Re: [PATCH] tree-optimization/104010 - fix SLP scalar costing with patterns

[Richard Biener <rguenther@suse.de>]

On Thu, 14 Apr 2022, Richard Sandiford wrote:

> Richard Biener <rguenther@suse.de> writes:
> > On Thu, 14 Apr 2022, Richard Sandiford wrote:
> >
> >> Richard Biener <rguenther@suse.de> writes:
> >> > When doing BB vectorization the scalar cost compute is derailed
> >> > by patterns, causing lanes to be considered live and thus not
> >> > costed on the scalar side.  For the testcase in PR104010 this
> >> > prevents vectorization which was done by GCC 11.  PR103941
> >> > shows similar cases of missed optimizations that are fixed by
> >> > this patch.
> >> >
> >> > Bootstrapped and tested on x86_64-unknown-linux-gnu.
> >> >
> >> > I'm only considering this now because PR104010 is identified
> >> > as regression on arm - Richards, what do you think?  I do think
> >> > this will enable vectorization of more stuff now which might
> >> > be good or bad - who knowns, but at least it needs to involve
> >> > patterns.
> >> >
> >> > Thanks,
> >> > Richard.
> >> >
> >> > 2022-04-13  Richard Biener  <rguenther@suse.de>
> >> >
> >> > 	PR tree-optimization/104010
> >> > 	PR tree-optimization/103941
> >> > 	* tree-vect-slp.cc (vect_bb_slp_scalar_cost): When
> >> > 	we run into stmts in patterns continue walking those
> >> > 	for uses outside of the vectorized region instead of
> >> > 	marking the lane live.
> >> >
> >> > 	* gcc.target/i386/pr103941-1.c: New testcase.
> >> > 	* gcc.target/i386/pr103941-2.c: Likewise.
> >> > ---
> >> >  gcc/testsuite/gcc.target/i386/pr103941-1.c | 14 +++++++
> >> >  gcc/testsuite/gcc.target/i386/pr103941-2.c | 12 ++++++
> >> >  gcc/tree-vect-slp.cc                       | 47 ++++++++++++++++------
> >> >  3 files changed, 61 insertions(+), 12 deletions(-)
> >> >  create mode 100644 gcc/testsuite/gcc.target/i386/pr103941-1.c
> >> >  create mode 100644 gcc/testsuite/gcc.target/i386/pr103941-2.c
> >> >
> >> > diff --git a/gcc/testsuite/gcc.target/i386/pr103941-1.c b/gcc/testsuite/gcc.target/i386/pr103941-1.c
> >> > new file mode 100644
> >> > index 00000000000..524fdd0b4b1
> >> > --- /dev/null
> >> > +++ b/gcc/testsuite/gcc.target/i386/pr103941-1.c
> >> > @@ -0,0 +1,14 @@
> >> > +/* { dg-do compile } */
> >> > +/* { dg-options "-O2 -msse2" } */
> >> > +
> >> > +unsigned char ur[16], ua[16], ub[16];
> >> > +
> >> > +void avgu_v2qi (void)
> >> > +{
> >> > +  int i;
> >> > +
> >> > +  for (i = 0; i < 2; i++)
> >> > +    ur[i] = (ua[i] + ub[i] + 1) >> 1;
> >> > +}
> >> > +
> >> > +/* { dg-final { scan-assembler "pavgb" } } */
> >> > diff --git a/gcc/testsuite/gcc.target/i386/pr103941-2.c b/gcc/testsuite/gcc.target/i386/pr103941-2.c
> >> > new file mode 100644
> >> > index 00000000000..972a32be997
> >> > --- /dev/null
> >> > +++ b/gcc/testsuite/gcc.target/i386/pr103941-2.c
> >> > @@ -0,0 +1,12 @@
> >> > +/* { dg-do compile } */
> >> > +/* { dg-options "-O2 -msse2" } */
> >> > +
> >> > +void foo (int *c, float *x, float *y)
> >> > +{
> >> > +  c[0] = x[0] < y[0];
> >> > +  c[1] = x[1] < y[1];
> >> > +  c[2] = x[2] < y[2];
> >> > +  c[3] = x[3] < y[3];
> >> > +}
> >> > +
> >> > +/* { dg-final { scan-assembler "cmpltps" } } */
> >> > diff --git a/gcc/tree-vect-slp.cc b/gcc/tree-vect-slp.cc
> >> > index 4ac2b70303c..c7687065374 100644
> >> > --- a/gcc/tree-vect-slp.cc
> >> > +++ b/gcc/tree-vect-slp.cc
> >> > @@ -5185,22 +5185,45 @@ vect_bb_slp_scalar_cost (vec_info *vinfo,
> >> >  	 the scalar cost.  */
> >> >        if (!STMT_VINFO_LIVE_P (stmt_info))
> >> >  	{
> >> > -	  FOR_EACH_PHI_OR_STMT_DEF (def_p, orig_stmt, op_iter, SSA_OP_DEF)
> >> > +	  auto_vec<gimple *, 8> worklist;
> >> > +	  hash_set<gimple *> *worklist_visited = NULL;
> >> > +	  worklist.quick_push (orig_stmt);
> >> > +	  do
> >> >  	    {
> >> > -	      imm_use_iterator use_iter;
> >> > -	      gimple *use_stmt;
> >> > -	      FOR_EACH_IMM_USE_STMT (use_stmt, use_iter, DEF_FROM_PTR (def_p))
> >> > -		if (!is_gimple_debug (use_stmt))
> >> > -		  {
> >> > -		    stmt_vec_info use_stmt_info = vinfo->lookup_stmt (use_stmt);
> >> > -		    if (!use_stmt_info
> >> > -			|| !vectorized_scalar_stmts.contains (use_stmt_info))
> >> > +	      gimple *work_stmt = worklist.pop ();
> >> > +	      FOR_EACH_PHI_OR_STMT_DEF (def_p, work_stmt, op_iter, SSA_OP_DEF)
> >> > +		{
> >> > +		  imm_use_iterator use_iter;
> >> > +		  gimple *use_stmt;
> >> > +		  FOR_EACH_IMM_USE_STMT (use_stmt, use_iter,
> >> > +					 DEF_FROM_PTR (def_p))
> >> > +		    if (!is_gimple_debug (use_stmt))
> >> >  		      {
> >> > -			(*life)[i] = true;
> >> > -			break;
> >> > +			stmt_vec_info use_stmt_info
> >> > +			  = vinfo->lookup_stmt (use_stmt);
> >> > +			if (!use_stmt_info
> >> > +			    || !vectorized_scalar_stmts.contains (use_stmt_info))
> >> > +			  {
> >> > +			    if (STMT_VINFO_IN_PATTERN_P (use_stmt_info))
> >> > +			      {
> >> 
> >> I guess I should walk through the testcase and figure it out for myself,
> >> but: I assume vectorized_scalar_stmts exists because not every statement
> >> we've considered vectorising has made the cut.  Isn't that also true
> >> of (original) scalar statements that would have been vectorised using
> >> patterns?
> >> 
> >> Does vectorized_scalar_stmts record original statements or statements to
> >> vectorise?  From its name I'd have assumed original statements, in which
> >> case I wouldn't have expected IN_PATTERN_P to need special handling.
> >
> > vectorized_scalar_stmts records statements to vectorize, and yes, that's
> > done because PURE_SLP isn't accurate when we promoted SLP nodes to
> > extern because they failed to vectorize.
> 
> Ah, OK.
> 
> > But it's true that in case a pattern is composed from multiple scalar
> > stmts we only cost the scalar root of the pattern.  As I said in the PR
> > the whole thing is up for another rewrite - we're basically trying to
> > compute the "scalar stmt cover" of the vectorized stmts and then
> > try to exclude those scalar stmts in the cover that have uses outside
> > of it.  Currently we're split this problem in an odd way that makes
> > doing it correctly difficult - I hope there's a better way but I
> > haven't yet found the time to do it.  Ideally we'd also merge it with
> > the copy that computes live lanes to be code generated as element
> > extracts.  And IIRC there's another variant somewhere, I think with
> > the SLP graph partitioning.
> >
> > If you have good ideas or want to give it a stab be welcome ;)
> 
> No, no good ideas.  I just wasn't sure why, if vectorized_scalar_stmts
> records statements to vectorize, we couldn't just use
> vect_stmt_to_vectorize (use_stmt_info).  But maybe that isn't
> reliable for SLP.

It could do that but I don't think that would change anything here
since we still have no good way to get at all scalar stmts the pattern
is composed of (and we don't record this set).

I've tried the "obvious", figuring scalar leafs of the SLP tree by
a SLP DFS walk and then simply pick the scalar SSA use-def chain until
I hit those but that's prone of derailing into address computation.
So it seems the "true" solution will involve mimicing
vect_get_and_check_slp_defs for this use-def chain walk at least.  Or
rather than trying to re-do this at costing time make sure to record
enough info when building the SLP tree (and when doing pattern recog).

> Anyway, you knoew this code much better than me :-), so I've certainly
> no objections to the patch for GCC 12.  We've done some aarch64 costing
> work recently but I don't remember a case in which we relied/worked
> around the specific behaviour being patched here.

OK, so for GCC 12 I think it's this patch or let it unfixed.  I'm going
to ponder for a few hours, maybe throw a dice.

Thanks,
Richard.