Function call CSE pass #45
Conversation
|
It would be nice if you could take a look @mscuttari and see what you think. As you asked it does not currently handle equations with induction variables. Additionally I was unsure whether calls in the initial equations had to be handled differently, so for not they are ignored. There is also just one test. I read the page on filecheck tests for mlir, and they mention that one should try to make the tests as isolated as possible. I tried to do that for the one I have made, so please let me know if something can be improved about that. I'll add some more before finishing the PR. |
arrangabriel
force-pushed
the
arrangabriel/function-cse
branch
5 times, most recently
from
1990b4d
to
a7398f8
Compare
arrangabriel
force-pushed
the
arrangabriel/function-cse
branch
from
aa3c4b2
to
f9d8329
Compare
arrangabriel
force-pushed
the
arrangabriel/function-cse
branch
from
5a8761a
to
1237c42
Compare
| for (EquationInstanceOp equationOp : dynamicEquationOps) { | ||
| EquationTemplateOp templateOp = equationOp.getTemplate(); | ||
| if (!templateOp.getInductionVariables().empty() || | ||
| llvm::is_contained(templateOps, templateOp)) { |
There was a problem hiding this comment.
This is O(n^2). It would be better to check for already visited templates by using a set, while at the same time keeping the vector for deterministic output.
There was a problem hiding this comment.
Yeah, I'll fix it
arrangabriel
force-pushed
the
arrangabriel/function-cse
branch
from
965fb28
to
da0eba0
Compare
Adds a pass that hoists repeated function calls with equivalent arguments in and between equations into new equations, so they are only computed once.