[OpenMP] Remove complex reduction variable support

[TIFitis]

This patch removes the complex reduction variables codegen. There are currently no tests for this, and from playing around with some complex reduction variable test cases the code seems unreachable.

The PR #80343 proposes to migrate reductions codegen to the OMPIRBuilder and it isn't currently well equipped to handle complex variables. This patch would enable PR #80343 to go forward.

[llvmbot]

@llvm/pr-subscribers-clang-codegen

@llvm/pr-subscribers-clang

Author: Akash Banerjee (TIFitis)

Changes

This patch removes the complex reduction variables codegen. There are currently no tests for this, and from playing around with some complex reduction variable test cases the code seems unreachable.

The PR #80343 proposes to migrate reductions codegen to the OMPIRBuilder and it isn't currently well equipped to handle complex variables. This patch would enable PR #80343 to go forward.

---

Full diff: https://github.com/llvm/llvm-project/pull/82497.diff

1 Files Affected:

• (modified) clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp (+3-12)

diff --git a/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp b/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp
index 299ee1460b3db0..a954ad0b3a5e2c 100644
--- a/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp
+++ b/clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp
@@ -1653,12 +1653,7 @@ static void emitReductionListCopy(
         break;
       }
       case TEK_Complex: {
-        CodeGenFunction::ComplexPairTy Elem = CGF.EmitLoadOfComplex(
-            CGF.MakeAddrLValue(SrcElementAddr, Private->getType()),
-            Private->getExprLoc());
-        CGF.EmitStoreOfComplex(
-            Elem, CGF.MakeAddrLValue(DestElementAddr, Private->getType()),
-            /*isInit=*/false);
+        llvm_unreachable("OpenMP Complex reduction not handled.");
         break;
       }
       case TEK_Aggregate:
@@ -2232,9 +2227,7 @@ static llvm::Value *emitListToGlobalCopyFunction(
       break;
     }
     case TEK_Complex: {
-      CodeGenFunction::ComplexPairTy V = CGF.EmitLoadOfComplex(
-          CGF.MakeAddrLValue(ElemPtr, Private->getType()), Loc);
-      CGF.EmitStoreOfComplex(V, GlobLVal, /*isInit=*/false);
+      llvm_unreachable("OpenMP Complex reduction not handled.");
       break;
     }
     case TEK_Aggregate:
@@ -2439,9 +2432,7 @@ static llvm::Value *emitGlobalToListCopyFunction(
       break;
     }
     case TEK_Complex: {
-      CodeGenFunction::ComplexPairTy V = CGF.EmitLoadOfComplex(GlobLVal, Loc);
-      CGF.EmitStoreOfComplex(V, CGF.MakeAddrLValue(ElemPtr, Private->getType()),
-                             /*isInit=*/false);
+      llvm_unreachable("OpenMP Complex reduction not handled.");
       break;
     }
     case TEK_Aggregate:

[TIFitis]

Ping for review.

[jsjodin]

@alexey-bataev do you have any concerns about removing this case?

[alexey-bataev]

@alexey-bataev do you have any concerns about removing this case?

I think better to add a test for complex reductions rather than removing its support. This simple test fails without complex supports.

#include <complex.h>
int foo() {
  int i;
  int j;
  complex float sum;

  #pragma omp target teams loop reduction(+:sum) collapse(2) \
      bind(parallel) order(concurrent) lastprivate(j) map(tofrom:sum)
  for(i=0; i<10; i++)
    for(j=0; j<10; j++)
      sum += i;

  return 0;
}

[TIFitis]

Thanks @alexey-bataev for the test. It does hit the complex scenario.

I'll close this PR and add that test in another PR.

[TIFitis]

@alexey-bataev I'm facing issues adding the test to Clang tests as it uses the complex.h header files. From what I can tell header files don't seem to be supported. Is there any way to add the test you provided to the Clang tests or a different one which can still check for the complex case?

[alexey-bataev]

You can try to run preprocessor and then manually reduce the expanded code after preprocessor