Fix PEELING_FOR_NITERS calculation (PR 87288)

PEELING_FOR_GAPS now means "peel one iteration for the epilogue",
in much the same way that PEELING_FOR_ALIGNMENT > 0 means
"peel that number of iterations for the prologue".  We weren't
taking this into account when deciding whether we needed to peel
further scalar iterations beyond the iterations for "gaps" and
"alignment".

Only the first test failed before the patch.  The other two
are just for completeness.

2018-09-20  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
	PR tree-optimization/87288
	* tree-vect-loop.c (vect_analyze_loop_2): Take PEELING_FOR_GAPS
	into account when determining PEELING_FOR_NITERS.

gcc/testsuite/
	PR tree-optimization/87288
	* gcc.dg/vect/pr87288-1.c: New test.
	* gcc.dg/vect/pr87288-2.c: Likewise,
	* gcc.dg/vect/pr87288-3.c: Likewise.

From-SVN: r264440

6 files changed, 201 insertions, 3 deletions

diff --git a/gcc/ChangeLog b/gcc/ChangeLog
index af5fddf1..f400284 100644
--- a/gcc/ChangeLog
+++ b/gcc/ChangeLog
@@ -1,5 +1,11 @@
 2018-09-20  Richard Sandiford  <richard.sandiford@arm.com>
 
+	PR tree-optimization/87288
+	* tree-vect-loop.c (vect_analyze_loop_2): Take PEELING_FOR_GAPS
+	into account when determining PEELING_FOR_NITERS.
+
+2018-09-20  Richard Sandiford  <richard.sandiford@arm.com>
+
 	PR tree-optimization/86877
 	* tree-vect-loop.c (vect_analyze_loop_2): Call
 	vect_verify_datarefs_alignment.
diff --git a/gcc/testsuite/ChangeLog b/gcc/testsuite/ChangeLog
index ade95db..c7c8e90 100644
--- a/gcc/testsuite/ChangeLog
+++ b/gcc/testsuite/ChangeLog
@@ -1,5 +1,12 @@
 2018-09-20  Richard Sandiford  <richard.sandiford@arm.com>
 
+	PR tree-optimization/87288
+	* gcc.dg/vect/pr87288-1.c: New test.
+	* gcc.dg/vect/pr87288-2.c: Likewise,
+	* gcc.dg/vect/pr87288-3.c: Likewise.
+
+2018-09-20  Richard Sandiford  <richard.sandiford@arm.com>
+
 	PR tree-optimization/86877
 	* gfortran.dg/vect/vect-8-epilogue.F90: New test.
 
diff --git a/gcc/testsuite/gcc.dg/vect/pr87288-1.c b/gcc/testsuite/gcc.dg/vect/pr87288-1.c
new file mode 100644
index 0000000..0d0a70d
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/vect/pr87288-1.c
@@ -0,0 +1,49 @@
+#include "tree-vect.h"
+
+#define N (VECTOR_BITS / 32)
+#define MAX_COUNT 4
+
+void __attribute__ ((noipa))
+run (int *restrict a, int *restrict b, int count)
+{
+  for (int i = 0; i < count * N; ++i)
+    {
+      a[i * 2] = b[i * 2] + count;
+      a[i * 2 + 1] = count;
+    }
+}
+
+void __attribute__ ((noipa))
+check (int *restrict a, int count)
+{
+  for (int i = 0; i < count * N; ++i)
+    if (a[i * 2] != i * 41 + count || a[i * 2 + 1] != count)
+      __builtin_abort ();
+  if (a[count * 2 * N] != 999)
+    __builtin_abort ();
+}
+
+int a[N * MAX_COUNT * 2 + 1], b[N * MAX_COUNT * 2];
+
+int
+main (void)
+{
+  check_vect ();
+
+  for (int i = 0; i < N * MAX_COUNT; ++i)
+    {
+      b[i * 2] = i * 41;
+      asm volatile ("" ::: "memory");
+    }
+
+  for (int i = 0; i <= MAX_COUNT; ++i)
+    {
+      a[i * 2 * N] = 999;
+      run (a, b, i);
+      check (a, i);
+    }
+
+  return 0;
+}
+
+/* { dg-final { scan-tree-dump-times {LOOP VECTORIZED} 1 "vect" { target { { vect_int && vect_perm } && vect_element_align } } } } */
diff --git a/gcc/testsuite/gcc.dg/vect/pr87288-2.c b/gcc/testsuite/gcc.dg/vect/pr87288-2.c
new file mode 100644
index 0000000..e9ff9a0
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/vect/pr87288-2.c
@@ -0,0 +1,64 @@
+#include "tree-vect.h"
+
+#define N (VECTOR_BITS / 32)
+#define MAX_COUNT 4
+
+#define RUN_COUNT(COUNT)				\
+  void __attribute__ ((noipa))				\
+  run_##COUNT (int *restrict a, int *restrict b)	\
+  {							\
+    for (int i = 0; i < N * COUNT; ++i)			\
+      {							\
+	a[i * 2] = b[i * 2] + COUNT;			\
+	a[i * 2 + 1] = COUNT;				\
+      }							\
+  }
+
+RUN_COUNT (1)
+RUN_COUNT (2)
+RUN_COUNT (3)
+RUN_COUNT (4)
+
+void __attribute__ ((noipa))
+check (int *restrict a, int count)
+{
+  for (int i = 0; i < count * N; ++i)
+    if (a[i * 2] != i * 41 + count || a[i * 2 + 1] != count)
+      __builtin_abort ();
+  if (a[count * 2 * N] != 999)
+    __builtin_abort ();
+}
+
+int a[N * MAX_COUNT * 2 + 1], b[N * MAX_COUNT * 2];
+
+int
+main (void)
+{
+  check_vect ();
+
+  for (int i = 0; i < N * MAX_COUNT; ++i)
+    {
+      b[i * 2] = i * 41;
+      asm volatile ("" ::: "memory");
+    }
+
+  a[N * 2] = 999;
+  run_1 (a, b);
+  check (a, 1);
+
+  a[N * 4] = 999;
+  run_2 (a, b);
+  check (a, 2);
+
+  a[N * 6] = 999;
+  run_3 (a, b);
+  check (a, 3);
+
+  a[N * 8] = 999;
+  run_4 (a, b);
+  check (a, 4);
+
+  return 0;
+}
+
+/* { dg-final { scan-tree-dump {LOOP VECTORIZED} "vect" { target { { vect_int && vect_perm } && vect_element_align } } } } */
diff --git a/gcc/testsuite/gcc.dg/vect/pr87288-3.c b/gcc/testsuite/gcc.dg/vect/pr87288-3.c
new file mode 100644
index 0000000..23f574c
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/vect/pr87288-3.c
@@ -0,0 +1,64 @@
+#include "tree-vect.h"
+
+#define N (VECTOR_BITS / 32)
+#define MAX_COUNT 4
+
+#define RUN_COUNT(COUNT)				\
+  void __attribute__ ((noipa))				\
+  run_##COUNT (int *restrict a, int *restrict b)	\
+  {							\
+    for (int i = 0; i < N * COUNT + 1; ++i)		\
+      {							\
+	a[i * 2] = b[i * 2] + COUNT;			\
+	a[i * 2 + 1] = COUNT;				\
+      }							\
+  }
+
+RUN_COUNT (1)
+RUN_COUNT (2)
+RUN_COUNT (3)
+RUN_COUNT (4)
+
+void __attribute__ ((noipa))
+check (int *restrict a, int count)
+{
+  for (int i = 0; i < count * N + 1; ++i)
+    if (a[i * 2] != i * 41 + count || a[i * 2 + 1] != count)
+      __builtin_abort ();
+  if (a[count * 2 * N + 2] != 999)
+    __builtin_abort ();
+}
+
+int a[N * MAX_COUNT * 2 + 3], b[N * MAX_COUNT * 2 + 2];
+
+int
+main (void)
+{
+  check_vect ();
+
+  for (int i = 0; i < N * MAX_COUNT + 1; ++i)
+    {
+      b[i * 2] = i * 41;
+      asm volatile ("" ::: "memory");
+    }
+
+  a[N * 2 + 2] = 999;
+  run_1 (a, b);
+  check (a, 1);
+
+  a[N * 4 + 2] = 999;
+  run_2 (a, b);
+  check (a, 2);
+
+  a[N * 6 + 2] = 999;
+  run_3 (a, b);
+  check (a, 3);
+
+  a[N * 8 + 2] = 999;
+  run_4 (a, b);
+  check (a, 4);
+
+  return 0;
+}
+
+/* { dg-final { scan-tree-dump {LOOP VECTORIZED} "vect" { target { { vect_int && vect_perm } && vect_element_align } } } } */
diff --git a/gcc/tree-vect-loop.c b/gcc/tree-vect-loop.c
index 70e9157..fdac10b 100644
--- a/gcc/tree-vect-loop.c
+++ b/gcc/tree-vect-loop.c
@@ -2074,14 +2074,22 @@ start_over:
     /* The main loop handles all iterations.  */
     LOOP_VINFO_PEELING_FOR_NITER (loop_vinfo) = false;
   else if (LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo)
-	   && LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo) > 0)
+	   && LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo) >= 0)
     {
-      if (!multiple_p (LOOP_VINFO_INT_NITERS (loop_vinfo)
-		       - LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo),
+      /* Work out the (constant) number of iterations that need to be
+	 peeled for reasons other than niters.  */
+      unsigned int peel_niter = LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo);
+      if (LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo))
+	peel_niter += 1;
+      if (!multiple_p (LOOP_VINFO_INT_NITERS (loop_vinfo) - peel_niter,
 		       LOOP_VINFO_VECT_FACTOR (loop_vinfo)))
 	LOOP_VINFO_PEELING_FOR_NITER (loop_vinfo) = true;
     }
   else if (LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo)
+	   /* ??? When peeling for gaps but not alignment, we could
+	      try to check whether the (variable) niters is known to be
+	      VF * N + 1.  That's something of a niche case though.  */
+	   || LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo)
 	   || !LOOP_VINFO_VECT_FACTOR (loop_vinfo).is_constant (&const_vf)
 	   || ((tree_ctz (LOOP_VINFO_NITERS (loop_vinfo))
 		< (unsigned) exact_log2 (const_vf))