[PATCH] avoid using masked vector epilogues when no scalar epilog is needed

[Richard Biener]

The following arranges for avoiding masked vector epilogues when we'll
eventually arrive at a vector epilogue with VF == 1 which implies no
scalar epilog will be necessary.

This avoids regressing performance in OpenColorIO when the
avx512_masked_epilogues tuning is enabled.  A testcase for one
example case is shown in PR122573.

Bootstrapped and tested on x86_64-unknown-linux-gnu.  The testcase
depends on the SLP patch posted earlier.

OK for trunk?

Thanks,
Richard.

        PR tree-optimization/122573
        * config/i386/i386.cc (ix86_vector_costs::finish_cost): Avoid
        using masked epilogues when an SSE epilogue would have a VF of one.

        * gcc.dg/vect/costmodel/x86_64/costmodel-pr122573.c: New testcase.
---
 gcc/config/i386/i386.cc                       |  5 ++++
 .../costmodel/x86_64/costmodel-pr122573.c     | 30 +++++++++++++++++++
 2 files changed, 35 insertions(+)
 create mode 100644 
gcc/testsuite/gcc.dg/vect/costmodel/x86_64/costmodel-pr122573.c

diff --git a/gcc/config/i386/i386.cc b/gcc/config/i386/i386.cc
index 6b6febc8870..8aac0820bc2 100644
--- a/gcc/config/i386/i386.cc
+++ b/gcc/config/i386/i386.cc
@@ -26609,6 +26609,11 @@ ix86_vector_costs::finish_cost (const vector_costs 
*scalar_costs)
   if (loop_vinfo
       && !LOOP_VINFO_EPILOGUE_P (loop_vinfo)
       && LOOP_VINFO_VECT_FACTOR (loop_vinfo).to_constant () > 2
+      /* Avoid a masked epilog if cascaded epilogues eventually get us
+        to one with VF 1 as that means no scalar epilog at all.  */
+      && !((GET_MODE_SIZE (loop_vinfo->vector_mode)
+           / LOOP_VINFO_VECT_FACTOR (loop_vinfo).to_constant () == 16)
+          && ix86_tune_features[X86_TUNE_AVX512_TWO_EPILOGUES])
       && ix86_tune_features[X86_TUNE_AVX512_MASKED_EPILOGUES]
       && !OPTION_SET_P (param_vect_partial_vector_usage))
     {
diff --git a/gcc/testsuite/gcc.dg/vect/costmodel/x86_64/costmodel-pr122573.c 
b/gcc/testsuite/gcc.dg/vect/costmodel/x86_64/costmodel-pr122573.c
new file mode 100644
index 00000000000..ca3294dca7a
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/vect/costmodel/x86_64/costmodel-pr122573.c
@@ -0,0 +1,30 @@
+/* { dg-do compile } */
+/* { dg-additional-options "-march=znver5" } */
+
+struct S {
+    float m_col1[4];
+    float m_col2[4];
+    float m_col3[4];
+    float m_col4[4];
+};
+
+void apply(struct S *s, const float *in, float *out, long numPixels)
+{
+  for (long idx = 0; idx < numPixels; ++idx)
+    {
+      const float r = in[0];
+      const float g = in[1];
+      const float b = in[2];
+      const float a = in[3];
+      out[0] = r*s->m_col1[0] + g*s->m_col2[0] + b*s->m_col3[0] + 
a*s->m_col4[0];
+      out[1] = r*s->m_col1[1] + g*s->m_col2[1] + b*s->m_col3[1] + 
a*s->m_col4[1];
+      out[2] = r*s->m_col1[2] + g*s->m_col2[2] + b*s->m_col3[2] + 
a*s->m_col4[2];
+      out[3] = r*s->m_col1[3] + g*s->m_col2[3] + b*s->m_col3[3] + 
a*s->m_col4[3];
+      in  += 4;
+      out += 4;
+    }
+}
+
+/* Check that we do not use a masked epilog but a SSE one with VF 1
+   (and possibly a AVX2 one as well).  */
+/* { dg-final { scan-tree-dump "optimized: epilogue loop vectorized using 16 
byte vectors and unroll factor 1" "vect" } } */
-- 
2.51.0