From: Richard Biener <rguenther@suse.de>
Date: Thu, 26 Mar 2020 07:33:57 +0000 (+0100)
Subject: widening_mul: restrict ops to be defined in the same basic-block when convert plusmin... 
X-Git-Url: https://git.libre-soc.org/?a=commitdiff_plain;h=d21dff5b4fee51ae432143065bededfc763dc344;p=gcc.git

widening_mul: restrict ops to be defined in the same basic-block when convert plusminus to widen

In the testcase for PR94269, widening_mul moves two multiply
instructions from outside the loop to inside
the loop, merging with two add instructions separately.  This
increases the cost of the loop.  Like FMA detection
in the same pass, simply restrict ops to be defined in the same
basic-block to avoid possibly moving multiply
to a different block with a higher execution frequency.

2020-03-26  Felix Yang  <felix.yang@huawei.com>

	PR tree-optimization/94269
	* tree-ssa-math-opts.c (convert_plusminus_to_widen): Restrict
	this
	operation to single basic block.

	* gcc.dg/pr94269.c: New test.
---

diff --git a/gcc/ChangeLog b/gcc/ChangeLog
index 3ad7a7aae0d..9f7927d748f 100644
--- a/gcc/ChangeLog
+++ b/gcc/ChangeLog
@@ -1,3 +1,10 @@
+2020-03-26  Felix Yang  <felix.yang@huawei.com>
+
+	PR tree-optimization/94269
+	* tree-ssa-math-opts.c (convert_plusminus_to_widen): Restrict
+	this
+	operation to single basic block.
+
 2020-03-25  Jeff Law  <law@redhat.com>
 
 	PR rtl-optimization/90275
diff --git a/gcc/testsuite/ChangeLog b/gcc/testsuite/ChangeLog
index e750dcbc790..f8512d3fb32 100644
--- a/gcc/testsuite/ChangeLog
+++ b/gcc/testsuite/ChangeLog
@@ -1,3 +1,8 @@
+2020-03-26  Felix Yang  <felix.yang@huawei.com>
+
+	PR tree-optimization/94269
+	* gcc.dg/pr94269.c: New test.
+
 2020-03-25  Andrew Stubbs  <ams@codesourcery.com>
 
 	* gcc.dg/vect/bb-slp-pr69907.c: Disable the dump scan for amdgcn.
diff --git a/gcc/testsuite/gcc.dg/pr94269.c b/gcc/testsuite/gcc.dg/pr94269.c
new file mode 100644
index 00000000000..49d57044f78
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/pr94269.c
@@ -0,0 +1,26 @@
+/* { dg-do compile { target aarch64*-*-* } } */
+/* { dg-options "-O2 -ftree-loop-vectorize -funsafe-math-optimizations -march=armv8.2-a+sve -msve-vector-bits=256" } */
+
+float
+foo(long n, float *x, int inc_x,
+            float *y, int inc_y)
+{
+  float dot = 0.0;
+  int ix = 0, iy = 0;
+
+  if (n < 0) {
+    return dot;
+  }
+
+  int i = 0;
+  while (i < n) {
+    dot += y[iy] * x[ix];
+    ix  += inc_x;
+    iy  += inc_y;
+    i++;
+  }
+
+  return dot;
+}
+
+/* { dg-final { scan-assembler-not "smaddl" { target aarch64*-*-* } } } */
diff --git a/gcc/tree-ssa-math-opts.c b/gcc/tree-ssa-math-opts.c
index 54ba035f5ee..969c1a6b6c6 100644
--- a/gcc/tree-ssa-math-opts.c
+++ b/gcc/tree-ssa-math-opts.c
@@ -2715,11 +2715,14 @@ convert_plusminus_to_widen (gimple_stmt_iterator *gsi, gimple *stmt,
      multiply-and-accumulate instructions.
 
      If the widened-multiplication result has more than one uses, it is
-     probably wiser not to do the conversion.  */
+     probably wiser not to do the conversion.  Also restrict this operation
+     to single basic block to avoid moving the multiply to a different block
+     with a higher execution frequency.  */
   if (code == PLUS_EXPR
       && (rhs1_code == MULT_EXPR || rhs1_code == WIDEN_MULT_EXPR))
     {
       if (!has_single_use (rhs1)
+	  || gimple_bb (rhs1_stmt) != gimple_bb (stmt)
 	  || !is_widening_mult_p (rhs1_stmt, &type1, &mult_rhs1,
 				  &type2, &mult_rhs2))
 	return false;
@@ -2729,6 +2732,7 @@ convert_plusminus_to_widen (gimple_stmt_iterator *gsi, gimple *stmt,
   else if (rhs2_code == MULT_EXPR || rhs2_code == WIDEN_MULT_EXPR)
     {
       if (!has_single_use (rhs2)
+	  || gimple_bb (rhs2_stmt) != gimple_bb (stmt)
 	  || !is_widening_mult_p (rhs2_stmt, &type1, &mult_rhs1,
 				  &type2, &mult_rhs2))
 	return false;