Fix double shufti's vector end false positive (#325)

* Add regression test for double shufti

It tests for false positive at vector edges.

Signed-off-by: Yoan Picchi <yoan.picchi@arm.com>

* Fix double shufti reporting false positives

Double shufti used to offset one vector, resulting in losing one character
at the end of every vector. This was replaced by a magic value indicating a
match. This meant that if the first char of a pattern fell on the last char of
a vector, double shufti would assume the second character is present and
report a match.
This patch fixes it by keeping the previous vector and feeding its data to the
new one when we shift it, preventing any loss of data.

Signed-off-by: Yoan Picchi <yoan.picchi@arm.com>

* vshl() will call the correct implementation

* implement missing vshr_512_imm(), simplifies caller x86 code

* Fix x86 case, use alignr instead

* it's the reverse, the avx512 alignr is incorrect, need to fix

* Make shufti's OR reduce size agnostic

Signed-off-by: Yoan Picchi <yoan.picchi@arm.com>

* Fix test's array size

Signed-off-by: Yoan Picchi <yoan.picchi@arm.com>

* Fix AVX2/AVX512 alignr implementations and unit tests

* Fix Power VSX alignr

---------

Signed-off-by: Yoan Picchi <yoan.picchi@arm.com>
Co-authored-by: Konstantinos Margaritis <konstantinos@vectorcamp.gr>

unit/internal/shufti.cpp

@@ -899,7 +899,40 @@ TEST(DoubleShufti, ExecMatchMixed3) {
         const u8 *rv = shuftiDoubleExec(lo1, hi1, lo2, hi2,
                                         reinterpret_cast<u8 *>(t2), reinterpret_cast<u8 *>(t2) + len);
 
-        ASSERT_EQ(reinterpret_cast<const u8 *>(&t2[len - i]), rv);
+        if(i < 2) {
+            // i=0 is "xy" out of buffer. i=1 is "x" in buffer but not "y"
+            ASSERT_EQ(reinterpret_cast<const u8 *>(t2 + len), rv);
+        }else {
+            ASSERT_EQ(reinterpret_cast<const u8 *>(&t2[len - i]), rv);
+        }
+    }
+}
+
+// Double shufti used to report matches when the first char of a pair landed at
+// the end of a vector. This test check for the regression.
+TEST(DoubleShufti, ExecNoMatchVectorEdge) {
+    m128 lo1, hi1, lo2, hi2;
+
+    flat_set<pair<u8, u8>> lits;
+
+    lits.insert(make_pair('a','c'));
+
+    bool ret = shuftiBuildDoubleMasks(CharReach(), lits, reinterpret_cast<u8 *>(&lo1), reinterpret_cast<u8 *>(&hi1),
+                                      reinterpret_cast<u8 *>(&lo2), reinterpret_cast<u8 *>(&hi2));
+    ASSERT_TRUE(ret);
+
+    const int len = 80;
+    char t1[len + 2];
+    memset(t1, 'b', len);
+
+    for (size_t i = 0; i < 70; i++) {
+        t1[len - i] = 'a';
+        t1[len - i + 1] = 'b';
+        DEBUG_PRINTF("i = %ld\n", i);
+        const u8 *rv = shuftiDoubleExec(lo1, hi1, lo2, hi2,
+                                        reinterpret_cast<u8 *>(t1), reinterpret_cast<u8 *>(t1) + len);
+
+        ASSERT_EQ(reinterpret_cast<const u8 *>(t1 + len), rv);
     }
 }