[flang-commits] [flang] [flang] Inline scalar-to-scalar TRANSFER for same-size trivial types (PR #191589)

[Zhen Wang via flang-commits]

================
@@ -54,3 +49,61 @@ integer function trans_test3(p)
     t = transfer(p, t)
     trans_test3 = t%x
   end function
+
+  ! Scalar same-size transfer (f64 -> i64) is inlined as fir.load + arith.bitcast.
+  subroutine trans_test_r8_to_i8(store, word)
+    ! CHECK-LABEL: func @_QPtrans_test_r8_to_i8(
+    ! CHECK-SAME:    %[[RES:.*]]: !fir.ref<i64>{{.*}}, %[[SRC:.*]]: !fir.ref<f64>{{.*}}) {
+    ! CHECK-DAG:     %[[store:.*]]:2 = hlfir.declare %[[RES]] {{.*}}{uniq_name = "_QFtrans_test_r8_to_i8Estore"}
+    ! CHECK-DAG:     %[[word:.*]]:2 = hlfir.declare %[[SRC]] {{.*}}{uniq_name = "_QFtrans_test_r8_to_i8Eword"}
+    ! CHECK:         %[[LOADED:.*]] = fir.load %[[word]]#0 : !fir.ref<f64>
+    ! CHECK:         %[[VAL:.*]] = arith.bitcast %[[LOADED]] : f64 to i64
+    ! CHECK:         hlfir.assign %[[VAL]] to %[[store]]#0 : i64, !fir.ref<i64>
+    ! CHECK-NOT:     fir.call @_FortranATransfer
+    ! CHECK:         return
+    ! CHECK:       }
+    integer(8) :: store
+    real(8) :: word
+    store = transfer(word, store)
+  end subroutine
+
+  ! BIND(C) struct (c_ptr) to integer(8): same byte size, inlined via
+  ! address-level reinterpret. Covers the c_devptr pattern on CUDA device code.
+  subroutine trans_test_cptr_to_i8(store, src)
+    ! CHECK-LABEL: func @_QPtrans_test_cptr_to_i8(
+    ! CHECK:         %[[srcDecl:.*]]:2 = hlfir.declare {{.*}}{uniq_name = "_QFtrans_test_cptr_to_i8Esrc"}
+    ! CHECK:         %[[storeDecl:.*]]:2 = hlfir.declare {{.*}}{uniq_name = "_QFtrans_test_cptr_to_i8Estore"}
+    ! CHECK:         %[[CAST:.*]] = fir.convert %[[srcDecl]]#0 : (!fir.ref<!fir.type<_QM__fortran_builtinsT__builtin_c_ptr{__address:i64}>>) -> !fir.ref<i64>
+    ! CHECK:         %[[VAL:.*]] = fir.load %[[CAST]] : !fir.ref<i64>
+    ! CHECK:         hlfir.assign %[[VAL]] to %[[storeDecl]]#0 : i64, !fir.ref<i64>
+    ! CHECK-NOT:     fir.call @_FortranATransfer
+    ! CHECK:         return
+    ! CHECK:       }
+    use iso_c_binding
+    integer(8) :: store
+    type(c_ptr) :: src
+    store = transfer(src, store)
+  end subroutine
+
+  ! Different-size scalar transfer (i32 -> i64) falls back to runtime.
+  subroutine trans_test_diff_size(store, src)
+    ! CHECK-LABEL: func @_QPtrans_test_diff_size(
+    ! CHECK:         fir.call @_FortranATransfer(
+    ! CHECK:         return
+    ! CHECK:       }
+    integer(8) :: store
+    integer(4) :: src
+    store = transfer(src, store)
+  end subroutine
+
+  ! Array mold without SIZE: result is rank-1 array, must use runtime.
+  subroutine trans_test_array_mold(src, result)
+    ! CHECK-LABEL: func @_QPtrans_test_array_mold(
+    ! CHECK:         fir.call @_FortranATransfer(
+    ! CHECK:         return
+    ! CHECK:       }
+    real :: src
+    integer, allocatable :: result(:)
+    integer :: mold(4)
+    result = transfer(src, mold)
+  end subroutine
----------------
wangzpgi wrote:

Thanks for the suggestion! Tests added.

https://github.com/llvm/llvm-project/pull/191589