Re: [google] pessimize stack accounting during inlining

[Richard Guenther <richard.guenther@gmail.com>]

On Fri, Jun 10, 2011 at 1:45 AM, Mark Heffernan <meheff@google.com> wrote:
> On Wed, Jun 8, 2011 at 1:54 AM, Richard Guenther
> <richard.guenther@gmail.com> wrote:
>> Well, it's still not a hard limit as we can't tell how many spill slots
>> or extra call argument or return value slots we need.
>
> Agreed.  It's not perfect.  But I've found this does a reasonable job
> of preventing the inliner from pushing the frame size much beyond the
> imposed limit especially if the limit is large (eg, many K) relative
> to the typical total size of spill slots, arguments, etc.

Do you have a testcase?

Richard.

> Mark
>
>>
>> Richard.
>>
>> > Thanks,
>> >
>> > David
>> >
>> > On Tue, Jun 7, 2011 at 4:29 PM, Mark Heffernan <meheff@google.com> wrote:
>> >> This patch pessimizes stack accounting during inlining.  This enables
>> >> setting a firm stack size limit (via parameters "large-stack-frame" and
>> >> "large-stack-frame-growth").  Without this patch the inliner is overly
>> >> optimistic about potential stack reuse resulting in actual stack frames much
>> >> larger than the parameterized limits.
>> >> Internal benchmarks show minor performance differences with non-fdo and
>> >> lipo, but overall neutral.  Tested/bootstrapped on x86-64.
>> >> Ok for google-main?
>> >> Mark
>> >>
>> >> 2011-06-07  Mark Heffernan  <meheff@google.com>
>> >>         * cgraph.h (cgraph_global_info): Remove field.
>> >>         * ipa-inline.c (cgraph_clone_inlined_nodes): Change
>> >>         stack frame computation.
>> >>         (cgraph_check_inline_limits): Ditto.
>> >>         (compute_inline_parameters): Remove dead initialization.
>> >>
>> >> Index: gcc/cgraph.h
>> >> ===================================================================
>> >> --- gcc/cgraph.h        (revision 174512)
>> >> +++ gcc/cgraph.h        (working copy)
>> >> @@ -136,8 +136,6 @@ struct GTY(()) cgraph_local_info {
>> >>  struct GTY(()) cgraph_global_info {
>> >>    /* Estimated stack frame consumption by the function.  */
>> >>    HOST_WIDE_INT estimated_stack_size;
>> >> -  /* Expected offset of the stack frame of inlined function.  */
>> >> -  HOST_WIDE_INT stack_frame_offset;
>> >>
>> >>    /* For inline clones this points to the function they will be
>> >>       inlined into.  */
>> >> Index: gcc/ipa-inline.c
>> >> ===================================================================
>> >> --- gcc/ipa-inline.c    (revision 174512)
>> >> +++ gcc/ipa-inline.c    (working copy)
>> >> @@ -229,8 +229,6 @@ void
>> >>  cgraph_clone_inlined_nodes (struct cgraph_edge *e, bool duplicate,
>> >>                             bool update_original)
>> >>  {
>> >> -  HOST_WIDE_INT peak;
>> >> -
>> >>    if (duplicate)
>> >>      {
>> >>        /* We may eliminate the need for out-of-line copy to be output.
>> >> @@ -279,13 +277,13 @@ cgraph_clone_inlined_nodes (struct cgrap
>> >>      e->callee->global.inlined_to = e->caller->global.inlined_to;
>> >>    else
>> >>      e->callee->global.inlined_to = e->caller;
>> >> -  e->callee->global.stack_frame_offset
>> >> -    = e->caller->global.stack_frame_offset
>> >> -      + inline_summary (e->caller)->estimated_self_stack_size;
>> >> -  peak = e->callee->global.stack_frame_offset
>> >> -      + inline_summary (e->callee)->estimated_self_stack_size;
>> >> -  if (e->callee->global.inlined_to->global.estimated_stack_size < peak)
>> >> -    e->callee->global.inlined_to->global.estimated_stack_size = peak;
>> >> +
>> >> +  /* Pessimistically assume no sharing of stack space.  That is, the
>> >> +     frame size of a function is estimated as the original frame size
>> >> +     plus the sum of the frame sizes of all inlined callees.  */
>> >> +  e->callee->global.inlined_to->global.estimated_stack_size +=
>> >> +    inline_summary (e->callee)->estimated_self_stack_size;
>> >> +
>> >>    cgraph_propagate_frequency (e->callee);
>> >>
>> >>    /* Recursively clone all bodies.  */
>> >> @@ -430,8 +428,7 @@ cgraph_check_inline_limits (struct cgrap
>> >>
>> >>    stack_size_limit += stack_size_limit * PARAM_VALUE
>> >> (PARAM_STACK_FRAME_GROWTH) / 100;
>> >>
>> >> -  inlined_stack = (to->global.stack_frame_offset
>> >> -                  + inline_summary (to)->estimated_self_stack_size
>> >> +  inlined_stack = (to->global.estimated_stack_size
>> >>                    + what->global.estimated_stack_size);
>> >>    if (inlined_stack  > stack_size_limit
>> >>        && inlined_stack > PARAM_VALUE (PARAM_LARGE_STACK_FRAME))
>> >> @@ -2064,7 +2061,6 @@ compute_inline_parameters (struct cgraph
>> >>    self_stack_size = optimize ? estimated_stack_frame_size (node) : 0;
>> >>    inline_summary (node)->estimated_self_stack_size = self_stack_size;
>> >>    node->global.estimated_stack_size = self_stack_size;
>> >> -  node->global.stack_frame_offset = 0;
>> >>
>> >>    /* Can this function be inlined at all?  */
>> >>    node->local.inlinable = tree_inlinable_function_p (node->decl);
>> >>
>> >
>