re PR target/70329 (wrong code with -mavx512bw and 64byte vectors)

	PR target/70329
	* config/i386/i386.c (ix86_expand_vecop_qihi): Don't bother computing
	d.perm[i] for i >= d.nelt.  If not full_interleave, compute d.perm[i]
	in a way that works also for AVX512BW.

	* gcc.target/i386/avx512bw-pr70329-1.c: New test.
	* gcc.target/i386/avx512bw-pr70329-2.c: New test.

From-SVN: r234394

5 files changed, 81 insertions, 4 deletions

diff --git a/gcc/ChangeLog b/gcc/ChangeLog
index f5687a5..6dc6643 100644
--- a/gcc/ChangeLog
+++ b/gcc/ChangeLog
@@ -1,5 +1,10 @@
 2016-03-22  Jakub Jelinek  <jakub@redhat.com>
 
+	PR target/70329
+	* config/i386/i386.c (ix86_expand_vecop_qihi): Don't bother computing
+	d.perm[i] for i >= d.nelt.  If not full_interleave, compute d.perm[i]
+	in a way that works also for AVX512BW.
+
 	PR target/70300
 	* config/i386/i386.md (cvtsd2ss splitter): Unpack in destination
 	instead of source if operands[1] is xmm16 and above and
diff --git a/gcc/config/i386/i386.c b/gcc/config/i386/i386.c
index 3d8dbc4..a4aad4f 100644
--- a/gcc/config/i386/i386.c
+++ b/gcc/config/i386/i386.c
@@ -51910,16 +51910,24 @@ ix86_expand_vecop_qihi (enum rtx_code code, rtx dest, rtx op1, rtx op2)
     {
       /* For SSE2, we used an full interleave, so the desired
 	 results are in the even elements.  */
-      for (i = 0; i < 64; ++i)
+      for (i = 0; i < d.nelt; ++i)
 	d.perm[i] = i * 2;
     }
   else
     {
       /* For AVX, the interleave used above was not cross-lane.  So the
 	 extraction is evens but with the second and third quarter swapped.
-	 Happily, that is even one insn shorter than even extraction.  */
-      for (i = 0; i < 64; ++i)
-	d.perm[i] = i * 2 + ((i & 24) == 8 ? 16 : (i & 24) == 16 ? -16 : 0);
+	 Happily, that is even one insn shorter than even extraction.
+	 For AVX512BW we have 4 lanes.  We extract evens from within a lane,
+	 always first from the first and then from the second source operand,
+	 the index bits above the low 4 bits remains the same.
+	 Thus, for d.nelt == 32 we want permutation
+	 0,2,4,..14, 32,34,36,..46, 16,18,20,..30, 48,50,52,..62
+	 and for d.nelt == 64 we want permutation
+	 0,2,4,..14, 64,66,68,..78, 16,18,20,..30, 80,82,84,..94,
+	 32,34,36,..46, 96,98,100,..110, 48,50,52,..62, 112,114,116,..126.  */
+      for (i = 0; i < d.nelt; ++i)
+	d.perm[i] = ((i * 2) & 14) + ((i & 8) ? d.nelt : 0) + (i & ~15);
     }
 
   ok = ix86_expand_vec_perm_const_1 (&d);
diff --git a/gcc/testsuite/ChangeLog b/gcc/testsuite/ChangeLog
index aad5d3f..73b4eb2 100644
--- a/gcc/testsuite/ChangeLog
+++ b/gcc/testsuite/ChangeLog
@@ -1,5 +1,9 @@
 2016-03-22  Jakub Jelinek  <jakub@redhat.com>
 
+	PR target/70329
+	* gcc.target/i386/avx512bw-pr70329-1.c: New test.
+	* gcc.target/i386/avx512bw-pr70329-2.c: New test.
+
 	PR target/70300
 	* gcc.target/i386/pr70300.c: New test.
 
diff --git a/gcc/testsuite/gcc.target/i386/avx512bw-pr70329-1.c b/gcc/testsuite/gcc.target/i386/avx512bw-pr70329-1.c
new file mode 100644
index 0000000..bb9a955
--- /dev/null
+++ b/gcc/testsuite/gcc.target/i386/avx512bw-pr70329-1.c
@@ -0,0 +1,27 @@
+/* PR target/70329 */
+/* { dg-do run } */
+/* { dg-options "-O0 -mavx512bw" } */
+/* { dg-require-effective-target avx512bw } */
+
+#define AVX512BW
+#include "avx512f-helper.h"
+
+typedef unsigned char A __attribute__ ((vector_size (64)));
+typedef unsigned int B __attribute__ ((vector_size (64)));
+
+unsigned __attribute__ ((noinline, noclone))
+foo (A a, A b, B c)
+{
+  a *= b;
+  c[1] += a[8];
+  return c[1];
+}
+
+void
+TEST (void)
+{
+  A a = (A) { 1, 2, 3, 4, 5, 6, 7, 8, 9 };
+  unsigned x = foo (a, a, (B) { 1, 2 });
+  if (x != 83)
+    abort ();
+}
diff --git a/gcc/testsuite/gcc.target/i386/avx512bw-pr70329-2.c b/gcc/testsuite/gcc.target/i386/avx512bw-pr70329-2.c
new file mode 100644
index 0000000..731b926
--- /dev/null
+++ b/gcc/testsuite/gcc.target/i386/avx512bw-pr70329-2.c
@@ -0,0 +1,33 @@
+/* PR target/70329 */
+/* { dg-do run } */
+/* { dg-options "-O2 -ftree-vectorize -mavx512bw" } */
+/* { dg-require-effective-target avx512bw } */
+
+#define AVX512BW
+#include "avx512f-helper.h"
+
+__attribute__((noinline, noclone)) void
+foo (unsigned char *src1, unsigned char *src2, unsigned char *dst)
+{
+  int i;
+
+  for (i = 0; i < 64; i++)
+    dst[i] = (unsigned char) ((int) src1[i] * (int) src2[i]);
+}
+
+void
+TEST (void)
+{
+  unsigned char a[64], b[64], c[64];
+  int i;
+
+  for (i = 0; i < 64; i++)
+    {
+      a[i] = i;
+      b[i] = (i + 1);
+    }
+  foo (a, b, c);
+  for (i = 0; i < 64; i++)
+    if (c[i] != (unsigned char) (i * (i + 1)))
+      abort ();
+}