Fix slice to empty not being converted to FullOp

[jacobhinkle]

This fixes an issue noticed by @naoyam, where a simple fusion containing a dynamic slice does not get properly converted to a NoOp in the case when the output size is zero. This was happening because the resized iterdomain was being concretized as Iteration first, after which the extent was replaced by zero. However, the extent that was replaced in that case was the old one, not the extent of the replaced IterDomain. The fix is to first concretize empty extents as zero before doing anything else, then taking care to respect mutated extents when doing other concretizations. In particular, we need to doctor the extents that get resized if their inputs' extents have already been concretized.

[jacobhinkle]

!test

[github-actions]

Review updated until commit ed89fbe

Description

• Fix slice to empty not being converted to FullOp

• Use mutated input in resize concretization

• Enhance concretization of empty extents

• Add test case for slice to zero

---

Changes walkthrough 📝

	Relevant files
Enhancement	dynamic_transform.cpp Enhance concretization of empty extents and resize              csrc/dynamic_transform.cpp Add concretizeEmptyExtents() call at the beginning of concretize() Use maybeMutated() for inputs in concretizeResize() Update concretizeExpand() to use mutated extent +30/-11
Bug fix	nodes.cpp Update resize method for symbolic inputs                                  csrc/ir/nodes.cpp Modify IterDomain::resize() to check if in is not symbolic +2/-1
Tests	test_resize.cpp Add test for slice to zero                                                              tests/cpp/test_resize.cpp Add test case SliceToZero for slice to zero conversion +32/-0

---

PR Reviewer Guide 🔍

Here are some key observations to aid the review process:

🧪 PR contains tests

⚡ Recommended focus areas for review Order of Operations The order of operations in the concretize method has been changed. Ensure that this change does not introduce any unintended side effects or dependencies between the concretization steps. // Registers replacement of all empty extents with zeroVal() concretizeEmptyExtents(); Extent Mutation Handling The handling of mutated extents in concretizeResize seems complex. Verify that the logic correctly handles all cases of extent mutation, especially when extents are concretized to zero. auto new_id = IterDomain::resize( def_in, maybeMutated(def->leftExpand()), maybeMutated(def->rightExpand()), id->isRFactorProduct(), iter_type); if (maybeMutated(def_in->extent()) != def_in->extent()) { // Note: if the extent of id is mutated, for example by the // concretizeEmptyExtents pass, we should also use the mutated extent to // compute the output extent. Val* new_extent = SimplifyingIrBuilder::addExpr( maybeMutated(def->leftExpand()), SimplifyingIrBuilder::addExpr( maybeMutated(def_in->extent()), maybeMutated(def->rightExpand()))); IterDomain* replacement = IterDomainBuilder(new_id).extent(new_extent).build(); ir_utils::transferDefinitionToNewOutputs( new_id->definition(), {replacement}); new_id = replacement; } Symbolic IterDomain Check The check if (!in->isSymbolic() && left_expansion->isConstInt() && right_expansion->isConstInt()) might exclude some valid cases. Ensure that the logic is correct and does not unintentionally skip valid resizes. if (!in->isSymbolic() && left_expansion->isConstInt() && right_expansion->isConstInt()) { auto left = left_expansion->evaluate();

[jacobhinkle]

Note that we could probably detect patterns like this in the slice op and just translate those to FullOp right away without relying on concretization. I think the attached fix should handle cases where the slice arguments are dynamic.

[jacobhinkle]

!test