[llvm-branch-commits] [mlir] [mlir][Transforms] Dialect Conversion: Add `replaceOpWithMultiple` (PR #115816)

[River Riddle via llvm-branch-commits]

River707 wrote:

> > The only very high level conern I have is the use and need of `InsertionPoint::after`. As I understand, this is a not-yet-implemented optimization for the future for the purpose of reusing materializations by calculating a better insertion point where it may dominate more operations which might need exactly that materialization after pattern application as well.
> 
> Yes, that's correct. But this optimization is already implemented. And it is actually quite deeply ingrained into the dialect conversion driver.
> 
> SSA values are not replaced directly, but in a delayed fashion. The replacement value is stored in the `ConversionValueMapping` (a kind of `IRMapping`). We store not only user-specified replacement values but also materializations in there. The data structure is kind of a linked list: for an original value, we store a list of replacement values: the specified replacement value and materializations to a different type. In the "finalize" phase, the driver goes through that list to find a replacement/materialization value with the correct type. We cannot store more than one replacement/materialization value with the same type; the driver would just pick the first one that matches. So we always reuse.
> 
> Note: Computing an insertion point for a 1:1 materialization (or a 1:N block signature conversion) is much easier because there is only one SSA value (or one block). That's why computing the insertion point was trivial until now.
> 
> > Is the current complexity of calculating the insertion point (and potentially calculatin the post-dom tree, a O(n log n) operation) worth the gain of the optimization? I am thinking that the better insertion point insertion logic should probably be post-poned until we can measure its effectivness (and avoid the risk of a premature optimization) and have something simpler and working that does not worsen the status-quo for now instead.
> 
> This PR also makes the `DominanceInfo` argument to `InsertPoint::after` optional in case of a single SSA value. (And no `DominanceInfo` is passed in that case.) For the most frequent case of 1:1 replacements, we do not compute a dominator tree at all. (And we are not doing any extra work.)
> 
> In case of N:1 materializations, the implementation uses `DominanceInfo` and we create it during `ConversionPatternRewriter::replaceOpWithMultiple`. Unfortunately, it is not safe to reuse the same `DominanceInfo` object because a pattern could have made IR changes that invalidate the dominator tree.
> 
> However, I believe `DominanceInfo` is quite "cheap" to use. The dominator tree is built lazily and it is built on a per-region basis. E.g., creating a new `DominanceInfo` and querying dominance for two ops in the same region will just build the dominator tree for that region (and only if the ops are in different blocks). In case of two ops from different regions (or different blocks in the same region), the implementation will find the common ancestors (in the same region) and then compute the dominator tree only for that region.
> 
> Long term, the `ConversionValueMapping` is going to disappear with the One-Shot Dialect Conversion. As part of that, I'm also thinking of removing the "materialization caching" mechanism and just building duplicate materializations (in the form of `unrealized_conversion_cast`, i.e., not calling the callback). These can then be CSE'd before calling the materialization callback. The CSE-ing will require `DominanceInfo`, but the same `DominanceInfo` can be reused for the entire dialect conversion because at this point of time we are done with pattern applications.

One question I have here is why not just insert right before the operation that is being replaced? What is the need for trying to insert after one of the values?

https://github.com/llvm/llvm-project/pull/115816