]> git.ipfire.org Git - thirdparty/gcc.git/commitdiff
aarch64: Restrict issue heuristics to inner vector loop
authorRichard Sandiford <richard.sandiford@arm.com>
Fri, 6 Aug 2021 14:36:57 +0000 (15:36 +0100)
committerRichard Sandiford <richard.sandiford@arm.com>
Fri, 6 Aug 2021 14:36:57 +0000 (15:36 +0100)
The AArch64 vector costs try to take issue rates into account.
However, when vectorising an outer loop, we lumped the inner
and outer operations together, which is somewhat meaningless.
This patch restricts the heuristic to the inner loop.

gcc/
* config/aarch64/aarch64.c (aarch64_add_stmt_cost): Only
record issue information for operations that occur in the
innermost loop.

(cherry picked from commit 9690309baf8294b0512b55b133bc102dc0dac5b5)

gcc/config/aarch64/aarch64.c

index fc9adefbe046c96f864d358ad8e3e43ecb2aa2d9..3113b3b339944981bc3745c6b0847b6c60c47fd1 100644 (file)
@@ -15394,6 +15394,10 @@ aarch64_add_stmt_cost (class vec_info *vinfo, void *data, int count,
       fractional_cost stmt_cost
        = aarch64_builtin_vectorization_cost (kind, vectype, misalign);
 
+      bool in_inner_loop_p = (where == vect_body
+                             && stmt_info
+                             && stmt_in_inner_loop_p (vinfo, stmt_info));
+
       /* Do one-time initialization based on the vinfo.  */
       loop_vec_info loop_vinfo = dyn_cast<loop_vec_info> (vinfo);
       bb_vec_info bb_vinfo = dyn_cast<bb_vec_info> (vinfo);
@@ -15440,14 +15444,15 @@ aarch64_add_stmt_cost (class vec_info *vinfo, void *data, int count,
          stmt_cost = aarch64_adjust_stmt_cost (kind, stmt_info, vectype,
                                                stmt_cost);
 
-         /* If we're recording a nonzero vector loop body cost, also estimate
-            the operations that would need to be issued by all relevant
-            implementations of the loop.  */
+         /* If we're recording a nonzero vector loop body cost for the
+            innermost loop, also estimate the operations that would need
+            to be issued by all relevant implementations of the loop.  */
          auto *issue_info = aarch64_tune_params.vec_costs->issue_info;
          if (loop_vinfo
              && issue_info
              && costs->vec_flags
              && where == vect_body
+             && (!LOOP_VINFO_LOOP (loop_vinfo)->inner || in_inner_loop_p)
              && vectype
              && stmt_cost != 0)
            {
@@ -15491,8 +15496,7 @@ aarch64_add_stmt_cost (class vec_info *vinfo, void *data, int count,
       /* Statements in an inner loop relative to the loop being
         vectorized are weighted more heavily.  The value here is
         arbitrary and could potentially be improved with analysis.  */
-      if (where == vect_body && stmt_info
-         && stmt_in_inner_loop_p (vinfo, stmt_info))
+      if (in_inner_loop_p)
        count *= 50; /*  FIXME  */
 
       retval = (count * stmt_cost).ceil ();