[llvm-branch-commits] [flang] [flang] optimize array function calls using hlfir.eval_in_mem (PR #118070)

[via llvm-branch-commits]

================
@@ -134,7 +134,7 @@ end function f
 !CHECK-NEXT: run 1 save    : where/mask
 !CHECK-NEXT: run 2 evaluate: where/region_assign1
 !CHECK-LABEL: ------------ scheduling where in _QPonly_once ------------
-!CHECK-NEXT: unknown effect: %{{[0-9]+}} = llvm.intr.stacksave : !llvm.ptr
+!CHECK-NEXT: unknown effect: %11 = fir.call @_QPcall_me_only_once() fastmath<contract> : () -> !fir.array<10x!fir.logical<4>>
----------------
jeanPerier wrote:

Because the llvm.intr.stacksave is not emitted anymore (at that point and in general).

At the point of the pipeline where the FORALL/WHERE logic is optimized/implemented, the call is now wrapped inside an hlfir.eval_in_mem, so it has not yet been decided if and how to allocate the storage, so there is no point emitting those ops in lowering anymore.

In general because hlfir.eval_in_mem now uses heap allocation for dynamically sized results like all other temps created in bufferization (it is more consistent and easier).

The analysis stops at the first unknow side effects, so the unknown side effect of the function call is not new, but it is now the first one encountered (which is an improvement if anything).

https://github.com/llvm/llvm-project/pull/118070