[llvm] [openmp] [OpenMP][DeviceRTL] Fix the chunked static distribution dropping iterations (PR #216117)

[Johannes Doerfert via llvm-commits]

================
@@ -0,0 +1,75 @@
+// Chunked static distribution: thread T owns iterations [T*chunk, T*chunk+chunk)
+// within each block chunk. The index used to start each thread has to account
+// for the chunk, and the block chunk has to cover one chunk per thread.
+//
+// The entry is called directly because no C/C++ construct reaches it: clang
+// emits __kmpc_for_static_init_4, only flang emits this one.
+//
+// RUN: %libomptarget-compile-run-and-check-generic
+// REQUIRES: gpu
+
+#include <omp.h>
+#include <stdio.h>
+
+#define N 32
+#define NT 8
+#define CHUNK 4
+
+struct Args {
+  int *Tid;
+};
+
+#pragma omp begin declare target
+extern void __kmpc_distribute_for_static_loop_4u(
+    void *Loc, void (*Fn)(unsigned, void *), void *Arg, unsigned NumIters,
+    unsigned NumThreads, unsigned BlockChunk, unsigned ThreadChunk,
+    unsigned char OneIterationPerThread);
+
+__attribute__((noinline)) static void body(unsigned I, void *A) {
+  ((struct Args *)A)->Tid[I] = omp_get_thread_num();
+}
+#pragma omp end declare target
+
+// For the host fallback copy only; the device uses the runtime's definition.
+#ifndef __AMDGCN__
----------------
jdoerfert wrote:

What about NVIDIA? Intel soon?
You can disable codegen for the fallback via a clang cmd arg, or always provide the host version here restricted to host only or made weak.

https://github.com/llvm/llvm-project/pull/216117