[PATCH] D110114: [OMPIRBuilder] Generate aggregate argument for parallel region outlined functions

[Kiran Chandramohan via Phabricator via llvm-commits]

kiranchandramohan added subscribers: Meinersbur, kiranchandramohan.
kiranchandramohan added a comment.
Herald added a project: All.

If I understand this correctly, filling the aggregate struct is only happening in the parallel case but not for the serialized parallel version. See example below, the call to `@sb_..omp_par` from kmpc_fork_call in block `omp_parallel` has the aggregate filled in while the call to `@sb_..omp_par` in the serialized parallel region in block 6 does not have this. I assume that the extractor creates these and it will only do it at the place that it is called. Would copying over the instructions that fill the aggregate to the serialized parallel region be a reasonable solution?

CC: @Meinersbur

  define void @sb_(ptr %0, ptr %1) !dbg !3 {
    %structArg = alloca { ptr }, align 8
    %tid.addr = alloca i32, align 4, !dbg !7
    %zero.addr = alloca i32, align 4, !dbg !7
    store i32 0, ptr %tid.addr, align 4, !dbg !7
    store i32 0, ptr %zero.addr, align 4, !dbg !7
    %3 = load i32, ptr %0, align 4, !dbg !7
    %4 = icmp ne i32 %3, 0, !dbg !7                 
    br label %entry, !dbg !7
  
  entry:                                            ; preds = %2
    %omp_global_thread_num = call i32 @__kmpc_global_thread_num(ptr @1), !dbg !7
    br i1 %4, label %5, label %6
  
  5:                                                ; preds = %entry
    br label %omp_parallel
  
  omp_parallel:                                     ; preds = %5
    %gep_ = getelementptr { ptr }, ptr %structArg, i32 0, i32 0
    store ptr %1, ptr %gep_, align 8
    call void (ptr, i32, ptr, ...) @__kmpc_fork_call(ptr @1, i32 1, ptr @sb_..omp_par, ptr %structArg), !dbg !9
    br label %omp.par.outlined.exit
  
  omp.par.outlined.exit:                            ; preds = %omp_parallel
    br label %omp.par.exit.split
  
  omp.par.exit.split:                               ; preds = %omp.par.outlined.exit
    br label %7
  
  6:                                                ; preds = %entry
    call void @__kmpc_serialized_parallel(ptr @1, i32 %omp_global_thread_num)
    call void @sb_..omp_par(ptr %tid.addr, ptr %zero.addr, ptr %structArg), !dbg !9
    call void @__kmpc_end_serialized_parallel(ptr @1, i32 %omp_global_thread_num)
    br label %7
  
  7:                                                ; preds = %6, %omp.par.exit.split
    ret void, !dbg !10
  }


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D110114/new/

https://reviews.llvm.org/D110114