Re: Question about ifcvt.c:noce_mem_write_may_trap_or_fault_p() vs. decl_readonly_section

[Ian Lance Taylor <iant at google dot com>]

Steven Bosscher <stevenb.gcc@gmail.com> writes:

> I am confused by the following code in
> ifcvt.c:noce_mem_write_may_trap_or_fault_p():
>
> static bool
> noce_mem_write_may_trap_or_fault_p (const_rtx mem)
> {
>   rtx addr;
>
>   if (MEM_READONLY_P (mem))
>     return true;
> (...)
>
>   addr = XEXP (mem, 0);
>
>   /* Call target hook to avoid the effects of -fpic etc....  */
>   addr = targetm.delegitimize_address (addr);
>
>   while (addr)
>     switch (GET_CODE (addr))
>       {
> (...)      case SYMBOL_REF:
>         if (SYMBOL_REF_DECL (addr)
>             && decl_readonly_section (SYMBOL_REF_DECL (addr), 0))
>           return true;
>
> If decl_readonly_section (SYMBOL_REF_DECL (addr), 0) is true, I would
> expect that MEM_READONLY_P is set, and the decl_readonly_section check
> is never true.
>
> Is there a difference between "could be read-only"
> (decl_readonly_section) and "is read-only" (MEM_READONLY_P)?

In principle, I agree that MEM_READONLY_P should be set if the address
is based on a SYMBOL_REF and decl_readonly_section is true for the
SYMBOL_REF.  In practice I would not be in the least surprised if the
two tests were not identical.  There is a lot of code in the compiler
that creates a MEM, and the functions to create a MEM do not check for a
SYMBOL_REF for which decl_readonly_section is true.  There are other
triggers for MEM_READONLY_P, so it could turn out to be identical, but I
would not be surprised if it isn't.

If you are really curious it's probably worth adding a test there to see
if it ever triggers--if the call to decl_readonly_section ever returns
true.

Ian