Re: [Patch, vrp] Allow VRP type conversion folding only for widenings upto word mode

[Senthil Kumar Selvaraj <senthil_kumar.selvaraj@atmel.com>]

On Mon, Nov 16, 2015 at 10:02:15AM +0100, Richard Biener wrote:
> On Sat, 14 Nov 2015, Senthil Kumar Selvaraj wrote:
> 
> > On Sat, Nov 14, 2015 at 09:57:40AM +0100, Richard Biener wrote:
> > > On November 14, 2015 9:49:28 AM GMT+01:00, Senthil Kumar Selvaraj <senthil_kumar.selvaraj@atmel.com> wrote:
> > > >On Sat, Nov 14, 2015 at 09:13:41AM +0100, Marc Glisse wrote:
> > > >> On Sat, 14 Nov 2015, Senthil Kumar Selvaraj wrote:
> > > >> 
> > > >> >This patch came out of a discussion held in the gcc mailing list
> > > >> >(https://gcc.gnu.org/ml/gcc/2015-11/msg00067.html).
> > > >> >
> > > >> >The patch restricts folding of conditional exprs with lhs previously
> > > >> >set by a type conversion to occur only if the source of the type
> > > >> >conversion's mode is word mode or smaller.
> > > >> >
> > > >> >Bootstrapped and reg tested on x86_64 (with
> > > >--enable-languages=c,c++).
> > > >> >
> > > >> >If ok, could you commit please? I don't have commit access.
> > > >> >
> > > >> >Regards
> > > >> >Senthil
> > > >> >
> > > >> >gcc/ChangeLog
> > > >> >
> > > >> >2015-11-11  Senthil Kumar Selvaraj 
> > > ><senthil_kumar.selvaraj@atmel.com>
> > > >> >
> > > >> >	* tree-vrp.c (simplify_cond_using_ranges): Fold only
> > > >> >	if innerop's mode is word_mode or smaller.
> > > >> >
> > > >> >
> > > >> >diff --git gcc/tree-vrp.c gcc/tree-vrp.c
> > > >> >index e2393e4..c139bc6 100644
> > > >> >--- gcc/tree-vrp.c
> > > >> >+++ gcc/tree-vrp.c
> > > >> >@@ -9467,6 +9467,8 @@ simplify_cond_using_ranges (gcond *stmt)
> > > >> >      innerop = gimple_assign_rhs1 (def_stmt);
> > > >> >
> > > >> >      if (TREE_CODE (innerop) == SSA_NAME
> > > >> >+         && (GET_MODE_SIZE(TYPE_MODE(TREE_TYPE(innerop)))
> > > >> >+           <= GET_MODE_SIZE(word_mode))
> > > >> >	  && !POINTER_TYPE_P (TREE_TYPE (innerop)))
> > > >> >	{
> > > >> >	  value_range *vr = get_value_range (innerop);
> > > >> 
> > > >> I thought the result of the discussion was that the transformation is
> > > >ok if
> > > >> either it is narrowing or it widens but to something no bigger than
> > > >> word_mode. So you should have 2 comparisons, or 1 with a max.
> > > >
> > > >Hmm, I came to the opposite conclusion - I thought Richard only okayed
> > > >"widening upto word-mode", not the narrowing. 
> > > 
> > > I didn't mean to suggest narrowing is not OK.  In fact narrowing is always OK.
> > 
> > My bad. Here's a revised patch that checks for both conditions, using
> > max as Marc suggested to limit to word_mode or narrowing conversions.
> > 
> > Bootstrapped and regtested for x86_64 with c and c++.
> > 
> > Is this ok? If yes, would you commit it
> > for me please? I don't have commit access.
> > 
> > gcc/ChangeLog
> > 2015-11-14  Senthil Kumar Selvaraj  <senthil_kumar.selvaraj@atmel.com>
> > 
> > 	* tree-vrp.c (simplify_cond_using_ranges): Fold only
> > 	if innerop's mode smaller or equal to word_mode or op0's mode.
> > 
> > 
> > diff --git a/gcc/tree-vrp.c b/gcc/tree-vrp.c
> > index e2393e4..cfd90e7 100644
> > --- a/gcc/tree-vrp.c
> > +++ b/gcc/tree-vrp.c
> > @@ -9467,7 +9467,10 @@ simplify_cond_using_ranges (gcond *stmt)
> >        innerop = gimple_assign_rhs1 (def_stmt);
> >  
> >        if (TREE_CODE (innerop) == SSA_NAME
> > -	  && !POINTER_TYPE_P (TREE_TYPE (innerop)))
> > +	  && !POINTER_TYPE_P (TREE_TYPE (innerop))
> > +         && (GET_MODE_SIZE (TYPE_MODE (TREE_TYPE (innerop)))
> > +           <= std::max (GET_MODE_SIZE (word_mode),
> > +                        GET_MODE_SIZE (TYPE_MODE (TREE_TYPE (op0))))))
> 
> Please use TYPE_PRECISION (...) and GET_MODE_PRECISION (word_mode) and
> add a comment as to what we are testing here and why.
> 
> Btw, ideally we'd factor out a
> 
> bool
> desired_pro_or_demotion_p (tree to_type, tree from_type) {}
> 
> function somewhere as we have similar tests throughout the compiler
> that we might want to unify (and also have a central place to
> eventually add a target hook if ever desired).
> 
> In fact in other places we also check that the type we promote/demote
> to matches its mode precision or the type we promote/demote from
> already does not.
> 
> I'd suggest tree.[ch] for that function.
> 
> Please also add a testcase.

How does the below patch look? Bootstrapped, but not regtested yet.

The testcase was rather tricky to write - I wasn't sure how to reliably
get a type bigger than a word for all targets. I resorted to __int128,
not sure it's a good idea though - I should probably add dg-skip-if for
targets that don't support that. Do you know of a better way to write
that?

Regards
Senthil

diff --git gcc/testsuite/gcc.dg/tree-ssa/vrp98.c gcc/testsuite/gcc.dg/tree-ssa/vrp98.c
new file mode 100644
index 0000000..448ceba
--- /dev/null
+++ gcc/testsuite/gcc.dg/tree-ssa/vrp98.c
@@ -0,0 +1,40 @@
+/* { dg-do compile } */
+/* { dg-options "-Os -fdump-tree-vrp1-details" } */
+
+#include <stdint.h>
+#include <limits.h>
+
+typedef unsigned int word __attribute__((mode(word)));
+typedef unsigned __int128 bigger_than_word;
+
+int
+foo (bigger_than_word a, word b, uint8_t c)
+{
+  /* Must fold use of t1 into use of b, as b is no wider than word_mode. */
+  const uint8_t t1 = b % UCHAR_MAX;
+
+  /* Must NOT fold use of t2 into use of a, as a is wider than word_mode. */
+  const uint8_t t2 = a % UCHAR_MAX;
+
+  /* Must fold use of t3 into use of c, as c is narrower than t3. */
+  const uint32_t t3 = (const uint32_t)(c >> 1);
+
+  uint16_t ret = 0;
+
+  if (t1 == 1)
+    ret = 20;
+  else if (t2 == 2)
+    ret = 30;
+  else if (t3 == 3)
+    ret = 40;
+  /* Th extra condition below is necessary to prevent a prior pass from
+     folding away the cast. Ignored in scan-tree-dump. */
+  else if (t3 == 4)
+    ret = 50;
+
+  return ret;
+}
+
+/* { dg-final { scan-tree-dump "Folded into: if \\(_\[0-9\]+ == 1\\)" "vrp1" } } */
+/* { dg-final { scan-tree-dump-not "Folded into: if \\(_\[0-9\]+ == 2\\)" "vrp1" } } */
+/* { dg-final { scan-tree-dump "Folded into: if \\(_\[0-9\]+ == 3\\)" "vrp1" } } */
diff --git gcc/tree-vrp.c gcc/tree-vrp.c
index 5d085b4..4513b88 100644
--- gcc/tree-vrp.c
+++ gcc/tree-vrp.c
@@ -9467,7 +9467,8 @@ simplify_cond_using_ranges (gcond *stmt)
       innerop = gimple_assign_rhs1 (def_stmt);
 
       if (TREE_CODE (innerop) == SSA_NAME
-	  && !POINTER_TYPE_P (TREE_TYPE (innerop)))
+	  && !POINTER_TYPE_P (TREE_TYPE (innerop))
+         && desired_pro_or_demotion_p (innerop, op0))
 	{
 	  value_range *vr = get_value_range (innerop);
 
diff --git gcc/tree.h gcc/tree.h
index 0b9c3b9..91099ed 100644
--- gcc/tree.h
+++ gcc/tree.h
@@ -5316,4 +5316,18 @@ get_decl_source_range (tree decl)
   return get_range_from_loc (line_table, loc);
 }
 
+/* Return true if it makes sense to promote/demote from_type to to_type. */
+inline bool
+desired_pro_or_demotion_p (tree to_type, tree from_type)
+{
+  unsigned int to_type_precision = TYPE_PRECISION (TREE_TYPE (to_type));
+
+  /* OK to promote if to_type is no bigger than word_mode. */
+  if (to_type_precision <= GET_MODE_PRECISION (word_mode))
+    return true;
+
+  /* Otherwise, allow only if narrowing or same precision conversions. */
+  return to_type_precision <= TYPE_PRECISION (TREE_TYPE (from_type));
+}
+
 #endif  /* GCC_TREE_H  */

> 
> Thanks,
> Richard.
> 
> >  	{
> >  	  value_range *vr = get_value_range (innerop);
> >  
> > 
> > Regards
> > Senthil
> > > 
> > > Richard.
> > > 
> > > >Richard?
> > > >
> > > >Regards
> > > >Senthil
> > > >> 
> > > >> -- 
> > > >> Marc Glisse
> > > 
> > > 
> > 
> > 
> 
> -- 
> Richard Biener <rguenther@suse.de>
> SUSE LINUX GmbH, GF: Felix Imendoerffer, Jane Smithard, Graham Norton, HRB 21284 (AG Nuernberg)