* [PATCH] tree-optimization/97307 - improve sinking of loads
@ 2020-10-07 14:55 Richard Biener
0 siblings, 0 replies; only message in thread
From: Richard Biener @ 2020-10-07 14:55 UTC (permalink / raw)
To: gcc-patches
This improves the heuristics finding a sink location for loads that does
not cross any store.
Bootstrapped and tested on x86_64-unknown-linux-gnu, pushed.
2020-10-07 Richard Biener <rguenther@suse.de>
PR tree-optimization/97307
* tree-ssa-sink.c (statement_sink_location): Change heuristic
for not skipping stores to look for virtual definitions
rather than uses.
* gcc.dg/tree-ssa/ssa-sink-17.c: New testcase.
* gcc.dg/vect/pr65947-3.c: XFAIL.
---
gcc/testsuite/gcc.dg/tree-ssa/ssa-sink-17.c | 15 +++++++
gcc/testsuite/gcc.dg/vect/pr65947-3.c | 5 ++-
gcc/tree-ssa-sink.c | 45 +++++++++++----------
3 files changed, 43 insertions(+), 22 deletions(-)
create mode 100644 gcc/testsuite/gcc.dg/tree-ssa/ssa-sink-17.c
diff --git a/gcc/testsuite/gcc.dg/tree-ssa/ssa-sink-17.c b/gcc/testsuite/gcc.dg/tree-ssa/ssa-sink-17.c
new file mode 100644
index 00000000000..cf2e2a0f766
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/tree-ssa/ssa-sink-17.c
@@ -0,0 +1,15 @@
+/* PR tree-optimization/97307 */
+/* { dg-do compile } */
+/* { dg-options "-O2 -fdump-tree-sink-details" } */
+
+int pure_f(int a, int b) __attribute__((pure));
+int my_f(int a, int b)
+{
+ int x = pure_f(a, b);
+ if (a > 0)
+ return x;
+ return a;
+}
+
+/* We should sink the call to pure_f to the if block. */
+/* { dg-final { scan-tree-dump "Sinking # VUSE" "sink" } } */
diff --git a/gcc/testsuite/gcc.dg/vect/pr65947-3.c b/gcc/testsuite/gcc.dg/vect/pr65947-3.c
index 6b4077e1a62..8a2608cf0f1 100644
--- a/gcc/testsuite/gcc.dg/vect/pr65947-3.c
+++ b/gcc/testsuite/gcc.dg/vect/pr65947-3.c
@@ -51,6 +51,9 @@ main (void)
return 0;
}
-/* { dg-final { scan-tree-dump-times "LOOP VECTORIZED" 2 "vect" } } */
+/* XFAILed because of the fix for PR97307 which sinks the load of a[i], preventing
+ if-conversion to happen. */
+/* { dg-final { scan-tree-dump-times "LOOP VECTORIZED" 2 "vect" { xfail *-*-* } } } */
+/* { dg-final { scan-tree-dump-times "LOOP VECTORIZED" 1 "vect" } } */
/* { dg-final { scan-tree-dump-times "optimizing condition reduction with FOLD_EXTRACT_LAST" 2 "vect" { target vect_fold_extract_last } } } */
/* { dg-final { scan-tree-dump-not "condition expression based on integer induction." "vect" } } */
diff --git a/gcc/tree-ssa-sink.c b/gcc/tree-ssa-sink.c
index 4cc5195f2f8..ba8e5577e9c 100644
--- a/gcc/tree-ssa-sink.c
+++ b/gcc/tree-ssa-sink.c
@@ -369,10 +369,9 @@ statement_sink_location (gimple *stmt, basic_block frombb,
return false;
/* If this is a load then do not sink past any stores.
- ??? This is overly simple but cheap. We basically look
- for an existing load with the same VUSE in the path to one
- of the sink candidate blocks and we adjust commondom to the
- nearest to commondom. */
+ Look for virtual definitions in the path from frombb to the sink
+ location computed from the real uses and if found, adjust
+ that it a common dominator. */
if (gimple_vuse (stmt))
{
/* Do not sink loads from hard registers. */
@@ -383,29 +382,33 @@ statement_sink_location (gimple *stmt, basic_block frombb,
imm_use_iterator imm_iter;
use_operand_p use_p;
- basic_block found = NULL;
FOR_EACH_IMM_USE_FAST (use_p, imm_iter, gimple_vuse (stmt))
{
gimple *use_stmt = USE_STMT (use_p);
basic_block bb = gimple_bb (use_stmt);
- /* For PHI nodes the block we know sth about
- is the incoming block with the use. */
+ /* For PHI nodes the block we know sth about is the incoming block
+ with the use. */
if (gimple_code (use_stmt) == GIMPLE_PHI)
- bb = EDGE_PRED (bb, PHI_ARG_INDEX_FROM_USE (use_p))->src;
- /* Any dominator of commondom would be ok with
- adjusting commondom to that block. */
- bb = nearest_common_dominator (CDI_DOMINATORS, bb, commondom);
- if (!found)
- found = bb;
- else if (dominated_by_p (CDI_DOMINATORS, bb, found))
- found = bb;
- /* If we can't improve, stop. */
- if (found == commondom)
- break;
+ {
+ /* In case the PHI node post-dominates the current insert location
+ we can disregard it. */
+ if (commondom != bb
+ && dominated_by_p (CDI_POST_DOMINATORS, commondom, bb))
+ continue;
+ bb = EDGE_PRED (bb, PHI_ARG_INDEX_FROM_USE (use_p))->src;
+ }
+ else if (!gimple_vdef (use_stmt))
+ continue;
+ /* If the use is not dominated by the path entry it is not on
+ the path. */
+ if (!dominated_by_p (CDI_DOMINATORS, bb, frombb))
+ continue;
+ /* There is no easy way to disregard defs not on the path from
+ frombb to commondom so just consider them all. */
+ commondom = nearest_common_dominator (CDI_DOMINATORS, bb, commondom);
+ if (commondom == frombb)
+ return false;
}
- commondom = found;
- if (commondom == frombb)
- return false;
}
/* Our common dominator has to be dominated by frombb in order to be a
--
2.26.2
^ permalink raw reply [flat|nested] only message in thread
only message in thread, other threads:[~2020-10-07 14:55 UTC | newest]
Thread overview: (only message) (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2020-10-07 14:55 [PATCH] tree-optimization/97307 - improve sinking of loads 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).