public inbox for gcc-patches@gcc.gnu.org
 help / color / mirror / Atom feed
* [PATCH] [i386] Prevent vectorization for load from parm_decl at O2 to avoid STF issue.
@ 2022-03-04  7:27 liuhongt
  2022-03-07  8:26 ` Hongtao Liu
  2022-03-07  9:37 ` Richard Biener
  0 siblings, 2 replies; 9+ messages in thread
From: liuhongt @ 2022-03-04  7:27 UTC (permalink / raw)
  To: gcc-patches

For parameter passing through stack, vectorized load from parm_decl
in callee may trigger serious STF issue. This is why GCC12 regresses
50% for cray at -O2 compared to GCC11.

The patch add an extremely large number to stmt_cost to prevent
vectorization for loads from parm_decl under very-cheap cost model,
this can at least prevent O2 regression due to STF issue, but may lose
some perf where there's no such issue(1 vector_load vs n scalar_load +
CTOR).

No impact for SPEC2017 for both plain O2 and native O2 on ICX.
Bootstrapped and regtested on x86_64-linux-gnu{-m32,}.
Ok for trunk?

gcc/ChangeLog:

	PR target/101908
	* config/i386/i386.cc (ix86_load_maybe_stfs_p): New.
	(ix86_vector_costs::add_stmt_cost): Add extra cost for
	vector_load/unsigned_load which may have stall forward issue.

gcc/testsuite/ChangeLog:

	* gcc.target/i386/pr101908-1.c: New test.
	* gcc.target/i386/pr101908-2.c: New test.
---
 gcc/config/i386/i386.cc                    | 31 ++++++++++++++++++++++
 gcc/testsuite/gcc.target/i386/pr101908-1.c | 12 +++++++++
 gcc/testsuite/gcc.target/i386/pr101908-2.c | 12 +++++++++
 3 files changed, 55 insertions(+)
 create mode 100644 gcc/testsuite/gcc.target/i386/pr101908-1.c
 create mode 100644 gcc/testsuite/gcc.target/i386/pr101908-2.c

diff --git a/gcc/config/i386/i386.cc b/gcc/config/i386/i386.cc
index b2bf90576d5..3bbaaf65ea8 100644
--- a/gcc/config/i386/i386.cc
+++ b/gcc/config/i386/i386.cc
@@ -22976,6 +22976,19 @@ ix86_noce_conversion_profitable_p (rtx_insn *seq, struct noce_if_info *if_info)
   return default_noce_conversion_profitable_p (seq, if_info);
 }
 
+/* Return true if REF may have STF issue, otherwise false.  */
+static bool
+ix86_load_maybe_stfs_p (tree ref)
+{
+  tree addr = get_base_address (ref);
+
+  if (TREE_CODE (addr) != PARM_DECL
+      || !tree_fits_uhwi_p (TYPE_SIZE (TREE_TYPE (addr)))
+      || tree_to_uhwi (TYPE_SIZE (TREE_TYPE (addr))) <= MAX_BITS_PER_WORD)
+    return false;
+  return true;
+}
+
 /* x86-specific vector costs.  */
 class ix86_vector_costs : public vector_costs
 {
@@ -23203,6 +23216,24 @@ ix86_vector_costs::add_stmt_cost (int count, vect_cost_for_stmt kind,
 	if (TREE_CODE (op) == SSA_NAME)
 	  TREE_VISITED (op) = 0;
     }
+
+  /* Prevent vectorization for load from parm_decl at O2 to avoid STF issue.
+     Performance may lose when there's no STF issue(1 vector_load vs n
+     scalar_load + CTOR).
+     TODO: both extra cost(2000) and ix86_load_maybe_stfs_p need to be fine
+     tuned.  */
+  if ((kind == vector_load || kind == unaligned_load)
+      && flag_vect_cost_model == VECT_COST_MODEL_VERY_CHEAP
+      && stmt_info
+      && stmt_info->slp_type == pure_slp
+      && stmt_info->stmt
+      && gimple_assign_load_p (stmt_info->stmt)
+      && ix86_load_maybe_stfs_p (gimple_assign_rhs1 (stmt_info->stmt)))
+    {
+      stmt_cost = ix86_builtin_vectorization_cost (kind, vectype, misalign);
+      stmt_cost += 2000;
+    }
+
   if (stmt_cost == -1)
     stmt_cost = ix86_builtin_vectorization_cost (kind, vectype, misalign);
 
diff --git a/gcc/testsuite/gcc.target/i386/pr101908-1.c b/gcc/testsuite/gcc.target/i386/pr101908-1.c
new file mode 100644
index 00000000000..f8e0f2e26bb
--- /dev/null
+++ b/gcc/testsuite/gcc.target/i386/pr101908-1.c
@@ -0,0 +1,12 @@
+/* { dg-do compile } */
+/* { dg-options "-O2 -fdump-tree-slp-details" } */
+/* { dg-final { scan-tree-dump {(?n)add new stmt:.*MEM \<vector\(2\) double\>} "slp2" } } */
+
+struct X { double x[2]; };
+typedef double v2df __attribute__((vector_size(16)));
+
+v2df __attribute__((noipa))
+foo (struct X* x, struct X* y)
+{
+  return (v2df) {x->x[1], x->x[0] } + (v2df) { y->x[1], y->x[0] };
+}
diff --git a/gcc/testsuite/gcc.target/i386/pr101908-2.c b/gcc/testsuite/gcc.target/i386/pr101908-2.c
new file mode 100644
index 00000000000..7f2f00cebab
--- /dev/null
+++ b/gcc/testsuite/gcc.target/i386/pr101908-2.c
@@ -0,0 +1,12 @@
+/* { dg-do compile } */
+/* { dg-options "-O2 -fdump-tree-slp-details" } */
+/* { dg-final { scan-tree-dump-not {(?n)add new stmt:.*MEM \<vector\(2\) double\>} "slp2" } } */
+
+struct X { double x[2]; };
+typedef double v2df __attribute__((vector_size(16)));
+
+v2df __attribute__((noipa))
+foo (struct X x, struct X y)
+{
+  return (v2df) {x.x[1], x.x[0] } + (v2df) { y.x[1], y.x[0] };
+}
-- 
2.18.1


^ permalink raw reply	[flat|nested] 9+ messages in thread
end of thread, other threads:[~2022-03-17  7:12 UTC | newest]

Thread overview: 9+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2022-03-04  7:27 [PATCH] [i386] Prevent vectorization for load from parm_decl at O2 to avoid STF issue liuhongt
2022-03-07  8:26 ` Hongtao Liu
2022-03-07  9:37 ` Richard Biener
2022-03-08  3:39   ` Hongtao Liu
2022-03-15  9:13     ` [PATCH] [i386] Add extra cost for unsigned_load which may have stall forward issue liuhongt
2022-03-16  1:03     ` liuhongt
2022-03-16  2:19     ` liuhongt
2022-03-16  9:54       ` Richard Biener
2022-03-17  7:12         ` Hongtao Liu

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).