[Bug rtl-optimization/103386] New: stage1 with PGO produces bad offsets in rtl-reload on aarch64

[Bug rtl-optimization/103386] New: stage1 with PGO produces bad offsets in rtl-reload on aarch64

[mittorn at sibmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103386

            Bug ID: 103386
           Summary: stage1 with PGO produces bad offsets in rtl-reload on
                    aarch64
           Product: gcc
           Version: 11.2.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: rtl-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: mittorn at sibmail dot com
  Target Milestone: ---

Created attachment 51862
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=51862&action=edit
pgo build logs and system environment

This result in failed conftest binary returning nonzero status when trying to
build libgomp.
Simple program to reproduce conftest behaviour:

int main()
{
return 0;
}

stage1 makes bad assembly on it:

        .cfi_startproc
        stp     x0, x1, [sp, -32]!
        .cfi_def_cfa_offset 32
        .cfi_offset 0, -32
        .cfi_offset 1, -24
        stp     x2, x3, [sp, 16]
        .cfi_offset 2, -16
        .cfi_offset 3, -8
        mov     w0, 0
        ldp     x2, x3, [sp, 16]
        ldp     x0, x1, [sp], 32
        .cfi_restore 1
        .cfi_restore 0
        .cfi_restore 2
        .cfi_restore 3
        .cfi_def_cfa_offset 0
        ret
        .cfi_endproc

instead of this one:
        .cfi_startproc
        mov     w0, 0
        ret
        .cfi_endproc


after dumping RTL stages, i found this (source files are identical, second is
stage1):

 ~ # diff 1.c.292r.reload 2.c.292r.reload
7a8,12
> New elimination table:
> Can eliminate 65 to 31 (offset=32, prev_offset=0)
> Can eliminate 65 to 29 (offset=32, prev_offset=0)
> Can eliminate 64 to 31 (offset=32, prev_offset=0)
> Can eliminate 64 to 29 (offset=32, prev_offset=0)
51,54c56,59
< Can eliminate 65 to 31 (offset=0, prev_offset=0)
< Can eliminate 65 to 29 (offset=0, prev_offset=0)
< Can eliminate 64 to 31 (offset=0, prev_offset=0)
< Can eliminate 64 to 29 (offset=0, prev_offset=0)
---
> Can eliminate 65 to 31 (offset=32, prev_offset=32)
> Can eliminate 65 to 29 (offset=32, prev_offset=0)
> Can eliminate 64 to 31 (offset=32, prev_offset=32)
> Can eliminate 64 to 29 (offset=32, prev_offset=0)
81c86
<         (const_int 0 [0])) "1.c":3:8 52 {*movsi_aarch64}
---
>         (const_int 0 [0])) "2.c":3:8 52 {*movsi_aarch64}
83c88
< (insn 13 5 15 2 (use (reg/i:SI 0 x0)) "1.c":4:1 -1
---
> (insn 13 5 15 2 (use (reg/i:SI 0 x0)) "2.c":4:1 -1


Full build tree temporary placed here:
http://mittorn.mentality.rip/
needed log and environment info in attachment

[Bug middle-end/103386] stage1 with PGO produces bad offsets in rtl-reload on aarch64

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103386

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |WAITING
     Ever confirmed|0                           |1
   Last reconfirmed|                            |2021-11-23
          Component|rtl-optimization            |middle-end
           Keywords|                            |wrong-code

--- Comment #1 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
What GCC are you starting with?
Because you are building stage 1 compiler with "-O3 -ftree-vectorize
-fomit-frame-pointer -mcpu=neoverse-n1 -mtune=neoverse-n1 -fno-plt -fipa-pta
-funsafe-math-optimizations -funsafe-loop-optimizations
-fdevirtualize-at-ltrans -floop-nest-optimize -fgraphite-identity
-floop-interchange"

I think both "-fipa-pta" and "-fgraphite-identity" have known issues which
could be causing issues here.  Plus GCC is not normally compiled with
"-funsafe-math-optimizations" or "-funsafe-loop-optimizations".

[Bug middle-end/103386] stage1 with PGO produces bad offsets in rtl-reload on aarch64

[mittorn at sibmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103386

--- Comment #2 from mittorn at sibmail dot com ---
Removing -floop-nest-optimize -floop-interchange -fgraphite-identity fixes
build, but removing only -fgraphite-identity not

[Bug middle-end/103386] -floop-nest-optimize -floop-interchange -fgraphite-identity miscompiles gcc

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103386

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
     Ever confirmed|1                           |0
            Summary|stage1 with PGO produces    |-floop-nest-optimize
                   |bad offsets in rtl-reload   |-floop-interchange
                   |on aarch64                  |-fgraphite-identity
                   |                            |miscompiles gcc
             Status|WAITING                     |UNCONFIRMED