Make sibcall argument overlap check less pessimistic (PR middle-end/50074)

Hi!

Kirill's recent change to mem_overlaps_already_clobbered_arg_p
resulted in various code quality regressions, many calls that used
to be tail call optimized no longer are.

Here is an attempt to make the check more complete (e.g.
the change wouldn't see overlap if addr was PLUS of two REGs,
where one of the REGs was based on internal_arg_pointer, etc.)
and less pessimistic.  As tree-tailcall.c doesn't allow tail calls
from functions that have address of any of the caller's parameters
taken, IMHO it is enough to look for internal_arg_pointer based
pseudos initialized in the tail call sequence.
This patch scans the tail call sequence and notes which pseudos
are based on internal_arg_pointer (and what offset from
that pointer they have) and uses that in
mem_overlaps_already_clobbered_arg_p.

Bootstrapped/regtested on x86_64-linux and i686-linux, tested on some
testcases on ia64-linux (as an example of target which doesn't have
reg + disp addressing and thus forces everything into registers).
Ok for trunk?

2011-11-25  Jakub Jelinek  <jakub@redhat.com>

	PR middle-end/50074
	* calls.c (internal_arg_pointer_seq_start,
	internal_arg_pointer_cache): New variables.
	(internal_arg_pointer_based_reg_1): New function.
	(internal_arg_pointer_based_reg): New function.
	(mem_overlaps_already_clobbered_arg_p): Use it.
	(expand_call): Free internal_arg_pointer_cache vector
	and clear internal_arg_pointer_seq_start.


	Jakub

> Here is an attempt to make the check more complete (e.g.
> the change wouldn't see overlap if addr was PLUS of two REGs,
> where one of the REGs was based on internal_arg_pointer, etc.)
> and less pessimistic.  As tree-tailcall.c doesn't allow tail calls
> from functions that have address of any of the caller's parameters
> taken, IMHO it is enough to look for internal_arg_pointer based
> pseudos initialized in the tail call sequence.
> This patch scans the tail call sequence and notes which pseudos
> are based on internal_arg_pointer (and what offset from
> that pointer they have) and uses that in
> mem_overlaps_already_clobbered_arg_p.

This looks reasonable, but the logic is a bit hard to follow, especially the 
double usage of internal_arg_pointer_based_reg depending on SCAN's value.
Would it be possible to split it into 2 functions that recursively call each 
other?

You should also make it clearer that internal_arg_pointer_seq_start is part of 
the caching mechanism (I wondered for some minutes whether it has anything to 
do with RTL sequences), maybe:

/* Last insn that has been scanned by internal_arg_pointer_based_reg, or
   NULL_RTX if none has been scanned yet.  */
static rtx internal_arg_pointer_insn_cache;

/* Vector indexed by REGNO - FIRST_PSEUDO_REGISTER, recording if a pseudo is
   based on crtl->args.internal_arg_pointer.  The element is NULL_RTX if the
   pseudo isn't based on it, a CONST_INT offset if the pseudo is based on it
   with fixed offset, or PC if this is with variable or unknown offset.  */
static VEC(rtx, heap) *internal_arg_pointer_pseudo_cache;


> +/* Helper function for internal_arg_pointer_based_reg, called through
> +   for_each_rtx.  Return 1 if a crtl->args.internal_arg_pointer based
> +   register is seen anywhere.  */
> +
> +static int
> +internal_arg_pointer_based_reg_1 (rtx *loc, void *data ATTRIBUTE_UNUSED)
> +{
> +  if (REG_P (*loc) && internal_arg_pointer_based_reg (*loc, false) !=
> NULL_RTX) +    return 1;
> +  if (MEM_P (*loc))
> +    return -1;
> +  return 0;
> +}

Missing comment for the MEM_P case.


> +/* If REG is based on crtl->args.internal_arg_pointer, return either
> +   a CONST_INT offset from crtl->args.internal_arg_pointer if
> +   offset from it is known constant, or PC if the offset is unknown.
> +   Return NULL_RTX if REG isn't based on crtl->args.internal_arg_pointer. 
> */ +
> +static rtx
> +internal_arg_pointer_based_reg (rtx reg, bool scan)
> +{
> +  rtx insn;
> +
> +  if (CONSTANT_P (reg))
> +    return NULL_RTX;
> +
> +  if (reg == crtl->args.internal_arg_pointer)
> +    return const0_rtx;
> +
> +  if (REG_P (reg) && REGNO (reg) < FIRST_PSEUDO_REGISTER)
> +    return NULL_RTX;

if (REG_P (reg) && HARD_REGISTER_P (reg))
  return NULL_RTX;

On Mon, Nov 28, 2011 at 11:10:56PM +0100, Eric Botcazou wrote:
> This looks reasonable, but the logic is a bit hard to follow, especially the 
> double usage of internal_arg_pointer_based_reg depending on SCAN's value.
> Would it be possible to split it into 2 functions that recursively call each 
> other?

Well, splitting off the scanning into separate function is quite easy, but
that would still mean SCAN parameter.  The thing is, for the common case
where ADDR is already internal_arg_pointer, or sum of that and CONST_INT, or
SYMBOL_REF, I'd prefer to avoid the scanning of the instructions and only do
that if really needed (i.e. when either looking for a pseudo or arbitrary
expression that needs to be for_each_rtx scanned).  To get rid of the SCAN
argument I'm afraid the first part of the function would need to be
duplicated (for the current SCAN and !SCAN cases), then the scanning phase
in another function and the last part in another one.

> You should also make it clearer that internal_arg_pointer_seq_start is part of 
> the caching mechanism (I wondered for some minutes whether it has anything to 
> do with RTL sequences), maybe:

Ok.

> > +static int
> > +internal_arg_pointer_based_reg_1 (rtx *loc, void *data ATTRIBUTE_UNUSED)
> > +{
> > +  if (REG_P (*loc) && internal_arg_pointer_based_reg (*loc, false) !=
> > NULL_RTX) +    return 1;
> > +  if (MEM_P (*loc))
> > +    return -1;
> > +  return 0;
> > +}
> 
> Missing comment for the MEM_P case.

Will add.
> 
> if (REG_P (reg) && HARD_REGISTER_P (reg))
>   return NULL_RTX;

Ok.  BTW, it fixes also wrong-code PR51323, so I'll be including new
testcase in the next patch iteration.

	Jakub