[Flang] Add flag dependent code to execute the loop-body atleast once

[kiranchandramohan]

Summary

Given the flag --always-execute-loop-body the compiler emits code
to execute the body of the loop atleast once.

Note: This is part of upstreaming from the fir-dev branch of
https://github.com/flang-compiler/f18-llvm-project.

Co-authored-by: Jean Perier <[e-mail]>
Co-authored-by: Eric Schweitz <[e-mail]>
Co-authored-by: V Donaldson <[e-mail]>
Co-authored-by: Valentin Clement <[e-mail]>
Co-authored-by: Sameeran Joshi <[e-mail]>

[awarzynski]

flang/test/Lower/always-execute-loop-body.f90

1

Would you mind testing with flang-new as well?

[schweitz]

LGTM. (The bots are complaining.)

[kiranchandramohan]

Run clang-format.

[kiranchandramohan]

flang/test/Lower/always-execute-loop-body.f90

1

This is one of the cl::opt options that is present in files in Lower directory. Do we support such flags currently in the Driver?
What is the right way to share the flag with the Driver?

../flang/lib/Lower/Bridge.cpp:static llvm::cl::opt<bool> dumpBeforeFir(
../flang/lib/Lower/Bridge.cpp:static llvm::cl::opt<bool> forceLoopToExecuteOnce(
../flang/lib/Lower/PFTBuilder.cpp:static llvm::cl::opt<bool> clDisableStructuredFir(
../flang/lib/Lower/PFTBuilder.cpp:static llvm::cl::opt<bool> nonRecursiveProcedures(
../flang/lib/Lower/Allocatable.cpp:static llvm::cl::opt<bool> useAllocateRuntime(
../flang/lib/Lower/Allocatable.cpp:static llvm::cl::opt<bool> useDescForMutableBox(
../flang/lib/Lower/Bridge.cpp.orig:static llvm::cl::opt<bool> dumpBeforeFir(
../flang/lib/Lower/IntrinsicCall.cpp:static llvm::cl::opt<bool> outlineAllIntrinsics(
../flang/lib/Lower/IntrinsicCall.cpp:llvm::cl::opt<MathRuntimeVersion> mathRuntimeVersion(
../flang/lib/Lower/ConvertExpr.cpp:static llvm::cl::opt<bool> generateArrayCoordinate(
../flang/lib/Lower/ConvertExpr.cpp:static llvm::cl::opt<unsigned> clInitialBufferSize(

[awarzynski]

flang/test/Lower/always-execute-loop-body.f90

1

flang-new -fc1 supports -mmlir:

bin/flang-new -fc1 -mmlir --help | grep -E "always-execute-loop-body"
  --always-execute-loop-body                         - force the body of a loop to execute at least once

So, flang-new -fc1 -mmlir --always-execute-loop-body --emit-fir %s -o - should work ™ .

[kiranchandramohan]

Test with the flang driver as well using suggesting by @awarzynski.

[kiranchandramohan]

flang/test/Lower/always-execute-loop-body.f90

1

Thanks @awarzynski. That worked.

I was kind of thinking that -mmlir is only for the MLIR passes/transformations. Technically this is still in the frontend (lowering from parse-tree to MLIR).

[awarzynski]

LGTM, thanks!

flang/test/Lower/always-execute-loop-body.f90

1

I was kind of thinking that -mmlir is only for the MLIR passes/transformations. Technically this is still in the frontend (lowering from parse-tree to MLIR).

That would be my intuition/expectations as well, but we haven't really defined where these options belong.

One of the main problems with llvm::cl is that it relies on there being global variables corresponding to various options. And since "global", these options are visible no matter what, i.e. it is tricky to separate LLVM options from MLIR and Flang from MLIR. In MLIR, they use registerMLIRContextCLOptions. We could look into something similar in Flang (in fact, we probably should). But for the time being, Flang options are bundled with MLIR options :)

As for flang-new, everything that's hidden behind -mllvm and -mmlir are internal compiler features/flags that are not intended for end users. So at least we are not exposing our users to this confusion.

flang/lib/Lower/Bridge.cpp

 //===-- Bridge.cpp -- bridge to lower to MLIR -----------------------------===//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
 //
for (IncrementLoopInfo &info : incrementLoopNestInfo) {
       } else {
         auto diff1 =
             builder->create<mlir::arith::SubIOp>(loc, upperValue, lowerValue);
         auto diff2 =
             builder->create<mlir::arith::AddIOp>(loc, diff1, info.stepValue);
         tripCount =
             builder->create<mlir::arith::DivSIOp>(loc, diff2, info.stepValue);
       }
+      if (forceLoopToExecuteOnce) { // minimum tripCount is 1
+        mlir::Value one =
+            builder->createIntegerConstant(loc, tripCount.getType(), 1);
+        auto cond = builder->create<mlir::arith::CmpIOp>(
+            loc, mlir::arith::CmpIPredicate::slt, tripCount, one);
+        tripCount = builder->create<mlir::arith::SelectOp>(loc, cond, one, tripCount);
+      }
       info.tripVariable = builder->createTemporary(loc, tripCount.getType());
       builder->create<fir::StoreOp>(loc, tripCount, info.tripVariable);
       builder->create<fir::StoreOp>(loc, lowerValue, info.loopVariable);
 
       // Unstructured loop header - generate loop condition and mask.
       startBlock(info.headerBlock);
       tripCount = builder->create<fir::LoadOp>(loc, info.tripVariable);
       mlir::Value zero =

flang/test/Lower/always-execute-loop-body.f90

+! RUN: bbc --always-execute-loop-body --emit-fir %s -o - | FileCheck %s
+
+! Given the flag `--always-execute-loop-body` the compiler emits an extra
+! code to change to tripcount, test tries to verify the extra emitted FIR.
+
+! CHECK-LABEL: func @_QPsome
+subroutine some()
+  integer :: i
+
+  ! CHECK: [[tripcount:%[0-9]+]] = arith.divsi
+  ! CHECK: [[one:%c1_i32[_0-9]*]] = arith.constant 1 : i32
+  ! CHECK: [[cmp:%[0-9]+]] = arith.cmpi slt, [[tripcount]], [[one]] : i32
+  ! CHECK: [[newtripcount:%[0-9]+]] = arith.select [[cmp]], [[one]], [[tripcount]] : i32
+  ! CHECK: fir.store [[newtripcount]] to %{{[0-9]+}} : !fir.ref<i32>
+  do i=4,1,1
+    stop 2
+  end do
+  return
+end