Change to workshare loop wrapper op

Author: ivanradanov

flang/lib/Optimizer/OpenMP/LowerWorkshare.cpp

@@ -21,6 +21,7 @@
 #include "llvm/ADT/STLExtras.h"
 #include "llvm/ADT/iterator_range.h"
 
+#include <mlir/Dialect/OpenMP/OpenMPClauseOperands.h>
 #include <variant>
 
 namespace flangomp {
@@ -73,7 +74,7 @@ static bool isSafeToParallelize(Operation *op) {
 ///
 /// omp.workshare {
 ///   %a = fir.allocmem
-///   omp.wsloop {}
+///   omp.workshare_loop_wrapper {}
 ///   fir.call Assign %b %a
 ///   fir.freemem %a
 /// }
@@ -85,7 +86,7 @@ static bool isSafeToParallelize(Operation *op) {
 ///   fir.store %a %tmp
 /// }
 /// %a_reloaded = fir.load %tmp
-/// omp.wsloop {}
+/// omp.workshare_loop_wrapper {}
 /// omp.single {
 ///   fir.call Assign %b %a_reloaded
 ///   fir.freemem %a_reloaded
@@ -180,20 +181,20 @@ void lowerWorkshare(mlir::omp::WorkshareOp wsOp) {
   assert(wsBlock->getTerminator()->getNumOperands() == 0);
   Operation *terminator = wsBlock->getTerminator();
 
-  SmallVector<std::variant<SingleRegion, omp::WsloopOp>> regions;
+  SmallVector<std::variant<SingleRegion, omp::WorkshareLoopWrapperOp>> regions;
 
   auto it = wsBlock->begin();
   auto getSingleRegion = [&]() {
     if (&*it == terminator)
       return false;
-    if (auto pop = dyn_cast<omp::WsloopOp>(&*it)) {
+    if (auto pop = dyn_cast<omp::WorkshareLoopWrapperOp>(&*it)) {
       regions.push_back(pop);
       it++;
       return true;
     }
     SingleRegion sr;
     sr.begin = it;
-    while (&*it != terminator && !isa<omp::WsloopOp>(&*it))
+    while (&*it != terminator && !isa<omp::WorkshareLoopWrapperOp>(&*it))
       it++;
     sr.end = it;
     assert(sr.begin != sr.end);
@@ -214,9 +215,16 @@ void lowerWorkshare(mlir::omp::WorkshareOp wsOp) {
       singleBuilder.createBlock(&singleOp.getRegion());
       moveToSingle(std::get<SingleRegion>(loopOrSingle), singleBuilder);
     } else {
-      rootBuilder.clone(*std::get<omp::WsloopOp>(loopOrSingle), rootMapping);
-      if (!isLast)
-        rootBuilder.create<omp::BarrierOp>(loc);
+      omp::WsloopOperands wsloopOperands;
+      if (isLast)
+        wsloopOperands.nowait = rootBuilder.getUnitAttr();
+      auto wsloop =
+          rootBuilder.create<mlir::omp::WsloopOp>(loc, wsloopOperands);
+      auto wslw = std::get<omp::WorkshareLoopWrapperOp>(loopOrSingle);
+      auto clonedWslw = cast<omp::WorkshareLoopWrapperOp>(
+          rootBuilder.clone(*wslw, rootMapping));
+      wsloop.getRegion().takeBody(clonedWslw.getRegion());
+      clonedWslw->erase();
     }
   }
 

flang/test/HLFIR/bufferize-workshare.fir

@@ -12,7 +12,7 @@
 // CHECK:               %[[VAL_6:.*]]:2 = hlfir.declare %[[VAL_5]](%[[VAL_3]]) {uniq_name = ".tmp.array"} : (!fir.heap<!fir.array<42xi32>>, !fir.shape<1>) -> (!fir.heap<!fir.array<42xi32>>, !fir.heap<!fir.array<42xi32>>)
 // CHECK:               %[[VAL_7:.*]] = arith.constant true
 // CHECK:               %[[VAL_8:.*]] = arith.constant 1 : index
-// CHECK:               omp.wsloop {
+// CHECK:               "omp.workshare_loop_wrapper"() ({
 // CHECK:                 omp.loop_nest (%[[VAL_9:.*]]) : index = (%[[VAL_8]]) to (%[[VAL_1]]) inclusive step (%[[VAL_8]]) {
 // CHECK:                   %[[VAL_10:.*]] = hlfir.designate %[[VAL_4]]#0 (%[[VAL_9]])  : (!fir.ref<!fir.array<42xi32>>, index) -> !fir.ref<i32>
 // CHECK:                   %[[VAL_11:.*]] = fir.load %[[VAL_10]] : !fir.ref<i32>
@@ -22,7 +22,7 @@
 // CHECK:                   omp.yield
 // CHECK:                 }
 // CHECK:                 omp.terminator
-// CHECK:               }
+// CHECK:               }) : () -> ()
 // CHECK:               %[[VAL_14:.*]] = fir.undefined tuple<!fir.heap<!fir.array<42xi32>>, i1>
 // CHECK:               %[[VAL_15:.*]] = fir.insert_value %[[VAL_14]], %[[VAL_7]], [1 : index] : (tuple<!fir.heap<!fir.array<42xi32>>, i1>, i1) -> tuple<!fir.heap<!fir.array<42xi32>>, i1>
 // CHECK:               %[[VAL_16:.*]] = fir.insert_value %[[VAL_15]], %[[VAL_6]]#0, [0 : index] : (tuple<!fir.heap<!fir.array<42xi32>>, i1>, !fir.heap<!fir.array<42xi32>>) -> tuple<!fir.heap<!fir.array<42xi32>>, i1>

flang/test/Transforms/OpenMP/lower-workshare.mlir

@@ -34,7 +34,6 @@ module {
 // CHECK:               }
 // CHECK:               omp.terminator
 // CHECK:             }
-// CHECK:             omp.barrier
 // CHECK:             omp.single nowait {
 // CHECK:               hlfir.assign %[[VAL_14]] to %[[VAL_13]] : !fir.heap<!fir.array<42xi32>>, !fir.ref<!fir.array<42xi32>>
 // CHECK:               fir.freemem %[[VAL_14]] : !fir.heap<!fir.array<42xi32>>
@@ -56,7 +55,7 @@ module {
         %3:2 = hlfir.declare %2(%0) {uniq_name = ".tmp.array"} : (!fir.heap<!fir.array<42xi32>>, !fir.shape<1>) -> (!fir.heap<!fir.array<42xi32>>, !fir.heap<!fir.array<42xi32>>)
         %true = arith.constant true
         %c1 = arith.constant 1 : index
-        omp.wsloop {
+        "omp.workshare_loop_wrapper"() ({
           omp.loop_nest (%arg1) : index = (%c1) to (%c42) inclusive step (%c1) {
             %7 = hlfir.designate %1#0 (%arg1)  : (!fir.ref<!fir.array<42xi32>>, index) -> !fir.ref<i32>
             %8 = fir.load %7 : !fir.ref<i32>
@@ -66,7 +65,7 @@ module {
             omp.yield
           }
           omp.terminator
-        }
+        }) : () -> ()
         %4 = fir.undefined tuple<!fir.heap<!fir.array<42xi32>>, i1>
         %5 = fir.insert_value %4, %true, [1 : index] : (tuple<!fir.heap<!fir.array<42xi32>>, i1>, i1) -> tuple<!fir.heap<!fir.array<42xi32>>, i1>
         %6 = fir.insert_value %5, %3#0, [0 : index] : (tuple<!fir.heap<!fir.array<42xi32>>, i1>, !fir.heap<!fir.array<42xi32>>) -> tuple<!fir.heap<!fir.array<42xi32>>, i1>