[Bug c/100491] New: Code generation get worse when including function prototype on ARM

[Bug c/100491] New: Code generation get worse when including function prototype on ARM

[fredrik.hederstierna@securitas-direct.com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100491

            Bug ID: 100491
           Summary: Code generation get worse when including function
                    prototype on ARM
           Product: gcc
           Version: 11.1.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c
          Assignee: unassigned at gcc dot gnu.org
          Reporter: fredrik.hederstierna@securitas-direct.com
  Target Milestone: ---

Created attachment 50779
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=50779&action=edit
sym.c

New in arm-none-eabi-gcc-11.1.0 I found that code generated when including a
function prototype was getting worse. I've in the past tested gcc 4-10, and
these didn't show this behavior.

Attach example from CSiBE Flex.
(Code was getting +17% size both with -O2 and -Os.)
Target: arm-none-eabi, cortex-m3, softfloat

GCC-10.2.0

00000000 <addsym.constprop.0.isra.0>:
   0:   e92d 43f8       stmdb   sp!, {r3, r4, r5, r6, r7, r8, r9, lr}
   4:   4689            mov     r9, r1
   6:   2165            movs    r1, #101        ; 0x65
   8:   461c            mov     r4, r3
   a:   4607            mov     r7, r0
   c:   4690            mov     r8, r2
   e:   f7ff fffe       bl      0 <hashfunct>
  12:   4605            mov     r5, r0
  14:   f854 6020       ldr.w   r6, [r4, r0, lsl #2]
  18:   b98e            cbnz    r6, 3e <addsym.constprop.0.isra.0+0x3e>
  1a:   2014            movs    r0, #20
  1c:   f7ff fffe       bl      0 <flex_alloc>
  20:   f854 3025       ldr.w   r3, [r4, r5, lsl #2]
  24:   b19b            cbz     r3, 4e <addsym.constprop.0.isra.0+0x4e>
  26:   6043            str     r3, [r0, #4]
  28:   6018            str     r0, [r3, #0]
  2a:   2300            movs    r3, #0
  2c:   e9c0 7902       strd    r7, r9, [r0, #8]
  30:   6003            str     r3, [r0, #0]
  32:   f8c0 8010       str.w   r8, [r0, #16]
  36:   f844 0025       str.w   r0, [r4, r5, lsl #2]
  3a:   e8bd 83f8       ldmia.w sp!, {r3, r4, r5, r6, r7, r8, r9, pc}
  3e:   4638            mov     r0, r7
  40:   68b1            ldr     r1, [r6, #8]
  42:   f7ff fffe       bl      0 <strcmp>
  46:   2800            cmp     r0, #0
  48:   d0f7            beq.n   3a <addsym.constprop.0.isra.0+0x3a>
  4a:   6876            ldr     r6, [r6, #4]
  4c:   e7e4            b.n     18 <addsym.constprop.0.isra.0+0x18>
  4e:   6046            str     r6, [r0, #4]
  50:   e7eb            b.n     2a <addsym.constprop.0.isra.0+0x2a>

00000052 <cclinstal>:
  52:   b510            push    {r4, lr}
  54:   460c            mov     r4, r1
  56:   f7ff fffe       bl      0 <copy_unsigned_string>
  5a:   4622            mov     r2, r4
  5c:   e8bd 4010       ldmia.w sp!, {r4, lr}
  60:   2100            movs    r1, #0
  62:   4b01            ldr     r3, [pc, #4]    ; (68 <cclinstal+0x16>)
  64:   e7cc            b.n     0 <addsym.constprop.0.isra.0>
  66:   bf00            nop
  68:   00000000        .word   0x00000000

0000006c <ndinstal>:
  6c:   b570            push    {r4, r5, r6, lr}
  6e:   460d            mov     r5, r1
  70:   f7ff fffe       bl      0 <copy_string>
  74:   4604            mov     r4, r0
  76:   4628            mov     r0, r5
  78:   f7ff fffe       bl      0 <copy_unsigned_string>
  7c:   2200            movs    r2, #0
  7e:   4601            mov     r1, r0
  80:   4b02            ldr     r3, [pc, #8]    ; (8c <ndinstal+0x20>)
  82:   4620            mov     r0, r4
  84:   e8bd 4070       ldmia.w sp!, {r4, r5, r6, lr}
  88:   e7ba            b.n     0 <addsym.constprop.0.isra.0>
  8a:   bf00            nop
  8c:   00000194        .word   0x00000194


GCC-11.1.0

00000000 <addsym.constprop.0>:
   0:   e92d 47f0       stmdb   sp!, {r4, r5, r6, r7, r8, r9, sl, lr}
   4:   4689            mov     r9, r1
   6:   2165            movs    r1, #101        ; 0x65
   8:   461d            mov     r5, r3
   a:   4607            mov     r7, r0
   c:   4690            mov     r8, r2
   e:   f7ff fffe       bl      0 <hashfunct>
  12:   4606            mov     r6, r0
  14:   f855 a020       ldr.w   sl, [r5, r0, lsl #2]
  18:   f1ba 0f00       cmp.w   sl, #0
  1c:   d112            bne.n   44 <addsym.constprop.0+0x44>
  1e:   2014            movs    r0, #20
  20:   f7ff fffe       bl      0 <flex_alloc>
  24:   f855 3026       ldr.w   r3, [r5, r6, lsl #2]
  28:   4604            mov     r4, r0
  2a:   b1a3            cbz     r3, 56 <addsym.constprop.0+0x56>
  2c:   6043            str     r3, [r0, #4]
  2e:   6018            str     r0, [r3, #0]
  30:   2000            movs    r0, #0
  32:   e9c4 7902       strd    r7, r9, [r4, #8]
  36:   6020            str     r0, [r4, #0]
  38:   f8c4 8010       str.w   r8, [r4, #16]
  3c:   f845 4026       str.w   r4, [r5, r6, lsl #2]
  40:   e8bd 87f0       ldmia.w sp!, {r4, r5, r6, r7, r8, r9, sl, pc}
  44:   4638            mov     r0, r7
  46:   f8da 1008       ldr.w   r1, [sl, #8]
  4a:   f7ff fffe       bl      0 <strcmp>
  4e:   b128            cbz     r0, 5c <addsym.constprop.0+0x5c>
  50:   f8da a004       ldr.w   sl, [sl, #4]
  54:   e7e0            b.n     18 <addsym.constprop.0+0x18>
  56:   f8c0 a004       str.w   sl, [r0, #4]
  5a:   e7e9            b.n     30 <addsym.constprop.0+0x30>
  5c:   f04f 30ff       mov.w   r0, #4294967295 ; 0xffffffff
  60:   e7ee            b.n     40 <addsym.constprop.0+0x40>

00000062 <cclinstal>:
  62:   b513            push    {r0, r1, r4, lr}
  64:   460c            mov     r4, r1
  66:   f7ff fffe       bl      0 <copy_unsigned_string>
  6a:   2365            movs    r3, #101        ; 0x65
  6c:   4622            mov     r2, r4
  6e:   9300            str     r3, [sp, #0]
  70:   2100            movs    r1, #0
  72:   4b02            ldr     r3, [pc, #8]    ; (7c <cclinstal+0x1a>)
  74:   f7ff ffc4       bl      0 <addsym.constprop.0>
  78:   b002            add     sp, #8
  7a:   bd10            pop     {r4, pc}
  7c:   00000000        .word   0x00000000

00000080 <ndinstal>:
  80:   b537            push    {r0, r1, r2, r4, r5, lr}
  82:   460d            mov     r5, r1
  84:   f7ff fffe       bl      0 <copy_string>
  88:   4604            mov     r4, r0
  8a:   4628            mov     r0, r5
  8c:   f7ff fffe       bl      0 <copy_unsigned_string>
  90:   2365            movs    r3, #101        ; 0x65
  92:   4601            mov     r1, r0
  94:   9300            str     r3, [sp, #0]
  96:   2200            movs    r2, #0
  98:   4620            mov     r0, r4
  9a:   4b02            ldr     r3, [pc, #8]    ; (a4 <ndinstal+0x24>)
  9c:   f7ff ffb0       bl      0 <addsym.constprop.0>
  a0:   b003            add     sp, #12
  a2:   bd30            pop     {r4, r5, pc}
  a4:   00000194        .word   0x00000194

[Bug c/100491] Code generation get worse when including function prototype on ARM

[fredrik.hederstierna@securitas-direct.com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100491

--- Comment #1 from Fredrik Hederstierna <fredrik.hederstierna@securitas-direct.com> ---
Created attachment 50780
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=50780&action=edit
makefile

Adding makefile to compile test example. See in makefile how to
trigger/untrigger the unexpected condition.

[Bug ipa/100491] [11 Regression] IPA-SRA is not happening any more

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100491

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
            Summary|Code generation get worse   |[11 Regression] IPA-SRA is
                   |when including function     |not happening any more
                   |prototype on ARM            |
                 CC|                            |marxin at gcc dot gnu.org
          Component|tree-optimization           |ipa

--- Comment #2 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
So it looks like IPA SRA did not happen any more in GCC 11 while it did in GCC
10.

[Bug ipa/100491] [11 Regression] IPA-SRA is not happening any more

[fredrik.hederstierna@securitas-direct.com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100491

--- Comment #3 from Fredrik Hederstierna <fredrik.hederstierna@securitas-direct.com> ---
It is strange that SRA is skipped only when adding the (unnecessary) function
prototype:

static int addsym (register char[], char *, int, hash_table_t, int);

If skipping the prototype, or adding 'wrong' prototype, then SRA runs
correctly.
Eg, this will also work adding

static int addsym ();

[Bug ipa/100491] [11 Regression] IPA-SRA is not happening any more

[fredrik.hederstierna@securitas-direct.com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100491

--- Comment #4 from Fredrik Hederstierna <fredrik.hederstierna@securitas-direct.com> ---
Created attachment 50782
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=50782&action=edit
092t.fixup_cfg3 working

Example where SRA is used, good example.

[Bug ipa/100491] [11 Regression] IPA-SRA is not happening any more

[fredrik.hederstierna@securitas-direct.com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100491

--- Comment #5 from Fredrik Hederstierna <fredrik.hederstierna@securitas-direct.com> ---
Created attachment 50783
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=50783&action=edit
092t.fixup_cfg3 not working

092t.fixup_cfg3 example SRA not used.

[Bug ipa/100491] [11/12 Regression] IPA-SRA is not happening any more

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100491

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |jamborm at gcc dot gnu.org
   Target Milestone|---                         |11.2
            Summary|[11 Regression] IPA-SRA is  |[11/12 Regression] IPA-SRA
                   |not happening any more      |is not happening any more

[Bug ipa/100491] [11/12 Regression] IPA-SRA is not happening any more

[jamborm at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100491

Martin Jambor <jamborm at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
     Ever confirmed|0                           |1
   Last reconfirmed|                            |2021-05-12
             Status|UNCONFIRMED                 |NEW

--- Comment #6 from Martin Jambor <jamborm at gcc dot gnu.org> ---
It is the char[] part of the declaration that causes this, changing it
to a pointer also avoids the issue.

In GCC 11, functions with parameters which are these length-less
arrays automagically receive an access function type attribute and
IPA-SRA does not understand them and so punts.

Understanding some attributes, eventually including access, is on my
TODO list, meanwhile please use a pointer type.

[Bug ipa/100491] [11/12 Regression] IPA-SRA is not happening any more

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100491

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|11.2                        |11.3

--- Comment #7 from Richard Biener <rguenth at gcc dot gnu.org> ---
GCC 11.2 is being released, retargeting bugs to GCC 11.3

[Bug ipa/100491] [11/12 Regression] IPA-SRA is not happening any more

[jamborm at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100491

--- Comment #8 from Martin Jambor <jamborm at gcc dot gnu.org> ---
I believe this has been fixed in November by r12-5223-gecdf414bd89e6b.
(At some point I'll verify it, unless someone is faster, for which I
would be grateful).

Unfortunately, I do not expect the commit to be backported to GCC 11.

[Bug ipa/100491] [11 Regression] IPA-SRA is not happening any more

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100491

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
            Summary|[11/12 Regression] IPA-SRA  |[11 Regression] IPA-SRA is
                   |is not happening any more   |not happening any more
           Priority|P3                          |P2
      Known to work|                            |12.0

[Bug ipa/100491] [11 Regression] IPA-SRA is not happening any more

[fredrik.hederstierna@securitas-direct.com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100491

--- Comment #9 from Fredrik Hederstierna <fredrik.hederstierna@securitas-direct.com> ---
I tested with gcc-12-20220123 snapshot, and lastest gcc now produces the same
result as gcc-10.2.0, so I agree I think this issue is resolved now.
Thanks! BR Fredrik

[Bug ipa/100491] [11 Regression] IPA-SRA is not happening any more

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100491

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|11.3                        |11.4

--- Comment #10 from Richard Biener <rguenth at gcc dot gnu.org> ---
GCC 11.3 is being released, retargeting bugs to GCC 11.4.

[Bug ipa/100491] [11 Regression] IPA-SRA is not happening any more

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100491

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|11.4                        |11.5

--- Comment #11 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
GCC 11.4 is being released, retargeting bugs to GCC 11.5.