On December 10, 2019 12:28:24 PM GMT+01:00, Richard Sandiford
<richard.sandif...@arm.com> wrote:
>Unlike most vector ops, extract-last reductions replace the original
>scalar code in-situ rather than adding an adjacent vector
>implementation.
>I.e.:
>
> dest_1 = COND_EXPR <...>;
>
>becomes:
>
> dest_1 = .EXTRACT_LAST (...);
>
>gcc.dg/vect/vect-cond-reduc-4.c was ICEing for SVE because we tried
>to replace the pattern statement in this way, rather than replacing
>the original scalar statement.
>
>Tested on aarch64-linux-gnu and x86_64-linux-gnu. OK to install?
Ok.
Richard.
>Richard
>
>
>2019-12-10 Richard Sandiford <richard.sandif...@arm.com>
>
>gcc/
> * tree-vect-stmts.c (vect_finish_replace_stmt): Always use the
> original scalar statement rather than a pattern statement.
> (vectorizable_condition): Likewise, in the handling of extract-last
> reductions.
>
>Index: gcc/tree-vect-stmts.c
>===================================================================
>--- gcc/tree-vect-stmts.c 2019-12-10 09:17:18.768002919 +0000
>+++ gcc/tree-vect-stmts.c 2019-12-10 11:27:23.619355131 +0000
>@@ -1779,9 +1779,10 @@ vect_finish_stmt_generation_1 (stmt_vec_
> stmt_vec_info
> vect_finish_replace_stmt (stmt_vec_info stmt_info, gimple *vec_stmt)
> {
>- gcc_assert (gimple_get_lhs (stmt_info->stmt) == gimple_get_lhs
>(vec_stmt));
>+ gimple *scalar_stmt = vect_orig_stmt (stmt_info)->stmt;
>+ gcc_assert (gimple_get_lhs (scalar_stmt) == gimple_get_lhs
>(vec_stmt));
>
>- gimple_stmt_iterator gsi = gsi_for_stmt (stmt_info->stmt);
>+ gimple_stmt_iterator gsi = gsi_for_stmt (scalar_stmt);
> gsi_replace (&gsi, vec_stmt, true);
>
> return vect_finish_stmt_generation_1 (stmt_info, vec_stmt);
>@@ -10324,20 +10325,21 @@ vectorizable_condition (stmt_vec_info st
>
> if (reduction_type == EXTRACT_LAST_REDUCTION)
> {
>+ gimple *old_stmt = vect_orig_stmt (stmt_info)->stmt;
>+ tree lhs = gimple_get_lhs (old_stmt);
> gcall *new_stmt = gimple_build_call_internal
> (IFN_FOLD_EXTRACT_LAST, 3, else_clause, vec_compare,
> vec_then_clause);
>- gimple_call_set_lhs (new_stmt, scalar_dest);
>- SSA_NAME_DEF_STMT (scalar_dest) = new_stmt;
>- if (stmt_info->stmt == gsi_stmt (*gsi))
>+ gimple_call_set_lhs (new_stmt, lhs);
>+ SSA_NAME_DEF_STMT (lhs) = new_stmt;
>+ if (old_stmt == gsi_stmt (*gsi))
> new_stmt_info = vect_finish_replace_stmt (stmt_info, new_stmt);
> else
> {
> /* In this case we're moving the definition to later in the
> block. That doesn't matter because the only uses of the
> lhs are in phi statements. */
>- gimple_stmt_iterator old_gsi
>- = gsi_for_stmt (stmt_info->stmt);
>+ gimple_stmt_iterator old_gsi = gsi_for_stmt (old_stmt);
> gsi_remove (&old_gsi, true);
> new_stmt_info
> = vect_finish_stmt_generation (stmt_info, new_stmt, gsi);
