[Bug target/112304] New: cinc is not being used for (small) constant

[Bug target/112304] New: cinc is not being used for (small) constant

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112304

            Bug ID: 112304
           Summary: cinc is not being used for (small) constant
           Product: gcc
           Version: 14.0
            Status: UNCONFIRMED
          Keywords: missed-optimization
          Severity: enhancement
          Priority: P3
         Component: target
          Assignee: pinskia at gcc dot gnu.org
          Reporter: pinskia at gcc dot gnu.org
  Target Milestone: ---
            Target: aarch64-linux-gnu

Take:
```
int f(int a)
{
        return (a!=0)+42;
}
int f1(int a, int b)
{
        return (a!=0)+b;
}
```
Currently we get:
```
f:
        cmp     w0, 0
        cset    w0, ne
        add     w0, w0, 42
        ret
f1:
        cmp     w0, 0
        cinc    w0, w1, ne
        ret
```

But f should really be:
```
        mov     w1, #42
        cmp     w0, #0
        cinc    w0, w1, ne
        ret
```

Because for many newer aarch64 cores the `mov` in this case is "free" and does
not take up an issue slot or even the for a dual issue, the mov and cmp could
be issued together.  Even though it is the same # of instruction, the second
case is going to be faster due to latencies.

[Bug target/112304] cinc is not being used for (small) constant

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112304

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
     Ever confirmed|0                           |1
   Last reconfirmed|                            |2023-10-31
             Status|UNCONFIRMED                 |ASSIGNED

--- Comment #1 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Looking into this one next week.

[Bug target/112304] cinc is not being used for (small) constant

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112304

--- Comment #2 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Trying 35 -> 16:
   35: r95:SI=cc:CC!=0
      REG_DEAD cc:CC
   16: x0:SI=r95:SI+0x2a
      REG_DEAD r95:SI
Failed to match this instruction:
(set (reg/i:SI 0 x0)
    (plus:SI (ne:SI (reg:CC 66 cc)
            (const_int 0 [0]))
        (const_int 42 [0x2a])))


vs:


Trying 9 -> 14:
    9: r97:SI=cc:CC!=0+r101:SI
      REG_DEAD cc:CC
      REG_DEAD r101:SI
   14: x0:SI=r97:SI
      REG_DEAD r97:SI
Successfully matched this instruction:
(set (reg/i:SI 0 x0)
    (plus:SI (ne:SI (reg:CC 66 cc)
            (const_int 0 [0]))
        (reg:SI 101)))

[Bug target/112304] cinc is not being used for (small) constant

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112304

--- Comment #3 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Most like the following pattern's predicates should be expanded to include
constants and not just registers.

(define_insn "*csinc2<mode>_insn"
  [(set (match_operand:GPI 0 "register_operand" "=r")
        (plus:GPI (match_operand 2 "aarch64_comparison_operation" "")
                  (match_operand:GPI 1 "register_operand" "r")))]


s/register_operand/general_operand/ here For the second predicate.

And maybe a few others.

It is similar what I just did for `*cmov<mode>_insn_insv` pattern.

[Bug target/112304] cinc is not being used for (small) constant

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112304

--- Comment #4 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Note my patch does not change the fact that we don't pull the mov outside of
the loop for:
```
void f(int *a, int l)
{
  for(int i = 0; i < l; i++)
        a[i]=(a[i]!=0)+42;
}
```
But that is for a different issue to fix. The issue there is we have a cset/add
during the (RTL) LIM and we don't pull the constant out from the add. And then
we go and do the combine and found it is an cinc and then 42 becomes part of
the cinc. This is pre-existing issue and most likely could be reproduce with
other testcases too.

[Bug target/112304] cinc is not being used for (small) constant

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112304

--- Comment #5 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Created attachment 58031
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=58031&action=edit
Patch which I am testing

Once testing is finished and GCC 15 stage 1 opens up, I will submit this.

[Bug target/112304] cinc is not being used for (small) constant

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112304

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |patch
                URL|                            |https://gcc.gnu.org/piperma
                   |                            |il/gcc-patches/2024-April/6
                   |                            |50081.html

--- Comment #6 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Patch posted:
https://gcc.gnu.org/pipermail/gcc-patches/2024-April/650081.html

[Bug target/112304] cinc is not being used for (small) constant

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112304

--- Comment #7 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
(In reply to Andrew Pinski from comment #6)
> Patch posted:
> https://gcc.gnu.org/pipermail/gcc-patches/2024-April/650081.html

Maybe it should be nonmemory_operand rather than general_operand ...

Or aarch64_reg_or_imm