[FLANG] Solving issue with adjustr intrinsic in where construct

[EbinJose2002]

Fixes #122823
An unnecessary convert instruction was generated in the case of reference to boxchar conversion. So excluded them in the condition. The output is generating correctly.

[llvmbot]

@llvm/pr-subscribers-flang-fir-hlfir

Author: Ebin-McW (EbinJose2002)

Changes

Solves #122823
An unnecessary convert instruction was generated in the case of reference to boxchar conversion. So excluded them in the condition. The output is generating correctly.

---

Full diff: https://github.com/llvm/llvm-project/pull/146851.diff

2 Files Affected:

• (modified) flang/lib/Optimizer/HLFIR/Transforms/BufferizeHLFIR.cpp (+2-1)
• (added) flang/test/Lower/adjustr.f90 (+14)

diff --git a/flang/lib/Optimizer/HLFIR/Transforms/BufferizeHLFIR.cpp b/flang/lib/Optimizer/HLFIR/Transforms/BufferizeHLFIR.cpp
index 58f2b57712974..240b79138b0e1 100644
--- a/flang/lib/Optimizer/HLFIR/Transforms/BufferizeHLFIR.cpp
+++ b/flang/lib/Optimizer/HLFIR/Transforms/BufferizeHLFIR.cpp
@@ -446,7 +446,8 @@ struct AssociateOpConversion
             ((mlir::isa<fir::BoxCharType>(sourceVar.getType()) &&
               !mlir::isa<fir::BoxCharType>(assocType)))) {
           sourceVar = builder.create<fir::BoxAddrOp>(loc, assocType, sourceVar);
-        } else {
+        } else if (!mlir::isa<fir::ReferenceType>(sourceVar.getType()) ||
+                   !mlir::isa<fir::BoxCharType>(assocType)) {
           sourceVar = builder.createConvert(loc, assocType, sourceVar);
         }
         return sourceVar;
diff --git a/flang/test/Lower/adjustr.f90 b/flang/test/Lower/adjustr.f90
new file mode 100644
index 0000000000000..4813834614bd6
--- /dev/null
+++ b/flang/test/Lower/adjustr.f90
@@ -0,0 +1,14 @@
+! RUN: bbc -emit-fir -o - %s | FileCheck %s
+
+! Checking no reference to boxchar conversion is in the emitted fir.
+! CHECK-NOT: (!fir.ref<!fir.char<1,4>>) -> !fir.boxchar<1>
+
+program main
+  logical         ,dimension(1):: mask = .true.
+  character(len=2),dimension(1):: d1 = "a "
+  character(len=4),dimension(1):: d4
+  where (mask)
+    d4 = adjustr(d1 // d1)
+  end where
+  write(6,*) "d4     =", d4
+end

[vzakhari]

Thank you for the changes, but I think we cannot just avoid generating proper "conversion" here. Maybe it is not the case currently, but imagine that both results of hlfir.associate actually have uses. In this case, the users that expect !fir.boxchar<1> operand will receive !fir.ref<.... I think we need to generate fir.emboxchar here to produce the new result-0 with the expected data type.

[vzakhari]

Please generate fir.emboxchar instead.

[github-actions]

✅ With the latest revision this PR passed the C/C++ code formatter.

[vzakhari]

Thanks for the update!

This will only work for the types with constant length. I do not know if we should be here for non-constant lengths, but I would just take the type parameters from the associate operation - this should work for all the cases.

[EbinJose2002]

Sorry for the delayed reply.
I have made the change to access the length from associate instruction itself.

[vzakhari]

Please update the checks to show what kind of emboxchar we generate after this change.

[EbinJose2002]

I have added check for the emboxchar that is generated.

[vzakhari]

Please add more context into the checks, e.g. the operands of fir.emboxchar, where they are coming from, and how the results of fir.emboxchar are used.

[vzakhari]

nit: please assert that getTypeparams() returns a vector of length one.

[vzakhari]

Please add more context into the checks, e.g. the operands of fir.emboxchar, where they are coming from, and how the results of fir.emboxchar are used.