From: Richard Sandiford <richard.sandiford@linaro.org>
Date: Sun, 18 Mar 2018 10:25:29 +0000 (+0000)
Subject: Don't try to vectorise COND_EXPR reduction chains (PR 84913)
X-Git-Url: https://git.libre-soc.org/?a=commitdiff_plain;h=928b965f29f587bb033555f0db646c94afe8e7ef;p=gcc.git

Don't try to vectorise COND_EXPR reduction chains (PR 84913)

The testcase ICEd for both SVE and AVX512 because we were trying
to vectorise a chain of COND_EXPRs as a reduction and getting
confused by reduc_index == -1.

2018-03-18  Richard Sandiford  <richard.sandiford@linaro.org>

gcc/
	PR tree-optimization/84913
	* tree-vect-loop.c (vectorizable_reduction): Don't try to
	vectorize chains of COND_EXPRs.

gcc/testsuite/
	PR tree-optimization/84913
	* gfortran.dg/vect/pr84913.f90: New test.

From-SVN: r258631
---

diff --git a/gcc/ChangeLog b/gcc/ChangeLog
index 601cf5962f3..2a77e3b9245 100644
--- a/gcc/ChangeLog
+++ b/gcc/ChangeLog
@@ -1,3 +1,9 @@
+2018-03-18  Richard Sandiford  <richard.sandiford@linaro.org>
+
+	PR tree-optimization/84913
+	* tree-vect-loop.c (vectorizable_reduction): Don't try to
+	vectorize chains of COND_EXPRs.
+
 2018-03-18  Chung-Ju Wu  <jasonwucj@gmail.com>
 
 	* config/nds32/nds32.h (MAX_REGS_PER_ADDRESS): Fix the value.
diff --git a/gcc/testsuite/ChangeLog b/gcc/testsuite/ChangeLog
index 098444aaf33..7d9a3fd4c2c 100644
--- a/gcc/testsuite/ChangeLog
+++ b/gcc/testsuite/ChangeLog
@@ -1,3 +1,8 @@
+2018-03-18  Richard Sandiford  <richard.sandiford@linaro.org>
+
+	PR tree-optimization/84913
+	* gfortran.dg/vect/pr84913.f90: New test.
+
 2018-03-18  Thomas Koenig  <tkoenig@gcc.gnu.org>
 
 	PR fortran/79929
diff --git a/gcc/testsuite/gfortran.dg/vect/pr84913.f90 b/gcc/testsuite/gfortran.dg/vect/pr84913.f90
new file mode 100644
index 00000000000..f2ec4d233bf
--- /dev/null
+++ b/gcc/testsuite/gfortran.dg/vect/pr84913.f90
@@ -0,0 +1,13 @@
+! { dg-do compile }
+
+function foo(a, b, c, n)
+  integer :: a(n), b(n), c(n), n, i, foo
+  foo = 0
+  do i = 1, n
+    if (a(i) .eq. b(i)) then
+      foo = 1
+    else if (a(i) .eq. c(i)) then
+      foo = 2
+    end if
+  end do
+end function foo
diff --git a/gcc/tree-vect-loop.c b/gcc/tree-vect-loop.c
index 6585c85cfe0..7b3009aed2d 100644
--- a/gcc/tree-vect-loop.c
+++ b/gcc/tree-vect-loop.c
@@ -6788,6 +6788,30 @@ vectorizable_reduction (gimple *stmt, gimple_stmt_iterator *gsi,
   /* If we have a condition reduction, see if we can simplify it further.  */
   if (v_reduc_type == COND_REDUCTION)
     {
+      /* TODO: We can't yet handle reduction chains, since we need to treat
+	 each COND_EXPR in the chain specially, not just the last one.
+	 E.g. for:
+
+	    x_1 = PHI <x_3, ...>
+	    x_2 = a_2 ? ... : x_1;
+	    x_3 = a_3 ? ... : x_2;
+
+	 we're interested in the last element in x_3 for which a_2 || a_3
+	 is true, whereas the current reduction chain handling would
+	 vectorize x_2 as a normal VEC_COND_EXPR and only treat x_3
+	 as a reduction operation.  */
+      if (reduc_index == -1)
+	{
+	  if (dump_enabled_p ())
+	    dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
+			     "conditional reduction chains not supported\n");
+	  return false;
+	}
+
+      /* vect_is_simple_reduction ensured that operand 2 is the
+	 loop-carried operand.  */
+      gcc_assert (reduc_index == 2);
+
       /* Loop peeling modifies initial value of reduction PHI, which
 	 makes the reduction stmt to be transformed different to the
 	 original stmt analyzed.  We need to record reduction code for