From ae96851db2dd9645a421bf0c83461333fccdab5d Mon Sep 17 00:00:00 2001
From: jsteube <jens.steube@gmail.com>
Date: Sun, 2 Jul 2017 18:08:27 +0200
Subject: [PATCH] Add code generators for
 switch_buffer_by_offset_4x4_carry_be_S()

---
 ..._switch_buffer_by_offset_4x4_carry_be_S.pl | 54 +++++++++++++++++++
 ..._switch_buffer_by_offset_4x4_carry_be_S.pl | 54 +++++++++++++++++++
 2 files changed, 108 insertions(+)
 create mode 100644 tools/code_generators/GEN_AMD_switch_buffer_by_offset_4x4_carry_be_S.pl
 create mode 100644 tools/code_generators/GEN_NV_switch_buffer_by_offset_4x4_carry_be_S.pl

diff --git a/tools/code_generators/GEN_AMD_switch_buffer_by_offset_4x4_carry_be_S.pl b/tools/code_generators/GEN_AMD_switch_buffer_by_offset_4x4_carry_be_S.pl
new file mode 100644
index 000000000..06586846d
--- /dev/null
+++ b/tools/code_generators/GEN_AMD_switch_buffer_by_offset_4x4_carry_be_S.pl
@@ -0,0 +1,54 @@
+#!/usr/bin/perl
+
+use strict;
+use warnings;
+
+for (my $i = 0, my $s = 0; $i < 16; $i++, $s++)
+{
+  printf ("    case %2d:\n", $i);
+  for (my $j = 32 - 1; $j >= 0; $j--)
+  {
+    my $jd4   = $j / 4;
+    my $jm4   = $j % 4;
+
+    my $js1d4 = ($j - $s - 1) / 4;
+    my $js1m4 = ($j - $s - 1) % 4;
+
+    my $js0d4 = ($j - $s - 0) / 4;
+    my $js0m4 = ($j - $s - 0) % 4;
+
+    next if (($j - $s) > 16);
+
+    my $c1 = "w";
+    my $c2 = "w";
+    my $c3 = "w";
+
+    if ($jd4 >= 4)
+    {
+      $jd4 -= 4;
+
+      $c1 = "c";
+    }
+
+    if ($js0d4 >= 4)
+    {
+      printf ("      %s%d[%d] = amd_bytealign_S (%s%d[%d],     0, offset);\n", $c1, $jd4, $jm4, $c2, $js1d4, $js1m4);
+    }
+    elsif ((($j - $s - 1) >= 0) && (($j - $s - 0) >= 0))
+    {
+      printf ("      %s%d[%d] = amd_bytealign_S (%s%d[%d], %s%d[%d], offset);\n", $c1, $jd4, $jm4, $c2, $js1d4, $js1m4, $c3, $js0d4, $js0m4);
+    }
+    elsif (($j - $s - 0) >= 0)
+    {
+      printf ("      %s%d[%d] = amd_bytealign_S (    0, %s%d[%d], offset);\n", $c1, $jd4, $jm4, $c2, $js0d4, $js0m4);
+    }
+    else
+    {
+      printf ("      %s%d[%d] = 0;\n", $c1, $jd4, $jm4);
+    }
+  }
+  printf ("\n");
+
+  printf ("      break;\n");
+  printf ("\n");
+}
diff --git a/tools/code_generators/GEN_NV_switch_buffer_by_offset_4x4_carry_be_S.pl b/tools/code_generators/GEN_NV_switch_buffer_by_offset_4x4_carry_be_S.pl
new file mode 100644
index 000000000..ac5041192
--- /dev/null
+++ b/tools/code_generators/GEN_NV_switch_buffer_by_offset_4x4_carry_be_S.pl
@@ -0,0 +1,54 @@
+#!/usr/bin/perl
+
+use strict;
+use warnings;
+
+for (my $i = 0, my $s = 0; $i < 16; $i++, $s++)
+{
+  printf ("    case %2d:\n", $i);
+  for (my $j = 32 - 1; $j >= 0; $j--)
+  {
+    my $jd4   = $j / 4;
+    my $jm4   = $j % 4;
+
+    my $js1d4 = ($j - $s - 1) / 4;
+    my $js1m4 = ($j - $s - 1) % 4;
+
+    my $js0d4 = ($j - $s - 0) / 4;
+    my $js0m4 = ($j - $s - 0) % 4;
+
+    next if (($j - $s) > 16);
+
+    my $c1 = "w";
+    my $c2 = "w";
+    my $c3 = "w";
+
+    if ($jd4 >= 4)
+    {
+      $jd4 -= 4;
+
+      $c1 = "c";
+    }
+
+    if ($js0d4 >= 4)
+    {
+      printf ("      %s%d[%d] = __byte_perm_S (    0, %s%d[%d], selector);\n", $c1, $jd4, $jm4, $c2, $js1d4, $js1m4);
+    }
+    elsif ((($j - $s - 1) >= 0) && (($j - $s - 0) >= 0))
+    {
+      printf ("      %s%d[%d] = __byte_perm_S (%s%d[%d], %s%d[%d], selector);\n", $c1, $jd4, $jm4, $c2, $js0d4, $js0m4, $c3, $js1d4, $js1m4);
+    }
+    elsif (($j - $s - 0) >= 0)
+    {
+      printf ("      %s%d[%d] = __byte_perm_S (%s%d[%d],     0, selector);\n", $c1, $jd4, $jm4, $c2, $js0d4, $js0m4);
+    }
+    else
+    {
+      printf ("      %s%d[%d] = 0;\n", $c1, $jd4, $jm4);
+    }
+  }
+  printf ("\n");
+
+  printf ("      break;\n");
+  printf ("\n");
+}