Avoid SCC hashing on unmergeable trees

This is new incarantion of patch to identify unmergeable tree at streaming out
time rather than streaming in and to avoid pickling them to sccs with with hash
codes.

Building cc1 plus this patch reduces:

[WPA] read 4452927 SCCs of average size 1.986030
[WPA] 8843646 tree bodies read in total
[WPA] tree SCC table: size 524287, 205158 elements, collision ratio: 0.505204
[WPA] tree SCC max chain length 43 (size 1)
[WPA] Compared 947551 SCCs, 780270 collisions (0.823460)
[WPA] Merged 944038 SCCs
[WPA] Merged 5253521 tree bodies
[WPA] Merged 590027 types
...
[WPA] Size of mmap'd section decls: 99229066 bytes
[WPA] Size of mmap'd section function_body: 18398837 bytes
[WPA] Size of mmap'd section refs: 733678 bytes
[WPA] Size of mmap'd section jmpfuncs: 2965981 bytes
[WPA] Size of mmap'd section pureconst: 170248 bytes
[WPA] Size of mmap'd section profile: 17985 bytes
[WPA] Size of mmap'd section symbol_nodes: 3392736 bytes
[WPA] Size of mmap'd section inline: 2693920 bytes
[WPA] Size of mmap'd section icf: 435557 bytes
[WPA] Size of mmap'd section offload_table: 0 bytes
[WPA] Size of mmap'd section lto: 4320 bytes
[WPA] Size of mmap'd section ipa_sra: 651660 bytes

... to ...

[WPA] read 3312246 unshared trees
[WPA] read 1144381 mergeable SCCs of average size 4.833785
[WPA] 8843938 tree bodies read in total
[WPA] tree SCC table: size 524287, 197767 elements, collision ratio: 0.506446
[WPA] tree SCC max chain length 43 (size 1)
[WPA] Compared 946614 SCCs, 775077 collisions (0.818789)
[WPA] Merged 943798 SCCs
[WPA] Merged 5253336 tree bodies
[WPA] Merged 590105 types
....
[WPA] Size of mmap'd section decls: 81262144 bytes
[WPA] Size of mmap'd section function_body: 14702611 bytes
[WPA] Size of mmap'd section ext_symtab: 0 bytes
[WPA] Size of mmap'd section refs: 733695 bytes
[WPA] Size of mmap'd section jmpfuncs: 2332150 bytes
[WPA] Size of mmap'd section pureconst: 170292 bytes
[WPA] Size of mmap'd section profile: 17986 bytes
[WPA] Size of mmap'd section symbol_nodes: 3393358 bytes
[WPA] Size of mmap'd section inline: 2567939 bytes
[WPA] Size of mmap'd section icf: 435633 bytes
[WPA] Size of mmap'd section lto: 4320 bytes
[WPA] Size of mmap'd section ipa_sra: 651824 bytes

so results in about 22% reduction in global decl stream and 24% reduction on
function bodies stream (which is read mostly by ICF)

Martin, the zstd compression breaks the compression statistics (it works when
GCC is configured for zlib)

At first ltrans I get:

[LTRANS] Size of mmap'd section decls: 3734248 bytes
[LTRANS] Size of mmap'd section function_body: 4895962 bytes

... to ...

[LTRANS] Size of mmap'd section decls: 3479850 bytes
[LTRANS] Size of mmap'd section function_body: 3722935 bytes

So 7% reduction of global stream and 31% reduction of function bodies.

Stream in seems to get about 3% faster and stream out about 5% but it is
close to noise factor of my experiment.  I expect bigger speedups on
Firefox but I did not test it today since my Firefox setup broke again.
GCC is not very good example on the problem with anonymous namespace
types since we do not have so many of them.

Sice of object files in gcc directory is reduced by 11% (because hash
numbers do not compress well I guess).

The patch makes DFS walk to recognize trees that are not merged (anonymous
namespace, local function/variable decls, anonymous types etc).  As discussed
on IRC this is now done during the SCC walk rather than during the hash
computation.  When local tree is discovered we know that SCC components of everything that is on
the stack reffers to it and thus is also local. Moreover we mark trees into hash set in output block
so if we get a cross edge referring to local tree it gets marked too.

Patch also takes care of avoiding SCC wrappers around some trees. In particular
 1) singleton unmergeable SCCs are now streamed inline in global decl stream
    This includes INTEGER_CSTs and IDENTIFIER_NODEs that are shared by different
    code than rest of tree merging.
 2) We use LTO_trees instead of LTO_tree_scc to wrap unmergeable SCC components.
    It is still necessary to mark them because of forward references.  LTO_trees
    has simple header with number of trees and then things are streamed same way
    as for LTO_tree_scc. That is tree headers first followed by pickled references
    so things may point to future.

    Of course it is not necessary for LTO_tree_scc to be single component and
    streamer out may group more components together, but I decided to not snowball
    the patch even more
 3) In local streams when lto_output_tree is called and the topmost SCC components
    turns out to be singleton we stream the tree directly
    instead of LTO_tree_scc, hash code, pickled tree, reference to just stremaed tree.

    LTO_trees is used to wrap all trees needed to represent tree being streamed.
    It would make sense again to use only one LTO_trees rather than one per SCC
    but I think this can be done incrementally.

In general local trees are now recognized by new predicate local_tree_p

Bit subtle is handing of TRANLSATION_UNIT_DECL, INTEGER_CST and
IDENTIFIER_NODE.

TRANSLATION_UNIT_DECL a local tree but references to it does not make
other trees local (because we also understand local decls now).
So I check for it later after localness propagation is done.

INTEGER_CST and IDENTIFIER_NODEs are merged but not via the tree merging
machinery. So it makes sense to stream them as unmergeable trees but we
still need to compute their hashes so SCCs referring them do not get too
large collision chains.  For this reason they are checked just prior
stream out.

lto-bootstrapped/regteted x86_64-linux, OK?

gcc/ChangeLog:

2020-05-19  Jan Hubicka  <hubicka@ucw.cz>

	* lto-streamer-in.c (lto_input_scc): Add SHARED_SCC parameter.
	(lto_input_tree_1): Strenghten sanity check.
	(lto_input_tree): Update call of lto_input_scc.
	* lto-streamer-out.c: Include ipa-utils.h
	(create_output_block): Initialize local_trees if merigng is going
	to happen.
	(destroy_output_block): Destroy local_trees.
	(DFS): Add max_local_entry.
	(local_tree_p): New function.
	(DFS::DFS): Initialize and maintain it.
	(DFS::DFS_write_tree): Decide on streaming format.
	(lto_output_tree): Stream inline singleton SCCs
	* lto-streamer.h (enum LTO_tags): Add LTO_trees.
	(struct output_block): Add local_trees.
	(lto_input_scc): Update prototype.

gcc/lto/ChangeLog:

2020-05-19  Jan Hubicka  <hubicka@ucw.cz>

	* lto-common.c (compare_tree_sccs_1): Sanity check that we never
	read TRANSLATION_UNIT_DECL.
	(process_dref): Break out from ...
	(unify_scc): ... here.
	(process_new_tree): Break out from ...
	(lto_read_decls): ... here; handle streaming of singleton trees.
	(print_lto_report_1): Update statistics.

2 files changed, 154 insertions, 96 deletions

diff --git a/gcc/lto/ChangeLog b/gcc/lto/ChangeLog
index dfcb9da..aedcba9 100644
--- a/gcc/lto/ChangeLog
+++ b/gcc/lto/ChangeLog
@@ -1,3 +1,13 @@
+2020-05-20  Jan Hubicka  <hubicka@ucw.cz>
+
+	* lto-common.c (compare_tree_sccs_1): Sanity check that we never
+	read TRANSLATION_UNIT_DECL.
+	(process_dref): Break out from ...
+	(unify_scc): ... here.
+	(process_new_tree): Break out from ...
+	(lto_read_decls): ... here; handle streaming of singleton trees.
+	(print_lto_report_1): Update statistics.
+
 2020-05-07  Richard Biener  <rguenther@suse.de>
 
 	PR middle-end/94703
diff --git a/gcc/lto/lto-common.c b/gcc/lto/lto-common.c
index 682a82d..d04b1c9 100644
--- a/gcc/lto/lto-common.c
+++ b/gcc/lto/lto-common.c
@@ -56,6 +56,7 @@ along with GCC; see the file COPYING3.  If not see
 #include "builtins.h"
 #include "lto-common.h"
 #include "tree-pretty-print.h"
+#include "print-tree.h"
 
 /* True when no new types are going to be streamd from the global stream.  */
 
@@ -1054,6 +1055,7 @@ static unsigned long num_prevailing_types;
 static unsigned long num_type_scc_trees;
 static unsigned long total_scc_size;
 static unsigned long num_sccs_read;
+static unsigned long num_unshared_trees_read;
 static unsigned long total_scc_size_merged;
 static unsigned long num_sccs_merged;
 static unsigned long num_scc_compares;
@@ -1088,6 +1090,10 @@ compare_tree_sccs_1 (tree t1, tree t2, tree **map)
   compare_values (TREE_CODE);
   code = TREE_CODE (t1);
 
+  /* If we end up comparing translation unit decls we either forgot to mark
+     some SCC as local or we compare too much.  */
+  gcc_checking_assert (code != TRANSLATION_UNIT_DECL);
+
   if (!TYPE_P (t1))
     {
       compare_values (TREE_SIDE_EFFECTS);
@@ -1626,6 +1632,28 @@ cmp_tree (const void *p1_, const void *p2_)
   return ((uintptr_t)p1[1] < (uintptr_t)p2[1]) ? -1 : 1;
 }
 
+/* New scc of size 1 containing T was streamed in from DATA_IN and not merged.
+   Register it to reader cache at index FROM.  */
+
+static void
+process_dref (class data_in *data_in, tree t, unsigned from)
+{
+  struct streamer_tree_cache_d *cache = data_in->reader_cache;
+  /* If we got a debug reference queued, see if the prevailing
+     tree has a debug reference and if not, register the one
+     for the tree we are about to throw away.  */
+  if (dref_queue.length () == 1)
+    {
+      dref_entry e = dref_queue.pop ();
+      gcc_assert (e.decl
+		  == streamer_tree_cache_get_tree (cache, from));
+      const char *sym;
+      unsigned HOST_WIDE_INT off;
+      if (!debug_hooks->die_ref_for_decl (t, &sym, &off))
+	debug_hooks->register_external_die (t, e.sym, e.off);
+    }
+}
+
 /* Try to unify the SCC with nodes FROM to FROM + LEN in CACHE and
    hash value SCC_HASH with an already recorded SCC.  Return true if
    that was successful, otherwise return false.  */
@@ -1646,22 +1674,16 @@ unify_scc (class data_in *data_in, unsigned from,
     {
       tree t = streamer_tree_cache_get_tree (cache, from + i);
       scc->entries[i] = t;
-      /* Do not merge SCCs with local entities inside them.  Also do
-	 not merge TRANSLATION_UNIT_DECLs and anonymous namespaces
-	 and types therein types.  */
-      if (TREE_CODE (t) == TRANSLATION_UNIT_DECL
-	  || (VAR_OR_FUNCTION_DECL_P (t)
-	      && !(TREE_PUBLIC (t) || DECL_EXTERNAL (t)))
-	  || TREE_CODE (t) == LABEL_DECL
-	  || (TREE_CODE (t) == NAMESPACE_DECL && !DECL_NAME (t))
-	  || (TYPE_P (t)
-	      && type_with_linkage_p (TYPE_MAIN_VARIANT (t))
-	      && type_in_anonymous_namespace_p (TYPE_MAIN_VARIANT (t))))
-	{
-	  /* Avoid doing any work for these cases and do not worry to
-	     record the SCCs for further merging.  */
-	  return false;
-	}
+      /* These types should be streamed as unshared.  */
+      gcc_checking_assert
+	 (!(TREE_CODE (t) == TRANSLATION_UNIT_DECL
+	    || (VAR_OR_FUNCTION_DECL_P (t)
+		&& !(TREE_PUBLIC (t) || DECL_EXTERNAL (t)))
+	    || TREE_CODE (t) == LABEL_DECL
+	    || (TREE_CODE (t) == NAMESPACE_DECL && !DECL_NAME (t))
+	    || (TYPE_P (t)
+		&& type_with_linkage_p (TYPE_MAIN_VARIANT (t))
+		&& type_in_anonymous_namespace_p (TYPE_MAIN_VARIANT (t)))));
     }
 
   /* Look for the list of candidate SCCs to compare against.  */
@@ -1712,21 +1734,7 @@ unify_scc (class data_in *data_in, unsigned from,
 	     to the tree node mapping computed by compare_tree_sccs.  */
 	  if (len == 1)
 	    {
-	      /* If we got a debug reference queued, see if the prevailing
-		 tree has a debug reference and if not, register the one
-		 for the tree we are about to throw away.  */
-	      if (dref_queue.length () == 1)
-		{
-		  dref_entry e = dref_queue.pop ();
-		  gcc_assert (e.decl
-			      == streamer_tree_cache_get_tree (cache, from));
-		  const char *sym;
-		  unsigned HOST_WIDE_INT off;
-		  if (!debug_hooks->die_ref_for_decl (pscc->entries[0], &sym,
-						      &off))
-		    debug_hooks->register_external_die (pscc->entries[0],
-							e.sym, e.off);
-		}
+	      process_dref (data_in, pscc->entries[0], from);
 	      lto_maybe_register_decl (data_in, pscc->entries[0], from);
 	      streamer_tree_cache_replace_tree (cache, pscc->entries[0], from);
 	    }
@@ -1785,7 +1793,65 @@ unify_scc (class data_in *data_in, unsigned from,
   return unified_p;
 }
 
+typedef int_hash<unsigned, 0, UINT_MAX> code_id_hash;
+
+/* Do registering necessary once new tree fully streamed in (including all
+   trees it reffers to).  */
+
+static void
+process_new_tree (tree t, hash_map <code_id_hash, unsigned> *hm,
+		  unsigned index, unsigned *total, class data_in *data_in)
+{
+  /* Reconstruct the type variant and pointer-to/reference-to
+     chains.  */
+  if (TYPE_P (t))
+    {
+      /* Map the tree types to their frequencies.  */
+      if (flag_lto_dump_type_stats)
+	{
+	  unsigned key = (unsigned) TREE_CODE (t);
+	  unsigned *countp = hm->get (key);
+	  hm->put (key, countp ? (*countp) + 1 : 1);
+	  (*total)++;
+	}
+
+      num_prevailing_types++;
+      lto_fixup_prevailing_type (t);
 
+      /* Compute the canonical type of all non-ODR types.
+	 Delay ODR types for the end of merging process - the canonical
+	 type for those can be computed using the (unique) name however
+	 we want to do this only if units in other languages do not
+	 contain structurally equivalent type.
+
+	 Because SCC components are streamed in random (hash) order
+	 we may have encountered the type before while registering
+	 type canonical of a derived type in the same SCC.  */
+      if (!TYPE_CANONICAL (t))
+	{
+	  if (!RECORD_OR_UNION_TYPE_P (t)
+	      || !TYPE_CXX_ODR_P (t))
+	    gimple_register_canonical_type (t);
+	  else if (COMPLETE_TYPE_P (t))
+	    vec_safe_push (types_to_register, t);
+	}
+      if (TYPE_MAIN_VARIANT (t) == t && odr_type_p (t))
+	register_odr_type (t);
+    }
+  /* Link shared INTEGER_CSTs into TYPE_CACHED_VALUEs of its
+     type which is also member of this SCC.  */
+  if (TREE_CODE (t) == INTEGER_CST
+      && !TREE_OVERFLOW (t))
+    cache_integer_cst (t);
+  if (!flag_ltrans)
+    {
+      lto_maybe_register_decl (data_in, t, index);
+      /* Scan the tree for references to global functions or
+	 variables and record those for later fixup.  */
+      if (mentions_vars_p (t))
+	vec_safe_push (tree_with_vars, t);
+    }
+}
 
 /* Read all the symbols from buffer DATA, using descriptors in DECL_DATA.
    RESOLUTIONS is the set of symbols picked by the linker (read from the
@@ -1813,7 +1879,6 @@ lto_read_decls (struct lto_file_decl_data *decl_data, const void *data,
   /* We do not uniquify the pre-loaded cache entries, those are middle-end
      internal types that should not be merged.  */
 
-  typedef int_hash<unsigned, 0, UINT_MAX> code_id_hash;
   hash_map <code_id_hash, unsigned> hm;
   unsigned total = 0;
 
@@ -1824,31 +1889,41 @@ lto_read_decls (struct lto_file_decl_data *decl_data, const void *data,
       unsigned from = data_in->reader_cache->nodes.length ();
       /* Read and uniquify SCCs as in the input stream.  */
       enum LTO_tags tag = streamer_read_record_start (&ib_main);
-      if (tag == LTO_tree_scc)
+      if (tag == LTO_tree_scc || tag == LTO_trees)
 	{
 	  unsigned len_;
 	  unsigned scc_entry_len;
+
+	  /* Because we stream in SCC order we know that all unshared trees
+	     are now fully streamed.  Process them.  */
 	  hashval_t scc_hash = lto_input_scc (&ib_main, data_in, &len_,
-					      &scc_entry_len);
+					      &scc_entry_len,
+					      tag == LTO_tree_scc);
 	  unsigned len = data_in->reader_cache->nodes.length () - from;
 	  gcc_assert (len == len_);
 
-	  total_scc_size += len;
-	  num_sccs_read++;
+	  if (tag == LTO_tree_scc)
+	    {
+	      total_scc_size += len;
+	      num_sccs_read++;
+	    }
+	  else
+	    num_unshared_trees_read += len;
 
 	  /* We have the special case of size-1 SCCs that are pre-merged
 	     by means of identifier and string sharing for example.
 	     ???  Maybe we should avoid streaming those as SCCs.  */
 	  tree first = streamer_tree_cache_get_tree (data_in->reader_cache,
 						     from);
-	  if (len == 1
-	      && (TREE_CODE (first) == IDENTIFIER_NODE
-		  || (TREE_CODE (first) == INTEGER_CST
-		      && !TREE_OVERFLOW (first))))
-	    continue;
+	  /* Identifier and integers are shared specially, they should never
+	     go by the tree merging path.  */
+	  gcc_checking_assert ((TREE_CODE (first) != IDENTIFIER_NODE
+				&& (TREE_CODE (first) != INTEGER_CST
+				    || TREE_OVERFLOW (first)))
+			       || len != 1);
 
 	  /* Try to unify the SCC with already existing ones.  */
-	  if (!flag_ltrans
+	  if (!flag_ltrans && tag != LTO_trees
 	      && unify_scc (data_in, from,
 			    len, scc_entry_len, scc_hash))
 	    continue;
@@ -1862,56 +1937,9 @@ lto_read_decls (struct lto_file_decl_data *decl_data, const void *data,
 	    {
 	      tree t = streamer_tree_cache_get_tree (data_in->reader_cache,
 						     from + i);
-	      /* Reconstruct the type variant and pointer-to/reference-to
-		 chains.  */
+	      process_new_tree (t, &hm, from + i, &total, data_in);
 	      if (TYPE_P (t))
-		{
-		  /* Map the tree types to their frequencies.  */
-		  if (flag_lto_dump_type_stats)
-		    {
-		      unsigned key = (unsigned) TREE_CODE (t);
-		      unsigned *countp = hm.get (key);
-		      hm.put (key, countp ? (*countp) + 1 : 1);
-		      total++;
-		    }
-
-		  seen_type = true;
-		  num_prevailing_types++;
-		  lto_fixup_prevailing_type (t);
-
-		  /* Compute the canonical type of all non-ODR types.
-		     Delay ODR types for the end of merging process - the canonical
-		     type for those can be computed using the (unique) name however
-		     we want to do this only if units in other languages do not
-		     contain structurally equivalent type.
-
-		     Because SCC components are streamed in random (hash) order
-		     we may have encountered the type before while registering
-		     type canonical of a derived type in the same SCC.  */
-		  if (!TYPE_CANONICAL (t))
-		    {
-		      if (!RECORD_OR_UNION_TYPE_P (t)
-			  || !TYPE_CXX_ODR_P (t))
-		        gimple_register_canonical_type (t);
-		      else if (COMPLETE_TYPE_P (t))
-			vec_safe_push (types_to_register, t);
-		    }
-		  if (TYPE_MAIN_VARIANT (t) == t && odr_type_p (t))
-		    register_odr_type (t);
-		}
-	      /* Link shared INTEGER_CSTs into TYPE_CACHED_VALUEs of its
-		 type which is also member of this SCC.  */
-	      if (TREE_CODE (t) == INTEGER_CST
-		  && !TREE_OVERFLOW (t))
-		cache_integer_cst (t);
-	      if (!flag_ltrans)
-		{
-		  lto_maybe_register_decl (data_in, t, from + i);
-		  /* Scan the tree for references to global functions or
-		     variables and record those for later fixup.  */
-		  if (mentions_vars_p (t))
-		    vec_safe_push (tree_with_vars, t);
-		}
+		seen_type = true;
 	    }
 
 	  /* Register DECLs with the debuginfo machinery.  */
@@ -1926,9 +1954,26 @@ lto_read_decls (struct lto_file_decl_data *decl_data, const void *data,
 	}
       else
 	{
-	  /* Pickle stray references.  */
 	  t = lto_input_tree_1 (&ib_main, data_in, tag, 0);
-	  gcc_assert (t && data_in->reader_cache->nodes.length () == from);
+	  /* We streamed in new tree.  Add it to cache and process dref.  */
+	  if (data_in->reader_cache->nodes.length () == from + 1)
+	    {
+	      num_unshared_trees_read++;
+	      data_in->location_cache.accept_location_cache ();
+	      process_dref (data_in, t, from);
+	      if (TREE_CODE (t) == IDENTIFIER_NODE
+		  || (TREE_CODE (t) == INTEGER_CST
+		      && !TREE_OVERFLOW (t)))
+		;
+	      else
+		{
+		  lto_maybe_register_decl (data_in, t, from);
+		  process_new_tree (t, &hm, from, &total, data_in);
+		}
+	    }
+	  else
+	    /* FIXME: It seems useless to pickle stray references.  */
+	    gcc_assert (data_in->reader_cache->nodes.length () == from);
 	}
     }
 
@@ -2953,10 +2998,13 @@ print_lto_report_1 (void)
   const char *pfx = (flag_lto) ? "LTO" : (flag_wpa) ? "WPA" : "LTRANS";
   fprintf (stderr, "%s statistics\n", pfx);
 
-  fprintf (stderr, "[%s] read %lu SCCs of average size %f\n",
+  fprintf (stderr, "[%s] read %lu unshared trees\n",
+	   pfx, num_unshared_trees_read);
+  fprintf (stderr, "[%s] read %lu mergeable SCCs of average size %f\n",
 	   pfx, num_sccs_read, total_scc_size / (double)num_sccs_read);
-  fprintf (stderr, "[%s] %lu tree bodies read in total\n", pfx, total_scc_size);
-  if (flag_wpa && tree_scc_hash)
+  fprintf (stderr, "[%s] %lu tree bodies read in total\n", pfx,
+	   total_scc_size + num_unshared_trees_read);
+  if (flag_wpa && tree_scc_hash && num_sccs_read)
     {
       fprintf (stderr, "[%s] tree SCC table: size %ld, %ld elements, "
 	       "collision ratio: %f\n", pfx,