i965/vec4: Add VEC4_OPCODE_PACK_4_BYTES.

Will be used by emit_pack_{s,u}norm_4x8().

src/mesa/drivers/dri/i965/brw_defines.h

@@ -908,6 +908,8 @@ enum opcode {
    SHADER_OPCODE_GEN4_SCRATCH_WRITE,
    SHADER_OPCODE_GEN7_SCRATCH_READ,
 
+   VEC4_OPCODE_PACK_BYTES,
+
    FS_OPCODE_DDX,
    FS_OPCODE_DDY,
    FS_OPCODE_PIXEL_X,

src/mesa/drivers/dri/i965/brw_shader.cpp

@@ -447,6 +447,9 @@ brw_instruction_name(enum opcode op)
    case SHADER_OPCODE_GEN7_SCRATCH_READ:
       return "gen7_scratch_read";
 
+   case VEC4_OPCODE_PACK_BYTES:
+      return "pack_bytes";
+
    case FS_OPCODE_DDX:
       return "ddx";
    case FS_OPCODE_DDY:

src/mesa/drivers/dri/i965/brw_vec4.cpp

@@ -352,6 +352,12 @@ vec4_visitor::opt_reduce_swizzle()
 
       /* Determine which channels of the sources are read. */
       switch (inst->opcode) {
+      case VEC4_OPCODE_PACK_BYTES:
+         swizzle[0] = 0;
+         swizzle[1] = 1;
+         swizzle[2] = 2;
+         swizzle[3] = 3;
+         break;
       case BRW_OPCODE_DP4:
       case BRW_OPCODE_DPH: /* FINISHME: DPH reads only three channels of src0,
                             *           but all four of src1.
@@ -1018,6 +1024,12 @@ vec4_instruction::reswizzle(int dst_writemask, int swizzle)
          if (src[i].file == BAD_FILE || src[i].file == IMM)
             continue;
 
+         /* Destination write mask doesn't correspond to source swizzle for the
+          * pack_bytes instruction.
+          */
+         if (opcode == VEC4_OPCODE_PACK_BYTES)
+            continue;
+
          for (int c = 0; c < 4; c++) {
             new_swizzle[c] = BRW_GET_SWZ(src[i].swizzle, BRW_GET_SWZ(swizzle, c));
          }

src/mesa/drivers/dri/i965/brw_vec4_generator.cpp

@@ -1495,6 +1495,41 @@ vec4_generator::generate_code(const cfg_t *cfg)
          generate_unpack_flags(inst, dst);
          break;
 
+      case VEC4_OPCODE_PACK_BYTES: {
+         /* Is effectively:
+          *
+          *   mov(8) dst<16,4,1>:UB src<4,1,0>:UB
+          *
+          * but destinations' only regioning is horizontal stride, so instead we
+          * have to use two instructions:
+          *
+          *   mov(4) dst<1>:UB     src<4,1,0>:UB
+          *   mov(4) dst.16<1>:UB  src.16<4,1,0>:UB
+          *
+          * where they pack the four bytes from the low and high four DW.
+          */
+         dst.type = BRW_REGISTER_TYPE_UB;
+
+         brw_set_default_access_mode(p, BRW_ALIGN_1);
+
+         src[0].type = BRW_REGISTER_TYPE_UB;
+         src[0].vstride = BRW_VERTICAL_STRIDE_4;
+         src[0].width = BRW_WIDTH_1;
+         src[0].hstride = BRW_HORIZONTAL_STRIDE_0;
+         struct brw_inst *insn = brw_MOV(p, dst, src[0]);
+         brw_inst_set_exec_size(brw, insn, BRW_EXECUTE_4);
+         brw_inst_set_no_dd_clear(brw, insn, true);
+
+         src[0].subnr = 16;
+         dst.subnr = 16;
+         insn = brw_MOV(p, dst, src[0]);
+         brw_inst_set_exec_size(brw, insn, BRW_EXECUTE_4);
+         brw_inst_set_no_dd_check(brw, insn, true);
+
+         brw_set_default_access_mode(p, BRW_ALIGN_16);
+         break;
+      }
+
       default:
          if (inst->opcode < (int) ARRAY_SIZE(opcode_descs)) {
             _mesa_problem(&brw->ctx, "Unsupported opcode in `%s' in vec4\n",