Patchwork fix for PR49888 var-tracking compile-time regression

login
register
mail settings
Submitter Alexandre Oliva
Date Jan. 16, 2013, 10:28 a.m.
Message ID <ord2x55slw.fsf@livre.localdomain>
Download mbox | patch
Permalink /patch/212477/
State New
Headers show

Comments

Alexandre Oliva - Jan. 16, 2013, 10:28 a.m.
PR49888 introduced clobber_overlapping_mems to detach VALUEs (and
variables bound to them) from MEMs as the MEMs are modified.  This
turned out to be quite expensive, particularly the computation of
canonical addresses passed to alias dependency.

This patch introduces caching of the canonical addresses to alleviate
this expensive operation.  We cache mappings that apply to the entire
function, from equivalences recorded in the global cselib table, and
mappings that apply only to the local basic block.  This cut down 2% of
a full regstrap cycle, and about 1% of the time it took to build target
libraries for C, C++ and Java.

Regstrapped on x86_64-linux-gnu and i686-linux-gnu.  Ok to install?
Jakub Jelinek - Jan. 16, 2013, 10:58 a.m.
On Wed, Jan 16, 2013 at 08:28:59AM -0200, Alexandre Oliva wrote:
> PR49888 introduced clobber_overlapping_mems to detach VALUEs (and
> variables bound to them) from MEMs as the MEMs are modified.  This
> turned out to be quite expensive, particularly the computation of
> canonical addresses passed to alias dependency.
> 
> This patch introduces caching of the canonical addresses to alleviate
> this expensive operation.  We cache mappings that apply to the entire
> function, from equivalences recorded in the global cselib table, and
> mappings that apply only to the local basic block.  This cut down 2% of
> a full regstrap cycle, and about 1% of the time it took to build target
> libraries for C, C++ and Java.

Can you safely cache the canon addresses already during vt_initialize
(when cselib_* is still processing new insns, cselib VALUEs contain
REGs and MEMs that are flushed at the end of processing the current bb
in vt_initialize)?  In my earlier attempts to cache something in
var-tracking, I've always started caching at the end of vt_initialize,
when the VALUEs should be (with one small exception of
variable_post_merge_new_vals created VALUEs) pretty much stable, everything
should be flushed that needs to be flushed, etc.

Also, what effects (if any) does the patch have on the
.debug_info/.debug_loc size and coverage?

	Jakub
Alexandre Oliva - Jan. 16, 2013, 1:25 p.m.
On Jan 16, 2013, Jakub Jelinek <jakub@redhat.com> wrote:

> On Wed, Jan 16, 2013 at 08:28:59AM -0200, Alexandre Oliva wrote:
>> PR49888 introduced clobber_overlapping_mems to detach VALUEs (and
>> variables bound to them) from MEMs as the MEMs are modified.  This
>> turned out to be quite expensive, particularly the computation of
>> canonical addresses passed to alias dependency.
>> 
>> This patch introduces caching of the canonical addresses to alleviate
>> this expensive operation.  We cache mappings that apply to the entire
>> function, from equivalences recorded in the global cselib table, and
>> mappings that apply only to the local basic block.  This cut down 2% of
>> a full regstrap cycle, and about 1% of the time it took to build target
>> libraries for C, C++ and Java.

> Can you safely cache the canon addresses already during vt_initialize
> (when cselib_* is still processing new insns, cselib VALUEs contain
> REGs and MEMs that are flushed at the end of processing the current bb
> in vt_initialize)?

No.  AFAICT we only call the address canonicalization function after
collecting all MOps, when the cselib table is fully constructed and
cleaned up from any local information, retaining only the global
equivalences.

> Also, what effects (if any) does the patch have on the
> .debug_info/.debug_loc size and coverage?

It shouldn't have any, since it's just caching results that would have
been recomputed over and over.  However, there's a possibility of being
“lucky” and recording an equivalence in the cache whose path would later
be removed from a dynamic set (say, if an incoming VALUE is reset and
re-bound within a block; I'm not sure this ever actually happens).  In
this case, these retained equivalences might enable alias analysis to
figure out that two memory refs do not overlap, and so one can be
retained in a dynamic equivalence list when we process a MOp that
modifies the other.  Or something ;-) It shouldn't really make any
difference, just speed things up a bit.  Paraphrasing Knuth, “I proved
it, but I didn't test it” ;-)
Jakub Jelinek - Jan. 16, 2013, 4:54 p.m.
On Wed, Jan 16, 2013 at 11:25:46AM -0200, Alexandre Oliva wrote:
> > Also, what effects (if any) does the patch have on the
> > .debug_info/.debug_loc size and coverage?
> 
> It shouldn't have any, since it's just caching results that would have
> been recomputed over and over.  However, there's a possibility of being
> “lucky” and recording an equivalence in the cache whose path would later
> be removed from a dynamic set (say, if an incoming VALUE is reset and
> re-bound within a block; I'm not sure this ever actually happens).  In
> this case, these retained equivalences might enable alias analysis to
> figure out that two memory refs do not overlap, and so one can be
> retained in a dynamic equivalence list when we process a MOp that
> modifies the other.  Or something ;-) It shouldn't really make any
> difference, just speed things up a bit.  Paraphrasing Knuth, “I proved
> it, but I didn't test it” ;-)

Ok, seems it is almost no change, but if I do between
--enable-languages=all,obj-c++,ada,lto,go --enable-checking=yes,rtl
x86_64-linux and i686-linux (the latter without ,ada) trees (once without
your var-tracking.c patch, once with) readelf -WS comparison of all the
gcc/*.o gcc/*/*.o files from stage3, I get differences beyond
var-tracking.o (which is of course expected):

for i686-linux tree-ssa-pre.o is different, and
for x86_64-linux go/export.o is different.
All other objects have the same readelf -WS output (not comparing .debug_*
section data, as that could be slightly different as I haven't done the
build with exactly the same object directory (different dirname, but same
length thereof)).  If I strip those two object files, then they are
identical between the two trees.  There are some differences in .debug_loc
in both cases.

	Jakub

Patch

Cache canonical addresses within VTA

From: Alexandre Oliva <aoliva@redhat.com>

for  gcc/ChangeLog

	PR debug/54114
	PR debug/54402
	PR debug/49888
	* var-tracking.c (negative_power_of_two_p): New.
	(global_get_addr_cache, local_get_addr_cache): New.
	(get_addr_from_global_cache, get_addr_from_local_cache): New.
	(vt_canonicalize_addr): Rewrite using the above.
	(local_get_addr_clear_given_value): New.
	(val_reset): Clear local cached entries.
	(compute_bb_dataflow): Create and release the local cache.
	(vt_emit_notes): Likewise.
	(vt_initialize, vt_finalize): Create and release the global
	cache, respectively.
---

 gcc/var-tracking.c |  243 +++++++++++++++++++++++++++++++++++++++++++---------
 1 files changed, 200 insertions(+), 43 deletions(-)


diff --git a/gcc/var-tracking.c b/gcc/var-tracking.c
index 8c76fe0..befde0a 100644
--- a/gcc/var-tracking.c
+++ b/gcc/var-tracking.c
@@ -1948,6 +1948,14 @@  var_regno_delete (dataflow_set *set, int regno)
   *reg = NULL;
 }
 
+/* Return true if I is the negated value of a power of two.  */
+static bool
+negative_power_of_two_p (HOST_WIDE_INT i)
+{
+  unsigned HOST_WIDE_INT x = -(unsigned HOST_WIDE_INT)i;
+  return x == (x & -x);
+}
+
 /* Strip constant offsets and alignments off of LOC.  Return the base
    expression.  */
 
@@ -1958,12 +1966,123 @@  vt_get_canonicalize_base (rtx loc)
 	  || GET_CODE (loc) == AND)
 	 && GET_CODE (XEXP (loc, 1)) == CONST_INT
 	 && (GET_CODE (loc) != AND
-	     || INTVAL (XEXP (loc, 1)) < 0))
+	     || negative_power_of_two_p (INTVAL (XEXP (loc, 1)))))
     loc = XEXP (loc, 0);
 
   return loc;
 }
 
+/* This caches canonicalized addresses for VALUEs, computed using
+   information in the global cselib table.  */
+static struct pointer_map_t *global_get_addr_cache;
+
+/* This caches canonicalized addresses for VALUEs, computed using
+   information from the global cache and information pertaining to a
+   basic block being analyzed.  */
+static struct pointer_map_t *local_get_addr_cache;
+
+static rtx vt_canonicalize_addr (dataflow_set *, rtx);
+
+/* Return the canonical address for LOC, that must be a VALUE, using a
+   cached global equivalence or computing it and storing it in the
+   global cache.  */
+
+static rtx
+get_addr_from_global_cache (rtx const loc)
+{
+  rtx x;
+  void **slot;
+
+  gcc_checking_assert (GET_CODE (loc) == VALUE);
+  
+  slot = pointer_map_insert (global_get_addr_cache, loc);
+  if (*slot)
+    return (rtx)*slot;
+
+  x = canon_rtx (get_addr (loc));
+
+  /* Tentative, avoiding infinite recursion.  */
+  *slot = x;
+
+  if (x != loc)
+    {
+      rtx nx = vt_canonicalize_addr (NULL, x);
+      if (nx != x)
+	{
+	  /* The table may have moved during recursion, recompute
+	     SLOT.  */
+	  slot = pointer_map_contains (global_get_addr_cache, loc);
+	  *slot = x = nx;
+	}
+    }
+
+  return x;
+}
+
+/* Return the canonical address for LOC, that must be a VALUE, using a
+   cached local equivalence or computing it and storing it in the
+   local cache.  */
+
+static rtx
+get_addr_from_local_cache (dataflow_set *set, rtx const loc)
+{
+  rtx x;
+  void **slot;
+  decl_or_value dv;
+  variable var;
+  location_chain l;
+
+  gcc_checking_assert (GET_CODE (loc) == VALUE);
+  
+  slot = pointer_map_insert (local_get_addr_cache, loc);
+  if (*slot)
+    return (rtx)*slot;
+
+  x = get_addr_from_global_cache (loc);
+  
+  /* Tentative, avoiding infinite recursion.  */
+  *slot = x;
+
+  /* Recurse to cache local expansion of X, or if we need to search
+     for a VALUE in the expansion.  */
+  if (x != loc)
+    {
+      rtx nx = vt_canonicalize_addr (set, x);
+      if (nx != x)
+	{
+	  slot = pointer_map_contains (local_get_addr_cache, loc);
+	  *slot = x = nx;
+	}
+      return x;
+    }
+
+  dv = dv_from_rtx (x);
+  var = (variable) htab_find_with_hash (shared_hash_htab (set->vars),
+					dv, dv_htab_hash (dv));
+  if (!var)
+    return x;
+
+  /* Look for an improved equivalent expression.  */
+  for (l = var->var_part[0].loc_chain; l; l = l->next)
+    {
+      rtx base = vt_get_canonicalize_base (l->loc);
+      if (GET_CODE (base) == REG
+	  || (GET_CODE (base) == VALUE
+	      && canon_value_cmp (base, loc)))
+	{
+	  rtx nx = vt_canonicalize_addr (set, l->loc);
+	  if (x != nx)
+	    {
+	      slot = pointer_map_contains (local_get_addr_cache, loc);
+	      *slot = x = nx;
+	    }
+	  break;
+	}
+    }
+
+  return x;
+}
+
 /* Canonicalize LOC using equivalences from SET in addition to those
    in the cselib static table.  */
 
@@ -1972,69 +2091,56 @@  vt_canonicalize_addr (dataflow_set *set, rtx oloc)
 {
   HOST_WIDE_INT ofst = 0;
   enum machine_mode mode = GET_MODE (oloc);
-  rtx loc = canon_rtx (get_addr (oloc));
+  rtx loc = oloc;
+  rtx x;
+  bool retry = true;
 
-  /* Try to substitute a base VALUE for equivalent expressions as much
-     as possible.  The goal here is to expand stack-related addresses
-     to one of the stack base registers, so that we can compare
-     addresses for overlaps.  */
-  while (GET_CODE (vt_get_canonicalize_base (loc)) == VALUE)
+  while (retry)
     {
-      rtx x;
-      decl_or_value dv;
-      variable var;
-      location_chain l;
-
-      while (GET_CODE (loc) == PLUS)
+      while (GET_CODE (loc) == PLUS
+	     && GET_CODE (XEXP (loc, 1)) == CONST_INT)
 	{
 	  ofst += INTVAL (XEXP (loc, 1));
 	  loc = XEXP (loc, 0);
-	  continue;
 	}
 
       /* Alignment operations can't normally be combined, so just
 	 canonicalize the base and we're done.  We'll normally have
 	 only one stack alignment anyway.  */
-      if (GET_CODE (loc) == AND)
+      if (GET_CODE (loc) == AND
+	  && GET_CODE (XEXP (loc, 1)) == CONST_INT
+	  && negative_power_of_two_p (INTVAL (XEXP (loc, 1))))
 	{
 	  x = vt_canonicalize_addr (set, XEXP (loc, 0));
 	  if (x != XEXP (loc, 0))
 	    loc = gen_rtx_AND (mode, x, XEXP (loc, 1));
-	  loc = canon_rtx (get_addr (loc));
-	  break;
+	  retry = false;
 	}
 
-      x = canon_rtx (get_addr (loc));
-
-      /* We've made progress!  Start over.  */
-      if (x != loc || GET_CODE (x) != VALUE)
+      if (GET_CODE (loc) == VALUE)
 	{
-	  loc = x;
-	  continue;
-	}
-
-      dv = dv_from_rtx (x);
-      var = (variable) htab_find_with_hash (shared_hash_htab (set->vars),
-					    dv, dv_htab_hash (dv));
-      if (!var)
-	break;
+	  if (set)
+	    loc = get_addr_from_local_cache (set, loc);
+	  else
+	    loc = get_addr_from_global_cache (loc);
 
-      /* Look for an improved equivalent expression.  */
-      for (l = var->var_part[0].loc_chain; l; l = l->next)
-	{
-	  rtx base = vt_get_canonicalize_base (l->loc);
-	  if (GET_CODE (base) == REG
-	      || (GET_CODE (base) == VALUE
-		  && canon_value_cmp (base, loc)))
+	  /* Consolidate plus_constants.  */
+	  while (ofst && GET_CODE (loc) == PLUS
+		 && GET_CODE (XEXP (loc, 1)) == CONST_INT)
 	    {
-	      loc = l->loc;
-	      break;
+	      ofst += INTVAL (XEXP (loc, 1));
+	      loc = XEXP (loc, 0);
 	    }
-	}
 
-      /* No luck with the dataflow set, so we're done.  */
-      if (!l)
-	break;
+	  retry = false;
+	}
+      else
+	{
+	  x = canon_rtx (loc);
+	  if (retry)
+	    retry = (x != loc);
+	  loc = x;
+	}
     }
 
   /* Add OFST back in.  */
@@ -2358,6 +2464,17 @@  val_store (dataflow_set *set, rtx val, rtx loc, rtx insn, bool modified)
   val_bind (set, val, loc, modified);
 }
 
+/* Clear (canonical address) slots that reference X.  */
+
+static bool
+local_get_addr_clear_given_value (const void *v ATTRIBUTE_UNUSED,
+				  void **slot, void *x)
+{
+  if (vt_get_canonicalize_base ((rtx)*slot) == x)
+    *slot = NULL;
+  return true;
+}
+
 /* Reset this node, detaching all its equivalences.  Return the slot
    in the variable hash table that holds dv, if there is one.  */
 
@@ -2373,6 +2490,28 @@  val_reset (dataflow_set *set, decl_or_value dv)
 
   gcc_assert (var->n_var_parts == 1);
 
+  if (var->onepart == ONEPART_VALUE)
+    {
+      rtx x = dv_as_value (dv);
+      void **slot;
+      
+      /* Relationships in the global cache don't change, so reset the
+	 local cache entry only.  */
+      slot = pointer_map_contains (local_get_addr_cache, x);
+      if (slot)
+	{
+	  /* If the value resolved back to itself, odds are that other
+	     values may have cached it too.  These entries now refer
+	     to the old X, so detach them too.  Entries that used the
+	     old X but resolved to something else remain ok as long as
+	     that something else isn't also reset.  */
+	  if (*slot == x)
+	    pointer_map_traverse (local_get_addr_cache,
+				  local_get_addr_clear_given_value, x);
+	  *slot = NULL;
+	}
+    }
+
   cval = NULL;
   for (node = var->var_part[0].loc_chain; node; node = node->next)
     if (GET_CODE (node->loc) == VALUE
@@ -6420,6 +6559,9 @@  compute_bb_dataflow (basic_block bb)
   dataflow_set_copy (&old_out, out);
   dataflow_set_copy (out, in);
 
+  if (MAY_HAVE_DEBUG_INSNS)
+    local_get_addr_cache = pointer_map_create ();
+
   FOR_EACH_VEC_ELT (VTI (bb)->mos, i, mo)
     {
       rtx insn = mo->insn;
@@ -6696,6 +6838,9 @@  compute_bb_dataflow (basic_block bb)
 
   if (MAY_HAVE_DEBUG_INSNS)
     {
+      pointer_map_destroy (local_get_addr_cache);
+      local_get_addr_cache = NULL;
+
       dataflow_set_equiv_regs (out);
       htab_traverse (shared_hash_htab (out->vars), canonicalize_values_mark,
 		     out);
@@ -9236,9 +9381,16 @@  vt_emit_notes (void)
 	 subsequent basic blocks.  */
       emit_notes_for_differences (BB_HEAD (bb), &cur, &VTI (bb)->in);
 
+      if (MAY_HAVE_DEBUG_INSNS)
+	local_get_addr_cache = pointer_map_create ();
+
       /* Emit the notes for the changes in the basic block itself.  */
       emit_notes_in_bb (bb, &cur);
 
+      if (MAY_HAVE_DEBUG_INSNS)
+	pointer_map_destroy (local_get_addr_cache);
+      local_get_addr_cache = NULL;
+
       /* Free memory occupied by the in hash table, we won't need it
 	 again.  */
       dataflow_set_clear (&VTI (bb)->in);
@@ -9607,11 +9759,13 @@  vt_initialize (void)
       valvar_pool = create_alloc_pool ("small variable_def pool",
 				       sizeof (struct variable_def), 256);
       preserved_values.create (256);
+      global_get_addr_cache = pointer_map_create ();
     }
   else
     {
       scratch_regs = NULL;
       valvar_pool = NULL;
+      global_get_addr_cache = NULL;
     }
 
   if (MAY_HAVE_DEBUG_INSNS)
@@ -9948,6 +10102,9 @@  vt_finalize (void)
 
   if (MAY_HAVE_DEBUG_INSNS)
     {
+      if (global_get_addr_cache)
+	pointer_map_destroy (global_get_addr_cache);
+      global_get_addr_cache = NULL;
       if (loc_exp_dep_pool)
 	free_alloc_pool (loc_exp_dep_pool);
       loc_exp_dep_pool = NULL;