i386: implement costs for float<->int conversions in ix86_vector_costs::add_stmt_cost

This patch adds pattern matching for float<->int conversions both as normal
statements and promote_demote.  While updating promote_demote I noticed that
in cleanups I turned "stmt_cost =" into "int stmt_cost = " which turned
the existing FP costing to NOOP. I also added comment on how demotes are done
when turning i.e. 32bit into 8bit value (which is the case of pr19919.c).

The patch disables vectorization in pr119919.c on generic tuning, but keeps
it at both zen and skylake+. The underlying problem is bad cost of open-coded
scatter which is tracked by 119902 so I simply added -mtune=znver1 so the testcase
keeps testing vectorization.

gcc/ChangeLog:

	* config/i386/i386.cc (ix86_vector_costs::add_stmt_cost): Add FLOAT_EXPR;
	FIX_TRUNC_EXPR and vec_promote_demote costs.

gcc/testsuite/ChangeLog:

	* gcc.target/i386/pr119919.c: Add -mtune=znver1

diff --git a/gcc/config/i386/i386.cc b/gcc/config/i386/i386.cc
index bef95ea18c87c29b6a3ebfdf1897d4a280829acb..fd36ea802c00d7191edb303c8531e6894fe477c7 100644 (file)
--- a/gcc/config/i386/i386.cc
+++ b/gcc/config/i386/i386.cc
@@ -25767,6 +25767,26 @@ ix86_vector_costs::add_stmt_cost (int count, vect_cost_for_stmt kind,
                          (ix86_tune_cost, GET_MODE_BITSIZE (mode));
          break;
 
+       case FLOAT_EXPR:
+           if (SSE_FLOAT_MODE_SSEMATH_OR_HFBF_P (mode))
+             stmt_cost = ix86_cost->cvtsi2ss;
+           else if (X87_FLOAT_MODE_P (mode))
+             /* TODO: We do not have cost tables for x87.  */
+             stmt_cost = ix86_cost->fadd;
+           else
+             stmt_cost = ix86_vec_cost (mode, ix86_cost->cvtpi2ps);
+           break;
+
+       case FIX_TRUNC_EXPR:
+           if (SSE_FLOAT_MODE_SSEMATH_OR_HFBF_P (mode))
+             stmt_cost = ix86_cost->cvtss2si;
+           else if (X87_FLOAT_MODE_P (mode))
+             /* TODO: We do not have cost tables for x87.  */
+             stmt_cost = ix86_cost->fadd;
+           else
+             stmt_cost = ix86_vec_cost (mode, ix86_cost->cvtps2pi);
+           break;
+
        case COND_EXPR:
          {
            /* SSE2 conditinal move sequence is:
@@ -25930,8 +25950,7 @@ ix86_vector_costs::add_stmt_cost (int count, vect_cost_for_stmt kind,
        break;
       }
 
-  if (kind == vec_promote_demote
-      && fp && FLOAT_TYPE_P (TREE_TYPE (gimple_assign_rhs1 (stmt_info->stmt))))
+  if (kind == vec_promote_demote)
     {
       int outer_size
        = tree_to_uhwi
@@ -25941,16 +25960,25 @@ ix86_vector_costs::add_stmt_cost (int count, vect_cost_for_stmt kind,
        = tree_to_uhwi
            (TYPE_SIZE
                (TREE_TYPE (gimple_assign_rhs1 (stmt_info->stmt))));
-      int stmt_cost = vec_fp_conversion_cost
-                       (ix86_tune_cost, GET_MODE_BITSIZE (mode));
-      /* VEC_PACK_TRUNC_EXPR: If inner size is greater than outer size we will end
-        up doing two conversions and packing them.  */
+      bool inner_fp = FLOAT_TYPE_P
+                       (TREE_TYPE (gimple_assign_rhs1 (stmt_info->stmt)));
+
+      if (fp && inner_fp)
+       stmt_cost = vec_fp_conversion_cost
+                         (ix86_tune_cost, GET_MODE_BITSIZE (mode));
+      else if (fp && !inner_fp)
+       stmt_cost = ix86_vec_cost (mode, ix86_cost->cvtpi2ps);
+      else if (!fp && inner_fp)
+       stmt_cost = ix86_vec_cost (mode, ix86_cost->cvtps2pi);
+      else
+       stmt_cost = ix86_vec_cost (mode, ix86_cost->sse_op);
+      /* VEC_PACK_TRUNC_EXPR and similar demote operations: If outer size is
+        greater than inner size we will end up doing two conversions and
+        packing them.  We always pack pairs; if the size difference is greater
+        it is split into multiple demote operations.  */
       if (inner_size > outer_size)
-       {
-         int n = inner_size / outer_size;
-         stmt_cost = stmt_cost * n
-                     + (n - 1) * ix86_vec_cost (mode, ix86_cost->sse_op);
-       }
+       stmt_cost = stmt_cost * 2
+                   + ix86_vec_cost (mode, ix86_cost->sse_op);
     }
 
   /* If we do elementwise loads into a vector then we are bound by

diff --git a/gcc/testsuite/gcc.target/i386/pr119919.c b/gcc/testsuite/gcc.target/i386/pr119919.c
index ed646561bd1f604aabbae723d9da20176e9ff561..e39819f682dbbd5573665716c315fbc59618fe9a 100644 (file)
--- a/gcc/testsuite/gcc.target/i386/pr119919.c
+++ b/gcc/testsuite/gcc.target/i386/pr119919.c
@@ -1,5 +1,5 @@
 /* { dg-do compile } */
-/* { dg-options "-O2 -msse2 -fdump-tree-vect-details" } */
+/* { dg-options "-O2 -msse2 -fdump-tree-vect-details -mtune=znver1" } */
 int a[9*9];
 bool b[9];
 void test()