[flang-commits] [flang] [flang] Apply nocapture attribute to dummy arguments (PR #116182)

[via flang-commits]

s-watanabe314 wrote:

> The change in the pass looks good to me, but you should probably enable this pass by default (or at least when O > O0, opinions welcomed here) because it is currently only run under a specific set of conditions (see
> 
> https://github.com/llvm/llvm-project/blob/40afff7bd95090a75bc68a0d26b8017cc0ae65c1/flang/lib/Optimizer/Passes/Pipelines.cpp#L278
> 
> ).
> Currently, if one run the patch with `flang -fc1 -emit-llvm -O2` on the code below the pass is not run and the attribute not added.
> 
> ```
> subroutine foo(x)
>   call bar(x)
> end subroutine
> ```



Thank you for your review. I'll modify the code so that the path is enabled by default. 
After making the changes below, a dozen or so regression tests failed, so I'll confirm them before committing.

```
  // Add function attributes
  mlir::LLVM::framePointerKind::FramePointerKind framePointerKind;

  if (config.FramePointerKind == llvm::FramePointerKind::NonLeaf)
    framePointerKind = mlir::LLVM::framePointerKind::FramePointerKind::NonLeaf;
  else if (config.FramePointerKind == llvm::FramePointerKind::All)
    framePointerKind = mlir::LLVM::framePointerKind::FramePointerKind::All;
  else
    framePointerKind = mlir::LLVM::framePointerKind::FramePointerKind::None;

  pm.addPass(fir::createFunctionAttr(
      {framePointerKind, config.NoInfsFPMath, config.NoNaNsFPMath,
       config.ApproxFuncFPMath, config.NoSignedZerosFPMath,
       config.UnsafeFPMath}));
```

https://github.com/llvm/llvm-project/pull/116182