[flang-commits] [flang] [flang][cuda] Lower simple host to device data transfer (PR #85960)

[Valentin Clement バレンタイン クレメン via flang-commits]

================
@@ -3772,9 +3797,13 @@ class FirConverter : public Fortran::lower::AbstractConverter {
       Fortran::lower::StatementContext localStmtCtx;
       hlfir::Entity rhs = evaluateRhs(localStmtCtx);
       hlfir::Entity lhs = evaluateLhs(localStmtCtx);
-      builder.create<hlfir::AssignOp>(loc, rhs, lhs,
-                                      isWholeAllocatableAssignment,
-                                      keepLhsLengthInAllocatableAssignment);
+      if (lhsIsDevice || rhsIsDevice) {
+        genCUDADataTransfer(builder, loc, lhsIsDevice, lhs, rhsIsDevice, rhs);
----------------
clementval wrote:

The spec is not clear on what happens with the assignment and if it follow the standard semantic while doing the transfer.

I'll make a survey of what is currently implemented in compilers implementing cuf. 

In the mean time I can error out with a TODO for all the non trivial case.  

https://github.com/llvm/llvm-project/pull/85960