[Bug c++/100322] New: Switching from std=c++17 to std=c++20 causes performance regression in relationals

[Bug c++/100322] New: Switching from std=c++17 to std=c++20 causes performance regression in relationals

[aaron at aarongraham dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100322

            Bug ID: 100322
           Summary: Switching from std=c++17 to std=c++20 causes
                    performance regression in relationals
           Product: gcc
           Version: 12.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c++
          Assignee: unassigned at gcc dot gnu.org
          Reporter: aaron at aarongraham dot com
  Target Milestone: ---

Experiment here: https://godbolt.org/z/PT73cn5e5

#include <chrono>

using clk = std::chrono::steady_clock;
bool compare_count(clk::duration a, clk::duration b) {
    return a.count() > b.count();
}
bool compare(clk::duration a, clk::duration b) {
    return a > b;
}

Compiling with -std=c++17 I get:

_Z13compare_countNSt6chrono8durationIxSt5ratioILx1ELx1000000000EEEES3_:
        cmp     r2, r0
        sbcs    r3, r3, r1
        ite     lt
        movlt   r0, #1
        movge   r0, #0
        bx      lr
_Z7compareNSt6chrono8durationIxSt5ratioILx1ELx1000000000EEEES3_:
        cmp     r2, r0
        sbcs    r3, r3, r1
        ite     lt
        movlt   r0, #1
        movge   r0, #0
        bx      lr

Compiling with -std=c++20 I get:

_Z13compare_countNSt6chrono8durationIxSt5ratioILx1ELx1000000000EEEES3_:
        cmp     r2, r0
        sbcs    r3, r3, r1
        ite     lt
        movlt   r0, #1
        movge   r0, #0
        bx      lr
_Z7compareNSt6chrono8durationIxSt5ratioILx1ELx1000000000EEEES3_:
        cmp     r1, r3
        it      eq
        cmpeq   r0, r2
        beq     .L4
        cmp     r0, r2
        sbcs    r3, r1, r3
        bge     .L5
        mov     r0, #-1
.L3:
        cmp     r0, #0
        ite     le
        movle   r0, #0
        movgt   r0, #1
        bx      lr
.L4:
        movs    r0, #0
        b       .L3
.L5:
        movs    r0, #1
        b       .L3

(Note that clang doesn't have this problem)

[Bug c++/100322] Switching from std=c++17 to std=c++20 causes performance regression in relationals

[redi at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100322

--- Comment #1 from Jonathan Wakely <redi at gcc dot gnu.org> ---
Maybe related to either PR 94006 or PR 94566

[Bug c++/100322] Switching from std=c++17 to std=c++20 causes performance regression in relationals

[redi at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100322

Jonathan Wakely <redi at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2021-04-28
     Ever confirmed|0                           |1

--- Comment #2 from Jonathan Wakely <redi at gcc dot gnu.org> ---
While it's not the cause of the bad codegen for the <=> comparisons, this test
shouldn't even be using <=>. The problem is that the synthesized operator> that
is based on <=> is a better match for non-const durations than the one that
would generate good code:

If you write it like this you get good codegen:

bool compare(const clk::duration& a, const clk::duration& b)
{
    return a > b;
}

In fact you don't even need to pass them by reference, just:

bool compare(const clk::duration a, const clk::duration b)
{
    return a > b;
}

This seems like a defect in the std::lib, but it's separate from the poor
codegen for the <=> comparison.

[Bug c++/100322] Switching from std=c++17 to std=c++20 causes performance regression in relationals

[redi at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100322

--- Comment #3 from Jonathan Wakely <redi at gcc dot gnu.org> ---
(In reply to Jonathan Wakely from comment #2)
> If you write it like this you get good codegen:

I think I messed up my testing, and it doesn't help. GCC always chooses the
synthesized operator> instead of the intended one, so always produces poor
code.

[Bug c++/100322] Switching from std=c++17 to std=c++20 causes performance regression in relationals

[barry.revzin at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100322

Barry Revzin <barry.revzin at gmail dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |barry.revzin at gmail dot com

--- Comment #4 from Barry Revzin <barry.revzin at gmail dot com> ---
Reduced to avoid chrono: https://godbolt.org/z/o96oz9d6K

[Bug c++/100322] Switching from std=c++17 to std=c++20 causes performance regression in relationals

[barry.revzin at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100322

--- Comment #5 from Barry Revzin <barry.revzin at gmail dot com> ---
Sorry meant to actually copy the reduction:

#include <compare> 

bool compare_count(int a, int b)
{
    return a > b;
}

bool compare(int a, int b)
{
    return (a <=> b) > 0;
}

which generates:

compare_count(int, int):
        cmp     r0, r1
        ite     le
        movle   r0, #0
        movgt   r0, #1
        bx      lr
compare(int, int):
        cmp     r0, r1
        beq     .L4
        blt     .L5
        movs    r0, #1
.L3:
        cmp     r0, #0
        ite     le
        movle   r0, #0
        movgt   r0, #1
        bx      lr
.L4:
        movs    r0, #0
        b       .L3
.L5:
        mov     r0, #-1
        b       .L3

[Bug c++/100322] Switching from std=c++17 to std=c++20 causes performance regression in relationals

[redi at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100322

--- Comment #6 from Jonathan Wakely <redi at gcc dot gnu.org> ---
So it's just bad codegen for <=> comparisons. Thanks, Barry.

The library part is a red herring, the operator> synthesized from <=> is fine
to use, it just produces bad code (with GCC, but not Clang).

[Bug c++/100322] Switching from std=c++17 to std=c++20 causes performance regression in relationals

[glisse at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100322

--- Comment #7 from Marc Glisse <glisse at gcc dot gnu.org> ---
PR94589 then.

[Bug c++/100322] Switching from std=c++17 to std=c++20 causes performance regression in relationals

[redi at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100322

Jonathan Wakely <redi at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
         Resolution|---                         |DUPLICATE
             Status|NEW                         |RESOLVED

--- Comment #8 from Jonathan Wakely <redi at gcc dot gnu.org> ---
Ah yes, thanks, Marc.

*** This bug has been marked as a duplicate of bug 94589 ***