Re: [PATCH] tree-optimization/109609 - correctly interpret arg size in fnspec

[Jan Hubicka <hubicka@ucw.cz>]

> By majority vote and a hint from the API name which is
> arg_max_access_size_given_by_arg_p this interprets a memory access
> size specified as given as other argument such as for strncpy
> in the testcase which has "1cO313" as specifying the _maximum_
> size read/written rather than the exact size.  There are two
> uses interpreting it that way already and one differing.  The
> following adjusts the differing and clarifies the documentation.

This makes sense to me. Bit sad is memcpy/memset may be common enough so
the extra disambiguation based on maximal object size may be useful.
> 
> Bootstrap & regtest running on x86_64-unknown-linux-gnu.
> 
> It looks like arg_access_size_given_by_type_p is interpreted in
> a similar way but the API doesn't reflect this (no _max).  On IRC
> we discussed it would be nice to have both, exact and non-exact
> size (for example use 'a'..'j' for the alternate variant).

Hmm 'a'...'j' may work to stick this information in....
> 
> I still wonder if my interpretation is correct in that '1'..'9'
> specify a _bound_ for the size.  Honza, do you remember why
> you wrote the IPA modref users this way?  Was that intentional?
> Thus is my reverse engineering of the desired semantics correct?

My fnspec implementation was based on what the code was doing
beforehand. In GCC 10 we have:

        case BUILT_IN_STRCPY:
        case BUILT_IN_STRNCPY:
        case BUILT_IN_MEMCPY:
        case BUILT_IN_MEMMOVE:
        case BUILT_IN_MEMPCPY:
        case BUILT_IN_STPCPY:
        case BUILT_IN_STPNCPY:
        case BUILT_IN_TM_MEMCPY:
        case BUILT_IN_TM_MEMMOVE:
          {
            ao_ref dref;
            tree size = NULL_TREE;
            if (gimple_call_num_args (call) == 3)
              size = gimple_call_arg (call, 2);
            ao_ref_init_from_ptr_and_size (&dref,
                                           gimple_call_arg (call, 1),
                                           size);
            return refs_may_alias_p_1 (&dref, ref, false);
          }
So it seem that we handled strncpy the way claiming full size access
there too and I simply copied it into the new implementation.

While adding fnspec I just used adapted this code, so the difference
between modref and tree-ssa-alias implementation was not quite
intentional.
> 
> Thanks,
> Richard.
> 
> 	PR tree-optimization/109609
> 	* attr-fnspec.h (arg_max_access_size_given_by_arg_p):
> 	Clarify semantics.
> 	* tree-ssa-alias.cc (check_fnspec): Correctly interpret
> 	the size given by arg_max_access_size_given_by_arg_p as
> 	maximum, not exact, size.
> 
> 	* gcc.dg/torture/pr109609.c: New testcase.
Patch looks good to me (modulo the fact that it would be nice to be also
able to handle memcpy/memset precisely).

Honza