public inbox for gcc-cvs@sourceware.org
help / color / mirror / Atom feed
From: Jakub Jelinek <jakub@gcc.gnu.org>
To: gcc-cvs@gcc.gnu.org
Subject: [gcc r12-7832] ecog: Return 1 from insn_invalid_p if REG_INC reg overlaps some stored reg [PR103775]
Date: Sat, 26 Mar 2022 15:23:19 +0000 (GMT)	[thread overview]
Message-ID: <20220326152319.9739D3858D28@sourceware.org> (raw)

https://gcc.gnu.org/g:6459e6537632bc06e04e6011ca7fb6488f0e8e7d

commit r12-7832-g6459e6537632bc06e04e6011ca7fb6488f0e8e7d
Author: Jakub Jelinek <jakub@redhat.com>
Date:   Sat Mar 26 16:21:36 2022 +0100

    ecog: Return 1 from insn_invalid_p if REG_INC reg overlaps some stored reg [PR103775]
    
    The following testcase ICEs on aarch64-linux with -g and
    assembles with a warning otherwise, because it emits
    ldrb w0,[x0,16]!
    instruction which sets the x0 register multiple times.
    Due to disabled DCE (from -Og) we end up before REE with:
    (insn 12 39 13 2 (set (reg:SI 1 x1 [orig:93 _2 ] [93])
            (zero_extend:SI (mem/c:QI (pre_modify:DI (reg/f:DI 0 x0 [114])
                        (plus:DI (reg/f:DI 0 x0 [114])
                            (const_int 16 [0x10]))) [1 u128_1+0 S1 A128]))) "pr103775.c":5:35 117 {*zero_extendqisi2_aarch64}
         (expr_list:REG_INC (reg/f:DI 0 x0 [114])
            (nil)))
    (insn 13 12 14 2 (set (reg:DI 0 x0 [orig:112 _2 ] [112])
            (zero_extend:DI (reg:SI 1 x1 [orig:93 _2 ] [93]))) "pr103775.c":5:16 111 {*zero_extendsidi2_aarch64}
         (nil))
    which is valid but not exactly efficient as x0 is dead after the
    insn that auto-increments it.  REE turns it into:
    (insn 12 39 44 2 (set (reg:DI 0 x0)
            (zero_extend:DI (mem/c:QI (pre_modify:DI (reg/f:DI 0 x0 [114])
                        (plus:DI (reg/f:DI 0 x0 [114])
                            (const_int 16 [0x10]))) [1 u128_1+0 S1 A128]))) "pr103775.c":5:35 119 {*zero_extendqidi2_aarch64}
         (expr_list:REG_INC (reg/f:DI 0 x0 [114])
            (nil)))
    (insn 44 12 14 2 (set (reg:DI 1 x1)
            (reg:DI 0 x0)) "pr103775.c":5:35 -1
         (nil))
    which is invalid because it sets x0 multiple times, one
    in SET_DEST of the PATTERN and once in PRE_MODIFY.
    As perhaps other passes than REE might suffer from it, IMHO it is better
    to reject this during change validation.
    
    2022-03-26  Jakub Jelinek  <jakub@redhat.com>
    
            PR rtl-optimization/103775
            * recog.cc (check_invalid_inc_dec): New function.
            (insn_invalid_p): Return 1 if REG_INC operand overlaps
            any stored REGs.
    
            * gcc.dg/pr103775.c: New test.

Diff:
---
 gcc/recog.cc                    | 22 ++++++++++++++++++++++
 gcc/testsuite/gcc.dg/pr103775.c | 12 ++++++++++++
 2 files changed, 34 insertions(+)

diff --git a/gcc/recog.cc b/gcc/recog.cc
index 1c9fe716322..cd2410ab2ae 100644
--- a/gcc/recog.cc
+++ b/gcc/recog.cc
@@ -329,6 +329,17 @@ canonicalize_change_group (rtx_insn *insn, rtx x)
     return false;
 }
 
+/* Check if REG_INC argument in *data overlaps a stored REG.  */
+
+static void
+check_invalid_inc_dec (rtx reg, const_rtx, void *data)
+{
+  rtx *pinc = (rtx *) data;
+  if (*pinc == NULL_RTX || MEM_P (reg))
+    return;
+  if (reg_overlap_mentioned_p (reg, *pinc))
+    *pinc = NULL_RTX;
+}
 
 /* This subroutine of apply_change_group verifies whether the changes to INSN
    were valid; i.e. whether INSN can still be recognized.
@@ -386,6 +397,17 @@ insn_invalid_p (rtx_insn *insn, bool in_group)
 	return 1;
     }
 
+  /* Punt if REG_INC argument overlaps some stored REG.  */
+  for (rtx link = FIND_REG_INC_NOTE (insn, NULL_RTX);
+       link; link = XEXP (link, 1))
+    if (REG_NOTE_KIND (link) == REG_INC)
+      {
+	rtx reg = XEXP (link, 0);
+	note_stores (insn, check_invalid_inc_dec, &reg);
+	if (reg == NULL_RTX)
+	  return 1;
+      }
+
   INSN_CODE (insn) = icode;
   return 0;
 }
diff --git a/gcc/testsuite/gcc.dg/pr103775.c b/gcc/testsuite/gcc.dg/pr103775.c
new file mode 100644
index 00000000000..4a8c0d6d145
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/pr103775.c
@@ -0,0 +1,12 @@
+/* PR rtl-optimization/103775 */
+/* { dg-do assemble { target int128 } } */
+/* { dg-options "-Og -fno-forward-propagate -free -g" } */
+
+int
+foo (char a, short b, int c, __int128 d, char e, short f, int g, __int128 h)
+{
+  long i = __builtin_clrsbll ((char) h);
+  __builtin_memset ((char *) &h + 4, d, 3);
+  c &= (char) h;
+  return c + i;
+}


                 reply	other threads:[~2022-03-26 15:23 UTC|newest]

Thread overview: [no followups] expand[flat|nested]  mbox.gz  Atom feed

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=20220326152319.9739D3858D28@sourceware.org \
    --to=jakub@gcc.gnu.org \
    --cc=gcc-cvs@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: 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).