[Bug target/109601] New: Useless branch not eliminated when writing to a union

[Bug target/109601] New: Useless branch not eliminated when writing to a union

[67 at mniip dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109601

            Bug ID: 109601
           Summary: Useless branch not eliminated when writing to a union
           Product: gcc
           Version: 12.2.1
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
          Assignee: unassigned at gcc dot gnu.org
          Reporter: 67 at mniip dot com
  Target Milestone: ---

Created attachment 54909
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=54909&action=edit
Minimal testcase

On x86_64-pc-linux-gnu the attached code compiles (with -O3) to rather
unsatisfying assembly with a branch that does the same thing whether taken or
not:

>   10:	85 c0                	test   %eax,%eax
>   12:	74 04                	je     18
>   14:	c3                   	ret
>   15:	0f 1f 00             	nopl   (%rax)
>   18:	c3                   	ret

I've tested on my machine in GCC 12.2.1, but Compiler Explorer (
https://godbolt.org/z/cW69rfrPc ) shows the same thing happens in a lot of
other versions, including trunk (6ab856aa49bef7c04efa6144a5048e129b3a058b).

Notably, this doesn't happen in GCC 4.5.3.

This seems to happen for a few other targets as well, e.g: arm, riscv32; but
not on some others, e.g. aarch64, mips.

[Bug rtl-optimization/109601] Useless branch not eliminated when writing to a union

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109601

--- Comment #1 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
jump2 (which is after RA) does this for aarch64:

Forwarding edge 2->3 to 5 failed.
verify found no changes in insn with uid = 11.
Edge 2->4 redirected to 5
scanning new insn with uid = 47.
Replacing insn 11 by jump 47
deleting insn with uid = 11.
deleting insn with uid = 38.
deleting block 3
deleting block 4
Removing jump 47.
deleting insn with uid = 47.
Merging block 5 into block 2...
changing bb of uid 43
changing bb of uid 44
  from 5 to 2
changing bb of uid 45
  from 5 to 2
Merged blocks 2 and 5.
Merged 2 and 5 without moving.

For x86_64 does not have this.

[Bug rtl-optimization/109601] Useless branch not eliminated when writing to a union

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109601

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Last reconfirmed|                            |2023-04-23
             Status|UNCONFIRMED                 |NEW
     Ever confirmed|0                           |1

--- Comment #2 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
x86 duplicated the simple return but never unduplicates it:
(jump_insn 27 12 28 3 (simple_return) 1030 {simple_return_internal}
     (nil)
 -> simple_return)
...

(jump_insn 32 31 33 4 (simple_return) "/app/example.cpp":29:1 -1
     (nil)
 -> simple_return)


But simple_return was only added in GCC 4.7.0 (r0-111280-g268987713c7105c691)
So something else changed.

There is also a missed optimization earlier because jump2 is way late to be
optimizing this case.

Also this is just an issue with return so it might not be as important as most
people think (the other missed optimization earlier is though).