf::ForOp makes this always returning the scf:::ForOp instead of the FuncOp.

This means that there are a lot of alloca that now happens inside the loops, that cannot be hosted out (side effects and all), and that make the stack explode.

from iree-llvm-sandbox.

I suppose we can extend the function that finds the appropriate alloca insertion point to something like "find the farthest allocation scope that is not isolated from above".

We also discussed with the author of D119743 the idea of having an "alloca hoisting" pass that would hoist allocas from nested scopes. They originally wanted it to be a canonicalization pattern, but this may have unintended consequences as the allocation lifetime is being extended and fake dependencies being introduced across different loop iterations. If you are willing to contribute such a transformation upstream, it would be very helpful.

from iree-llvm-sandbox.

This trivial fix lets the search step over loops. @hanhanW, please check if this helps with your issue.

--- a/mlir/lib/Dialect/Vector/Transforms/VectorTransferSplitRewritePatterns.cpp
+++ b/mlir/lib/Dialect/Vector/Transforms/VectorTransferSplitRewritePatterns.cpp
@@ -439,8 +439,14 @@ static void createFullPartialVectorTrans
 
 // TODO: Parallelism and threadlocal considerations with a ParallelScope trait.
 static Operation *getAutomaticAllocationScope(Operation *op) {
-  Operation *scope =
-      op->getParentWithTrait<OpTrait::AutomaticAllocationScope>();
+  Operation *scope = nullptr;
+  for (Operation *parent = op->getParentOp(); parent != nullptr;
+       parent = parent->getParentOp()) {
+    if (parent->hasTrait<OpTrait::AutomaticAllocationScope>())
+      scope = parent;
+    if (!isa<scf::ForOp>(parent))
+      break;
+  }
   assert(scope && "Expected op to be inside automatic allocation scope");
   return scope;
 }

from iree-llvm-sandbox.

Yes, it helps the issue. I'm able to run these benchmarks with the change.

from iree-llvm-sandbox.

Fix applied in llvm/llvm-project@4c807f2

from iree-llvm-sandbox.

Thanks for the fix!

from iree-llvm-sandbox.