[flang] Add -fstack-arrays flag

[tblah]

Summary

The implementation of -fstack-arrays was added in https://reviews.llvm.org/D140415

I stored the option in CodeGenOpts because this seemed like the most relevant of what is available and creating a new options structure for a single boolean flag seemed like overkill.

-fstack-arrays is implied by -Ofast to match behavior by other fortran compilers.

The MemoryAllocationOpt pass is disabled when stack arrays is enabled because its purpose is the opposite: to move array allocations to the heap (although by default it won't move any allocations).

[tblah]

Updated to construct StackArrays as a ModuleOp pass

[awarzynski]

clang/include/clang/Driver/Options.td

5056–5060

We should avoid duplicating options like this and use multiclasses instead. For example, see how debug_pass_manager is defined.

flang/docs/FlangDriver.md

573

[nit] "Clang" is a sub-project. It's not clear what "clang" would be - also a sub-project or the binary?

flang/include/flang/Tools/CLOptions.inc

159

CLANG-FORMAT-ME :) Same for other long lines here.

216

[nit] To me, it would make more sense to put stackArrays at the end. optLevelis a more powerful flag.

flang/test/Transforms/stack-arrays.f90

3

Also, I don't quite follow this comment:

We have to check llvm ir here to force flang to run the whole mlir pipeline

Why is checking LLVM IR going to force Flang to run anything?

6

flang/tools/bbc/bbc.cpp

276

Similar suggestion below.

[tblah]

Update to address review comments

flang/include/flang/Tools/CLOptions.inc

159

It seems clang-format is not run on this file. I have fixed the lines changed in my patch. Should I clang-format the whole file in a separate patch?

flang/test/Transforms/stack-arrays.f90

3

Just running flang-new -fc1 -emit-fir outputs the FIR before all of the transformation passes run (which is why I have to pipe the result through fir-opt to run the array value copy and stack arrays passes on the first line).

Outputting LLVM IR requires all of the MLIR transformation passes to be run to transform all of the higher level dialects into the LLVM dialect. Stack arrays is run as part of that same pipeline (fir::createMLIRToLLVMPassPipeline). I want to test that -fstack-arrays does cause the stack arrays pass to run as part of that pipeline, which is why I am checking the LLVM-IR.

One alternative would be to print out which passes have run and check that, but I felt this way fits more with the spirit of the other tests.

6

Lots of other tests do not follow this convention. It would change the FileCheck comments to look like ! LLVM: array_value_copy_simple.

I think it is good to require CHECK-FEATURE: so that any mention of FEATURE in comments does not confuse FileCheck. Especially for a name like LLVM which is likely to be written in capitals.

[awarzynski]

flang/include/flang/Tools/CLOptions.inc

159

+1

flang/test/Transforms/stack-arrays.f90

3

Thanks for the explanation. So you want to verify this functionality by investigating two pipelines:

• source --> FIR
• source --> LLVM IR

You could convey that by using CHECK-FIR and CHECK-LLVM-IR (or just FIR and LLVM-IR). I "complained", because this is unclear:

We have to check llvm ir here to force flang to run the whole mlir pipeline

Perhaps:

In order to verify the whole MLIR pipeline, make the driver generate LLVM IR.

6

Lots of other tests do not follow this convention.

I think that we would easily find examples for either approach.

I think it is good to require CHECK-FEATURE:

Well, LLVM is not a feature ;-) Also, most folks working with LLVM are familiar with the LIT syntax - the presence of CHECK in CHECK-FEATURE is just unnecessary noise to me. But I am happy either way. But I would appreciate replacing with LLVM with something a bit more meaningful - perhaps LLVM-IR?

[tblah]

Clarified a comment in stack-arrays.f90 and clarified the choice of FileCheck
prefix.

[awarzynski]

LGTM, thanks!

flang/test/Transforms/stack-arrays.f90

22–29

Hopefully I got this right, it's been a while :) https://llvm.org/docs/CommandGuide/FileCheck.html#the-check-dag-directive

[tblah]

Update to ensure that generated code for Options.td does not try to reference
clang::CodegenOpts::StackArrays.

[awarzynski]

clang/include/clang/Driver/Options.td

485–486

[nit] To me KeyPathAndMacro is just an implementation detail. The key part is that BoolOption does use the marshaling infra and BoolFlangOnlyOption does not. Perhaps rename as BoolOptionWithoutMarshalling? There isn't anything Flang specific here, is there?

[tblah]

• Rename option macro to make it clear that it doesn't do marshalling

[awarzynski]

Thanks for the updates, LGTM!