(0.40) Avoid generating store of uninitialized auto when reducing TRT2 #17605
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
In the case of a single non-constant delimiter, it's possible for the
pattern to match when the original load and comparison look as follows:
n56n istore elem
n55n b2i
n54n bloadi <array-shadow>
...
n60n istore tmpDelim
n59n b2i
n58n i2b
n57n iload delim
n65n ificmpeq --> block_EXIT
n55n ==>b2i
n59n ==>b2i
With such a match, the booltable node of the pattern graph corresponds
to the ificmpeq node of the target graph. The transformer function for
TRT2 (CISCTransform2FindBytes) finds this target node and from there it
uses the second child as the delimiter, but instead of the b2i that one
might expect, that second child is a variable (tmpDelim) that has been
matched up with the store node (n60n in this example). This situation
results in the transformation generating a load of tmpDelim while at the
same time removing the definition (n60n) that provides the value that is
expected at that load.
This problem should be a pretty rare occurrence. If the tmpDelim store
is dead, it should usually have been eliminated. If OTOH it isn't dead,
then the pattern doesn't match. As such, it would probably be reasonable
to detect this case and simply refuse to transform. However, it's not
necessarily straightforward to detect the problem. I believe that
tableCISCNode->getHeadOfTrNodeInfo()->_node being a store indicates that
the problem is occurring, but it's not obvious that we couldn't see the
same fundamental problem with a load node, or some other node that has a
load as a descendant. With this uncertainty, reliably detecting the
problem case requires walking a subtree and looking for auto loads that
aren't loop-invariant.
Since detecting the problem is already that complex, and since even in
the presence of a store the delimiter might be loop-invariant anyway,
this commit goes slightly further and chases down single definitions to
construct an expression that does not load from autos that are defined
in the loop (if possible). In the example above, that means that we
still remove the tmpDelim store, but now the arraytranslateAndTest node
uses b2i (i2b (iload delim)) instead of iload tmpDelim.
pshipton
approved these changes
Jun 15, 2023
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This is #17546 for 0.40, with the identical commit d33313e.