[flang] missed optimization of product intrinsic

[foxtran]

For the following code:

subroutine sum_and_prod(v, n, s, p)
  implicit none (type, external)
  integer, intent(in) :: n
  real, intent(in) :: v(n)
  real, intent(out) :: s, p
  s = sum(v)
  p = product(v)
end subroutine sum_and_prod

flang produces native ASM code for summation, while for product there is a call to runtime library.

Even if length of vector is known (and small):

subroutine sum_and_prod_vec3(v, s, p)
  implicit none (type, external)
  real, intent(in) :: v(3)
  real, intent(out) :: s, p
  s = sum(v)
  p = product(v)
end subroutine sum_and_prod_vec3

ASM still have call of runtime library.

It would be nice to avoid any calls especially for short arrays.

ASM output:
https://godbolt.org/z/d3Edr9GG4

[llvmbot]

@llvm/issue-subscribers-flang-ir

Author: foxtran (foxtran)

For the following code: ```fortran subroutine sum_and_prod(v, n, s, p) implicit none (type, external) integer, intent(in) :: n real, intent(in) :: v(n) real, intent(out) :: s, p s = sum(v) p = product(v) end subroutine sum_and_prod ``` flang produces native ASM code for summation, while for product there is a call to runtime library.

Even if length of vector is known (and small):

subroutine sum_and_prod_vec3(v, s, p)
  implicit none (type, external)
  real, intent(in) :: v(3)
  real, intent(out) :: s, p
  s = sum(v)
  p = product(v)
end subroutine sum_and_prod_vec3

ASM still have call of runtime library.

It would be nice to avoid any calls especially for short arrays.

ASM output:
https://godbolt.org/z/d3Edr9GG4

[stomfaig]

I can give this a go.

[eugeneepshteyn]

Note that HLFIR has correct call to hlfir.product:

$ flang -fc1 -emit-hlfir repro-2.f90 -O1
...
  9     %5 = hlfir.sum %4#0 {fastmath = #arith.fastmath<contract>} : (!fir.ref<!fir.array<3xf32>>) -> f32
 10     hlfir.assign %5 to %2#0 : f32, !fir.ref<f32>
 11     %6 = hlfir.product %4#0 {fastmath = #arith.fastmath<contract>} : (!fir.ref<!fir.array<3xf32>>) -> f32
 12     hlfir.assign %6 to %1#0 : f32, !fir.ref<f32>

Yet, when emitting FIR, the sum operation is inlined, but not the product:

$ flang -fc1 -emit-fir repro-2.f90 -O1
...
 13     %5 = fir.do_loop %arg3 = %c1 to %c3 step %c1 iter_args(%arg4 = %cst) -> (f32) {
 14       %15 = fir.array_coor %4(%3) %arg3 : (!fir.ref<!fir.array<3xf32>>, !fir.shape<1>, index) -> !fir.ref<f32>
 15       %16 = fir.load %15 : !fir.ref<f32>
 16       %17 = arith.addf %arg4, %16 fastmath<contract> : f32
 17       fir.result %17 : f32
 18     }
...
 28     %14 = fir.call @_FortranAProductReal4(%10, %11, %c6_i32, %12, %13) fastmath<contract> : (!fir.box<none>, !fir.ref<i8>, i3    2, i32, !fir.box<none>) -> f32
 29     fir.store %14 to %1 : !fir.ref<f32>

(You need to use at least -O1 to cause the inlining to happen. At default -O0, the runtime functions will be called for both sum and product.)

This tells me that somehow during HLFIR to FIR lowering the inlining doesn't happen for the product operation.

The code for this conversion seems to exist: using ProductOpConversion = HlfirReductionIntrinsicConversion<hlfir::ProductOp>;
... so either it doesn't support inlining or for some reason inlining is not kicking in.