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. |
| bmodelica.variable @y : !bmodelica.variable<f64> | ||
| bmodelica.variable @z : !bmodelica.variable<f64> | ||
|
|
||
| //%t0 = bmodelica.equation_template inductions = [] { |
There was a problem hiding this comment.
Missing CHECK commands for FileCheck
|
|
||
| //%t0 = bmodelica.equation_template inductions = [] { | ||
| // %0 = bmodelica.variable_get @x : f64 | ||
| // %lhs = bmodelica.equation_side %0 : tuple<f64> |
There was a problem hiding this comment.
Don't rely on some numbering or naming. Capture the actual SSA variable names using the double square bracket operator.
| public: | ||
| using CallCSEPassBase<CallCSEPass>::CallCSEPassBase; | ||
| using CallCSEPassBase::CallCSEPassBase; |
There was a problem hiding this comment.
Keep the template argument explicit so to be resilient to any possible change in the code automatically generated with Tablegen.
| @@ -1,6 +1,8 @@ | |||
| #include "marco/Dialect/BaseModelica/Transforms/CallCSE.h" | |||
| #include "marco/Dialect/BaseModelica/IR/BaseModelica.h" | |||
|
|
|||
| #include <marco/AST/Node/Operation.h> | |||
There was a problem hiding this comment.
Not clear why is the AST library is being imported here. At this point in the pipeline the AST is already gone.
| @@ -29,16 +53,16 @@ void CallCSEPass::runOnOperation() { | |||
| } | |||
| }); | |||
|
|
|||
| if (mlir::failed(mlir::failableParallelForEach( | |||
| if (failed(failableParallelForEach( | |||
There was a problem hiding this comment.
Keep the mlir namespace explicit
|
|
||
| /// Get all callOps in the model. | ||
| static void collectCallOps(ModelOp modelOp, | ||
| llvm::SmallVector<CallOp> &callOps); |
There was a problem hiding this comment.
Use llvm::SmallVectorImpl so not to force a particular amount of elements on the stack.
|
|
||
| auto *equivalenceGroup = | ||
| find_if(callEquivalenceGroups, [&](llvm::SmallVector<CallOp> &group) { | ||
| // front() is safe as there are no empty groups |
There was a problem hiding this comment.
Still, an assert would be nice to have.
| mlir::IRMapping &mapping, | ||
| mlir::RewriterBase &rewriter) { | ||
| std::vector<mlir::Operation *> toClone; | ||
| std::vector worklist({op}); |
| rewriter.setInsertionPointToStart(modelOp.getBody()); | ||
| auto cseVariable = rewriter.create<VariableOp>( | ||
| rewriter.getUnknownLoc(), "_cse" + std::to_string(emittedCSEs), | ||
| VariableType::wrap(representative.getResult(0).getType())); |
There was a problem hiding this comment.
What happens if the function has multiple results?
| templateOp->walk([&](CallOp callOp) { | ||
| callOps.push_back(callOp); | ||
| }); | ||
| int emittedCSEs = 0; |
There was a problem hiding this comment.
It should be possible - and would be nice - to add this measurement to the pass statistics.
Adds a pass that hoists repeated function calls with equivalent arguments in and between equations into new equations, so they are only computed once.