[Flang][OpenMP] Add host_eval clause lowering support (#180)

This patch updates Flang lowering to use the `host_eval` clause in `omp.target` operations to pass host information into the applicable clauses inside of the target region, instead of the previous approach where these clauses were attached to the `omp.target` operation itself.
ROCm · Nov 7, 2024 · 2fae4cc · 2fae4cc
1 parent 5654efd
commit 2fae4cc
Show file tree

Hide file tree

Showing 10 changed files with 332 additions and 466 deletions.
diff --git a/flang/include/flang/Lower/OpenMP/Utils.h b/flang/include/flang/Lower/OpenMP/Utils.h
@@ -174,15 +174,6 @@ void genObjectList(const ObjectList &objects,
 void lastprivateModifierNotSupported(const omp::clause::Lastprivate &lastp,
                                      mlir::Location loc);
 
-// TODO: consider moving this to the `omp.loop_nest` op. Would be something like
-// this:
-//
-// ```
-// mlir::Value LoopNestOp::calculateTripCount(mlir::OpBuilder &builder,
-// mlir::OpBuilder::InsertPoint ip)
-// ```
-mlir::Value calculateTripCount(fir::FirOpBuilder &builder, mlir::Location loc,
-                               const mlir::omp::LoopRelatedClauseOps &ops);
 } // namespace omp
 } // namespace lower
 } // namespace Fortran

diff --git a/flang/lib/Lower/OpenMP/OpenMP.cpp b/flang/lib/Lower/OpenMP/OpenMP.cpp
diff --git a/flang/lib/Lower/OpenMP/Utils.cpp b/flang/lib/Lower/OpenMP/Utils.cpp
@@ -538,84 +538,6 @@ void lastprivateModifierNotSupported(const omp::clause::Lastprivate &lastp,
   }
 }
 
-mlir::Value calculateTripCount(fir::FirOpBuilder &builder, mlir::Location loc,
-                               const mlir::omp::LoopRelatedClauseOps &ops) {
-  using namespace mlir::arith;
-  assert(ops.loopLowerBounds.size() == ops.loopUpperBounds.size() &&
-         ops.loopLowerBounds.size() == ops.loopSteps.size() &&
-         !ops.loopLowerBounds.empty() && "Invalid bounds or step");
-
-  // Get the bit width of an integer-like type.
-  auto widthOf = [](mlir::Type ty) -> unsigned {
-    if (mlir::isa<mlir::IndexType>(ty)) {
-      return mlir::IndexType::kInternalStorageBitWidth;
-    }
-    if (auto tyInt = mlir::dyn_cast<mlir::IntegerType>(ty)) {
-      return tyInt.getWidth();
-    }
-    llvm_unreachable("Unexpected type");
-  };
-
-  // For a type that is either IntegerType or IndexType, return the
-  // equivalent IntegerType. In the former case this is a no-op.
-  auto asIntTy = [&](mlir::Type ty) -> mlir::IntegerType {
-    if (ty.isIndex()) {
-      return mlir::IntegerType::get(ty.getContext(), widthOf(ty));
-    }
-    assert(ty.isIntOrIndex() && "Unexpected type");
-    return mlir::cast<mlir::IntegerType>(ty);
-  };
-
-  // For two given values, establish a common signless IntegerType
-  // that can represent any value of type of x and of type of y,
-  // and return the pair of x, y converted to the new type.
-  auto unifyToSignless =
-      [&](fir::FirOpBuilder &b, mlir::Value x,
-          mlir::Value y) -> std::pair<mlir::Value, mlir::Value> {
-    auto tyX = asIntTy(x.getType()), tyY = asIntTy(y.getType());
-    unsigned width = std::max(widthOf(tyX), widthOf(tyY));
-    auto wideTy = mlir::IntegerType::get(b.getContext(), width,
-                                         mlir::IntegerType::Signless);
-    return std::make_pair(b.createConvert(loc, wideTy, x),
-                          b.createConvert(loc, wideTy, y));
-  };
-
-  // Start with signless i32 by default.
-  auto tripCount = builder.createIntegerConstant(loc, builder.getI32Type(), 1);
-
-  for (auto [origLb, origUb, origStep] :
-       llvm::zip(ops.loopLowerBounds, ops.loopUpperBounds, ops.loopSteps)) {
-    auto tmpS0 = builder.createIntegerConstant(loc, origStep.getType(), 0);
-    auto [step, step0] = unifyToSignless(builder, origStep, tmpS0);
-    auto reverseCond =
-        builder.create<CmpIOp>(loc, CmpIPredicate::slt, step, step0);
-    auto negStep = builder.create<SubIOp>(loc, step0, step);
-    mlir::Value absStep =
-        builder.create<SelectOp>(loc, reverseCond, negStep, step);
-
-    auto [lb, ub] = unifyToSignless(builder, origLb, origUb);
-    auto start = builder.create<SelectOp>(loc, reverseCond, ub, lb);
-    auto end = builder.create<SelectOp>(loc, reverseCond, lb, ub);
-
-    mlir::Value range = builder.create<SubIOp>(loc, end, start);
-    auto rangeCond =
-        builder.create<CmpIOp>(loc, CmpIPredicate::slt, end, start);
-    std::tie(range, absStep) = unifyToSignless(builder, range, absStep);
-    // numSteps = (range /u absStep) + 1
-    auto numSteps = builder.create<AddIOp>(
-        loc, builder.create<DivUIOp>(loc, range, absStep),
-        builder.createIntegerConstant(loc, range.getType(), 1));
-
-    auto trip0 = builder.createIntegerConstant(loc, numSteps.getType(), 0);
-    auto loopTripCount =
-        builder.create<SelectOp>(loc, rangeCond, trip0, numSteps);
-    auto [totalTC, thisTC] = unifyToSignless(builder, tripCount, loopTripCount);
-    tripCount = builder.create<MulIOp>(loc, totalTC, thisTC);
-  }
-
-  return tripCount;
-}
-
 } // namespace omp
 } // namespace lower
 } // namespace Fortran