]> git.ipfire.org Git - thirdparty/gcc.git/commitdiff
vect: Don't retry if the previous analysis fails
authorKewen Lin <linkw@linux.ibm.com>
Wed, 17 May 2023 07:48:40 +0000 (02:48 -0500)
committerKewen Lin <linkw@linux.ibm.com>
Wed, 17 May 2023 07:48:40 +0000 (02:48 -0500)
When working on a cost tweaking patch, I found that a newly
added test case has different dumpings with stage-1 and
bootstrapped gcc.  By looking into it, the apparent reason
is vect_analyze_loop_2 doesn't get slp_done_for_suggested_uf
set expectedly, the following retrying will use the garbage
slp_done_for_suggested_uf instead.  In fact, the setting of
slp_done_for_suggested_uf only happens when the previous
analysis succeeds, for the mentioned test case, its previous
analysis does fail, it's unexpected to use the value of
slp_done_for_suggested_uf any more.

In function vect_analyze_loop_1, we only return success when
res is true, which is the result of 1st analysis.  It means
we never try to vectorize with unroll_vinfo if the previous
analysis fails.  So this patch shouldn't break anything, and
just stop some useless analysis early.

gcc/ChangeLog:

* tree-vect-loop.cc (vect_analyze_loop_1): Don't retry analysis with
suggested unroll factor once the previous analysis fails.

gcc/tree-vect-loop.cc

index ed0166fedab356d3eef546a059477e17fab27fa5..905145ae97be235adff835b4659fa0f6ec4f04d5 100644 (file)
@@ -3044,7 +3044,7 @@ vect_analyze_loop_1 (class loop *loop, vec_info_shared *shared,
                     res ? "succeeded" : " failed",
                     GET_MODE_NAME (loop_vinfo->vector_mode));
 
-  if (!main_loop_vinfo && suggested_unroll_factor > 1)
+  if (res && !main_loop_vinfo && suggested_unroll_factor > 1)
     {
       if (dump_enabled_p ())
        dump_printf_loc (MSG_NOTE, vect_location,