This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
[PATCH] Fix bdverN vector cost of cond_[not_]taken_branch_cost
- From: Richard Biener <rguenther at suse dot de>
- To: gcc-patches at gcc dot gnu dot org
- Cc: ubizjak at gmail dot com
- Date: Tue, 7 Apr 2015 10:19:12 +0200 (CEST)
- Subject: [PATCH] Fix bdverN vector cost of cond_[not_]taken_branch_cost
- Authentication-results: sourceware.org; auth=none
They are suspiciously low (compared to say scalar_stmt_cost) and with
them and the fix for the vectorizer cost model to properly account
scalar stmt costs (and thus correctly dealing with odd costs as bdverN
have) we regress 252.eon because we consider a loop vectorized and
peeled for alignment loop profitable which clearly isn't.
Bootstrap and regtest running on x86_64-unknown-linux-gnu. I've
tested with all b[dt]verN -marchs and the slp-pr56812.cc testcase
(yes, we've run into a similar issue earlier). I've also put the
patch on our SPEC tester to look for fallout.
It really looks like the costs were derived by some automatic
searching of the parameter space and thus "optimizing" for bugs
in the vectorizer cost model that have meanwhile been fixed
(scalar stmt cost == 6 but scalar load/store cost == 4!?). It is
not a good idea to put in paramters that you can't make sense of
from an architectural point of view (yes, taken/not-taken branch
is somewhat bogus kinds, I'd like to change that to correctly
predicted / wrongly predicted for GCC 6).
Ok for trunk and 4.9 branch?
Thanks,
Richard.
2015-04-07 Richard Biener <rguenther@suse.de>
PR target/65660
* config/i386/i386.c (bdver1_cost): Double cond_taken_branch_cost
and cond_not_taken_branch_cost to 4 and 2.
(bdver2_cost): Likewise.
(bdver3_cost): Likewise.
(bdver4_cost): Likewise.
Index: gcc/config/i386/i386.c
===================================================================
*** gcc/config/i386/i386.c (revision 221888)
--- gcc/config/i386/i386.c (working copy)
*************** const struct processor_costs bdver1_cost
*** 1025,1032 ****
4, /* vec_align_load_cost. */
4, /* vec_unalign_load_cost. */
4, /* vec_store_cost. */
! 2, /* cond_taken_branch_cost. */
! 1, /* cond_not_taken_branch_cost. */
};
/* BDVER2 has optimized REP instruction for medium sized blocks, but for
--- 1025,1032 ----
4, /* vec_align_load_cost. */
4, /* vec_unalign_load_cost. */
4, /* vec_store_cost. */
! 4, /* cond_taken_branch_cost. */
! 2, /* cond_not_taken_branch_cost. */
};
/* BDVER2 has optimized REP instruction for medium sized blocks, but for
*************** const struct processor_costs bdver2_cost
*** 1121,1128 ****
4, /* vec_align_load_cost. */
4, /* vec_unalign_load_cost. */
4, /* vec_store_cost. */
! 2, /* cond_taken_branch_cost. */
! 1, /* cond_not_taken_branch_cost. */
};
--- 1121,1128 ----
4, /* vec_align_load_cost. */
4, /* vec_unalign_load_cost. */
4, /* vec_store_cost. */
! 4, /* cond_taken_branch_cost. */
! 2, /* cond_not_taken_branch_cost. */
};
*************** struct processor_costs bdver3_cost = {
*** 1208,1215 ****
4, /* vec_align_load_cost. */
4, /* vec_unalign_load_cost. */
4, /* vec_store_cost. */
! 2, /* cond_taken_branch_cost. */
! 1, /* cond_not_taken_branch_cost. */
};
/* BDVER4 has optimized REP instruction for medium sized blocks, but for
--- 1208,1215 ----
4, /* vec_align_load_cost. */
4, /* vec_unalign_load_cost. */
4, /* vec_store_cost. */
! 4, /* cond_taken_branch_cost. */
! 2, /* cond_not_taken_branch_cost. */
};
/* BDVER4 has optimized REP instruction for medium sized blocks, but for
*************** struct processor_costs bdver4_cost = {
*** 1294,1301 ****
4, /* vec_align_load_cost. */
4, /* vec_unalign_load_cost. */
4, /* vec_store_cost. */
! 2, /* cond_taken_branch_cost. */
! 1, /* cond_not_taken_branch_cost. */
};
/* BTVER1 has optimized REP instruction for medium sized blocks, but for
--- 1294,1301 ----
4, /* vec_align_load_cost. */
4, /* vec_unalign_load_cost. */
4, /* vec_store_cost. */
! 4, /* cond_taken_branch_cost. */
! 2, /* cond_not_taken_branch_cost. */
};
/* BTVER1 has optimized REP instruction for medium sized blocks, but for