public inbox for gcc-bugs@sourceware.org help / color / mirror / Atom feed
From: "olegendo at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org> To: gcc-bugs@gcc.gnu.org Subject: [Bug target/54089] [SH] Refactor shift patterns Date: Wed, 07 Nov 2012 23:32:00 -0000 [thread overview] Message-ID: <bug-54089-4-bzU4JykAir@http.gcc.gnu.org/bugzilla/> (raw) In-Reply-To: <bug-54089-4@http.gcc.gnu.org/bugzilla/> http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54089 --- Comment #25 from Oleg Endo <olegendo at gcc dot gnu.org> 2012-11-07 23:31:20 UTC --- Created attachment 28633 --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=28633 Arithmetic right shift rework 2 This could be an alternative approach for the arith right shifts. Applied to rev 193240: CSiBE -m4-single -O2 -ml -mpretend-cmove sum: 3159747 -> 3160047 +300 / +0.009494 % CSiBE -m2 -O2 -ml sum: 3314907 -> 3313319 -1588 / -0.047905 % The idea here is to let arith shifts initially go through a slightly more complex pattern 'ashrsi3_bridge_const', which allows the lib function address or shad constant to be CSE'ed early on. As mentioned in comment #18, the arith -> logical shift conversion is done by combine only under certain circumstances. After some trial and error I arrived at the 'ashrsi3_bridge_const' pattern, which makes combine convert most of the arith shifts into logical shifts. In some cases it also finds more opportunities, but regressions like this one still remain: void MC_put_xy_16_c (uint8_t * dest, const uint8_t * ref, const int stride, int height) { // must see a shlr instead of 2x shar. // somehow the loop prevents this... do { dest[0] = (((ref[0]+ref[0 +1]+(ref+stride)[0]+(ref+stride)[0 +1]+2)>>2)); ref += stride; dest += stride; } while (--height); } In addition to that, some unlucky register allocations appear, which cause unneeded reg-reg copies (it seems something has trouble utilizing commutative insns such as 'add'). The other problem is that patterns such as: (define_insn_and_split "*lshrsi3" [(set (match_operand:SI 0 "arith_reg_dest") (lshiftrt:SI (match_operand:SI 1 "arith_reg_operand") (match_operand:SI 2 "shift_count_operand"))) (clobber (reg:SI T_REG)) (use (match_operand:SI 3 "general_operand")) (use (match_operand:SI 4 "general_operand"))] "TARGET_SH1 && can_create_pseudo_p () && CONST_INT_P (operands[2])" "#" "&& 1" [(const_int 0)] { emit_insn (gen_lshrsi3 (operands[0], operands[1], operands[2])); DONE; }) have to be added for combine to convert more arith -> logical shifts. In this patch I've added only one such pattern. However, if for example zero_extract patterns are added, variations such as above would be required, too. This looks discomforting to me. Maybe the better solution would indeed be to add a arith -> logical shift conversion pass before combine, or try to convert arith shifts in the split pass after combine by looking at the data flow of the shifted values.
next prev parent reply other threads:[~2012-11-07 23:32 UTC|newest] Thread overview: 102+ messages / expand[flat|nested] mbox.gz Atom feed top 2012-07-24 23:42 [Bug target/54089] New: " olegendo at gcc dot gnu.org 2012-07-24 23:46 ` [Bug target/54089] " olegendo at gcc dot gnu.org 2012-07-25 23:03 ` olegendo at gcc dot gnu.org 2012-07-27 17:36 ` olegendo at gcc dot gnu.org 2012-07-30 6:43 ` olegendo at gcc dot gnu.org 2012-08-09 21:55 ` olegendo at gcc dot gnu.org 2012-08-09 23:18 ` olegendo at gcc dot gnu.org 2012-08-09 23:28 ` olegendo at gcc dot gnu.org 2012-08-09 23:36 ` olegendo at gcc dot gnu.org 2012-08-09 23:43 ` olegendo at gcc dot gnu.org 2012-08-10 0:40 ` kkojima at gcc dot gnu.org 2012-08-10 14:25 ` rmansfield at qnx dot com 2012-08-10 15:40 ` olegendo at gcc dot gnu.org 2012-08-11 20:26 ` olegendo at gcc dot gnu.org 2012-08-16 23:13 ` olegendo at gcc dot gnu.org 2012-08-20 21:29 ` olegendo at gcc dot gnu.org 2012-08-22 22:52 ` olegendo at gcc dot gnu.org 2012-09-10 20:35 ` olegendo at gcc dot gnu.org 2012-09-10 21:27 ` olegendo at gcc dot gnu.org 2012-09-17 23:30 ` olegendo at gcc dot gnu.org 2012-09-19 17:49 ` olegendo at gcc dot gnu.org 2012-09-25 19:07 ` olegendo at gcc dot gnu.org 2012-09-30 18:46 ` olegendo at gcc dot gnu.org 2012-10-16 10:53 ` amylaar at gcc dot gnu.org 2012-10-16 11:49 ` olegendo at gcc dot gnu.org 2012-11-06 11:56 ` olegendo at gcc dot gnu.org 2012-11-07 23:32 ` olegendo at gcc dot gnu.org [this message] 2012-11-09 10:48 ` olegendo at gcc dot gnu.org 2012-11-09 13:29 ` amylaar at gcc dot gnu.org 2012-11-13 1:14 ` olegendo at gcc dot gnu.org 2013-02-16 11:36 ` olegendo at gcc dot gnu.org 2013-12-17 11:11 ` olegendo at gcc dot gnu.org 2014-05-16 23:12 ` olegendo at gcc dot gnu.org 2014-12-16 21:38 ` olegendo at gcc dot gnu.org 2015-05-30 13:51 ` bugdal at aerifal dot cx 2015-05-30 14:11 ` olegendo at gcc dot gnu.org 2015-05-30 14:35 ` bugdal at aerifal dot cx 2015-05-30 14:42 ` olegendo at gcc dot gnu.org 2015-05-30 14:46 ` bugdal at aerifal dot cx 2015-05-30 14:57 ` olegendo at gcc dot gnu.org 2015-08-11 3:56 ` bugdal at aerifal dot cx 2015-08-11 14:17 ` olegendo at gcc dot gnu.org 2015-08-11 21:05 ` segher at gcc dot gnu.org 2015-08-12 1:30 ` segher at gcc dot gnu.org 2023-06-01 7:32 ` klepikov.alex+bugs at gmail dot com 2023-06-01 8:21 ` olegendo at gcc dot gnu.org 2023-06-01 11:18 ` klepikov.alex+bugs at gmail dot com 2023-06-01 12:08 ` olegendo at gcc dot gnu.org 2023-06-01 17:45 ` segher at gcc dot gnu.org 2023-06-01 20:46 ` olegendo at gcc dot gnu.org 2023-06-03 7:24 ` klepikov.alex+bugs at gmail dot com 2023-06-03 8:50 ` olegendo at gcc dot gnu.org 2023-06-03 15:43 ` klepikov.alex+bugs at gmail dot com 2023-06-03 16:09 ` olegendo at gcc dot gnu.org 2023-06-04 18:35 ` klepikov.alex+bugs at gmail dot com 2023-06-05 0:03 ` olegendo at gcc dot gnu.org 2023-06-06 10:30 ` klepikov.alex+bugs at gmail dot com 2023-06-06 10:51 ` olegendo at gcc dot gnu.org 2023-06-06 12:17 ` klepikov.alex+bugs at gmail dot com 2023-06-06 12:39 ` olegendo at gcc dot gnu.org 2023-06-06 12:55 ` klepikov.alex+bugs at gmail dot com 2023-06-06 16:10 ` klepikov.alex+bugs at gmail dot com 2023-06-06 19:07 ` klepikov.alex+bugs at gmail dot com 2023-06-06 23:11 ` olegendo at gcc dot gnu.org 2023-06-08 12:07 ` klepikov.alex+bugs at gmail dot com 2023-06-08 12:09 ` klepikov.alex+bugs at gmail dot com 2023-06-08 13:22 ` olegendo at gcc dot gnu.org 2023-06-08 15:22 ` klepikov.alex+bugs at gmail dot com 2023-06-08 20:17 ` olegendo at gcc dot gnu.org 2023-06-14 9:30 ` klepikov.alex+bugs at gmail dot com 2023-06-14 9:31 ` klepikov.alex+bugs at gmail dot com 2023-06-16 13:57 ` klepikov.alex+bugs at gmail dot com 2023-06-16 21:58 ` olegendo at gcc dot gnu.org 2023-06-17 6:08 ` klepikov.alex+bugs at gmail dot com 2023-06-17 7:06 ` olegendo at gcc dot gnu.org 2023-06-17 8:24 ` klepikov.alex+bugs at gmail dot com 2023-06-17 9:46 ` olegendo at gcc dot gnu.org 2023-06-17 11:00 ` klepikov.alex+bugs at gmail dot com 2023-06-17 17:57 ` klepikov.alex+bugs at gmail dot com 2023-06-20 11:28 ` klepikov.alex+bugs at gmail dot com 2023-06-20 17:03 ` klepikov.alex+bugs at gmail dot com 2023-06-20 23:46 ` olegendo at gcc dot gnu.org 2023-06-21 11:51 ` klepikov.alex+bugs at gmail dot com 2023-06-21 20:13 ` segher at gcc dot gnu.org 2023-06-21 20:32 ` segher at gcc dot gnu.org 2023-06-22 11:09 ` klepikov.alex+bugs at gmail dot com 2023-06-22 11:44 ` olegendo at gcc dot gnu.org 2023-06-22 12:34 ` klepikov.alex+bugs at gmail dot com 2023-06-23 6:02 ` klepikov.alex+bugs at gmail dot com 2023-06-23 6:06 ` olegendo at gcc dot gnu.org 2023-06-23 14:11 ` segher at gcc dot gnu.org 2023-07-06 14:16 ` olegendo at gcc dot gnu.org 2023-07-07 11:10 ` klepikov.alex+bugs at gmail dot com 2023-07-07 11:45 ` olegendo at gcc dot gnu.org 2023-07-08 7:56 ` klepikov.alex+bugs at gmail dot com 2023-07-09 13:55 ` olegendo at gcc dot gnu.org 2023-07-10 14:02 ` klepikov.alex+bugs at gmail dot com 2023-07-13 23:40 ` olegendo at gcc dot gnu.org 2023-07-14 14:31 ` klepikov.alex+bugs at gmail dot com 2023-10-13 5:02 ` olegendo at gcc dot gnu.org 2023-10-14 15:20 ` klepikov.alex+bugs at gmail dot com 2023-10-15 1:06 ` olegendo at gcc dot gnu.org
Reply instructions: You may reply publicly to this message via plain-text email using any one of the following methods: * Save the following mbox file, import it into your mail client, and reply-to-all from there: mbox Avoid top-posting and favor interleaved quoting: https://en.wikipedia.org/wiki/Posting_style#Interleaved_style * Reply using the --to, --cc, and --in-reply-to switches of git-send-email(1): git send-email \ --in-reply-to=bug-54089-4-bzU4JykAir@http.gcc.gnu.org/bugzilla/ \ --to=gcc-bugzilla@gcc.gnu.org \ --cc=gcc-bugs@gcc.gnu.org \ /path/to/YOUR_REPLY https://kernel.org/pub/software/scm/git/docs/git-send-email.html * If your mail client supports setting the In-Reply-To header via mailto: links, try the mailto: linkBe sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions for how to clone and mirror all data and code used for this inbox; as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).