[Bug target/63483] Scheduler performs Invalid move of aliased memory reference

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63483

--- Comment #9 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to Richard Biener from comment #8)
> How can I reproduce this?  Even with -mcpu=ev4 I get
> 
> $foo..ng:
> foo:
>         .frame $30,0,$26,0
> $LFB0:
>         .cfi_startproc
>         .prologue 0
>         mov $31,$2
>         ldq_u $1,1($2)
>         zapnot $1,253,$1
>         stq_u $1,1($2)
>         call_pal 0x81
> 
> I configured with --target=alphaev68-linux-gnu
> 
> And used -O2 -mcpu=ev4 (sounds "earliest").
> 
> Note that the pointers 'a' and 'b' are likely made constant by IPA reference
> and thus loads of the pointers can validly get /u, but not loads from the
> pointers (well, we may even figure out their value is always zero...).

Indeed.  From my .optimized dump:

foo ()
{
  <bb 2>:
  MEM[(char *)0B + 1B] ={v} 0;
  __builtin_trap ();


> Does it "fix" the testcase if you have a function writing to a and b or
> making a and b not static?

Then I get

$foo..ng:
        .prologue 1
        ldq $1,a($29)           !literal
        ldq $2,0($1)
        lda $1,1($31)
        cpys $f31,$f31,$f31
        lda $4,1($2)
        ldq_u $3,1($2)
        insbl $1,$4,$1
        mskbl $3,$4,$3
        bis $1,$3,$1
        stq_u $1,1($2)
        lda $2,1($31)
        ldq $1,b($29)           !literal
        ldq $1,0($1)
        stl $2,8($1)
        ret $31,($26),1

not sure if that's correct.  Adjusted testcase:

char *a;
int *b;

void foo (void)
{
  a[1] = 1;
  b[2] = 1;
}

int bar (void)
{
  return a && b;
}