[Bug tree-optimization/113434] New: [13/14 Regression] Missed optimization for Loop Unswitch

[Bug tree-optimization/113434] New: [13/14 Regression] Missed optimization for Loop Unswitch

[652023330028 at smail dot nju.edu.cn]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113434

            Bug ID: 113434
           Summary: [13/14 Regression] Missed optimization for Loop
                    Unswitch
           Product: gcc
           Version: 14.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: 652023330028 at smail dot nju.edu.cn
  Target Milestone: ---

Hello, we noticed that maybe there is a missed optimization for Loop Unswitch.

In the following code, if(a) is actually always true.

https://godbolt.org/z/rr5M77sEG

int a, b;
void test() {
    a = 0;
    for (int c = 0; c < 100; c += 1) {
        a += 2;
        if (a)
            b += 3;
    }
}

But GCC (trunk) -O3 -fwrapv:
test():
        mov     edx, DWORD PTR b[rip]
        mov     eax, 2
.L3:
        test    eax, eax
        jne     .L2
        mov     eax, 2
.L2:
        add     eax, 2
        add     edx, 3
        cmp     eax, 202
        jne     .L3
        mov     DWORD PTR a[rip], 200
        mov     DWORD PTR b[rip], edx
        ret

Expected code (GCC-12.3 -O3 -fwrapv):
test():
        mov     edx, DWORD PTR b[rip]
        xor     eax, eax
.L2:
        add     eax, 2
        add     edx, 3
        cmp     eax, 200
        jne     .L2
        mov     DWORD PTR a[rip], 200
        mov     DWORD PTR b[rip], edx
        ret

Thank you very much for your time and effort! We look forward to hearing from
you.

[Bug tree-optimization/113434] [13/14 Regression] VRP misses conditional in loop is always false sometimes

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113434

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Target Milestone|---                         |13.3
   Last reconfirmed|                            |2024-01-17
     Ever confirmed|0                           |1
            Summary|[13/14 Regression] Missed   |[13/14 Regression] VRP
                   |optimization for Loop       |misses conditional in loop
                   |Unswitch                    |is always false sometimes

--- Comment #1 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Confirmed. Note it is not loop switching but rather VRP which should have
figured out a bounds are [2,202] inside the loop but for some reason it is not
able to.

[Bug tree-optimization/113434] [13/14 Regression] VRP misses conditional in loop is always false sometimes

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113434

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |rguenth at gcc dot gnu.org
           Priority|P3                          |P2

--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
It needs LIM (store-motion) to make 'a' a register inside the loop but that
runs after VRP1, the next VRP only runs after SCCP and also IVOPTs which
confuses it.  With -fno-ivopts it's OK but still too late for SCCP.

Thus, a pass ordering issue between LIM / VRP and a missed VRP after IVOPTs
obfuscation.

We might want to consider doing store-motion even earlier ...

[Bug tree-optimization/113434] [13/14/15 Regression] VRP misses conditional in loop is always false sometimes

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113434

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|13.3                        |13.4

--- Comment #3 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
GCC 13.3 is being released, retargeting bugs to GCC 13.4.