From: Uros Bizjak <ubizjak@gmail.com>
To: Hongtao Liu <crazylht@gmail.com>
Cc: GCC Patches <gcc-patches@gcc.gnu.org>,
Jakub Jelinek <jakub@redhat.com>,
"H. J. Lu" <hjl.tools@gmail.com>
Subject: Re: [PATCH] [i386] Fix _mm256_zeroupper to notify LRA that vzeroupper will kill sse registers. [PR target/82735]
Date: Thu, 13 May 2021 11:40:55 +0200 [thread overview]
Message-ID: <CAFULd4aUuYVQ90VkAZp4N1rsCSsJaokHmPeD_fnqdqCeLDS1ZA@mail.gmail.com> (raw)
In-Reply-To: <CAMZc-bwzYQWvQH7sbDczJg6hWktBqA7Pdcbp-FtAnza1By8TAQ@mail.gmail.com>
On Thu, May 13, 2021 at 11:18 AM Hongtao Liu <crazylht@gmail.com> wrote:
>
> Hi:
> When __builtin_ia32_vzeroupper is called explicitly, the corresponding
> vzeroupper pattern does not carry any CLOBBERS or SETs before LRA,
> which leads to incorrect optimization in pass_reload.
> In order to solve this problem, this patch introduces a pre_reload
> splitter which adds CLOBBERS to vzeroupper's pattern, it can solve the
> problem in pr.
>
> At the same time, in order to optimize the low 128 bits in
> post_reload CSE, this patch also transforms those CLOBBERS to SETs in
> pass_vzeroupper.
>
> It works fine except for TARGET_64BIT_MS_ABI, under which xmm6-xmm15
> are callee-saved, so even if there're no other uses of xmm6-xmm15 in the
> function, because of vzeroupper's pattern, pro_epilog will save and
> restore those registers, which is obviously redundant. In order to
> eliminate this redundancy, a post_reload splitter is introduced, which
> drops those SETs, until epilogue_completed splitter adds those SETs
> back, it looks to be safe since there's no CSE between post_reload
> split2 and epilogue_completed split3??? Also frame info needs to be
> updated in pro_epilog, which saves and restores xmm6-xmm15 only if
> there's usage other than explicit vzeroupper pattern.
>
> Bootstrapped and regtested on X86_64-linux-gnu{-m32,}
> Ok for trunk?
Some time ago a support for CLOBBER_HIGH RTX was added (and later
removed for some reason). Perhaps we could resurrect the patch for the
purpose of ferrying 128bit modes via vzeroupper RTX?
+(define_split
+ [(match_parallel 0 "vzeroupper_pattern"
+ [(unspec_volatile [(const_int 0)] UNSPECV_VZEROUPPER)])]
+ "TARGET_AVX && ix86_pre_reload_split ()"
+ [(match_dup 0)]
+{
+ /* When vzeroupper is explictly used, for LRA purpose, make it clear
+ the instruction kills sse registers. */
+ gcc_assert (cfun->machine->has_explicit_vzeroupper);
+ unsigned int nregs = TARGET_64BIT ? 16 : 8;
+ rtvec vec = rtvec_alloc (nregs + 1);
+ RTVEC_ELT (vec, 0) = gen_rtx_UNSPEC_VOLATILE (VOIDmode,
+ gen_rtvec (1, const1_rtx),
+ UNSPECV_VZEROUPPER);
+ for (unsigned int i = 0; i < nregs; ++i)
+ {
+ unsigned int regno = GET_SSE_REGNO (i);
+ rtx reg = gen_rtx_REG (V2DImode, regno);
+ RTVEC_ELT (vec, i + 1) = gen_rtx_CLOBBER (VOIDmode, reg);
+ }
+ operands[0] = gen_rtx_PARALLEL (VOIDmode, vec);
+})
Wouldn't this also kill lower 128bit values that are not touched by
vzeroupper? A CLOBBER_HIGH would be more appropriate here.
Uros.
> gcc/ChangeLog:
>
> PR target/82735
> * config/i386/i386-expand.c (ix86_expand_builtin): Count
> number of __builtin_ia32_vzeroupper.
> * config/i386/i386-features.c (ix86_add_reg_usage_to_vzerouppers):
> Transform CLOBBERs to SETs for explicit vzeroupper pattern so
> that CSE can optimize lower 128 bits.
> * config/i386/i386.c (ix86_handle_explicit_vzeroupper_in_pro_epilog):
> New.
> (ix86_save_reg): If there's no use of xmm6~xmm15 other than
> explicit vzeroupper under TARGET_64BIT_MS_ABI, no need to save
> REGNO.
> (ix86_finalize_stack_frame_flags): Recompute frame layout if
> there's explicit vzeroupper under TARGET_64BIT_MS_ABI.
> * config/i386/i386.h (struct machine_function): Change type of
> has_explicit_vzeroupper from BOOL_BITFILED to unsigned int.
> * config/i386/sse.md (*avx_vzeroupper_2): New post-reload
> splitter which will drop all SETs for explicit vzeroupper
> patterns.
> (*avx_vzeroupper_1): Generate SET reg to reg instead of
> CLOBBER, and add pre-reload splitter after it.
>
> gcc/testsuite/ChangeLog:
>
> PR target/82735
> * gcc.target/i386/pr82735-1.c: New test.
> * gcc.target/i386/pr82735-2.c: New test.
> * gcc.target/i386/pr82735-3.c: New test.
> * gcc.target/i386/pr82735-4.c: New test.
> * gcc.target/i386/pr82735-5.c: New test.
>
>
> --
> BR,
> Hongtao
next prev parent reply other threads:[~2021-05-13 9:41 UTC|newest]
Thread overview: 45+ messages / expand[flat|nested] mbox.gz Atom feed top
2021-05-13 9:23 Hongtao Liu
2021-05-13 9:40 ` Uros Bizjak [this message]
2021-05-13 9:43 ` Uros Bizjak
2021-05-13 9:54 ` Jakub Jelinek
2021-05-13 11:32 ` Richard Sandiford
2021-05-13 11:37 ` Jakub Jelinek
2021-05-13 11:52 ` Richard Sandiford
2021-05-14 2:27 ` Hongtao Liu
2021-05-17 8:44 ` Hongtao Liu
2021-05-17 9:56 ` Richard Sandiford
2021-05-18 13:12 ` Hongtao Liu
2021-05-18 15:18 ` Richard Sandiford
2021-05-25 6:04 ` Hongtao Liu
2021-05-25 6:30 ` Hongtao Liu
2021-05-27 5:07 ` Hongtao Liu
2021-05-27 7:05 ` Uros Bizjak
2021-06-01 2:24 ` Hongtao Liu
2021-06-03 6:54 ` [PATCH 1/2] CALL_INSN may not be a real function call liuhongt
2021-06-03 6:54 ` [PATCH 2/2] Fix _mm256_zeroupper by representing the instructions as call_insns in which the call has a special vzeroupper ABI liuhongt
2021-06-04 2:56 ` Hongtao Liu
2021-06-04 6:26 ` Uros Bizjak
2021-06-04 6:34 ` Hongtao Liu
2021-06-07 19:04 ` [PATCH] x86: Don't compile pr82735-[345].c for x32 H.J. Lu
2021-06-04 2:55 ` [PATCH 1/2] CALL_INSN may not be a real function call Hongtao Liu
2021-06-04 7:50 ` Jakub Jelinek
2021-07-05 23:30 ` Segher Boessenkool
2021-07-06 0:03 ` Jeff Law
2021-07-06 1:49 ` Hongtao Liu
2021-07-07 14:55 ` Segher Boessenkool
2021-07-07 17:56 ` Jeff Law
2021-07-06 1:37 ` Hongtao Liu
2021-07-07 2:44 ` Hongtao Liu
2021-07-07 8:15 ` Richard Biener
2021-07-07 14:52 ` Segher Boessenkool
2021-07-07 15:23 ` Hongtao Liu
2021-07-07 23:42 ` Segher Boessenkool
2021-07-08 4:14 ` Hongtao Liu
2021-07-07 15:32 ` Hongtao Liu
2021-07-07 23:54 ` Segher Boessenkool
2021-07-09 7:20 ` Hongtao Liu
2021-07-07 15:52 ` Hongtao Liu
2021-05-27 7:20 ` [PATCH] [i386] Fix _mm256_zeroupper to notify LRA that vzeroupper will kill sse registers. [PR target/82735] Jakub Jelinek
2021-05-27 10:50 ` Richard Sandiford
2021-06-01 2:22 ` Hongtao Liu
2021-06-01 2:25 ` Hongtao Liu
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=CAFULd4aUuYVQ90VkAZp4N1rsCSsJaokHmPeD_fnqdqCeLDS1ZA@mail.gmail.com \
--to=ubizjak@gmail.com \
--cc=crazylht@gmail.com \
--cc=gcc-patches@gcc.gnu.org \
--cc=hjl.tools@gmail.com \
--cc=jakub@redhat.com \
/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: link
Be 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).