AArch64: update costing for combining vector conditionals

boolean comparisons have different cost depending on the mode. e.g.
for SVE, a && b doesn't require an additional instruction when a or b
is predicated by combining the predicate of the one operation into the
second one.  At the moment though we only fuse compares so this update
requires one of the operands to be a comparison.

Scalars also don't require this because the non-ifcvt variant is a series of
branches where following the branch sequences themselves are natural ANDs.

Advanced SIMD however does require an actual AND to combine the boolean values.

As such this patch discounts Scalar and SVE boolean operation latency and
throughput.

With this patch comparison heavy code prefers SVE as it should, especially in
cases with SVE VL == Advanced SIMD VL where previously the SVE prologue costs
would tip it towards Advanced SIMD.

gcc/ChangeLog:

	* config/aarch64/aarch64.cc (aarch64_bool_compound_p): New.
	(aarch64_adjust_stmt_cost, aarch64_vector_costs::count_ops): Use it.

diff --git a/gcc/config/aarch64/aarch64.cc b/gcc/config/aarch64/aarch64.cc
index 53fbecbbb5d45d6f8d34ec68c4b94cd586dcc9e9..7cd230c4602a15980016bdc92e80579be0c07094 100644 (file)
--- a/gcc/config/aarch64/aarch64.cc
+++ b/gcc/config/aarch64/aarch64.cc
@@ -16447,6 +16447,49 @@ aarch64_multiply_add_p (vec_info *vinfo, stmt_vec_info stmt_info,
   return false;
 }
 
+/* Return true if STMT_INFO is the second part of a two-statement boolean AND
+   expression sequence that might be suitable for fusing into a
+   single instruction.  If VEC_FLAGS is zero, analyze the operation as
+   a scalar one, otherwise analyze it as an operation on vectors with those
+   VEC_* flags.  */
+
+static bool
+aarch64_bool_compound_p (vec_info *vinfo, stmt_vec_info stmt_info,
+                        unsigned int vec_flags)
+{
+  gassign *assign = dyn_cast<gassign *> (stmt_info->stmt);
+  if (!assign
+      || gimple_assign_rhs_code (assign) != BIT_AND_EXPR
+      || !STMT_VINFO_VECTYPE (stmt_info)
+      || !VECTOR_BOOLEAN_TYPE_P (STMT_VINFO_VECTYPE (stmt_info)))
+    return false;
+
+  for (int i = 1; i < 3; ++i)
+    {
+      tree rhs = gimple_op (assign, i);
+
+      if (TREE_CODE (rhs) != SSA_NAME)
+       continue;
+
+      stmt_vec_info def_stmt_info = vinfo->lookup_def (rhs);
+      if (!def_stmt_info
+         || STMT_VINFO_DEF_TYPE (def_stmt_info) != vect_internal_def)
+       continue;
+
+      gassign *rhs_assign = dyn_cast<gassign *> (def_stmt_info->stmt);
+      if (!rhs_assign
+         || TREE_CODE_CLASS (gimple_assign_rhs_code (rhs_assign))
+               != tcc_comparison)
+       continue;
+
+      if (vec_flags & VEC_ADVSIMD)
+       return false;
+
+      return true;
+    }
+  return false;
+}
+
 /* We are considering implementing STMT_INFO using SVE.  If STMT_INFO is an
    in-loop reduction that SVE supports directly, return its latency in cycles,
    otherwise return zero.  SVE_COSTS specifies the latencies of the relevant
@@ -16744,10 +16787,16 @@ aarch64_adjust_stmt_cost (vec_info *vinfo, vect_cost_for_stmt kind,
        }
 
       gassign *assign = dyn_cast<gassign *> (STMT_VINFO_STMT (stmt_info));
-      if (assign && !vect_is_reduction (stmt_info))
+      if (assign)
        {
          /* For MLA we need to reduce the cost since MLA is 1 instruction.  */
-         if (aarch64_multiply_add_p (vinfo, stmt_info, vec_flags))
+         if (!vect_is_reduction (stmt_info)
+             && aarch64_multiply_add_p (vinfo, stmt_info, vec_flags))
+           return 0;
+
+         /* For vector boolean ANDs with a compare operand we just need
+            one insn.  */
+         if (aarch64_bool_compound_p (vinfo, stmt_info, vec_flags))
            return 0;
        }
 
@@ -16823,6 +16872,12 @@ aarch64_vector_costs::count_ops (unsigned int count, vect_cost_for_stmt kind,
   if (stmt_info && aarch64_multiply_add_p (m_vinfo, stmt_info, m_vec_flags))
     return;
 
+  /* Assume that bool AND with compare operands will become a single
+     operation.  */
+  if (stmt_info
+      && aarch64_bool_compound_p (m_vinfo, stmt_info, m_vec_flags))
+    return;
+
   /* Count the basic operation cost associated with KIND.  */
   switch (kind)
     {