================
@@ -2479,6 +2479,39 @@ OpFoldResult cir::UnaryOp::fold(FoldAdaptor adaptor) {
if (isBoolNot(previous))
return previous.getInput();
+ // Fold constant unary operations.
+ if (auto srcConst = getInput().getDefiningOp<cir::ConstantOp>()) {
----------------
Lancern wrote:
> That said, WHICH direction we go (switch on kind, then type, vs type then
> kind) doesn't seem particularly motivated in 1 way or the other?
I didn't quite know how MLIR would actually fold a sequence of operations. My
guess was, given a sequence of CIR operations such as:
```
%0 = cir.const #cir.int<1>
%1 = cir.unary(neg, %0)
%2 = cir.unary(neg, %1)
```
If we fold by testing whether the input comes from a constant operation, we
might fail to fold `%2` because, well, its input does not come directly from a
constant operation.
To verify this guess, I just go through several places in MLIR that folds
operations [^1][^2]. It turns out that after successfully folding an operation,
MLIR would always first materialize the folded result into a constant operation
before folding the next. So after folding `%1`, the current MLIR implementation
would always transform the operation sequence to something like:
```
%0 = cir.const #cir.int<1>
%1 = cir.const #cir.int<-1>
%2 = cir.unary(neg, %1)
```
before it starts to attempt folding `%2`. So Andy's implementation should work.
But I'm not sure whether we should rely on this behavior. Taking the folded
input from the adaptor could avoid these problems.
[^1]: In function
[`OpBuilder::tryFold`](https://github.com/llvm/llvm-project/blob/012097d4fe7f300a59f9d2b89e528e1bd449d4c1/mlir/lib/IR/Builders.cpp#L472-L474)
[^2]: In function
[`GreedyPatternRewriteDriver::processWorklist`](https://github.com/llvm/llvm-project/blob/012097d4fe7f300a59f9d2b89e528e1bd449d4c1/mlir/lib/Transforms/Utils/GreedyPatternRewriteDriver.cpp#L492-L496)
https://github.com/llvm/llvm-project/pull/174882
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
