[Bug c++/115351] New: [14 regression] pointless movs when passing by value on x86-64

[Bug c++/115351] New: [14 regression] pointless movs when passing by value on x86-64

[michael.kenzel at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115351

            Bug ID: 115351
           Summary: [14 regression] pointless movs when passing by value
                    on x86-64
           Product: gcc
           Version: 14.1.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c++
          Assignee: unassigned at gcc dot gnu.org
          Reporter: michael.kenzel at gmail dot com
  Target Milestone: ---

When passing structs of certain shape by value on x86-64 (seems to be specific
to this target), gcc 14 emits (fails to optimize away?) a number of redundant
moves:

        struct complex
        {
                double real;
                double imag;
        };

        complex blub(complex z)
        {
                return {
                        z.real * z.real - z.imag * z.imag,
                        2 * z.real * z.imag
                };
        }

on gcc 13 with -O3 results in

        blub(complex):
                movapd  xmm3, xmm1
                movapd  xmm2, xmm0
                mulsd   xmm3, xmm1
                addsd   xmm2, xmm2
                mulsd   xmm0, xmm0
                mulsd   xmm1, xmm2
                subsd   xmm0, xmm3
                ret

gcc 14 and later with -O3 emit

        blub(complex):
                        movq    rax, xmm1
                        movq    rdx, xmm0
                        xchg    rdx, rax
                        movq    xmm2, rax
                        movq    xmm1, rdx
                        movapd  xmm0, xmm2
                        movapd  xmm3, xmm1
                        mulsd   xmm3, xmm1
                        mulsd   xmm0, xmm2
                        addsd   xmm2, xmm2
                        mulsd   xmm1, xmm2
                        subsd   xmm0, xmm3
                        ret

godbolt: https://godbolt.org/z/hzEfs3ob4

[Bug rtl-optimization/115351] [14/15 regression] pointless movs when passing by value on x86-64

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115351

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Target|                            |x86_64-*-*
            Summary|[14 regression] pointless   |[14/15 regression]
                   |movs when passing by value  |pointless movs when passing
                   |on x86-64                   |by value on x86-64
   Target Milestone|---                         |14.2
     Ever confirmed|0                           |1
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2024-06-05
          Component|c++                         |rtl-optimization
           Keywords|                            |missed-optimization,
                   |                            |needs-bisection

--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
Confirmed.  The IL we expand from is the same.

[Bug rtl-optimization/115351] [14/15 regression] pointless movs when passing by value on x86-64

[liuhongt at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115351

Hongtao Liu <liuhongt at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |liuhongt at gcc dot gnu.org

--- Comment #2 from Hongtao Liu <liuhongt at gcc dot gnu.org> ---
There're 

(insn 5 4 6 2 (set (reg:TI 110)
        (ior:TI (and:TI (reg:TI 110)
                (const_wide_int 0xffffffffffffffff0000000000000000))
            (zero_extend:TI (subreg:DI (reg:DF 111) 0))))
"/app/example.cpp":8:1 136 {*insvti_lowpart_1}
     (nil))
(insn 6 5 7 2 (set (reg:TI 110)
        (ior:TI (and:TI (reg:TI 110)
                (const_wide_int 0x0ffffffffffffffff))
            (ashift:TI (zero_extend:TI (subreg:DI (reg:DF 112) 0))
                (const_int 64 [0x40])))) "/app/example.cpp":8:1 133
{*insvti_highpart_1}
     (nil))
(insn 7 6 8 2 (set (reg/v:TI 109 [ z ])

in GCC14's rtl dump, guess related to r14-589-g1e3054d27c83ee?

[Bug target/115351] [14/15 regression] pointless movs when passing by value on x86-64

[roger at nextmovesoftware dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115351

Roger Sayle <roger at nextmovesoftware dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Assignee|unassigned at gcc dot gnu.org      |roger at nextmovesoftware dot com

--- Comment #3 from Roger Sayle <roger at nextmovesoftware dot com> ---
I have a fix (to ix86_rtx_costs) that I'm bootstrapping and regression testing.
By making concatditi3 slightly cheaper than *insvti_highpart (COSTS_N_INSN(2)
vs. COSTS_N_INSNS(2)+1), fwprop is able to do the right.

[Bug target/115351] [14/15 regression] pointless movs when passing by value on x86-64

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115351

--- Comment #4 from GCC Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Roger Sayle <sayle@gcc.gnu.org>:

https://gcc.gnu.org/g:fb3e4c549d16d5050e10114439ad77149f33c597

commit r15-1101-gfb3e4c549d16d5050e10114439ad77149f33c597
Author: Roger Sayle <roger@nextmovesoftware.com>
Date:   Fri Jun 7 14:03:20 2024 +0100

    i386: PR target/115351: RTX costs for *concatditi3 and *insvti_highpart.

    This patch addresses PR target/115351, which is a code quality regression
    on x86 when passing floating point complex numbers.  The ABI considers
    these arguments to have TImode, requiring interunit moves to place the
    FP values (which are actually passed in SSE registers) into the upper
    and lower parts of a TImode pseudo, and then similar moves back again
    before they can be used.

    The cause of the regression is that changes in how TImode initialization
    is represented in RTL now prevents the RTL optimizers from eliminating
    these redundant moves.  The specific cause is that the *concatditi3
    pattern, (zext(hi)<<64)|zext(lo), has an inappropriately high (default)
    rtx_cost, preventing fwprop1 from propagating it.  This pattern just
    sets the hipart and lopart of a double-word register, typically two
    instructions (less if reload can allocate things appropriately) but
    the current ix86_rtx_costs actually returns INSN_COSTS(13), i.e. 52.

    propagating insn 5 into insn 6, replacing:
    (set (reg:TI 110)
        (ior:TI (and:TI (reg:TI 110)
                (const_wide_int 0x0ffffffffffffffff))
            (ashift:TI (zero_extend:TI (subreg:DI (reg:DF 112 [ zD.2796+8 ])
0))
                (const_int 64 [0x40]))))
    successfully matched this instruction to *concatditi3_3:
    (set (reg:TI 110)
        (ior:TI (ashift:TI (zero_extend:TI (subreg:DI (reg:DF 112 [ zD.2796+8
]) 0))
                (const_int 64 [0x40]))
            (zero_extend:TI (subreg:DI (reg:DF 111 [ zD.2796 ]) 0))))
    change not profitable (cost 50 -> cost 52)

    This issue is resolved by having ix86_rtx_costs return more reasonable
    values for these (place-holder) patterns.

    2024-06-07  Roger Sayle  <roger@nextmovesoftware.com>

    gcc/ChangeLog
            PR target/115351
            * config/i386/i386.cc (ix86_rtx_costs): Provide estimates for
            the *concatditi3 and *insvti_highpart patterns, about two insns.

    gcc/testsuite/ChangeLog
            PR target/115351
            * g++.target/i386/pr115351.C: New test case.

[Bug target/115351] [14 regression] pointless movs when passing by value on x86-64

[roger at nextmovesoftware dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115351

Roger Sayle <roger at nextmovesoftware dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Assignee|roger at nextmovesoftware dot com  |unassigned at gcc dot gnu.org
            Summary|[14/15 regression]          |[14 regression] pointless
                   |pointless movs when passing |movs when passing by value
                   |by value on x86-64          |on x86-64
      Known to work|                            |15.0

--- Comment #5 from Roger Sayle <roger at nextmovesoftware dot com> ---
This has been fixed on mainline (for GCC 15).