public inbox for gcc-patches@gcc.gnu.org
 help / color / mirror / Atom feed
* [COMMITTED 2/5] Fix ranger when called from SCEV.
@ 2024-04-30 21:24 Andrew MacLeod
  2024-05-13 18:19 ` Jan-Benedict Glaw
  0 siblings, 1 reply; 4+ messages in thread
From: Andrew MacLeod @ 2024-04-30 21:24 UTC (permalink / raw)
  To: gcc-patches; +Cc: hernandez, aldy

[-- Attachment #1: Type: text/plain, Size: 705 bytes --]

Also during stage 3/4, we discovered that it is unsafe to call ranger 
from within SCEV.  This is because ranger uses SCEV to resolve PHIS, and 
we can end up in a bad loop under the right conditions.

The fix is for ranger's cache to NOT try to pre-evaluate PHIs (which is 
kind of waste anyway since they rarely resolve based on known incoming 
ranges to anything interesting).

Combined with this, it is now safe to make filling the block cache 
re-entrant (which can also happen if called from SCEV).

Turns out not wasting time pre-evaluating PHIs was also a time win.. so 
this patch also provides some modest speedups.

Bootstrapped on x86_64-pc-linux-gnu with no regressions.  pushed.

Andrew



[-- Attachment #2: 0002-Fix-ranger-when-called-from-SCEV.patch --]
[-- Type: text/x-patch, Size: 2431 bytes --]

From 83e95c10ed822270e39cb8da8c09f607ad65abbd Mon Sep 17 00:00:00 2001
From: Andrew MacLeod <amacleod@redhat.com>
Date: Wed, 13 Mar 2024 14:10:41 -0400
Subject: [PATCH 2/9] Fix ranger when called from SCEV.

Do not pre-evaluate PHIs in the cache, and allow fill_block_cache to
be re-entrant.  This allows SCEV to call into ranger with a context
and not produce cycles or loops.

	* gimple-range-cache.cc (ranger_cache::get_global_range): Do not
	pre-evaluate PHI nodes from the cache.
	(ranger_cache::fill_block_cache): Make re-entrant.
---
 gcc/gimple-range-cache.cc | 10 ++++++----
 1 file changed, 6 insertions(+), 4 deletions(-)

diff --git a/gcc/gimple-range-cache.cc b/gcc/gimple-range-cache.cc
index a33b7a73872..72ac2552311 100644
--- a/gcc/gimple-range-cache.cc
+++ b/gcc/gimple-range-cache.cc
@@ -1047,7 +1047,9 @@ ranger_cache::get_global_range (vrange &r, tree name, bool &current_p)
       if (r.varying_p () && !cfun->after_inlining)
 	{
 	  gimple *s = SSA_NAME_DEF_STMT (name);
-	  if (gimple_get_lhs (s) == name)
+	  // Do not process PHIs as SCEV may be in use and it can
+	  // spawn cyclic lookups.
+	  if (gimple_get_lhs (s) == name && !is_a<gphi *> (s))
 	    {
 	      if (!fold_range (r, s, get_global_range_query ()))
 		gimple_range_global (r, name);
@@ -1413,7 +1415,7 @@ ranger_cache::fill_block_cache (tree name, basic_block bb, basic_block def_bb)
 
   // At this point we shouldn't be looking at the def, entry block.
   gcc_checking_assert (bb != def_bb && bb != ENTRY_BLOCK_PTR_FOR_FN (cfun));
-  gcc_checking_assert (m_workback.length () == 0);
+  unsigned start_length = m_workback.length ();
 
   // If the block cache is set, then we've already visited this block.
   if (m_on_entry.bb_range_p (name, bb))
@@ -1500,7 +1502,7 @@ ranger_cache::fill_block_cache (tree name, basic_block bb, basic_block def_bb)
 	}
 
       m_on_entry.set_bb_range (name, bb, block_result);
-      gcc_checking_assert (m_workback.length () == 0);
+      gcc_checking_assert (m_workback.length () == start_length);
       return;
     }
 
@@ -1512,7 +1514,7 @@ ranger_cache::fill_block_cache (tree name, basic_block bb, basic_block def_bb)
   m_on_entry.set_bb_range (name, bb, undefined);
   gcc_checking_assert (m_update->empty_p ());
 
-  while (m_workback.length () > 0)
+  while (m_workback.length () > start_length)
     {
       basic_block node = m_workback.pop ();
       if (DEBUG_RANGE_CACHE)
-- 
2.41.0


^ permalink raw reply	[flat|nested] 4+ messages in thread
end of thread, other threads:[~2024-05-14 12:47 UTC | newest]

Thread overview: 4+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2024-04-30 21:24 [COMMITTED 2/5] Fix ranger when called from SCEV Andrew MacLeod
2024-05-13 18:19 ` Jan-Benedict Glaw
2024-05-13 18:27   ` Jan-Benedict Glaw
2024-05-14 12:47     ` Richard Biener

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).