[Bug c/100163] New: -falign-loops sometimes produces invalid code for SH-2

[Bug c/100163] New: -falign-loops sometimes produces invalid code for SH-2

[vluchits at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100163

            Bug ID: 100163
           Summary: -falign-loops sometimes produces invalid code for SH-2
           Product: gcc
           Version: 9.3.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c
          Assignee: unassigned at gcc dot gnu.org
          Reporter: vluchits at gmail dot com
  Target Milestone: ---

Hello,

relocating functions to .data segment causes gcc to produce invalid code for
'nop' instruction on SH2 architecture when -O2 optimization flag is specified
or -falign-loops is enabled explicitly.

Here's disassembly of a C function R_SegCommand that demonstrates the code that
works:
0000050c <_R_SegCommands>:
 50c:   2f 86           mov.l   r8,@-r15
 50e:   2f 96           mov.l   r9,@-r15
 510:   2f a6           mov.l   r10,@-r15
 512:   2f b6           mov.l   r11,@-r15
 514:   2f c6           mov.l   r12,@-r15
 516:   2f d6           mov.l   r13,@-r15
 518:   2f e6           mov.l   r14,@-r15
 51a:   91 b1           mov.w   680 <_R_SegCommands+0x174>,r1   ! 258
 51c:   4f 22           sts.l   pr,@-r15
 51e:   d2 5a           mov.l   688 <_R_SegCommands+0x17c>,r2   ! 20004024
 520:   3f 18           sub     r1,r15
 522:   00 09           nop
 524:   60 21           mov.w   @r2,r0
 526:   64 0d           extu.w  r0,r4
 528:   24 48           tst     r4,r4
 52a:   8f fb           bf.s    524 <_R_SegCommands+0x18>
 52c:   e7 03           mov     #3,r7
...

Please note the 'nop' command at offset 522, which is encoded as 00 09, which
is the correct opcode for nop on SH2.

Now if -O2 or -Os -align-loops are specified and R_SegCommands is relocated to
the .data segment and aligned to 16-byte boundary:
void R_SegCommands(void) __attribute__((section(".data"), aligned(16)));

the following code is produced:

000000d0 <_R_SegCommands>:
  d0:   2f 86           mov.l   r8,@-r15
  d2:   2f 96           mov.l   r9,@-r15
  d4:   2f a6           mov.l   r10,@-r15
  d6:   2f b6           mov.l   r11,@-r15
  d8:   2f c6           mov.l   r12,@-r15
  da:   2f d6           mov.l   r13,@-r15
  dc:   2f e6           mov.l   r14,@-r15
  de:   91 b1           mov.w   244 <_R_SegCommands+0x174>,r1   ! 258
  e0:   4f 22           sts.l   pr,@-r15
  e2:   d2 5a           mov.l   24c <_R_SegCommands+0x17c>,r2   ! 20004024
  e4:   3f 18           sub     r1,r15
  e6:   00 00           .word 0x0000
  e8:   60 21           mov.w   @r2,r0
  ea:   64 0d           extu.w  r0,r4
  ec:   24 48           tst     r4,r4
  ee:   8f fb           bf.s    e8 <_R_SegCommands+0x18>
  f0:   e7 03           mov     #3,r7
  f2:   d5 57           mov.l   250 <_R_SegCommands+0x180>,r5   ! 0
<_R_DrawTexture>
  f4:   e3 01           mov     #1,r3
...

Note the opcode at offset e6, which is a nop instruction and was 00 09 in the
previous version and which is 00 00 now. When this opcode is encountered during
the program's execution, it an causes immediate crash.

Replacing all occurrences of 00 00 in the binary code with 00 09 makes the
makes the program run fine again.

[Bug target/100163] -falign-loops sometimes produces invalid code for SH-2

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100163

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |WAITING
     Ever confirmed|0                           |1
   Last reconfirmed|                            |2021-04-20

--- Comment #1 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
>relocating functions to .data segment

This is more likely an assembly issue or rather an user issue.
.align inside the text section uses nops but inside non-text sections uses 0s.
What does GCC output -S show?

[Bug target/100163] -falign-loops sometimes produces invalid code for SH-2

[vluchits at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100163

--- Comment #2 from Victor Luchitz <vluchits at gmail dot com> ---
Here's relevant -S output for -O2:
...
        sub     r1,r15
        .align 2
.L186:
        mov.w   @r2,r0
        extu.w  r0,r4

And for -O2 -fno-align-loops:
...
        sub     r1,r15
.L186:
        mov.w   @r2,r0
        extu.w  r0,r4

I'm not sure what's workarounds are available for an end user here other than
to explicitly disable all kinds of alignments.

[Bug target/100163] -falign-loops sometimes produces invalid code for SH-2

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100163

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
         Resolution|---                         |INVALID
             Status|WAITING                     |RESOLVED

--- Comment #3 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Yes this is not a bug.  You can't use .align in the data section and expect
nops to happen.  You marking a function in the data section makes this invalid.

See
https://sourceware.org/binutils/docs-2.36/as/Align.html#Align
"However, on most systems, if the section is marked as containing code and the
fill value is omitted, the space is filled with no-op instructions."

GCC is assuming you are using a section which is marked as containing code
which .data is not one of those sections.