Re: [RFC][IPA-VRP] Add support for IPA VRP in ipa-cp/ipa-prop

[Prathamesh Kulkarni <prathamesh.kulkarni@linaro.org>]

On 30 August 2016 at 10:50, Kugan Vivekanandarajah
<kugan.vivekanandarajah@linaro.org> wrote:
> Hi Honza,
>
> Here is a re-based version which also addresses the review comments.
>
> On 21/07/16 22:54, Jan Hubicka wrote:
>>> Maybe it is better to separate value range and alignment summary
>>> writing/reading to different functions. Here is another updated
>>> version which does this.
>>
>> Makes sense to me. Note that the alignment summary propagation can be either
>> handled by doing bitwise constant propagation and/or extending our value ranges
>> by stride (as described in
>> http://www.lighterra.com/papers/valuerangeprop/Patterson1995-ValueRangeProp.pdf
>> I would like it to go eventually away in favour of more generic solution.
>>
>>> -/* If DEST_PLATS already has aggregate items, check that aggs_by_ref matches
>>> +/* Propagate value range across jump function JFUNC that is associated with
>>> +   edge CS and update DEST_PLATS accordingly.  */
>>> +
>>> +static bool
>>> +propagate_vr_accross_jump_function (cgraph_edge *cs,
>>> +    ipa_jump_func *jfunc,
>>> +    struct ipcp_param_lattices *dest_plats)
>>> +{
>>> +  struct ipcp_param_lattices *src_lats;
>>> +  ipcp_vr_lattice *dest_lat = &dest_plats->m_value_range;
>>> +
>>> +  if (dest_lat->bottom_p ())
>>> +    return false;
>>> +
>>> +  if (jfunc->type == IPA_JF_PASS_THROUGH)
>>> +    {
>>> +      struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
>>> +      int src_idx = ipa_get_jf_pass_through_formal_id (jfunc);
>>> +      src_lats = ipa_get_parm_lattices (caller_info, src_idx);
>>> +
>>> +      if (ipa_get_jf_pass_through_operation (jfunc) == NOP_EXPR)
>>> + return dest_lat->meet_with (src_lats->m_value_range);
>>
>> Clearly we can propagate thorugh expressions here (PLUS_EXPR). I have run
>> into similar issue in loop code that builds simple generic expresisons
>> (like (int)ssa_name+10) and it would be nice to have easy way to deterine
>> their value range based on the knowledge of SSA_NAME's valur range.
>>
>> Bit this is fine for initial implementaiotn for sure.
>
> Indeed. I will do this as a follow up.
>
>>>
>>> +/* Look up all VR information that we have discovered and copy it over
>>> +   to the transformation summary.  */
>>> +
>>> +static void
>>> +ipcp_store_vr_results (void)
>>> +{
>>> +  cgraph_node *node;
>>> +
>>> +  FOR_EACH_FUNCTION_WITH_GIMPLE_BODY (node)
>>> +  {
>>> +    ipa_node_params *info = IPA_NODE_REF (node);
>>> +    bool found_useful_result = false;
>>> +
>>> +    if (!opt_for_fn (node->decl, flag_ipa_vrp))
>>> +      {
>>> + if (dump_file)
>>> +  fprintf (dump_file, "Not considering %s for VR discovery "
>>> +   "and propagate; -fipa-ipa-vrp: disabled.\n",
>>> +   node->name ());
>>> + continue;
>>
>> I belive you need to also prevent propagation through functions copmiled with
>> -fno-ipa-vrp, not only prevent any transformations.
>
> Do you mean the following, I was following other implementations.
>
> @@ -2264,6 +2430,11 @@ propagate_constants_accross_call (struct
> cgraph_edge *cs)
>         &dest_plats->bits_lattice);
>    ret |= propagate_aggs_accross_jump_function (cs, jump_func,
>         dest_plats);
> +  if (opt_for_fn (callee->decl, flag_ipa_vrp))
> +    ret |= propagate_vr_accross_jump_function (cs,
> +       jump_func, dest_plats);
> +  else
> +    ret |= dest_plats->m_value_range.set_to_bottom ();
>
>>> +/* Update value range of formal parameters as described in
>>> +   ipcp_transformation_summary.  */
>>> +
>>> +static void
>>> +ipcp_update_vr (struct cgraph_node *node)
>>> +{
>>> +  tree fndecl = node->decl;
>>> +  tree parm = DECL_ARGUMENTS (fndecl);
>>> +  tree next_parm = parm;
>>> +  ipcp_transformation_summary *ts = ipcp_get_transformation_summary (node);
>>> +  if (!ts || vec_safe_length (ts->m_vr) == 0)
>>> +    return;
>>> +  const vec<ipa_vr, va_gc> &vr = *ts->m_vr;
>>> +  unsigned count = vr.length ();
>>> +
>>> +  for (unsigned i = 0; i < count; ++i, parm = next_parm)
>>> +    {
>>> +      if (node->clone.combined_args_to_skip
>>> +  && bitmap_bit_p (node->clone.combined_args_to_skip, i))
>>> + continue;
>>> +      gcc_checking_assert (parm);
>>> +      next_parm = DECL_CHAIN (parm);
>>> +      tree ddef = ssa_default_def (DECL_STRUCT_FUNCTION (node->decl), parm);
>>> +
>>> +      if (!ddef || !is_gimple_reg (parm))
>>> + continue;
>>> +
>>> +      if (cgraph_local_p (node)
>> The test of cgraph_local_p seems redundant here. The analysis phase should not determine
>> anything if function is reachable non-locally.
>
> Removed it.
>
>>> +/* Info about value ranges.  */
>>> +
>>> +struct GTY(()) ipa_vr
>>> +{
>>> +  /* The data fields below are valid only if known is true.  */
>>> +  bool known;
>>> +  enum value_range_type type;
>>> +  tree min;
>>> +  tree max;
>> What is the point of representing range as trees rather than wide ints. Can they
>> be non-constant integer?
>
> Changed to wide_int after adding that support.
>
> LTO Bootstrapped and regression tested on x86_64-linux-gnu with no new
> regressions, is this OK?
Hi Kugan,
Just noticed a small nit - why not reuse ipa_vr in ipa_jump_func
instead of adding vr_known and m_vr ?

Thanks,
Prathamesh
>
> Thanks
> Kugan