[PATCH] Adjust param_vect_max_version_for_alias_checks

This patch sets param_vect_max_version_for_alias_checks to 15.
This was causing GCC to miss vectorization opportunities for an application,
making it slower than LLVM by about ~14%.

Original default of 10 itself is arbitary. Given that, GCC's vectoriser does
consideres cost of alias checks, increasing this param is reasonable.

In this case we need a value of at teast 11 whereas the current
default is 10.

gcc/ChangeLog:

	* params.opt: Adjust param_vect_max_version_for_alias_checks

gcc/testsuite/ChangeLog:

	* g++.dg/alias-checks.C: New test.

Signed-off-by: Kugan Vivekanandarajah <kvivekananda@nvidia.com>

diff --git a/gcc/params.opt b/gcc/params.opt
index 24f440bbe71f6e5a4d81b77aca779d5310464552..7c572774df24b64024a87c1c4ef4891f1f1defd9 100644 (file)
--- a/gcc/params.opt
+++ b/gcc/params.opt
@@ -1187,7 +1187,7 @@ Common Joined UInteger Var(param_vect_max_peeling_for_alignment) Init(-1) Intege
 Maximum number of loop peels to enhance alignment of data references in a loop.
 
 -param=vect-max-version-for-alias-checks=
-Common Joined UInteger Var(param_vect_max_version_for_alias_checks) Init(10) Param Optimization
+Common Joined UInteger Var(param_vect_max_version_for_alias_checks) Init(15) Param Optimization
 Bound on number of runtime checks inserted by the vectorizer's loop versioning for alias check.
 
 -param=vect-max-version-for-alignment-checks=

diff --git a/gcc/testsuite/g++.dg/alias-checks.C b/gcc/testsuite/g++.dg/alias-checks.C
new file mode 100644 (file)
index 0000000..ee0ac53
--- /dev/null
+++ b/gcc/testsuite/g++.dg/alias-checks.C
@@ -0,0 +1,85 @@
+
+// { dg-do compile }
+// { dg-require-effective-target c++17 }
+// { dg-require-effective-target vect_double }
+// { dg-options "-O3 -fdump-tree-vect-all" }
+
+#include <cstdlib>
+#include <cmath>
+#include <iostream>
+void DoIfClause(double *l, double *r, const double *m2, const double *m1,
+        const double *c, const double *p1, const double *p2, int istart, int iend)
+{
+    constexpr double w5alpha[3][3] = {{1.0 / 3.0, -7.0 / 6.0, 11.0 / 6.0},
+        {-1.0 / 6.0, 5.0 / 6.0, 1.0 / 3.0},
+        {1.0 / 3.0, 5.0 / 6.0, -1.0 / 6.0}};
+    constexpr double w5gamma[3] = {0.1, 0.6, 0.3};
+    constexpr double eps = 1e-100;
+    constexpr double thirteen_thirds = 13.0 / 3.0;
+
+    for (int i = istart; i <= iend; ++i) {
+        const double q0 = m2[i];
+        const double q1 = m1[i];
+        const double q2 = c[i];
+        const double q3 = p1[i];
+        const double q4 = p2[i];
+        double &ql = l[i];
+        double &qr = r[i];
+        double a = q0 - 2 * q1 + q2;
+        double b = q0 - 4.0 * q1 + 3.0 * q2;
+        double beta0 = thirteen_thirds * a * a + b * b + eps;
+        a = q1 - 2.0 * q2 + q3;
+        b = q3 - q1;
+        double beta1 = thirteen_thirds * a * a + b * b + eps;
+        a = q2 - 2.0 * q3 + q4;
+        b = q4 - 4.0 * q3 + 3.0 * q2;
+        double beta2 = thirteen_thirds * a * a + b * b + eps;
+        const double tau5 = std::abs(beta2 - beta0);
+
+        beta0 = (beta0 + tau5) / beta0;
+        beta1 = (beta1 + tau5) / beta1;
+        beta2 = (beta2 + tau5) / beta2;
+
+        double w0 = w5gamma[0] * beta0 + eps;
+        double w1 = w5gamma[1] * beta1 + eps;
+        double w2 = w5gamma[2] * beta2 + eps;
+        double wsum = 1.0 / (w0 + w1 + w2);
+        ql = w0 * (w5alpha[0][0] * q0 + w5alpha[0][1] * q1 + w5alpha[0][2] * q2);
+        ql += w1 * (w5alpha[1][0] * q1 + w5alpha[1][1] * q2 + w5alpha[1][2] * q3);
+        ql += w2 * (w5alpha[2][0] * q2 + w5alpha[2][1] * q3 + w5alpha[2][2] * q4);
+        ql *= wsum;
+        const double alpha_l = 3.0 * wsum * w0 * w1 * w2 /
+            (w5gamma[2] * w0 * w1 + w5gamma[1] * w0 * w2 +
+             w5gamma[0] * w1 * w2) +
+            eps;
+
+        w0 = w5gamma[0] * beta2 + eps;
+        w1 = w5gamma[1] * beta1 + eps;
+        w2 = w5gamma[2] * beta0 + eps;
+        wsum = 1.0 / (w0 + w1 + w2);
+        qr = w0 * (w5alpha[0][0] * q4 + w5alpha[0][1] * q3 + w5alpha[0][2] * q2);
+        qr += w1 * (w5alpha[1][0] * q3 + w5alpha[1][1] * q2 + w5alpha[1][2] * q1);
+        qr += w2 * (w5alpha[2][0] * q2 + w5alpha[2][1] * q1 + w5alpha[2][2] * q0);
+        qr *= wsum;
+        const double alpha_r = 3.0 * wsum * w0 * w1 * w2 /
+            (w5gamma[2] * w0 * w1 + w5gamma[1] * w0 * w2 +
+             w5gamma[0] * w1 * w2) +
+            eps;
+
+        double dq = q3 - q2;
+        {
+            const double dm = q2 - q1;
+            const double dp = dq;
+            const double dc = (dm * dp > 0.0) * 0.5 * (dm + dp);
+            dq = 0.5 * std::copysign(
+                    std::min(std::fabs(dc),
+                        2.0 * std::min(std::fabs(dm), std::fabs(dp))),
+                    dc);
+        }
+
+        const double alpha_lin = 2.0 * alpha_l * alpha_r / (alpha_l + alpha_r);
+        ql = alpha_lin * ql + (1.0 - alpha_lin) * (q2 + dq);
+        qr = alpha_lin * qr + (1.0 - alpha_lin) * (q2 - dq);
+    }
+}
+/* { dg-final { scan-tree-dump "LOOP VECTORIZED" "vect" } } */