[Bug tree-optimization/114061] GCC fails vectorization when using __builtin_prefetch

["cvs-commit at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114061

--- Comment #6 from GCC Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Victor Do Nascimento
<victorldn@gcc.gnu.org>:

https://gcc.gnu.org/g:adcc815a01ae009d2768b6afb546e357bd37bbd2

commit r15-1211-gadcc815a01ae009d2768b6afb546e357bd37bbd2
Author: Victor Do Nascimento <victor.donascimento@arm.com>
Date:   Wed May 22 12:14:11 2024 +0100

    middle-end: Drop __builtin_prefetch calls in autovectorization [PR114061]

    At present the autovectorizer fails to vectorize simple loops
    involving calls to `__builtin_prefetch'.  A simple example of such
    loop is given below:

    void foo(double * restrict a, double * restrict b, int n){
      int i;
      for(i=0; i<n; ++i){
        a[i] = a[i] + b[i];
        __builtin_prefetch(&(b[i+8]));
      }
    }

    The failure stems from two issues:

    1. Given that it is typically not possible to fully reason about a
       function call due to the possibility of side effects, the
       autovectorizer does not attempt to vectorize loops which make such
       calls.

       Given the memory reference passed to `__builtin_prefetch', in the
       absence of assurances about its effect on the passed memory
       location the compiler deems the function unsafe to vectorize,
       marking it as clobbering memory in `vect_find_stmt_data_reference'.
       This leads to the failure in autovectorization.

    2. Notwithstanding the above issue, though the prefetch statement
       would be classed as `vect_unused_in_scope', the loop invariant that
       is used in the address of the prefetch is the scalar loop's and not
       the vector loop's IV. That is, it still uses `i' and not `vec_iv'
       because the instruction wasn't vectorized, causing DCE to think the
       value is live, such that we now have both the vector and scalar loop
       invariant actively used in the loop.

    This patch addresses both of these:

    1. About the issue regarding the memory clobber, data prefetch does
       not generate faults if its address argument is invalid and does not
       write to memory.  Therefore, it does not alter the internal state
       of the program or its control flow under any circumstance.  As
       such, it is reasonable that the function be marked as not affecting
       memory contents.

       To achieve this, we add the necessary logic to
       `get_references_in_stmt' to ensure that builtin functions are given
       given the same treatment as internal functions.  If the gimple call
       is to a builtin function and its function code is
       `BUILT_IN_PREFETCH', we mark `clobbers_memory' as false.

    2. Finding precedence in the way clobber statements are handled,
       whereby the vectorizer drops these from both the scalar and
       vectorized versions of a given loop, we choose to drop prefetch
       hints in a similar fashion.  This seems appropriate given how
       software prefetch hints are typically ignored by processors across
       architectures, as they seldom lead to performance gain over their
       hardware counterparts.

    gcc/ChangeLog:

            PR tree-optimization/114061
            * tree-data-ref.cc (get_references_in_stmt): set
            `clobbers_memory' to false for __builtin_prefetch.
            * tree-vect-loop.cc (vect_transform_loop): Drop all
            __builtin_prefetch calls from loops.

    gcc/testsuite/ChangeLog:

            * gcc.dg/vect/vect-prefetch-drop.c: New test.
            * gcc.target/aarch64/vect-prefetch-drop.c: Likewise.