Re: move increase_alignment from simple to regular ipa pass

[Prathamesh Kulkarni <prathamesh dot kulkarni at linaro dot org>]

On 17 June 2016 at 19:52, Prathamesh Kulkarni
<prathamesh.kulkarni@linaro.org> wrote:
> On 14 June 2016 at 18:31, Prathamesh Kulkarni
> <prathamesh.kulkarni@linaro.org> wrote:
>> On 13 June 2016 at 16:13, Jan Hubicka <hubicka@ucw.cz> wrote:
>>>> diff --git a/gcc/cgraph.h b/gcc/cgraph.h
>>>> index ecafe63..41ac408 100644
>>>> --- a/gcc/cgraph.h
>>>> +++ b/gcc/cgraph.h
>>>> @@ -1874,6 +1874,9 @@ public:
>>>>       if we did not do any inter-procedural code movement.  */
>>>>    unsigned used_by_single_function : 1;
>>>>
>>>> +  /* Set if -fsection-anchors is set.  */
>>>> +  unsigned section_anchor : 1;
>>>> +
>>>>  private:
>>>>    /* Assemble thunks and aliases associated to varpool node.  */
>>>>    void assemble_aliases (void);
>>>> diff --git a/gcc/cgraphunit.c b/gcc/cgraphunit.c
>>>> index 4bfcad7..e75d5c0 100644
>>>> --- a/gcc/cgraphunit.c
>>>> +++ b/gcc/cgraphunit.c
>>>> @@ -800,6 +800,9 @@ varpool_node::finalize_decl (tree decl)
>>>>       it is available to notice_global_symbol.  */
>>>>    node->definition = true;
>>>>    notice_global_symbol (decl);
>>>> +
>>>> +  node->section_anchor = flag_section_anchors;
>>>> +
>>>>    if (TREE_THIS_VOLATILE (decl) || DECL_PRESERVE_P (decl)
>>>>        /* Traditionally we do not eliminate static variables when not
>>>>        optimizing and when not doing toplevel reoder.  */
>>>> diff --git a/gcc/common.opt b/gcc/common.opt
>>>> index f0d7196..e497795 100644
>>>> --- a/gcc/common.opt
>>>> +++ b/gcc/common.opt
>>>> @@ -1590,6 +1590,10 @@ fira-algorithm=
>>>>  Common Joined RejectNegative Enum(ira_algorithm) Var(flag_ira_algorithm) Init(IRA_ALGORITHM_CB) Optimization
>>>>  -fira-algorithm=[CB|priority] Set the used IRA algorithm.
>>>>
>>>> +fipa-increase_alignment
>>>> +Common Report Var(flag_ipa_increase_alignment) Init(0) Optimization
>>>> +Option to gate increase_alignment ipa pass.
>>>> +
>>>>  Enum
>>>>  Name(ira_algorithm) Type(enum ira_algorithm) UnknownError(unknown IRA algorithm %qs)
>>>>
>>>> @@ -2133,7 +2137,7 @@ Common Report Var(flag_sched_dep_count_heuristic) Init(1) Optimization
>>>>  Enable the dependent count heuristic in the scheduler.
>>>>
>>>>  fsection-anchors
>>>> -Common Report Var(flag_section_anchors) Optimization
>>>> +Common Report Var(flag_section_anchors)
>>>>  Access data in the same section from shared anchor points.
>>>>
>>>>  fsee
>>>> diff --git a/gcc/config/aarch64/aarch64.c b/gcc/config/aarch64/aarch64.c
>>>> index a0db3a4..1482566 100644
>>>> --- a/gcc/config/aarch64/aarch64.c
>>>> +++ b/gcc/config/aarch64/aarch64.c
>>>> @@ -8252,6 +8252,8 @@ aarch64_override_options (void)
>>>>
>>>>    aarch64_register_fma_steering ();
>>>>
>>>> +  /* Enable increase_alignment pass.  */
>>>> +  flag_ipa_increase_alignment = 1;
>>>
>>> I would rather enable it always on targets that do support anchors.
>> AFAIK aarch64 supports section anchors.
>>>> diff --git a/gcc/lto/lto-symtab.c b/gcc/lto/lto-symtab.c
>>>> index ce9e146..7f09f3a 100644
>>>> --- a/gcc/lto/lto-symtab.c
>>>> +++ b/gcc/lto/lto-symtab.c
>>>> @@ -342,6 +342,13 @@ lto_symtab_merge (symtab_node *prevailing, symtab_node *entry)
>>>>       The type compatibility checks or the completing of types has properly
>>>>       dealt with most issues.  */
>>>>
>>>> +  /* ??? is this assert necessary ?  */
>>>> +  varpool_node *v_prevailing = dyn_cast<varpool_node *> (prevailing);
>>>> +  varpool_node *v_entry = dyn_cast<varpool_node *> (entry);
>>>> +  gcc_assert (v_prevailing && v_entry);
>>>> +  /* section_anchor of prevailing_decl wins.  */
>>>> +  v_entry->section_anchor = v_prevailing->section_anchor;
>>>> +
>>> Other flags are merged in lto_varpool_replace_node so please move this there.
>> Ah indeed, thanks for the pointers.
>> I wonder though if we need to set
>> prevailing_node->section_anchor = vnode->section_anchor ?
>> IIUC, the function merges flags from vnode into prevailing_node
>> and removes vnode. However we want prevailing_node->section_anchor
>> to always take precedence.
>>>> +/* Return true if alignment should be increased for this vnode.
>>>> +   This is done if every function that references/referring to vnode
>>>> +   has flag_tree_loop_vectorize set.  */
>>>> +
>>>> +static bool
>>>> +increase_alignment_p (varpool_node *vnode)
>>>> +{
>>>> +  ipa_ref *ref;
>>>> +
>>>> +  for (int i = 0; vnode->iterate_reference (i, ref); i++)
>>>> +    if (cgraph_node *cnode = dyn_cast<cgraph_node *> (ref->referred))
>>>> +      {
>>>> +     struct cl_optimization *opts = opts_for_fn (cnode->decl);
>>>> +     if (!opts->x_flag_tree_loop_vectorize)
>>>> +       return false;
>>>> +      }
>>>
>>> If you take address of function that has vectorizer enabled probably doesn't
>>> imply need to increase alignment of that var. So please drop the loop.
>>>
>>> You only want function that read/writes or takes address of the symbol. But
>>> onthe other hand, you need to walk all aliases of the symbol by
>>> call_for_symbol_and_aliases
>>>> +
>>>> +  for (int i = 0; vnode->iterate_referring (i, ref); i++)
>>>> +    if (cgraph_node *cnode = dyn_cast<cgraph_node *> (ref->referring))
>>>> +      {
>>>> +     struct cl_optimization *opts = opts_for_fn (cnode->decl);
>>>> +     if (!opts->x_flag_tree_loop_vectorize)
>>>> +       return false;
>>>> +      }
>>>> +
>>>> +  return true;
>>>> +}
>>>> +
>>>>  /* Entry point to increase_alignment pass.  */
>>>>  static unsigned int
>>>>  increase_alignment (void)
>>>> @@ -914,9 +942,12 @@ increase_alignment (void)
>>>>        tree decl = vnode->decl;
>>>>        unsigned int alignment;
>>>>
>>>> -      if ((decl_in_symtab_p (decl)
>>>> -       && !symtab_node::get (decl)->can_increase_alignment_p ())
>>>> -       || DECL_USER_ALIGN (decl) || DECL_ARTIFICIAL (decl))
>>>> +      if (!vnode->section_anchor
>>>> +       || (decl_in_symtab_p (decl)
>>>> +           && !symtab_node::get (decl)->can_increase_alignment_p ())
>>>> +       || DECL_USER_ALIGN (decl)
>>>> +       || DECL_ARTIFICIAL (decl)
>>>> +       || !increase_alignment_p (vnode))
>>>
>>> Incrementally we probably should do more testing whether the variable looks like
>>> someting that can be vectorized, i.e. it contains array, has address taken or the
>>> accesses are array accesses within loop.
>>> This can be done by the analysis phase of the IPA pass inspecting the function
>>> bodies.
>> Thanks, I will try to check for array accesses are within a loop in
>> followup patch.
>> I was wondering if we could we treat a homogeneous global struct
>> (having members of one type),
>> as a global array of that type and increase it's alignment if required ?
>>>
>>> I think it is important waste to bump up everything including error messages etc.
>>> At least on i386 the effect on firefox datasegment of various alignment setting is
>>> very visible.
>> Um for a start, would it be OK to check if all functions referencing variable
>> have attribute noreturn, and in that case we skip increasing the alignment ?
>> I suppose that error functions would be having attribute noreturn set ?
>>>
>>> Looks OK to me otherwise. please send updated patch.
>> I have done the changes in the attached patch (stage-1 built).
>> I am not sure what to return from the callback function and
>> arbitrarily chose to return true.
> Hi,
> In this version (stage-1 built), I added read/write summaries which
> stream those variables
> which we want to increase alignment for.
>
> I have a few questions:
>
> a) To check if global var is used inside a loop, I obtained
> corresponding ipa_ref
> and checked loop_containing_stmt (ref->stmt), however it returned non-NULL
> even when ref->stmt was not inside a loop.
> for instance:
> int a[32];
> int f() { int x = a[0]; return x; }
> Um how to check if stmt is used inside a loop ?
>
> b) Is it necessary during WPA to check if function has
> flag_tree_vectorize_set since
> during analysis phase in increase_alignment_generate_summary() I check
> if cnode has flag_tree_loop_vectorize_set ?
>
> c) In LTO_increase_alignment_section, the following is streamed:
> i) a "count" of type uwhi, to represent number of variables
> ii) decls corresponding to the variables
> The variables are then obtained during read_summay using
> symtab_node::get_create().
> I suppose since decls for varpool_nodes would already be streamed in
> LTO_section_decls, I was wondering if I
> could somehow refer to the decls in that section to avoid duplicate
> streaming of decls ?
Hi,
In this version, the variable is streamed if it occurs within a loop
or it's address is taken.
To check if stmt is inside a loop I am using:

struct loop *l = loop_containing_stmt (ref->stmt);
if (l != DECL_STRUCT_FUNCTION (cnode->decl)->x_current_loops->tree_root)
  vars->add (vnode);
Is this correct ?

I came across an unexpected issue in my previous patch with -ffat-lto-objects.
I am allocating vars = new hash_set<varpool_node *> () in
generate_summary() and freeing it in write_summary().
However with -ffat-lto-objects, it appears execute() gets called after
write_summary()
during LGEN and we hit segfault in execute() at:
for (hash_set<varpool_node *>::iterator it = vars.begin (); it !=
vars.end (); it++)
  { ... }
because write_summary() has freed vars.
To workaround the issue, I gated call to free vars in write_summary on
flag_fat_lto_objects,
I am not sure if that's a good solution.

Cross tested on arm*-*-*, aarch64*-*-*.

Thanks,
Prathamesh
>
> Thanks,
> Prathamesh
>>
>> Thanks,
>> Prathamesh
>>>
>>> Honza

Attachment: ias2r-9.diff
Description: Text document