From 40e1992d0897c6730a48faf30f7c6b5b990a8e5a Mon Sep 17 00:00:00 2001
From: jsteube <jens.steube@gmail.com>
Date: Sun, 2 Jul 2017 15:03:33 +0200
Subject: [PATCH] Rename code generators for switch_buffer_by_offset_le_S to
 switch_buffer_by_offset_64x1_le_S() and add code generators for
 switch_buffer_by_offset_16x4_le_S()

---
 ...N_AMD_switch_buffer_by_offset_16x4_le_S.pl | 53 +++++++++++++++
 ..._AMD_switch_buffer_by_offset_64x1_le_S.pl} | 65 +++++++++++--------
 ...EN_NV_switch_buffer_by_offset_16x4_le_S.pl | 38 +++++++++++
 ...N_NV_switch_buffer_by_offset_64x1_le_S.pl} | 56 ++++++++--------
 4 files changed, 158 insertions(+), 54 deletions(-)
 create mode 100644 tools/code_generators/GEN_AMD_switch_buffer_by_offset_16x4_le_S.pl
 rename tools/code_generators/{GEN_AMD_switch_buffer_by_offset_le_S.pl => GEN_AMD_switch_buffer_by_offset_64x1_le_S.pl} (68%)
 create mode 100644 tools/code_generators/GEN_NV_switch_buffer_by_offset_16x4_le_S.pl
 rename tools/code_generators/{GEN_NV_switch_buffer_by_offset_le_S.pl => GEN_NV_switch_buffer_by_offset_64x1_le_S.pl} (94%)

diff --git a/tools/code_generators/GEN_AMD_switch_buffer_by_offset_16x4_le_S.pl b/tools/code_generators/GEN_AMD_switch_buffer_by_offset_16x4_le_S.pl
new file mode 100644
index 000000000..d9d897054
--- /dev/null
+++ b/tools/code_generators/GEN_AMD_switch_buffer_by_offset_16x4_le_S.pl
@@ -0,0 +1,53 @@
+#!/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 = 16 - 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 - $s - 0) % 4;
+
+    if ((($j - $s - 1) >= 0) && (($j - $s - 0) >= 0))
+    {
+      printf ("      w%d[%d] = amd_bytealign_S (w%d[%d], w%d[%d], offset_minus_4);\n", $jd4, $jm4, $js0d4, $js0m4, $js1d4, $js1m4);
+    }
+    elsif (($j - $s - 0) >= 0)
+    {
+      printf ("      w%d[%d] = amd_bytealign_S (w%d[%d],     0, offset_minus_4);\n", $jd4, $jm4, $js0d4, $js0m4);
+    }
+    else
+    {
+      printf ("      w%d[%d] = 0;\n", $jd4, $jm4);
+    }
+  }
+  printf ("\n");
+
+  printf ("      if (offset_mod_4 == 0)\n");
+  printf ("      {\n");
+  for (my $j = $i; $j < 16 - 1; $j++)
+  {
+    my $jd4   = $j / 4;
+    my $jm4   = $j % 4;
+
+    my $ja1d4 = ($j + 1) / 4;
+    my $ja1m4 = ($j + 1) % 4;
+
+    printf ("        w%d[%d] = w%d[%d];\n", $jd4, $jm4, $ja1d4, $ja1m4);
+  }
+  printf ("        w3[3] = 0;\n");
+  printf ("      }\n");
+  printf ("\n");
+
+  printf ("      break;\n");
+  printf ("\n");
+}
diff --git a/tools/code_generators/GEN_AMD_switch_buffer_by_offset_le_S.pl b/tools/code_generators/GEN_AMD_switch_buffer_by_offset_64x1_le_S.pl
similarity index 68%
rename from tools/code_generators/GEN_AMD_switch_buffer_by_offset_le_S.pl
rename to tools/code_generators/GEN_AMD_switch_buffer_by_offset_64x1_le_S.pl
index 143342238..294b51f73 100644
--- a/tools/code_generators/GEN_AMD_switch_buffer_by_offset_le_S.pl
+++ b/tools/code_generators/GEN_AMD_switch_buffer_by_offset_64x1_le_S.pl
@@ -1,27 +1,38 @@
-#!/usr/bin/perl
-
-use strict;
-use warnings;
-
-for (my $i = 0, my $s = 0; $i < 64; $i++, $s++)
-{
-  printf ("    case %2d:\n", $i);
-
-  for (my $j = 64 - 1; $j >= 0; $j--)
-  {
-    if ((($j - $s - 1) >= 0) && (($j - $s - 0) >= 0))
-    {
-      printf ("      w[%2d] = amd_bytealign_S (w[%2d], w[%2d], offset_minus_4);\n", $j, $j - $s - 0, $j - $s - 1);
-    }
-    elsif (($j - $s - 0) >= 0)
-    {
-      printf ("      w[%2d] = amd_bytealign_S (w[%2d],     0, offset_minus_4);\n", $j, $j - $s - 0);
-    }
-    else
-    {
-      printf ("      w[%2d] = 0;\n", $j);
-    }
-  }
-
-  printf ("      break;\n");
-}
+#!/usr/bin/perl
+
+use strict;
+use warnings;
+
+for (my $i = 0, my $s = 0; $i < 64; $i++, $s++)
+{
+  printf ("    case %2d:\n", $i);
+  for (my $j = 64 - 1; $j >= 0; $j--)
+  {
+    if ((($j - $s - 1) >= 0) && (($j - $s - 0) >= 0))
+    {
+      printf ("      w[%2d] = amd_bytealign_S (w[%2d], w[%2d], offset_minus_4);\n", $j, $j - $s - 0, $j - $s - 1);
+    }
+    elsif (($j - $s - 0) >= 0)
+    {
+      printf ("      w[%2d] = amd_bytealign_S (w[%2d],     0, offset_minus_4);\n", $j, $j - $s - 0);
+    }
+    else
+    {
+      printf ("      w[%2d] = 0;\n", $j);
+    }
+  }
+  printf ("\n");
+
+  printf ("      if (offset_mod_4 == 0)\n");
+  printf ("      {\n");
+  for (my $j = $i; $j < 64 - 1; $j++)
+  {
+    printf ("        w[%2d] = w[%2d];\n", $j, $j + 1);
+  }
+  printf ("        w[63] = 0;\n");
+  printf ("      }\n");
+  printf ("\n");
+
+  printf ("      break;\n");
+  printf ("\n");
+}
diff --git a/tools/code_generators/GEN_NV_switch_buffer_by_offset_16x4_le_S.pl b/tools/code_generators/GEN_NV_switch_buffer_by_offset_16x4_le_S.pl
new file mode 100644
index 000000000..bcdcb962b
--- /dev/null
+++ b/tools/code_generators/GEN_NV_switch_buffer_by_offset_16x4_le_S.pl
@@ -0,0 +1,38 @@
+#!/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 = 16 - 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 - $s - 0) % 4;
+
+    if ((($j - $s - 1) >= 0) && (($j - $s - 0) >= 0))
+    {
+      printf ("      w%d[%d] = __byte_perm_S (w%d[%d], w%d[%d], selector);\n", $jd4, $jm4, $js1d4, $js1m4, $js0d4, $js0m4);
+    }
+    elsif (($j - $s - 0) >= 0)
+    {
+      printf ("      w%d[%d] = __byte_perm_S (    0, w%d[%d], selector);\n", $jd4, $jm4, $js0d4, $js0m4);
+    }
+    else
+    {
+      printf ("      w%d[%d] = 0;\n", $jd4, $jm4, );
+    }
+  }
+  printf ("\n");
+
+  printf ("      break;\n");
+  printf ("\n");
+}
diff --git a/tools/code_generators/GEN_NV_switch_buffer_by_offset_le_S.pl b/tools/code_generators/GEN_NV_switch_buffer_by_offset_64x1_le_S.pl
similarity index 94%
rename from tools/code_generators/GEN_NV_switch_buffer_by_offset_le_S.pl
rename to tools/code_generators/GEN_NV_switch_buffer_by_offset_64x1_le_S.pl
index c9228ce0a..87e4ac527 100644
--- a/tools/code_generators/GEN_NV_switch_buffer_by_offset_le_S.pl
+++ b/tools/code_generators/GEN_NV_switch_buffer_by_offset_64x1_le_S.pl
@@ -1,27 +1,29 @@
-#!/usr/bin/perl
-
-use strict;
-use warnings;
-
-for (my $i = 0, my $s = 0; $i < 64; $i++, $s++)
-{
-  printf ("    case %2d:\n", $i);
-
-  for (my $j = 64 - 1; $j >= 0; $j--)
-  {
-    if ((($j - $s - 1) >= 0) && (($j - $s - 0) >= 0))
-    {
-      printf ("      w[%2d] = __byte_perm_S (w[%2d], w[%2d], selector);\n", $j, $j - $s - 1, $j - $s - 0);
-    }
-    elsif (($j - $s - 0) >= 0)
-    {
-      printf ("      w[%2d] = __byte_perm_S (    0, w[%2d], selector);\n", $j, $j - $s - 0);
-    }
-    else
-    {
-      printf ("      w[%2d] = 0;\n", $j);
-    }
-  }
-
-  printf ("      break;\n");
-}
+#!/usr/bin/perl
+
+use strict;
+use warnings;
+
+for (my $i = 0, my $s = 0; $i < 64; $i++, $s++)
+{
+  printf ("    case %2d:\n", $i);
+
+  for (my $j = 64 - 1; $j >= 0; $j--)
+  {
+    if ((($j - $s - 1) >= 0) && (($j - $s - 0) >= 0))
+    {
+      printf ("      w[%2d] = __byte_perm_S (w[%2d], w[%2d], selector);\n", $j, $j - $s - 1, $j - $s - 0);
+    }
+    elsif (($j - $s - 0) >= 0)
+    {
+      printf ("      w[%2d] = __byte_perm_S (    0, w[%2d], selector);\n", $j, $j - $s - 0);
+    }
+    else
+    {
+      printf ("      w[%2d] = 0;\n", $j);
+    }
+  }
+  printf ("\n");
+
+  printf ("      break;\n");
+  printf ("\n");
+}