[OpenMP][OMPT] fix reduction test for 32-bit x86

[protze.joachim]

Summary

fix for: https://bugs.llvm.org/show_bug.cgi?id=44733

function __kmp_determine_reduction_method has strange architecture specific optimizations to determine the reduction algorithm. For 32-bit we need at least 3 variables to avoid atomic reduction.

This patch adds reduction variables to the testcase.

see: llvm-project/openmp/runtime/src/kmp_runtime.cpp:8143

[mgorny]

While it fixes i386 tests, it seems to break amd64 now. Do you need me to paste the output or can you reproduce?

[Hahnfeld]

@protze.joachim for 64 bit, we still need at least 5 threads IIRC

openmp/runtime/test/ompt/synchronization/reduction/tree_reduce.c

1–2

Do we need %sort-threads? It would be good to have either have both sorted, or none.

[protze.joachim]

Thanks @Hahnfeld for pointing into the right direction.

[Hahnfeld]

LGTM from the code, I didn't test that the issue is indeed fixed on i386 though

[mgorny]

I can confirm the tests pass both on i386 and amd64 for me, with this patch.

openmp/runtime/test/ompt/synchronization/reduction/tree_reduce.c

 // RUN: %libomp-compile-and-run | %sort-threads | FileCheck %s
+// RUN: %libomp-compile -DNOWAIT && %libomp-run | FileCheck %s
 // REQUIRES: ompt
 // UNSUPPORTED: gcc
 #include "callback.h"
 #include <omp.h>
 
 #ifdef NOWAIT
 #define FOR_CLAUSE nowait
 #else
 #define FOR_CLAUSE
 #endif
 
 int main() {
-  int sum = 0;
+  int sum = 0, a = 0, b = 0;
   int i;
-#pragma omp parallel num_threads(5)
-#pragma omp for reduction(+ : sum) FOR_CLAUSE
+#pragma omp parallel num_threads(4)
+// for 32-bit architecture we need at least 3 variables to trigger tree
+#pragma omp for reduction(+ : sum, a, b) FOR_CLAUSE
   for (i = 0; i < 10000; i++) {
-    sum += i;
+    a = b = sum += i;
   }
 
+
+  printf("%i\n", sum);
   // CHECK: 0: NULL_POINTER=[[NULL:.*$]]
 
   // CHECK: {{^}}[[MASTER_ID:[0-9]+]]: ompt_event_parallel_begin:
   // CHECK-SAME: parallel_id=[[PARALLEL_ID:[0-9]+]]
   // CHECK: {{^}}[[MASTER_ID]]: ompt_event_implicit_task_begin:
   // CHECK-SAME: parallel_id=[[PARALLEL_ID]], task_id=[[TASK_ID:[0-9]+]]
 
   // order and distribution to threads not determined