[PATCH 7/7] Plug ipa-prop escape analysis into gimple_call_arg_flags

[Richard Biener]

On Thu, May 22, 2014 at 2:49 PM, Martin Jambor <mjambor@suse.cz> wrote:
> Hi,
>
> On Wed, May 21, 2014 at 04:27:32PM +0200, Richard Biener wrote:
>> On Wed, May 21, 2014 at 3:16 PM, Martin Jambor <mjambor@suse.cz> wrote:
>> > Hi,
>> >
>> > this demonstrates how results of ipa-prop escape analysis from
>> > previous patches can be used at a later stage of compilation by
>> > directly returning them from gimple_call_arg_flags which currently
>> > relies on fnspec annotations.
>> >
>> > Bootstrapped and tested on x86_64-linux and also passes LTO bootstrap.
>> > I have only had a brief look at behavior of this in SPEC 2006 and for
>> > example in astar 1.19% of invocations of gimple_call_arg_flags return
>> > noescape where we previously never did and in calculix this increases
>> > from 15.62% (from annotations) to 18.14%.  Noclobber flag is reported
>> > far less often still but for example in gamess that number raises from
>> > 5.21% to 7.66%.
>> >
>> > Thanks,
>> >
>> > Martin
>> >
>> >
>> > 2014-04-30  Martin Jambor  <mjambor@suse.cz>
>> >
>> >         * gimple.c: Include cgraph.h.
>> >         (gimple_call_arg_flags): Also query bitmaps in cgraph_node.
>> >
>> > Index: src/gcc/gimple.c
>> > ===================================================================
>> > --- src.orig/gcc/gimple.c
>> > +++ src/gcc/gimple.c
>> > @@ -47,7 +47,7 @@ along with GCC; see the file COPYING3.
>> >  #include "demangle.h"
>> >  #include "langhooks.h"
>> >  #include "bitmap.h"
>> > -
>> > +#include "cgraph.h"
>> >
>> >  /* All the tuples have their operand vector (if present) at the very bottom
>> >     of the structure.  Therefore, the offset required to find the
>> > @@ -1349,32 +1349,50 @@ int
>> >  gimple_call_arg_flags (const_gimple stmt, unsigned arg)
>> >  {
>> >    tree attr = gimple_call_fnspec (stmt);
>> > +  int ret;
>> >
>> > -  if (!attr || 1 + arg >= (unsigned) TREE_STRING_LENGTH (attr))
>> > -    return 0;
>> > -
>> > -  switch (TREE_STRING_POINTER (attr)[1 + arg])
>> > +  if (attr && 1 + arg < (unsigned) TREE_STRING_LENGTH (attr))
>> >      {
>> > -    case 'x':
>> > -    case 'X':
>> > -      return EAF_UNUSED;
>> > -
>> > -    case 'R':
>> > -      return EAF_DIRECT | EAF_NOCLOBBER | EAF_NOESCAPE;
>> > -
>> > -    case 'r':
>> > -      return EAF_NOCLOBBER | EAF_NOESCAPE;
>> > -
>> > -    case 'W':
>> > -      return EAF_DIRECT | EAF_NOESCAPE;
>> > -
>> > -    case 'w':
>> > -      return EAF_NOESCAPE;
>> > +      switch (TREE_STRING_POINTER (attr)[1 + arg])
>> > +       {
>> > +       case 'x':
>> > +       case 'X':
>> > +         ret = EAF_UNUSED;
>> > +         break;
>> > +       case 'R':
>> > +         ret = EAF_DIRECT | EAF_NOCLOBBER | EAF_NOESCAPE;
>> > +         break;
>> > +       case 'r':
>> > +         ret = EAF_NOCLOBBER | EAF_NOESCAPE;
>> > +         break;
>> > +       case 'W':
>> > +         ret = EAF_DIRECT | EAF_NOESCAPE;
>> > +         break;
>> > +       case 'w':
>> > +         ret = EAF_NOESCAPE;
>> > +         break;
>> > +       case '.':
>> > +       default:
>> > +         ret = 0;
>> > +       }
>> > +    }
>> > +  else
>> > +    ret = 0;
>> >
>> > -    case '.':
>> > -    default:
>> > -      return 0;
>> > +  tree callee_decl = gimple_call_fndecl (stmt);
>> > +  if (callee_decl)
>> > +    {
>> > +      cgraph_node *callee_node = cgraph_get_node (callee_decl);
>> > +      if (callee_node)
>> > +       {
>> > +         if (cgraph_param_noescape_p (callee_node, arg))
>> > +           ret |= EAF_NOESCAPE;
>> > +         if (cgraph_param_noclobber_p (callee_node, arg))
>> > +           ret |= EAF_NOCLOBBER;
>>
>> That's quite expensive.  I guess we need a better way to store
>> those?
>
> if we want to avoid the cgraph_node lookup, then I think we need to
> store this information in the decl or struct function.  That is
> certainly possible and might even be more appropriate.

Can we?  If the body is not readily available we only have decl and
cgraph-node, not struct function.

I suppose we could exchange the struct function pointer in
tree_function_decl for a cgraph_node pointer and put
the struct function pointer into the cgraph_node.

Of course that may have impacts on FEs who might create
struct function before creating a cgraph node.  But at least
it would avoid enlarging FUNCTION_DECL.

In the end most of the tree_decl_with_vis stuff should move over to symtab
and var-decls should get a varpool_node pointer as well.

Back to the call flags stuff - I also meant the representation of the
"fn spec" attribute.  Rather than parsing that again and again move
it to a better place (which you seem to invent?) and better unified
representation.

Can you try if removing the cgraph hash is possible with the
struct function pointer idea?

Thanks,
Richard.

> Thanks,
>
> Martin
>
>>
>> > +       }
>> >      }
>> > +
>> > +  return ret;
>> >  }
>> >
>> >  /* Detects return flags for the call STMT.  */
>> >