[Bug target/101529] New: wrong code with __builtin_convertvector() and __builtin_shufflevector() at -Og

[Bug target/101529] New: wrong code with __builtin_convertvector() and __builtin_shufflevector() at -Og

[zsojka at seznam dot cz]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101529

            Bug ID: 101529
           Summary: wrong code with __builtin_convertvector() and
                    __builtin_shufflevector() at -Og
           Product: gcc
           Version: 12.0
            Status: UNCONFIRMED
          Keywords: wrong-code
          Severity: normal
          Priority: P3
         Component: target
          Assignee: unassigned at gcc dot gnu.org
          Reporter: zsojka at seznam dot cz
  Target Milestone: ---
              Host: x86_64-pc-linux-gnu
            Target: aarch64-unknown-linux-gnu

Created attachment 51177
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=51177&action=edit
reduced testcase

Output:
$ aarch64-unknown-linux-gnu-gcc -Og testcase.c -static
$ qemu-aarch64 -- ./a.out 
qemu: uncaught target signal 6 (Aborted) - core dumped
Aborted

The function just returns 0:

foo:
        mov     w0, 0   //,
        ret

$ aarch64-unknown-linux-gnu-gcc -v
Using built-in specs.
COLLECT_GCC=/repo/gcc-trunk/binary-latest-aarch64/bin/aarch64-unknown-linux-gnu-gcc
COLLECT_LTO_WRAPPER=/repo/gcc-trunk/binary-trunk-r12-2414-20210720154202-g1ef9b135793-checking-yes-rtl-df-extra-aarch64/bin/../libexec/gcc/aarch64-unknown-linux-gnu/12.0.0/lto-wrapper
Target: aarch64-unknown-linux-gnu
Configured with: /repo/gcc-trunk//configure --enable-languages=c,c++
--enable-valgrind-annotations --disable-nls --enable-checking=yes,rtl,df,extra
--with-cloog --with-ppl --with-isl
--with-sysroot=/usr/aarch64-unknown-linux-gnu --build=x86_64-pc-linux-gnu
--host=x86_64-pc-linux-gnu --target=aarch64-unknown-linux-gnu
--with-ld=/usr/bin/aarch64-unknown-linux-gnu-ld
--with-as=/usr/bin/aarch64-unknown-linux-gnu-as --disable-libstdcxx-pch
--prefix=/repo/gcc-trunk//binary-trunk-r12-2414-20210720154202-g1ef9b135793-checking-yes-rtl-df-extra-aarch64
Thread model: posix
Supported LTO compression algorithms: zlib zstd
gcc version 12.0.0 20210720 (experimental) (GCC)

[Bug target/101529] wrong code with __builtin_convertvector() and __builtin_shufflevector() at -Og

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101529

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Last reconfirmed|                            |2021-07-20
             Status|UNCONFIRMED                 |NEW
     Ever confirmed|0                           |1

--- Comment #1 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Confirmed.

-Og:
  _13 = BIT_FIELD_REF <_12, 64, 0>;
  _4 = VEC_PERM_EXPR <_13, { 0, 0, 0, 0, 0, 0, 0, 0 }, { 3, 9, 10, 11, 12, 13,
14, 15 }>;
  _14 = BIT_FIELD_REF <_4, 8, 0>;


vs -O2:
  _14 = BIT_FIELD_REF <_12, 8, 24>;


Huh:

;; _4 = VEC_PERM_EXPR <_13, { 0, 0, 0, 0, 0, 0, 0, 0 }, { 3, 9, 10, 11, 12, 13,
14, 15 }>;


(insn 16 15 17 (set (reg:V8QI 95 [ _4 ])
        (vec_select:V8QI (reg:V16QI 98 [ _12 ])
            (parallel [
                    (const_int 0 [0])
                    (const_int 1 [0x1])
                    (const_int 2 [0x2])
                    (const_int 3 [0x3])
                    (const_int 4 [0x4])
                    (const_int 5 [0x5])
                    (const_int 6 [0x6])
                    (const_int 7 [0x7])
                ]))) "t5.c":11:10 -1
     (nil))

;;; Select the lower part aka _13

(insn 17 16 18 (set (reg:V8QI 112)
        (const_vector:V8QI [
                (const_int 0 [0]) repeated x8
            ])) "t5.c":11:10 -1
     (nil))

(insn 18 17 19 (set (reg:V8QI 95 [ _4 ])
        (reg:V8QI 112)) "t5.c":11:10 -1
     (nil))
;;; BUT reuse 95/_4 for the zero constant!!!!!!!!!!!!!!!

(insn 19 18 0 (set (reg:V8QI 95 [ _4 ])
        (vec_merge:V8QI (vec_duplicate:V8QI (vec_select:QI (reg:V8QI 95 [ _4 ])
                    (parallel [
                            (const_int 3 [0x3])
                        ])))
            (reg:V8QI 95 [ _4 ])
            (const_int 1 [0x1]))) "t5.c":11:10 -1
     (nil))

[Bug target/101529] wrong code with __builtin_convertvector() and __builtin_shufflevector() at -Og

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101529

--- Comment #2 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
*** Bug 101845 has been marked as a duplicate of this bug. ***

[Bug target/101529] [11/12 Regression] wrong code with __builtin_convertvector() and __builtin_shufflevector() at -Og

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101529

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |ASSIGNED
           Assignee|unassigned at gcc dot gnu.org      |pinskia at gcc dot gnu.org
            Summary|wrong code with             |[11/12 Regression] wrong
                   |__builtin_convertvector()   |code with
                   |and                         |__builtin_convertvector()
                   |__builtin_shufflevector()   |and
                   |at -Og                      |__builtin_shufflevector()
                   |                            |at -Og

--- Comment #3 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Mine, a regression from GCC 10 really.
aarch64_evpc_ins is where the problem is.
I have a fix.

[Bug target/101529] [11/12 Regression] wrong code with __builtin_convertvector() and __builtin_shufflevector() at -Og

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101529

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|---                         |11.3

[Bug target/101529] [11/12 Regression] wrong code with __builtin_convertvector() and __builtin_shufflevector() at -Og

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101529

--- Comment #4 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Patch which I am testing:
diff --git a/gcc/config/aarch64/aarch64.c b/gcc/config/aarch64/aarch64.c
index 2c00583e12c..d5b0cb7ae2f 100644
--- a/gcc/config/aarch64/aarch64.c
+++ b/gcc/config/aarch64/aarch64.c
@@ -23084,11 +23084,15 @@ aarch64_evpc_ins (struct expand_vec_perm_d *d)
     }
   gcc_assert (extractindex < nelt);

-  emit_move_insn (d->target, insv);
+  /* Use a new reg instead of target as one of the
+     operands might be target. */
+  rtx original = gen_reg_rtx (GET_MODE (insv));
+
+  emit_move_insn (original, insv);
   insn_code icode = code_for_aarch64_simd_vec_copy_lane (mode);
   expand_operand ops[5];
   create_output_operand (&ops[0], d->target, mode);
-  create_input_operand (&ops[1], d->target, mode);
+  create_input_operand (&ops[1], original, mode);
   create_integer_operand (&ops[2], 1 << idx);
   create_input_operand (&ops[3], extractv, mode);
   create_integer_operand (&ops[4], extractindex);

[Bug target/101529] [11/12 Regression] wrong code with __builtin_convertvector() and __builtin_shufflevector() at -Og caused by r11-2192

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101529

--- Comment #5 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
(In reply to Andrew Pinski from comment #4)
> Patch which I am testing:

My original patch before Dmitrij took it over, actually had the correct
behavior and the used a gen_reg_rtx. :(

[Bug target/101529] [11/12 Regression] wrong code with __builtin_convertvector() and __builtin_shufflevector() at -Og caused by r11-2192

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101529

--- Comment #6 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Created attachment 51741
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=51741&action=edit
patch which is testing including testcases

Full patch which is in testing right now. Added the testcases for this bug and
the dup bug too.

[Bug target/101529] [11/12 Regression] wrong code with __builtin_convertvector() and __builtin_shufflevector() at -Og caused by r11-2192

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101529

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                URL|                            |https://gcc.gnu.org/piperma
                   |                            |il/gcc-patches/2021-Novembe
                   |                            |r/583584.html
           Keywords|                            |patch

--- Comment #7 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Patch submitted:
https://gcc.gnu.org/pipermail/gcc-patches/2021-November/583584.html

[Bug target/101529] [11/12 Regression] wrong code with __builtin_convertvector() and __builtin_shufflevector() at -Og caused by r11-2192

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101529

--- Comment #8 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The trunk branch has been updated by Andrew Pinski <pinskia@gcc.gnu.org>:

https://gcc.gnu.org/g:52fa771758635d9c53cddb9116e5a66fae592230

commit r12-5078-g52fa771758635d9c53cddb9116e5a66fae592230
Author: Andrew Pinski <apinski@marvell.com>
Date:   Sat Nov 6 06:29:13 2021 +0000

    aarch64: [PR101529] Fix vector shuffle insertion expansion

    The function aarch64_evpc_ins would reuse the target even though
    it might be the same register as the two inputs.
    Instead of checking to see if we can reuse the target, just use the
    original input directly.

    Committed as approved after bootstrapped and tested on
    aarch64-linux-gnu with no regressions.

            PR target/101529

    gcc/ChangeLog:

            * config/aarch64/aarch64.c (aarch64_evpc_ins): Don't use target
            as an input, use original one.

    gcc/testsuite/ChangeLog:

            * c-c++-common/torture/builtin-convertvector-2.c: New test.
            * c-c++-common/torture/builtin-shufflevector-2.c: New test.

[Bug target/101529] [11 Regression] wrong code with __builtin_convertvector() and __builtin_shufflevector() at -Og caused by r11-2192

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101529

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
      Known to work|                            |12.0
            Summary|[11/12 Regression] wrong    |[11 Regression] wrong code
                   |code with                   |with
                   |__builtin_convertvector()   |__builtin_convertvector()
                   |and                         |and
                   |__builtin_shufflevector()   |__builtin_shufflevector()
                   |at -Og caused by r11-2192   |at -Og caused by r11-2192
      Known to fail|12.0                        |

--- Comment #9 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Fixed on the trunk, Will wait a week or so to backport.

[Bug target/101529] [11 Regression] wrong code with __builtin_convertvector() and __builtin_shufflevector() at -Og caused by r11-2192

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101529

--- Comment #10 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The releases/gcc-11 branch has been updated by Andrew Pinski
<pinskia@gcc.gnu.org>:

https://gcc.gnu.org/g:193e5ee46e5a4f0650dc1ccbdaec5d2e84f2a451

commit r11-9515-g193e5ee46e5a4f0650dc1ccbdaec5d2e84f2a451
Author: Andrew Pinski <apinski@marvell.com>
Date:   Sat Nov 6 06:29:13 2021 +0000

    aarch64: [PR101529] Fix vector shuffle insertion expansion

    The function aarch64_evpc_ins would reuse the target even though
    it might be the same register as the two inputs.
    Instead of checking to see if we can reuse the target, just use the
    original input directly.

    Committed as approved after bootstrapped and tested on
    aarch64-linux-gnu with no regressions.
    Note the testcases are not backported as __builtin_shufflevector
    does not exist in GCC 11.

            PR target/101529

    gcc/ChangeLog:

            * config/aarch64/aarch64.c (aarch64_evpc_ins): Don't use target
            as an input, use original one.

    (cherry picked from commit 52fa771758635d9c53cddb9116e5a66fae592230)

[Bug target/101529] [11 Regression] wrong code with __builtin_convertvector() and __builtin_shufflevector() at -Og caused by r11-2192

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101529

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|ASSIGNED                    |RESOLVED
         Resolution|---                         |FIXED

--- Comment #11 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Fixed.