From 453f4f23ad9c444e1725516da400af7a8a6da1a5 Mon Sep 17 00:00:00 2001
From: Samuel Iglesias Gonsalvez <siglesias@igalia.com>
Date: Thu, 24 Apr 2014 10:02:35 +0200
Subject: [PATCH 1/2] i965: Support instruction compaction on Gen < 6

Gen < 6 also supports instruction compaction. The i965 and ILK docs seem
to be missing the compaction tables, but the GM45 docs do contain them.
It seems likely that they're identical.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=77742

Signed-off-by: Samuel Iglesias Gonsalvez <siglesias@igalia.com>
---
 src/mesa/drivers/dri/i965/brw_eu_compact.c | 154 ++++++++++++++++++++++++++++-
 1 file changed, 151 insertions(+), 3 deletions(-)

diff --git a/src/mesa/drivers/dri/i965/brw_eu_compact.c b/src/mesa/drivers/dri/i965/brw_eu_compact.c
index 3004843..1d453f3 100644
--- a/src/mesa/drivers/dri/i965/brw_eu_compact.c
+++ b/src/mesa/drivers/dri/i965/brw_eu_compact.c
@@ -40,6 +40,146 @@
 #include "brw_context.h"
 #include "brw_eu.h"
 
+static const uint32_t gen4_control_index_table[32] = {
+   0b00000000000000000,
+   0b01000000000000000,
+   0b00110000000000000,
+   0b00000000000000010,
+   0b00100000000000000,
+   0b00010000000000000,
+   0b01000000000100000,
+   0b01000000100000000,
+   0b01010000000100000,
+   0b00000000100000010,
+   0b11000000000000000,
+   0b00001000100000010,
+   0b01001000100000000,
+   0b00000000100000000,
+   0b11000000000100000,
+   0b00001000100000000,
+   0b10110000000000000,
+   0b11010000000100000,
+   0b00110000100000000,
+   0b00100000100000000,
+   0b01000000000001000,
+   0b01000000000000100,
+   0b00111100000000000,
+   0b00101011000000000,
+   0b00110000000000010,
+   0b00010000100000000,
+   0b01000000000100100,
+   0b01000000000101000,
+   0b00110000000000110,
+   0b00000000000001010,
+   0b01010000000101000,
+   0b01010000000100100
+};
+
+static const uint32_t gen4_datatype_table[32] = {
+   0b001000000000100001,
+   0b001011010110101101,
+   0b001000001000110001,
+   0b001111011110111101,
+   0b001011010110101100,
+   0b001000000110101101,
+   0b001000000000100000,
+   0b010100010110110001,
+   0b001100011000101101,
+   0b001000000000100010,
+   0b001000001000110110,
+   0b010000001000110001,
+   0b001000001000110010,
+   0b011000001000110010,
+   0b001111011110111100,
+   0b001000000100101000,
+   0b010100011000110001,
+   0b001010010100101001,
+   0b001000001000101001,
+   0b010000001000110110,
+   0b101000001000110001,
+   0b001011011000101101,
+   0b001000000100001001,
+   0b001011011000101100,
+   0b110100011000110001,
+   0b001000001110111101,
+   0b110000001000110001,
+   0b011000000100101010,
+   0b101000001000101001,
+   0b001011010110001100,
+   0b001000000110100001,
+   0b001010010100001000
+};
+
+static const uint16_t gen4_subreg_table[32] = {
+   0b000000000000000,
+   0b000000010000000,
+   0b000001000000000,
+   0b000100000000000,
+   0b000000000100000,
+   0b100000000000000,
+   0b000000000010000,
+   0b001100000000000,
+   0b001010000000000,
+   0b000000100000000,
+   0b001000000000000,
+   0b000000000001000,
+   0b000000001000000,
+   0b000000000000001,
+   0b000010000000000,
+   0b000000010100000,
+   0b000000000000111,
+   0b000001000100000,
+   0b011000000000000,
+   0b000000110000000,
+   0b000000000000010,
+   0b000000000000100,
+   0b000000001100000,
+   0b000100000000010,
+   0b001110011000110,
+   0b001110100001000,
+   0b000110011000110,
+   0b000001000011000,
+   0b000110010000100,
+   0b001100000000110,
+   0b000000010000110,
+   0b000001000110000
+};
+
+static const uint16_t gen4_src_index_table[32] = {
+   0b000000000000,
+   0b010001101000,
+   0b010110001000,
+   0b011010010000,
+   0b001101001000,
+   0b010110001010,
+   0b010101110000,
+   0b011001111000,
+   0b001000101000,
+   0b000000101000,
+   0b010001010000,
+   0b111101101100,
+   0b010110001100,
+   0b010001101100,
+   0b011010010100,
+   0b010001001100,
+   0b001100101000,
+   0b000000000010,
+   0b111101001100,
+   0b011001101000,
+   0b010101001000,
+   0b000000000100,
+   0b000000101100,
+   0b010001101010,
+   0b000000111000,
+   0b010101011000,
+   0b000100100000,
+   0b010110000000,
+   0b010000000100,
+   0b010000111000,
+   0b000101100000,
+   0b111101110100
+};
+
 static const uint32_t gen6_control_index_table[32] = {
    0b00000000000000000,
    0b01000000000000000,
@@ -633,6 +773,10 @@ update_uip_jip(struct brw_instruction *insn, int this_old_ip,
 void
 brw_init_compaction_tables(struct brw_context *brw)
 {
+   assert(gen4_control_index_table[ARRAY_SIZE(gen4_control_index_table) - 1] != 0);
+   assert(gen4_datatype_table[ARRAY_SIZE(gen4_datatype_table) - 1] != 0);
+   assert(gen4_subreg_table[ARRAY_SIZE(gen4_subreg_table) - 1] != 0);
+   assert(gen4_src_index_table[ARRAY_SIZE(gen4_src_index_table) - 1] != 0);
    assert(gen6_control_index_table[ARRAY_SIZE(gen6_control_index_table) - 1] != 0);
    assert(gen6_datatype_table[ARRAY_SIZE(gen6_datatype_table) - 1] != 0);
    assert(gen6_subreg_table[ARRAY_SIZE(gen6_subreg_table) - 1] != 0);
@@ -655,6 +799,13 @@ brw_init_compaction_tables(struct brw_context *brw)
       subreg_table = gen6_subreg_table;
       src_index_table = gen6_src_index_table;
       break;
+   case 5:
+   case 4:
+      control_index_table = gen4_control_index_table;
+      datatype_table = gen4_datatype_table;
+      subreg_table = gen4_subreg_table;
+      src_index_table = gen4_src_index_table;
+      break;
    default:
       return;
    }
@@ -674,9 +825,6 @@ brw_compact_instructions(struct brw_compile *p)
     */
    int old_ip[p->next_insn_offset / 8];
 
-   if (brw->gen < 6)
-      return;
-
    int src_offset;
    int offset = 0;
    int compacted_count = 0;
-- 
1.9.2