Select Case constructs with character selector expressions

[vdonaldson]

Add the capability to lower select case constructs of any type to
either FIR SelectCaseOp's or a sequence of comparisons and branches.
Actually generate SelectCaseOp's for integer type constructs, but use
comparisons and branches for logical types (the code is better) and
for character types (it isn't possible yet to handle character
SelectCaseOp's downstream).

[jeanPerier]

The code looks good to me.
I am more in favor of having this logic as a FIR to FIR pass operating from a select-case op, which would both move some complexity outside of lowering and allow some more flexibility for future optimizations. Have you tried a bit this option ?

[schweitzpgi]

I'm approving this so you can commit it.

The plan is for the Burnside bridge to lower to fir.select_case unconditionally. A rewrite pass will lower these high-level control ops to some primitive implementation(s) (perhaps a choice from a menu of possibilities) before everything is converted to LLVM-IR dialect. But I can do that after this code is merged.

Thanks for pulling this together, Val. It's great to have CHARACTER support.

[schweitzpgi]

Why not use mlir's splitBlock function?

[vdonaldson]

mlir functions splitBlock and (one overload of) createBlock are complementary. createBlock inserts a new block before a given block, and splitBlock inserts the new block after the given block. Since "the given block" may actually be the first of multiple blocks, createBlock makes much more sense here. Of course, where blocks are inserted doesn't affect actual code relationships between blocks. But it is much easier to read and debug IR when blocks are placed in more natural locations.

[schweitzpgi]

I think we have moved to using the FIR type for these tests. @jeanPerier ?

[jeanPerier]

In general yes, when there is a mlir::Type available working with it is the preferred way. But Here I think it would make the code more complex given this is needed before producing any mlir values, and working with mlir::Type would not be 100% straightforward given logicals are mlir::IntegerType. So what Val has here looks clean to me.

[vdonaldson]

Right, that was the motivation for doing it this way.

[vdonaldson]

I am more in favor of having this logic as a FIR to FIR pass operating from a select-case op, which would both move some complexity outside of lowering and allow some more flexibility for future optimizations. Have you tried a bit this option ?

Although the current lack of support for generating runtime support calls in bbc is a significant motivation for doing this conversion in bbc now, as the code illustrates, the conversion is fairly simple, so it may be reasonable to have a 3-tier lowering/translation process even in the long run. Lowering early enables earlier optimization of lowered code. I can envision doing some of the lowering in bbc (as here) to take advantage of fir::IfOp, for example. Any fir::SelectCaseOp that reaches the beginning of bbc can then be considered for lowering at that point. And if some optimizations in bbc are best done directly on a fir::SelectCaseOp, then those cases can be retained, and any surviving fir::SelectCaseOp can finally be converted at the end of bbc. That allows a lot of flexibility for generating the best code.

[schweitzpgi]

I'm seeing some warnings here:

/local/home/eschweitz/f18-llvm-project/flang/lib/Lower/Bridge.cpp:1274:16: warning: variable 'pred' is used uninitialized whenever 'if' condition is false [-Wsometimes-uninitialized]
      else if (attr.isa<fir::UpperBoundAttr>())
               ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
/local/home/eschweitz/f18-llvm-project/flang/lib/Lower/Bridge.cpp:1276:41: note: uninitialized use occurs here
      auto cond = genCond(*caseValue++, pred);
                                        ^~~~
/local/home/eschweitz/f18-llvm-project/flang/lib/Lower/Bridge.cpp:1274:12: note: remove the 'if' if its condition is always true
      else if (attr.isa<fir::UpperBoundAttr>())
           ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
/local/home/eschweitz/f18-llvm-project/flang/lib/Lower/Bridge.cpp:1269:7: note: variable 'pred' is declared here
      mlir::CmpIPredicate pred;
      ^