[Bug target/58110] New: Useless GPR push and pop when only xmm registers are used.

[Bug target/58110] New: Useless GPR push and pop when only xmm registers are used.

[neleai at seznam dot cz]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=58110

            Bug ID: 58110
           Summary: Useless GPR push and pop when only xmm registers are
                    used.
           Product: gcc
           Version: 4.9.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
          Assignee: unassigned at gcc dot gnu.org
          Reporter: neleai at seznam dot cz

Hi,attached code generates extra push/pop rbx pair while there is no gpr
register assigned in segment between them.
This was generated by head xgcc -O3. A gcc-4.7 has similar issues with attached
program.

  pushq   %rbx
        .cfi_def_cfa_offset 16
        .cfi_offset 3, -16
        movdqu  (%rsi), %xmm15
        movdqu  -16(%rsi,%rdx), %xmm14
        movdqu  16(%rsi), %xmm13
        movdqu  -32(%rsi,%rdx), %xmm12
        movdqu  32(%rsi), %xmm11
        movdqu  -48(%rsi,%rdx), %xmm10
        movdqu  48(%rsi), %xmm9
        movdqu  -64(%rsi,%rdx), %xmm8
        movdqu  64(%rsi), %xmm7
        movdqu  -80(%rsi,%rdx), %xmm6
        movdqu  80(%rsi), %xmm5
        movdqu  -96(%rsi,%rdx), %xmm4
        movdqu  96(%rsi), %xmm3
        movdqu  -112(%rsi,%rdx), %xmm2
        movdqu  112(%rsi), %xmm1
        movdqu  -128(%rsi,%rdx), %xmm0
        movdqu  %xmm15, (%rdi)
        movdqu  %xmm14, -16(%rdi,%rdx)
        movdqu  %xmm13, 16(%rdi)
        movdqu  %xmm12, -32(%rdi,%rdx)
        movdqu  %xmm11, 32(%rdi)
        movdqu  %xmm10, -48(%rdi,%rdx)
        movdqu  %xmm9, 48(%rdi)
        movdqu  %xmm8, -64(%rdi,%rdx)
        movdqu  %xmm7, 64(%rdi)
        movdqu  %xmm6, -80(%rdi,%rdx)
        movdqu  %xmm5, 80(%rdi)
        movdqu  %xmm4, -96(%rdi,%rdx)
        movdqu  %xmm3, 96(%rdi)
        movdqu  %xmm2, -112(%rdi,%rdx)
        movdqu  %xmm1, 112(%rdi)
        movdqu  %xmm0, -128(%rdi,%rdx)
        popq    %rbx

[Bug target/58110] Useless GPR push and pop when only xmm registers are used.

[neleai at seznam dot cz]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=58110

--- Comment #1 from Ondrej Bilka <neleai at seznam dot cz> ---
Created attachment 30627
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=30627&action=edit
testcase

[Bug target/58110] Useless GPR push and pop when only xmm registers are used.

[hubicka at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=58110

Jan Hubicka <hubicka at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2013-08-09
                 CC|                            |hubicka at gcc dot gnu.org,
                   |                            |vmakarov at redhat dot com
     Ever confirmed|0                           |1
           Severity|normal                      |enhancement

--- Comment #2 from Jan Hubicka <hubicka at gcc dot gnu.org> ---
Here BX is used after reload for:

(insn 118 116 119 9 (parallel [
            (set (reg:DI 3 bx [orig:111 D.4408 ] [111])
                (plus:DI (reg/v:DI 1 dx [orig:242 n ] [242])
                    (const_int -16 [0xfffffffffffffff0])))
            (clobber (reg:CC 17 flags))
        ]) memcpy.c:96 249 {*adddi_1}
     (nil))

...

(insn 120 119 122 9 (set (reg:V16QI 51 xmm14 [281])
        (unspec:V16QI [
                (mem:V16QI (plus:DI (reg/v/f:DI 4 si [orig:241 src ] [241])
                        (reg:DI 3 bx [orig:111 D.4408 ] [111])) [0 S16 A8])
            ] UNSPEC_LOADU))
/usr/lib64/gcc/x86_64-suse-linux/4.7/include/emmintrin.h:685 1107
{sse2_loaddqu}
     (nil))
...
(insn 175 173 178 9 (set (mem:V16QI (plus:DI (reg/v/f:DI 0 ax [orig:240 dest ]
[240])
                (reg:DI 3 bx [orig:111 D.4408 ] [111])) [0 S16 A8])
        (unspec:V16QI [
                (reg:V16QI 51 xmm14 [281])
            ] UNSPEC_STOREU))
/usr/lib64/gcc/x86_64-suse-linux/4.7/include/emmintrin.h:703 1109
{sse2_storedqu}
     (nil))
....

and this gets updated by postrealod to

(insn 120 119 122 9 (set (reg:V16QI 51 xmm14 [281])
        (unspec:V16QI [
                (mem:V16QI (plus:DI (plus:DI (reg/v/f:DI 4 si [orig:241 src ]
[241])
                            (reg/v:DI 1 dx [orig:242 n ] [242]))
                        (const_int -16 [0xfffffffffffffff0])) [0 S16 A8])
            ] UNSPEC_LOADU))
/usr/lib64/gcc/x86_64-suse-linux/4.7/include/emmintrin.h:685 1107
{sse2_loaddqu}
     (nil))

So RBX become unused.

Vladimir, perhaps you can take a look?

[Bug target/58110] Useless GPR push and pop when only xmm registers are used.

[vmakarov at redhat dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=58110

--- Comment #3 from Vladimir Makarov <vmakarov at redhat dot com> ---
Thanks, Ondrej and Jan.  GCC with reload generates code with the same problem.

I mentioned on RA BOF that we should look at postreload.c and postreload-gcse.c
to figure out what should and can be removed as redundant and what can be
integrated with IRA/LRA.  This PR is just a good illustration of why it should
be done.

I don't think this work will be done soon but it is good to have the PR to
remember this.

[Bug target/58110] Useless GPR push and pop when only xmm registers are used.

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=58110

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
         Resolution|---                         |FIXED
             Status|NEW                         |RESOLVED
   Target Milestone|---                         |7.0

--- Comment #4 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Fixed by LRA in GCC 7.