[Mlir-commits] [clang] [llvm] [mlir] [OpenMP] Reserve the main thread's warp for generic mode kernels (PR #218790)

[Larry Meadows]

================
@@ -8582,6 +8596,15 @@ OpenMPIRBuilder::InsertPointTy OpenMPIRBuilder::createTargetInit(
     }
   }
 
+  // Generic mode runs the main thread on a warp of its own, past thread_limit.
+  if (MaxThreadsVal > 0 && Attrs.ExecFlags == omp::OMP_TGT_EXEC_MODE_GENERIC &&
+      hasGridValue(T)) {
+    // An out-of-range bound is dropped rather than clamped, so clamp it here.
+    const omp::GV &GridValue = getGridValue(T, Kernel);
+    MaxThreadsVal = std::min(MaxThreadsVal + int32_t(GridValue.GV_Warp_Size),
+                             int32_t(GridValue.GV_Max_WG_Size));
----------------
lfmeadow wrote:

Fixed: the addition is done in `int64_t` and clamped before narrowing back. Nothing about the computed values changes, so the clang, MLIR and OMPIRBuilder expectations are untouched.

https://github.com/llvm/llvm-project/pull/218790