i965/fs: Split the GLSL IR -> FS LIR visitor to brw_fs_visitor.cpp.
authorEric Anholt <eric@anholt.net>
Tue, 24 May 2011 23:45:17 +0000 (16:45 -0700)
committerEric Anholt <eric@anholt.net>
Fri, 27 May 2011 15:51:40 +0000 (08:51 -0700)
We now have:
brw_fs.cpp handles calling out to everything and optimization.
brw_fs_visitor.cpp handles translating to our LIR.
brw_fs_emit.cpp handles emitting from our LIR to native code.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
src/mesa/drivers/dri/i965/Makefile
src/mesa/drivers/dri/i965/brw_fs.cpp
src/mesa/drivers/dri/i965/brw_fs.h
src/mesa/drivers/dri/i965/brw_fs_visitor.cpp [new file with mode: 0644]

index 9f27b8f98eccbf1c5ff13653bc28792a6c28b6d0..d80628b1e776ffc96f0aed7b14531c8e3d0288bc 100644 (file)
@@ -117,6 +117,7 @@ CXX_SOURCES = \
        brw_cubemap_normalize.cpp \
        brw_fs.cpp \
        brw_fs_emit.cpp \
+       brw_fs_visitor.cpp \
        brw_fs_channel_expressions.cpp \
        brw_fs_reg_allocate.cpp \
        brw_fs_schedule_instructions.cpp \
index 9bc7b444657bd2a7a9ba59fe528cb6d7f11c3160..bec43b439471f3436fb8f2c8ce7d4c3eb84ab3da 100644 (file)
  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  * IN THE SOFTWARE.
+ */
+
+/** @file brw_fs.cpp
  *
- * Authors:
- *    Eric Anholt <eric@anholt.net>
- *
+ * This file drives the GLSL IR -> LIR translation, contains the
+ * optimizations on the LIR, and drives the generation of native code
+ * from the LIR.
  */
 
 extern "C" {
@@ -48,8 +51,8 @@ extern "C" {
 
 #define MAX_INSTRUCTION (1 << 30)
 
-static int
-type_size(const struct glsl_type *type)
+int
+fs_visitor::type_size(const struct glsl_type *type)
 {
    unsigned int size, i;
 
@@ -204,7 +207,7 @@ fs_reg::fs_reg(class fs_visitor *v, const struct glsl_type *type)
    init();
 
    this->file = GRF;
-   this->reg = v->virtual_grf_alloc(type_size(type));
+   this->reg = v->virtual_grf_alloc(v->type_size(type));
    this->reg_offset = 0;
    this->type = brw_type_for_base_type(type);
 }
@@ -582,1679 +585,6 @@ fs_visitor::emit_math(fs_opcodes opcode, fs_reg dst, fs_reg src0, fs_reg src1)
    return inst;
 }
 
-void
-fs_visitor::visit(ir_variable *ir)
-{
-   fs_reg *reg = NULL;
-
-   if (variable_storage(ir))
-      return;
-
-   if (strcmp(ir->name, "gl_FragColor") == 0) {
-      this->frag_color = ir;
-   } else if (strcmp(ir->name, "gl_FragData") == 0) {
-      this->frag_data = ir;
-   } else if (strcmp(ir->name, "gl_FragDepth") == 0) {
-      this->frag_depth = ir;
-   }
-
-   if (ir->mode == ir_var_in) {
-      if (!strcmp(ir->name, "gl_FragCoord")) {
-        reg = emit_fragcoord_interpolation(ir);
-      } else if (!strcmp(ir->name, "gl_FrontFacing")) {
-        reg = emit_frontfacing_interpolation(ir);
-      } else {
-        reg = emit_general_interpolation(ir);
-      }
-      assert(reg);
-      hash_table_insert(this->variable_ht, reg, ir);
-      return;
-   }
-
-   if (ir->mode == ir_var_uniform) {
-      int param_index = c->prog_data.nr_params;
-
-      if (c->dispatch_width == 16) {
-        if (!variable_storage(ir)) {
-           fail("Failed to find uniform '%s' in 16-wide\n", ir->name);
-        }
-        return;
-      }
-
-      if (!strncmp(ir->name, "gl_", 3)) {
-        setup_builtin_uniform_values(ir);
-      } else {
-        setup_uniform_values(ir->location, ir->type);
-      }
-
-      reg = new(this->mem_ctx) fs_reg(UNIFORM, param_index);
-      reg->type = brw_type_for_base_type(ir->type);
-   }
-
-   if (!reg)
-      reg = new(this->mem_ctx) fs_reg(this, ir->type);
-
-   hash_table_insert(this->variable_ht, reg, ir);
-}
-
-void
-fs_visitor::visit(ir_dereference_variable *ir)
-{
-   fs_reg *reg = variable_storage(ir->var);
-   this->result = *reg;
-}
-
-void
-fs_visitor::visit(ir_dereference_record *ir)
-{
-   const glsl_type *struct_type = ir->record->type;
-
-   ir->record->accept(this);
-
-   unsigned int offset = 0;
-   for (unsigned int i = 0; i < struct_type->length; i++) {
-      if (strcmp(struct_type->fields.structure[i].name, ir->field) == 0)
-        break;
-      offset += type_size(struct_type->fields.structure[i].type);
-   }
-   this->result.reg_offset += offset;
-   this->result.type = brw_type_for_base_type(ir->type);
-}
-
-void
-fs_visitor::visit(ir_dereference_array *ir)
-{
-   ir_constant *index;
-   int element_size;
-
-   ir->array->accept(this);
-   index = ir->array_index->as_constant();
-
-   element_size = type_size(ir->type);
-   this->result.type = brw_type_for_base_type(ir->type);
-
-   if (index) {
-      assert(this->result.file == UNIFORM ||
-            (this->result.file == GRF &&
-             this->result.reg != 0));
-      this->result.reg_offset += index->value.i[0] * element_size;
-   } else {
-      assert(!"FINISHME: non-constant array element");
-   }
-}
-
-/* Instruction selection: Produce a MOV.sat instead of
- * MIN(MAX(val, 0), 1) when possible.
- */
-bool
-fs_visitor::try_emit_saturate(ir_expression *ir)
-{
-   ir_rvalue *sat_val = ir->as_rvalue_to_saturate();
-
-   if (!sat_val)
-      return false;
-
-   this->result = reg_undef;
-   sat_val->accept(this);
-   fs_reg src = this->result;
-
-   this->result = fs_reg(this, ir->type);
-   fs_inst *inst = emit(BRW_OPCODE_MOV, this->result, src);
-   inst->saturate = true;
-
-   return true;
-}
-
-void
-fs_visitor::visit(ir_expression *ir)
-{
-   unsigned int operand;
-   fs_reg op[2], temp;
-   fs_inst *inst;
-
-   assert(ir->get_num_operands() <= 2);
-
-   if (try_emit_saturate(ir))
-      return;
-
-   /* This is where our caller would like us to put the result, if possible. */
-   fs_reg saved_result_storage = this->result;
-
-   for (operand = 0; operand < ir->get_num_operands(); operand++) {
-      this->result = reg_undef;
-      ir->operands[operand]->accept(this);
-      if (this->result.file == BAD_FILE) {
-        ir_print_visitor v;
-        fail("Failed to get tree for expression operand:\n");
-        ir->operands[operand]->accept(&v);
-      }
-      op[operand] = this->result;
-
-      /* Matrix expression operands should have been broken down to vector
-       * operations already.
-       */
-      assert(!ir->operands[operand]->type->is_matrix());
-      /* And then those vector operands should have been broken down to scalar.
-       */
-      assert(!ir->operands[operand]->type->is_vector());
-   }
-
-   /* Inherit storage from our parent if possible, and otherwise we
-    * alloc a temporary.
-    */
-   if (saved_result_storage.file == BAD_FILE) {
-      this->result = fs_reg(this, ir->type);
-   } else {
-      this->result = saved_result_storage;
-   }
-
-   switch (ir->operation) {
-   case ir_unop_logic_not:
-      /* Note that BRW_OPCODE_NOT is not appropriate here, since it is
-       * ones complement of the whole register, not just bit 0.
-       */
-      emit(BRW_OPCODE_XOR, this->result, op[0], fs_reg(1));
-      break;
-   case ir_unop_neg:
-      op[0].negate = !op[0].negate;
-      this->result = op[0];
-      break;
-   case ir_unop_abs:
-      op[0].abs = true;
-      op[0].negate = false;
-      this->result = op[0];
-      break;
-   case ir_unop_sign:
-      temp = fs_reg(this, ir->type);
-
-      /* Unalias the destination.  (imagine a = sign(a)) */
-      this->result = fs_reg(this, ir->type);
-
-      emit(BRW_OPCODE_MOV, this->result, fs_reg(0.0f));
-
-      inst = emit(BRW_OPCODE_CMP, reg_null_f, op[0], fs_reg(0.0f));
-      inst->conditional_mod = BRW_CONDITIONAL_G;
-      inst = emit(BRW_OPCODE_MOV, this->result, fs_reg(1.0f));
-      inst->predicated = true;
-
-      inst = emit(BRW_OPCODE_CMP, reg_null_f, op[0], fs_reg(0.0f));
-      inst->conditional_mod = BRW_CONDITIONAL_L;
-      inst = emit(BRW_OPCODE_MOV, this->result, fs_reg(-1.0f));
-      inst->predicated = true;
-
-      break;
-   case ir_unop_rcp:
-      emit_math(FS_OPCODE_RCP, this->result, op[0]);
-      break;
-
-   case ir_unop_exp2:
-      emit_math(FS_OPCODE_EXP2, this->result, op[0]);
-      break;
-   case ir_unop_log2:
-      emit_math(FS_OPCODE_LOG2, this->result, op[0]);
-      break;
-   case ir_unop_exp:
-   case ir_unop_log:
-      assert(!"not reached: should be handled by ir_explog_to_explog2");
-      break;
-   case ir_unop_sin:
-   case ir_unop_sin_reduced:
-      emit_math(FS_OPCODE_SIN, this->result, op[0]);
-      break;
-   case ir_unop_cos:
-   case ir_unop_cos_reduced:
-      emit_math(FS_OPCODE_COS, this->result, op[0]);
-      break;
-
-   case ir_unop_dFdx:
-      emit(FS_OPCODE_DDX, this->result, op[0]);
-      break;
-   case ir_unop_dFdy:
-      emit(FS_OPCODE_DDY, this->result, op[0]);
-      break;
-
-   case ir_binop_add:
-      emit(BRW_OPCODE_ADD, this->result, op[0], op[1]);
-      break;
-   case ir_binop_sub:
-      assert(!"not reached: should be handled by ir_sub_to_add_neg");
-      break;
-
-   case ir_binop_mul:
-      emit(BRW_OPCODE_MUL, this->result, op[0], op[1]);
-      break;
-   case ir_binop_div:
-      assert(!"not reached: should be handled by ir_div_to_mul_rcp");
-      break;
-   case ir_binop_mod:
-      assert(!"ir_binop_mod should have been converted to b * fract(a/b)");
-      break;
-
-   case ir_binop_less:
-   case ir_binop_greater:
-   case ir_binop_lequal:
-   case ir_binop_gequal:
-   case ir_binop_equal:
-   case ir_binop_all_equal:
-   case ir_binop_nequal:
-   case ir_binop_any_nequal:
-      temp = this->result;
-      /* original gen4 does implicit conversion before comparison. */
-      if (intel->gen < 5)
-        temp.type = op[0].type;
-
-      inst = emit(BRW_OPCODE_CMP, temp, op[0], op[1]);
-      inst->conditional_mod = brw_conditional_for_comparison(ir->operation);
-      emit(BRW_OPCODE_AND, this->result, this->result, fs_reg(0x1));
-      break;
-
-   case ir_binop_logic_xor:
-      emit(BRW_OPCODE_XOR, this->result, op[0], op[1]);
-      break;
-
-   case ir_binop_logic_or:
-      emit(BRW_OPCODE_OR, this->result, op[0], op[1]);
-      break;
-
-   case ir_binop_logic_and:
-      emit(BRW_OPCODE_AND, this->result, op[0], op[1]);
-      break;
-
-   case ir_binop_dot:
-   case ir_unop_any:
-      assert(!"not reached: should be handled by brw_fs_channel_expressions");
-      break;
-
-   case ir_unop_noise:
-      assert(!"not reached: should be handled by lower_noise");
-      break;
-
-   case ir_quadop_vector:
-      assert(!"not reached: should be handled by lower_quadop_vector");
-      break;
-
-   case ir_unop_sqrt:
-      emit_math(FS_OPCODE_SQRT, this->result, op[0]);
-      break;
-
-   case ir_unop_rsq:
-      emit_math(FS_OPCODE_RSQ, this->result, op[0]);
-      break;
-
-   case ir_unop_i2f:
-   case ir_unop_b2f:
-   case ir_unop_b2i:
-   case ir_unop_f2i:
-      emit(BRW_OPCODE_MOV, this->result, op[0]);
-      break;
-   case ir_unop_f2b:
-   case ir_unop_i2b:
-      temp = this->result;
-      /* original gen4 does implicit conversion before comparison. */
-      if (intel->gen < 5)
-        temp.type = op[0].type;
-
-      inst = emit(BRW_OPCODE_CMP, temp, op[0], fs_reg(0.0f));
-      inst->conditional_mod = BRW_CONDITIONAL_NZ;
-      inst = emit(BRW_OPCODE_AND, this->result, this->result, fs_reg(1));
-      break;
-
-   case ir_unop_trunc:
-      emit(BRW_OPCODE_RNDZ, this->result, op[0]);
-      break;
-   case ir_unop_ceil:
-      op[0].negate = !op[0].negate;
-      inst = emit(BRW_OPCODE_RNDD, this->result, op[0]);
-      this->result.negate = true;
-      break;
-   case ir_unop_floor:
-      inst = emit(BRW_OPCODE_RNDD, this->result, op[0]);
-      break;
-   case ir_unop_fract:
-      inst = emit(BRW_OPCODE_FRC, this->result, op[0]);
-      break;
-   case ir_unop_round_even:
-      emit(BRW_OPCODE_RNDE, this->result, op[0]);
-      break;
-
-   case ir_binop_min:
-      /* Unalias the destination */
-      this->result = fs_reg(this, ir->type);
-
-      inst = emit(BRW_OPCODE_CMP, this->result, op[0], op[1]);
-      inst->conditional_mod = BRW_CONDITIONAL_L;
-
-      inst = emit(BRW_OPCODE_SEL, this->result, op[0], op[1]);
-      inst->predicated = true;
-      break;
-   case ir_binop_max:
-      /* Unalias the destination */
-      this->result = fs_reg(this, ir->type);
-
-      inst = emit(BRW_OPCODE_CMP, this->result, op[0], op[1]);
-      inst->conditional_mod = BRW_CONDITIONAL_G;
-
-      inst = emit(BRW_OPCODE_SEL, this->result, op[0], op[1]);
-      inst->predicated = true;
-      break;
-
-   case ir_binop_pow:
-      emit_math(FS_OPCODE_POW, this->result, op[0], op[1]);
-      break;
-
-   case ir_unop_bit_not:
-      inst = emit(BRW_OPCODE_NOT, this->result, op[0]);
-      break;
-   case ir_binop_bit_and:
-      inst = emit(BRW_OPCODE_AND, this->result, op[0], op[1]);
-      break;
-   case ir_binop_bit_xor:
-      inst = emit(BRW_OPCODE_XOR, this->result, op[0], op[1]);
-      break;
-   case ir_binop_bit_or:
-      inst = emit(BRW_OPCODE_OR, this->result, op[0], op[1]);
-      break;
-
-   case ir_unop_u2f:
-   case ir_binop_lshift:
-   case ir_binop_rshift:
-      assert(!"GLSL 1.30 features unsupported");
-      break;
-   }
-}
-
-void
-fs_visitor::emit_assignment_writes(fs_reg &l, fs_reg &r,
-                                  const glsl_type *type, bool predicated)
-{
-   switch (type->base_type) {
-   case GLSL_TYPE_FLOAT:
-   case GLSL_TYPE_UINT:
-   case GLSL_TYPE_INT:
-   case GLSL_TYPE_BOOL:
-      for (unsigned int i = 0; i < type->components(); i++) {
-        l.type = brw_type_for_base_type(type);
-        r.type = brw_type_for_base_type(type);
-
-        if (predicated || !l.equals(&r)) {
-           fs_inst *inst = emit(BRW_OPCODE_MOV, l, r);
-           inst->predicated = predicated;
-        }
-
-        l.reg_offset++;
-        r.reg_offset++;
-      }
-      break;
-   case GLSL_TYPE_ARRAY:
-      for (unsigned int i = 0; i < type->length; i++) {
-        emit_assignment_writes(l, r, type->fields.array, predicated);
-      }
-      break;
-
-   case GLSL_TYPE_STRUCT:
-      for (unsigned int i = 0; i < type->length; i++) {
-        emit_assignment_writes(l, r, type->fields.structure[i].type,
-                               predicated);
-      }
-      break;
-
-   case GLSL_TYPE_SAMPLER:
-      break;
-
-   default:
-      assert(!"not reached");
-      break;
-   }
-}
-
-void
-fs_visitor::visit(ir_assignment *ir)
-{
-   struct fs_reg l, r;
-   fs_inst *inst;
-
-   /* FINISHME: arrays on the lhs */
-   this->result = reg_undef;
-   ir->lhs->accept(this);
-   l = this->result;
-
-   /* If we're doing a direct assignment, an RHS expression could
-    * drop its result right into our destination.  Otherwise, tell it
-    * not to.
-    */
-   if (ir->condition ||
-       !(ir->lhs->type->is_scalar() ||
-        (ir->lhs->type->is_vector() &&
-         ir->write_mask == (1 << ir->lhs->type->vector_elements) - 1))) {
-      this->result = reg_undef;
-   }
-
-   ir->rhs->accept(this);
-   r = this->result;
-
-   assert(l.file != BAD_FILE);
-   assert(r.file != BAD_FILE);
-
-   if (ir->condition) {
-      emit_bool_to_cond_code(ir->condition);
-   }
-
-   if (ir->lhs->type->is_scalar() ||
-       ir->lhs->type->is_vector()) {
-      for (int i = 0; i < ir->lhs->type->vector_elements; i++) {
-        if (ir->write_mask & (1 << i)) {
-           if (ir->condition) {
-              inst = emit(BRW_OPCODE_MOV, l, r);
-              inst->predicated = true;
-           } else if (!l.equals(&r)) {
-              inst = emit(BRW_OPCODE_MOV, l, r);
-           }
-
-           r.reg_offset++;
-        }
-        l.reg_offset++;
-      }
-   } else {
-      emit_assignment_writes(l, r, ir->lhs->type, ir->condition != NULL);
-   }
-}
-
-fs_inst *
-fs_visitor::emit_texture_gen4(ir_texture *ir, fs_reg dst, fs_reg coordinate,
-                             int sampler)
-{
-   int mlen;
-   int base_mrf = 1;
-   bool simd16 = false;
-   fs_reg orig_dst;
-
-   /* g0 header. */
-   mlen = 1;
-
-   if (ir->shadow_comparitor) {
-      for (int i = 0; i < ir->coordinate->type->vector_elements; i++) {
-        fs_inst *inst = emit(BRW_OPCODE_MOV,
-                             fs_reg(MRF, base_mrf + mlen + i), coordinate);
-        if (i < 3 && c->key.gl_clamp_mask[i] & (1 << sampler))
-           inst->saturate = true;
-
-        coordinate.reg_offset++;
-      }
-      /* gen4's SIMD8 sampler always has the slots for u,v,r present. */
-      mlen += 3;
-
-      if (ir->op == ir_tex) {
-        /* There's no plain shadow compare message, so we use shadow
-         * compare with a bias of 0.0.
-         */
-        emit(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen), fs_reg(0.0f));
-        mlen++;
-      } else if (ir->op == ir_txb) {
-        this->result = reg_undef;
-        ir->lod_info.bias->accept(this);
-        emit(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen), this->result);
-        mlen++;
-      } else {
-        assert(ir->op == ir_txl);
-        this->result = reg_undef;
-        ir->lod_info.lod->accept(this);
-        emit(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen), this->result);
-        mlen++;
-      }
-
-      this->result = reg_undef;
-      ir->shadow_comparitor->accept(this);
-      emit(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen), this->result);
-      mlen++;
-   } else if (ir->op == ir_tex) {
-      for (int i = 0; i < ir->coordinate->type->vector_elements; i++) {
-        fs_inst *inst = emit(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen + i),
-                             coordinate);
-        if (i < 3 && c->key.gl_clamp_mask[i] & (1 << sampler))
-           inst->saturate = true;
-        coordinate.reg_offset++;
-      }
-      /* gen4's SIMD8 sampler always has the slots for u,v,r present. */
-      mlen += 3;
-   } else if (ir->op == ir_txd) {
-      assert(!"TXD isn't supported on gen4 yet.");
-   } else {
-      /* Oh joy.  gen4 doesn't have SIMD8 non-shadow-compare bias/lod
-       * instructions.  We'll need to do SIMD16 here.
-       */
-      assert(ir->op == ir_txb || ir->op == ir_txl);
-
-      for (int i = 0; i < ir->coordinate->type->vector_elements; i++) {
-        fs_inst *inst = emit(BRW_OPCODE_MOV, fs_reg(MRF,
-                                                    base_mrf + mlen + i * 2),
-                             coordinate);
-        if (i < 3 && c->key.gl_clamp_mask[i] & (1 << sampler))
-           inst->saturate = true;
-        coordinate.reg_offset++;
-      }
-
-      /* lod/bias appears after u/v/r. */
-      mlen += 6;
-
-      if (ir->op == ir_txb) {
-        this->result = reg_undef;
-        ir->lod_info.bias->accept(this);
-        emit(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen), this->result);
-        mlen++;
-      } else {
-        this->result = reg_undef;
-        ir->lod_info.lod->accept(this);
-        emit(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen), this->result);
-        mlen++;
-      }
-
-      /* The unused upper half. */
-      mlen++;
-
-      /* Now, since we're doing simd16, the return is 2 interleaved
-       * vec4s where the odd-indexed ones are junk. We'll need to move
-       * this weirdness around to the expected layout.
-       */
-      simd16 = true;
-      orig_dst = dst;
-      dst = fs_reg(this, glsl_type::get_array_instance(glsl_type::vec4_type,
-                                                      2));
-      dst.type = BRW_REGISTER_TYPE_F;
-   }
-
-   fs_inst *inst = NULL;
-   switch (ir->op) {
-   case ir_tex:
-      inst = emit(FS_OPCODE_TEX, dst);
-      break;
-   case ir_txb:
-      inst = emit(FS_OPCODE_TXB, dst);
-      break;
-   case ir_txl:
-      inst = emit(FS_OPCODE_TXL, dst);
-      break;
-   case ir_txd:
-      inst = emit(FS_OPCODE_TXD, dst);
-      break;
-   case ir_txf:
-      assert(!"GLSL 1.30 features unsupported");
-      break;
-   }
-   inst->base_mrf = base_mrf;
-   inst->mlen = mlen;
-   inst->header_present = true;
-
-   if (simd16) {
-      for (int i = 0; i < 4; i++) {
-        emit(BRW_OPCODE_MOV, orig_dst, dst);
-        orig_dst.reg_offset++;
-        dst.reg_offset += 2;
-      }
-   }
-
-   return inst;
-}
-
-/* gen5's sampler has slots for u, v, r, array index, then optional
- * parameters like shadow comparitor or LOD bias.  If optional
- * parameters aren't present, those base slots are optional and don't
- * need to be included in the message.
- *
- * We don't fill in the unnecessary slots regardless, which may look
- * surprising in the disassembly.
- */
-fs_inst *
-fs_visitor::emit_texture_gen5(ir_texture *ir, fs_reg dst, fs_reg coordinate,
-                             int sampler)
-{
-   int mlen = 0;
-   int base_mrf = 2;
-   int reg_width = c->dispatch_width / 8;
-   bool header_present = false;
-
-   if (ir->offset) {
-      /* The offsets set up by the ir_texture visitor are in the
-       * m1 header, so we can't go headerless.
-       */
-      header_present = true;
-      mlen++;
-      base_mrf--;
-   }
-
-   for (int i = 0; i < ir->coordinate->type->vector_elements; i++) {
-      fs_inst *inst = emit(BRW_OPCODE_MOV,
-                          fs_reg(MRF, base_mrf + mlen + i * reg_width),
-                          coordinate);
-      if (i < 3 && c->key.gl_clamp_mask[i] & (1 << sampler))
-        inst->saturate = true;
-      coordinate.reg_offset++;
-   }
-   mlen += ir->coordinate->type->vector_elements * reg_width;
-
-   if (ir->shadow_comparitor) {
-      mlen = MAX2(mlen, header_present + 4 * reg_width);
-
-      this->result = reg_undef;
-      ir->shadow_comparitor->accept(this);
-      emit(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen), this->result);
-      mlen += reg_width;
-   }
-
-   fs_inst *inst = NULL;
-   switch (ir->op) {
-   case ir_tex:
-      inst = emit(FS_OPCODE_TEX, dst);
-      break;
-   case ir_txb:
-      this->result = reg_undef;
-      ir->lod_info.bias->accept(this);
-      mlen = MAX2(mlen, header_present + 4 * reg_width);
-      emit(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen), this->result);
-      mlen += reg_width;
-
-      inst = emit(FS_OPCODE_TXB, dst);
-
-      break;
-   case ir_txl:
-      this->result = reg_undef;
-      ir->lod_info.lod->accept(this);
-      mlen = MAX2(mlen, header_present + 4 * reg_width);
-      emit(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen), this->result);
-      mlen += reg_width;
-
-      inst = emit(FS_OPCODE_TXL, dst);
-      break;
-   case ir_txd:
-   case ir_txf:
-      assert(!"GLSL 1.30 features unsupported");
-      break;
-   }
-   inst->base_mrf = base_mrf;
-   inst->mlen = mlen;
-   inst->header_present = header_present;
-
-   if (mlen > 11) {
-      fail("Message length >11 disallowed by hardware\n");
-   }
-
-   return inst;
-}
-
-fs_inst *
-fs_visitor::emit_texture_gen7(ir_texture *ir, fs_reg dst, fs_reg coordinate,
-                             int sampler)
-{
-   int mlen = 0;
-   int base_mrf = 2;
-   int reg_width = c->dispatch_width / 8;
-   bool header_present = false;
-
-   if (ir->offset) {
-      /* The offsets set up by the ir_texture visitor are in the
-       * m1 header, so we can't go headerless.
-       */
-      header_present = true;
-      mlen++;
-      base_mrf--;
-   }
-
-   if (ir->shadow_comparitor) {
-      ir->shadow_comparitor->accept(this);
-      emit(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen), this->result);
-      mlen += reg_width;
-   }
-
-   /* Set up the LOD info */
-   switch (ir->op) {
-   case ir_tex:
-      break;
-   case ir_txb:
-      ir->lod_info.bias->accept(this);
-      emit(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen), this->result);
-      mlen += reg_width;
-      break;
-   case ir_txl:
-      ir->lod_info.lod->accept(this);
-      emit(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen), this->result);
-      mlen += reg_width;
-      break;
-   case ir_txd:
-   case ir_txf:
-      assert(!"GLSL 1.30 features unsupported");
-      break;
-   }
-
-   /* Set up the coordinate */
-   for (int i = 0; i < ir->coordinate->type->vector_elements; i++) {
-      fs_inst *inst = emit(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen),
-                          coordinate);
-      if (i < 3 && c->key.gl_clamp_mask[i] & (1 << sampler))
-        inst->saturate = true;
-      coordinate.reg_offset++;
-      mlen += reg_width;
-   }
-
-   /* Generate the SEND */
-   fs_inst *inst = NULL;
-   switch (ir->op) {
-   case ir_tex: inst = emit(FS_OPCODE_TEX, dst); break;
-   case ir_txb: inst = emit(FS_OPCODE_TXB, dst); break;
-   case ir_txl: inst = emit(FS_OPCODE_TXL, dst); break;
-   case ir_txd: inst = emit(FS_OPCODE_TXD, dst); break;
-   case ir_txf: assert(!"TXF unsupported.");
-   }
-   inst->base_mrf = base_mrf;
-   inst->mlen = mlen;
-   inst->header_present = header_present;
-
-   if (mlen > 11) {
-      fail("Message length >11 disallowed by hardware\n");
-   }
-
-   return inst;
-}
-
-void
-fs_visitor::visit(ir_texture *ir)
-{
-   int sampler;
-   fs_inst *inst = NULL;
-
-   this->result = reg_undef;
-   ir->coordinate->accept(this);
-   fs_reg coordinate = this->result;
-
-   if (ir->offset != NULL) {
-      ir_constant *offset = ir->offset->as_constant();
-      assert(offset != NULL);
-
-      signed char offsets[3];
-      for (unsigned i = 0; i < ir->offset->type->vector_elements; i++)
-        offsets[i] = (signed char) offset->value.i[i];
-
-      /* Combine all three offsets into a single unsigned dword:
-       *
-       *    bits 11:8 - U Offset (X component)
-       *    bits  7:4 - V Offset (Y component)
-       *    bits  3:0 - R Offset (Z component)
-       */
-      unsigned offset_bits = 0;
-      for (unsigned i = 0; i < ir->offset->type->vector_elements; i++) {
-        const unsigned shift = 4 * (2 - i);
-        offset_bits |= (offsets[i] << shift) & (0xF << shift);
-      }
-
-      /* Explicitly set up the message header by copying g0 to msg reg m1. */
-      emit(BRW_OPCODE_MOV, fs_reg(MRF, 1, BRW_REGISTER_TYPE_UD),
-          fs_reg(GRF, 0, BRW_REGISTER_TYPE_UD));
-
-      /* Then set the offset bits in DWord 2 of the message header. */
-      emit(BRW_OPCODE_MOV,
-          fs_reg(retype(brw_vec1_reg(BRW_MESSAGE_REGISTER_FILE, 1, 2),
-                        BRW_REGISTER_TYPE_UD)),
-          fs_reg(brw_imm_uw(offset_bits)));
-   }
-
-   /* Should be lowered by do_lower_texture_projection */
-   assert(!ir->projector);
-
-   sampler = _mesa_get_sampler_uniform_value(ir->sampler,
-                                            ctx->Shader.CurrentFragmentProgram,
-                                            &brw->fragment_program->Base);
-   sampler = c->fp->program.Base.SamplerUnits[sampler];
-
-   /* The 965 requires the EU to do the normalization of GL rectangle
-    * texture coordinates.  We use the program parameter state
-    * tracking to get the scaling factor.
-    */
-   if (ir->sampler->type->sampler_dimensionality == GLSL_SAMPLER_DIM_RECT) {
-      struct gl_program_parameter_list *params = c->fp->program.Base.Parameters;
-      int tokens[STATE_LENGTH] = {
-        STATE_INTERNAL,
-        STATE_TEXRECT_SCALE,
-        sampler,
-        0,
-        0
-      };
-
-      if (c->dispatch_width == 16) {
-        fail("rectangle scale uniform setup not supported on 16-wide\n");
-        this->result = fs_reg(this, ir->type);
-        return;
-      }
-
-      c->prog_data.param_convert[c->prog_data.nr_params] =
-        PARAM_NO_CONVERT;
-      c->prog_data.param_convert[c->prog_data.nr_params + 1] =
-        PARAM_NO_CONVERT;
-
-      fs_reg scale_x = fs_reg(UNIFORM, c->prog_data.nr_params);
-      fs_reg scale_y = fs_reg(UNIFORM, c->prog_data.nr_params + 1);
-      GLuint index = _mesa_add_state_reference(params,
-                                              (gl_state_index *)tokens);
-
-      this->param_index[c->prog_data.nr_params] = index;
-      this->param_offset[c->prog_data.nr_params] = 0;
-      c->prog_data.nr_params++;
-      this->param_index[c->prog_data.nr_params] = index;
-      this->param_offset[c->prog_data.nr_params] = 1;
-      c->prog_data.nr_params++;
-
-      fs_reg dst = fs_reg(this, ir->coordinate->type);
-      fs_reg src = coordinate;
-      coordinate = dst;
-
-      emit(BRW_OPCODE_MUL, dst, src, scale_x);
-      dst.reg_offset++;
-      src.reg_offset++;
-      emit(BRW_OPCODE_MUL, dst, src, scale_y);
-   }
-
-   /* Writemasking doesn't eliminate channels on SIMD8 texture
-    * samples, so don't worry about them.
-    */
-   fs_reg dst = fs_reg(this, glsl_type::vec4_type);
-
-   if (intel->gen >= 7) {
-      inst = emit_texture_gen7(ir, dst, coordinate, sampler);
-   } else if (intel->gen >= 5) {
-      inst = emit_texture_gen5(ir, dst, coordinate, sampler);
-   } else {
-      inst = emit_texture_gen4(ir, dst, coordinate, sampler);
-   }
-
-   /* If there's an offset, we already set up m1.  To avoid the implied move,
-    * use the null register.  Otherwise, we want an implied move from g0.
-    */
-   if (ir->offset != NULL || !inst->header_present)
-      inst->src[0] = reg_undef;
-   else
-      inst->src[0] = fs_reg(retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UW));
-
-   inst->sampler = sampler;
-
-   this->result = dst;
-
-   if (ir->shadow_comparitor)
-      inst->shadow_compare = true;
-
-   if (ir->type == glsl_type::float_type) {
-      /* Ignore DEPTH_TEXTURE_MODE swizzling. */
-      assert(ir->sampler->type->sampler_shadow);
-   } else if (c->key.tex_swizzles[inst->sampler] != SWIZZLE_NOOP) {
-      fs_reg swizzle_dst = fs_reg(this, glsl_type::vec4_type);
-
-      for (int i = 0; i < 4; i++) {
-        int swiz = GET_SWZ(c->key.tex_swizzles[inst->sampler], i);
-        fs_reg l = swizzle_dst;
-        l.reg_offset += i;
-
-        if (swiz == SWIZZLE_ZERO) {
-           emit(BRW_OPCODE_MOV, l, fs_reg(0.0f));
-        } else if (swiz == SWIZZLE_ONE) {
-           emit(BRW_OPCODE_MOV, l, fs_reg(1.0f));
-        } else {
-           fs_reg r = dst;
-           r.reg_offset += GET_SWZ(c->key.tex_swizzles[inst->sampler], i);
-           emit(BRW_OPCODE_MOV, l, r);
-        }
-      }
-      this->result = swizzle_dst;
-   }
-}
-
-void
-fs_visitor::visit(ir_swizzle *ir)
-{
-   this->result = reg_undef;
-   ir->val->accept(this);
-   fs_reg val = this->result;
-
-   if (ir->type->vector_elements == 1) {
-      this->result.reg_offset += ir->mask.x;
-      return;
-   }
-
-   fs_reg result = fs_reg(this, ir->type);
-   this->result = result;
-
-   for (unsigned int i = 0; i < ir->type->vector_elements; i++) {
-      fs_reg channel = val;
-      int swiz = 0;
-
-      switch (i) {
-      case 0:
-        swiz = ir->mask.x;
-        break;
-      case 1:
-        swiz = ir->mask.y;
-        break;
-      case 2:
-        swiz = ir->mask.z;
-        break;
-      case 3:
-        swiz = ir->mask.w;
-        break;
-      }
-
-      channel.reg_offset += swiz;
-      emit(BRW_OPCODE_MOV, result, channel);
-      result.reg_offset++;
-   }
-}
-
-void
-fs_visitor::visit(ir_discard *ir)
-{
-   assert(ir->condition == NULL); /* FINISHME */
-
-   emit(FS_OPCODE_DISCARD);
-   kill_emitted = true;
-}
-
-void
-fs_visitor::visit(ir_constant *ir)
-{
-   /* Set this->result to reg at the bottom of the function because some code
-    * paths will cause this visitor to be applied to other fields.  This will
-    * cause the value stored in this->result to be modified.
-    *
-    * Make reg constant so that it doesn't get accidentally modified along the
-    * way.  Yes, I actually had this problem. :(
-    */
-   const fs_reg reg(this, ir->type);
-   fs_reg dst_reg = reg;
-
-   if (ir->type->is_array()) {
-      const unsigned size = type_size(ir->type->fields.array);
-
-      for (unsigned i = 0; i < ir->type->length; i++) {
-        this->result = reg_undef;
-        ir->array_elements[i]->accept(this);
-        fs_reg src_reg = this->result;
-
-        dst_reg.type = src_reg.type;
-        for (unsigned j = 0; j < size; j++) {
-           emit(BRW_OPCODE_MOV, dst_reg, src_reg);
-           src_reg.reg_offset++;
-           dst_reg.reg_offset++;
-        }
-      }
-   } else if (ir->type->is_record()) {
-      foreach_list(node, &ir->components) {
-        ir_instruction *const field = (ir_instruction *) node;
-        const unsigned size = type_size(field->type);
-
-        this->result = reg_undef;
-        field->accept(this);
-        fs_reg src_reg = this->result;
-
-        dst_reg.type = src_reg.type;
-        for (unsigned j = 0; j < size; j++) {
-           emit(BRW_OPCODE_MOV, dst_reg, src_reg);
-           src_reg.reg_offset++;
-           dst_reg.reg_offset++;
-        }
-      }
-   } else {
-      const unsigned size = type_size(ir->type);
-
-      for (unsigned i = 0; i < size; i++) {
-        switch (ir->type->base_type) {
-        case GLSL_TYPE_FLOAT:
-           emit(BRW_OPCODE_MOV, dst_reg, fs_reg(ir->value.f[i]));
-           break;
-        case GLSL_TYPE_UINT:
-           emit(BRW_OPCODE_MOV, dst_reg, fs_reg(ir->value.u[i]));
-           break;
-        case GLSL_TYPE_INT:
-           emit(BRW_OPCODE_MOV, dst_reg, fs_reg(ir->value.i[i]));
-           break;
-        case GLSL_TYPE_BOOL:
-           emit(BRW_OPCODE_MOV, dst_reg, fs_reg((int)ir->value.b[i]));
-           break;
-        default:
-           assert(!"Non-float/uint/int/bool constant");
-        }
-        dst_reg.reg_offset++;
-      }
-   }
-
-   this->result = reg;
-}
-
-void
-fs_visitor::emit_bool_to_cond_code(ir_rvalue *ir)
-{
-   ir_expression *expr = ir->as_expression();
-
-   if (expr) {
-      fs_reg op[2];
-      fs_inst *inst;
-
-      assert(expr->get_num_operands() <= 2);
-      for (unsigned int i = 0; i < expr->get_num_operands(); i++) {
-        assert(expr->operands[i]->type->is_scalar());
-
-        this->result = reg_undef;
-        expr->operands[i]->accept(this);
-        op[i] = this->result;
-      }
-
-      switch (expr->operation) {
-      case ir_unop_logic_not:
-        inst = emit(BRW_OPCODE_AND, reg_null_d, op[0], fs_reg(1));
-        inst->conditional_mod = BRW_CONDITIONAL_Z;
-        break;
-
-      case ir_binop_logic_xor:
-        inst = emit(BRW_OPCODE_XOR, reg_null_d, op[0], op[1]);
-        inst->conditional_mod = BRW_CONDITIONAL_NZ;
-        break;
-
-      case ir_binop_logic_or:
-        inst = emit(BRW_OPCODE_OR, reg_null_d, op[0], op[1]);
-        inst->conditional_mod = BRW_CONDITIONAL_NZ;
-        break;
-
-      case ir_binop_logic_and:
-        inst = emit(BRW_OPCODE_AND, reg_null_d, op[0], op[1]);
-        inst->conditional_mod = BRW_CONDITIONAL_NZ;
-        break;
-
-      case ir_unop_f2b:
-        if (intel->gen >= 6) {
-           inst = emit(BRW_OPCODE_CMP, reg_null_d, op[0], fs_reg(0.0f));
-        } else {
-           inst = emit(BRW_OPCODE_MOV, reg_null_f, op[0]);
-        }
-        inst->conditional_mod = BRW_CONDITIONAL_NZ;
-        break;
-
-      case ir_unop_i2b:
-        if (intel->gen >= 6) {
-           inst = emit(BRW_OPCODE_CMP, reg_null_d, op[0], fs_reg(0));
-        } else {
-           inst = emit(BRW_OPCODE_MOV, reg_null_d, op[0]);
-        }
-        inst->conditional_mod = BRW_CONDITIONAL_NZ;
-        break;
-
-      case ir_binop_greater:
-      case ir_binop_gequal:
-      case ir_binop_less:
-      case ir_binop_lequal:
-      case ir_binop_equal:
-      case ir_binop_all_equal:
-      case ir_binop_nequal:
-      case ir_binop_any_nequal:
-        inst = emit(BRW_OPCODE_CMP, reg_null_cmp, op[0], op[1]);
-        inst->conditional_mod =
-           brw_conditional_for_comparison(expr->operation);
-        break;
-
-      default:
-        assert(!"not reached");
-        fail("bad cond code\n");
-        break;
-      }
-      return;
-   }
-
-   this->result = reg_undef;
-   ir->accept(this);
-
-   if (intel->gen >= 6) {
-      fs_inst *inst = emit(BRW_OPCODE_AND, reg_null_d, this->result, fs_reg(1));
-      inst->conditional_mod = BRW_CONDITIONAL_NZ;
-   } else {
-      fs_inst *inst = emit(BRW_OPCODE_MOV, reg_null_d, this->result);
-      inst->conditional_mod = BRW_CONDITIONAL_NZ;
-   }
-}
-
-/**
- * Emit a gen6 IF statement with the comparison folded into the IF
- * instruction.
- */
-void
-fs_visitor::emit_if_gen6(ir_if *ir)
-{
-   ir_expression *expr = ir->condition->as_expression();
-
-   if (expr) {
-      fs_reg op[2];
-      fs_inst *inst;
-      fs_reg temp;
-
-      assert(expr->get_num_operands() <= 2);
-      for (unsigned int i = 0; i < expr->get_num_operands(); i++) {
-        assert(expr->operands[i]->type->is_scalar());
-
-        this->result = reg_undef;
-        expr->operands[i]->accept(this);
-        op[i] = this->result;
-      }
-
-      switch (expr->operation) {
-      case ir_unop_logic_not:
-        inst = emit(BRW_OPCODE_IF, temp, op[0], fs_reg(0));
-        inst->conditional_mod = BRW_CONDITIONAL_Z;
-        return;
-
-      case ir_binop_logic_xor:
-        inst = emit(BRW_OPCODE_IF, reg_null_d, op[0], op[1]);
-        inst->conditional_mod = BRW_CONDITIONAL_NZ;
-        return;
-
-      case ir_binop_logic_or:
-        temp = fs_reg(this, glsl_type::bool_type);
-        emit(BRW_OPCODE_OR, temp, op[0], op[1]);
-        inst = emit(BRW_OPCODE_IF, reg_null_d, temp, fs_reg(0));
-        inst->conditional_mod = BRW_CONDITIONAL_NZ;
-        return;
-
-      case ir_binop_logic_and:
-        temp = fs_reg(this, glsl_type::bool_type);
-        emit(BRW_OPCODE_AND, temp, op[0], op[1]);
-        inst = emit(BRW_OPCODE_IF, reg_null_d, temp, fs_reg(0));
-        inst->conditional_mod = BRW_CONDITIONAL_NZ;
-        return;
-
-      case ir_unop_f2b:
-        inst = emit(BRW_OPCODE_IF, reg_null_f, op[0], fs_reg(0));
-        inst->conditional_mod = BRW_CONDITIONAL_NZ;
-        return;
-
-      case ir_unop_i2b:
-        inst = emit(BRW_OPCODE_IF, reg_null_d, op[0], fs_reg(0));
-        inst->conditional_mod = BRW_CONDITIONAL_NZ;
-        return;
-
-      case ir_binop_greater:
-      case ir_binop_gequal:
-      case ir_binop_less:
-      case ir_binop_lequal:
-      case ir_binop_equal:
-      case ir_binop_all_equal:
-      case ir_binop_nequal:
-      case ir_binop_any_nequal:
-        inst = emit(BRW_OPCODE_IF, reg_null_d, op[0], op[1]);
-        inst->conditional_mod =
-           brw_conditional_for_comparison(expr->operation);
-        return;
-      default:
-        assert(!"not reached");
-        inst = emit(BRW_OPCODE_IF, reg_null_d, op[0], fs_reg(0));
-        inst->conditional_mod = BRW_CONDITIONAL_NZ;
-        fail("bad condition\n");
-        return;
-      }
-      return;
-   }
-
-   this->result = reg_undef;
-   ir->condition->accept(this);
-
-   fs_inst *inst = emit(BRW_OPCODE_IF, reg_null_d, this->result, fs_reg(0));
-   inst->conditional_mod = BRW_CONDITIONAL_NZ;
-}
-
-void
-fs_visitor::visit(ir_if *ir)
-{
-   fs_inst *inst;
-
-   if (intel->gen != 6 && c->dispatch_width == 16) {
-      fail("Can't support (non-uniform) control flow on 16-wide\n");
-   }
-
-   /* Don't point the annotation at the if statement, because then it plus
-    * the then and else blocks get printed.
-    */
-   this->base_ir = ir->condition;
-
-   if (intel->gen == 6) {
-      emit_if_gen6(ir);
-   } else {
-      emit_bool_to_cond_code(ir->condition);
-
-      inst = emit(BRW_OPCODE_IF);
-      inst->predicated = true;
-   }
-
-   foreach_iter(exec_list_iterator, iter, ir->then_instructions) {
-      ir_instruction *ir = (ir_instruction *)iter.get();
-      this->base_ir = ir;
-      this->result = reg_undef;
-      ir->accept(this);
-   }
-
-   if (!ir->else_instructions.is_empty()) {
-      emit(BRW_OPCODE_ELSE);
-
-      foreach_iter(exec_list_iterator, iter, ir->else_instructions) {
-        ir_instruction *ir = (ir_instruction *)iter.get();
-        this->base_ir = ir;
-        this->result = reg_undef;
-        ir->accept(this);
-      }
-   }
-
-   emit(BRW_OPCODE_ENDIF);
-}
-
-void
-fs_visitor::visit(ir_loop *ir)
-{
-   fs_reg counter = reg_undef;
-
-   if (c->dispatch_width == 16) {
-      fail("Can't support (non-uniform) control flow on 16-wide\n");
-   }
-
-   if (ir->counter) {
-      this->base_ir = ir->counter;
-      ir->counter->accept(this);
-      counter = *(variable_storage(ir->counter));
-
-      if (ir->from) {
-        this->result = counter;
-
-        this->base_ir = ir->from;
-        this->result = counter;
-        ir->from->accept(this);
-
-        if (!this->result.equals(&counter))
-           emit(BRW_OPCODE_MOV, counter, this->result);
-      }
-   }
-
-   emit(BRW_OPCODE_DO);
-
-   if (ir->to) {
-      this->base_ir = ir->to;
-      this->result = reg_undef;
-      ir->to->accept(this);
-
-      fs_inst *inst = emit(BRW_OPCODE_CMP, reg_null_cmp, counter, this->result);
-      inst->conditional_mod = brw_conditional_for_comparison(ir->cmp);
-
-      inst = emit(BRW_OPCODE_BREAK);
-      inst->predicated = true;
-   }
-
-   foreach_iter(exec_list_iterator, iter, ir->body_instructions) {
-      ir_instruction *ir = (ir_instruction *)iter.get();
-
-      this->base_ir = ir;
-      this->result = reg_undef;
-      ir->accept(this);
-   }
-
-   if (ir->increment) {
-      this->base_ir = ir->increment;
-      this->result = reg_undef;
-      ir->increment->accept(this);
-      emit(BRW_OPCODE_ADD, counter, counter, this->result);
-   }
-
-   emit(BRW_OPCODE_WHILE);
-}
-
-void
-fs_visitor::visit(ir_loop_jump *ir)
-{
-   switch (ir->mode) {
-   case ir_loop_jump::jump_break:
-      emit(BRW_OPCODE_BREAK);
-      break;
-   case ir_loop_jump::jump_continue:
-      emit(BRW_OPCODE_CONTINUE);
-      break;
-   }
-}
-
-void
-fs_visitor::visit(ir_call *ir)
-{
-   assert(!"FINISHME");
-}
-
-void
-fs_visitor::visit(ir_return *ir)
-{
-   assert(!"FINISHME");
-}
-
-void
-fs_visitor::visit(ir_function *ir)
-{
-   /* Ignore function bodies other than main() -- we shouldn't see calls to
-    * them since they should all be inlined before we get to ir_to_mesa.
-    */
-   if (strcmp(ir->name, "main") == 0) {
-      const ir_function_signature *sig;
-      exec_list empty;
-
-      sig = ir->matching_signature(&empty);
-
-      assert(sig);
-
-      foreach_iter(exec_list_iterator, iter, sig->body) {
-        ir_instruction *ir = (ir_instruction *)iter.get();
-        this->base_ir = ir;
-        this->result = reg_undef;
-        ir->accept(this);
-      }
-   }
-}
-
-void
-fs_visitor::visit(ir_function_signature *ir)
-{
-   assert(!"not reached");
-   (void)ir;
-}
-
-fs_inst *
-fs_visitor::emit(fs_inst inst)
-{
-   fs_inst *list_inst = new(mem_ctx) fs_inst;
-   *list_inst = inst;
-
-   if (force_uncompressed_stack > 0)
-      list_inst->force_uncompressed = true;
-   else if (force_sechalf_stack > 0)
-      list_inst->force_sechalf = true;
-
-   list_inst->annotation = this->current_annotation;
-   list_inst->ir = this->base_ir;
-
-   this->instructions.push_tail(list_inst);
-
-   return list_inst;
-}
-
-/** Emits a dummy fragment shader consisting of magenta for bringup purposes. */
-void
-fs_visitor::emit_dummy_fs()
-{
-   /* Everyone's favorite color. */
-   emit(BRW_OPCODE_MOV, fs_reg(MRF, 2), fs_reg(1.0f));
-   emit(BRW_OPCODE_MOV, fs_reg(MRF, 3), fs_reg(0.0f));
-   emit(BRW_OPCODE_MOV, fs_reg(MRF, 4), fs_reg(1.0f));
-   emit(BRW_OPCODE_MOV, fs_reg(MRF, 5), fs_reg(0.0f));
-
-   fs_inst *write;
-   write = emit(FS_OPCODE_FB_WRITE, fs_reg(0), fs_reg(0));
-   write->base_mrf = 0;
-}
-
-/* The register location here is relative to the start of the URB
- * data.  It will get adjusted to be a real location before
- * generate_code() time.
- */
-struct brw_reg
-fs_visitor::interp_reg(int location, int channel)
-{
-   int regnr = urb_setup[location] * 2 + channel / 2;
-   int stride = (channel & 1) * 4;
-
-   assert(urb_setup[location] != -1);
-
-   return brw_vec1_grf(regnr, stride);
-}
-
-/** Emits the interpolation for the varying inputs. */
-void
-fs_visitor::emit_interpolation_setup_gen4()
-{
-   this->current_annotation = "compute pixel centers";
-   this->pixel_x = fs_reg(this, glsl_type::uint_type);
-   this->pixel_y = fs_reg(this, glsl_type::uint_type);
-   this->pixel_x.type = BRW_REGISTER_TYPE_UW;
-   this->pixel_y.type = BRW_REGISTER_TYPE_UW;
-
-   emit(FS_OPCODE_PIXEL_X, this->pixel_x);
-   emit(FS_OPCODE_PIXEL_Y, this->pixel_y);
-
-   this->current_annotation = "compute pixel deltas from v0";
-   if (brw->has_pln) {
-      this->delta_x = fs_reg(this, glsl_type::vec2_type);
-      this->delta_y = this->delta_x;
-      this->delta_y.reg_offset++;
-   } else {
-      this->delta_x = fs_reg(this, glsl_type::float_type);
-      this->delta_y = fs_reg(this, glsl_type::float_type);
-   }
-   emit(BRW_OPCODE_ADD, this->delta_x,
-       this->pixel_x, fs_reg(negate(brw_vec1_grf(1, 0))));
-   emit(BRW_OPCODE_ADD, this->delta_y,
-       this->pixel_y, fs_reg(negate(brw_vec1_grf(1, 1))));
-
-   this->current_annotation = "compute pos.w and 1/pos.w";
-   /* Compute wpos.w.  It's always in our setup, since it's needed to
-    * interpolate the other attributes.
-    */
-   this->wpos_w = fs_reg(this, glsl_type::float_type);
-   emit(FS_OPCODE_LINTERP, wpos_w, this->delta_x, this->delta_y,
-       interp_reg(FRAG_ATTRIB_WPOS, 3));
-   /* Compute the pixel 1/W value from wpos.w. */
-   this->pixel_w = fs_reg(this, glsl_type::float_type);
-   emit_math(FS_OPCODE_RCP, this->pixel_w, wpos_w);
-   this->current_annotation = NULL;
-}
-
-/** Emits the interpolation for the varying inputs. */
-void
-fs_visitor::emit_interpolation_setup_gen6()
-{
-   struct brw_reg g1_uw = retype(brw_vec1_grf(1, 0), BRW_REGISTER_TYPE_UW);
-
-   /* If the pixel centers end up used, the setup is the same as for gen4. */
-   this->current_annotation = "compute pixel centers";
-   fs_reg int_pixel_x = fs_reg(this, glsl_type::uint_type);
-   fs_reg int_pixel_y = fs_reg(this, glsl_type::uint_type);
-   int_pixel_x.type = BRW_REGISTER_TYPE_UW;
-   int_pixel_y.type = BRW_REGISTER_TYPE_UW;
-   emit(BRW_OPCODE_ADD,
-       int_pixel_x,
-       fs_reg(stride(suboffset(g1_uw, 4), 2, 4, 0)),
-       fs_reg(brw_imm_v(0x10101010)));
-   emit(BRW_OPCODE_ADD,
-       int_pixel_y,
-       fs_reg(stride(suboffset(g1_uw, 5), 2, 4, 0)),
-       fs_reg(brw_imm_v(0x11001100)));
-
-   /* As of gen6, we can no longer mix float and int sources.  We have
-    * to turn the integer pixel centers into floats for their actual
-    * use.
-    */
-   this->pixel_x = fs_reg(this, glsl_type::float_type);
-   this->pixel_y = fs_reg(this, glsl_type::float_type);
-   emit(BRW_OPCODE_MOV, this->pixel_x, int_pixel_x);
-   emit(BRW_OPCODE_MOV, this->pixel_y, int_pixel_y);
-
-   this->current_annotation = "compute pos.w";
-   this->pixel_w = fs_reg(brw_vec8_grf(c->source_w_reg, 0));
-   this->wpos_w = fs_reg(this, glsl_type::float_type);
-   emit_math(FS_OPCODE_RCP, this->wpos_w, this->pixel_w);
-
-   this->delta_x = fs_reg(brw_vec8_grf(2, 0));
-   this->delta_y = fs_reg(brw_vec8_grf(3, 0));
-
-   this->current_annotation = NULL;
-}
-
-void
-fs_visitor::emit_color_write(int index, int first_color_mrf, fs_reg color)
-{
-   int reg_width = c->dispatch_width / 8;
-
-   if (c->dispatch_width == 8 || intel->gen == 6) {
-      /* SIMD8 write looks like:
-       * m + 0: r0
-       * m + 1: r1
-       * m + 2: g0
-       * m + 3: g1
-       *
-       * gen6 SIMD16 DP write looks like:
-       * m + 0: r0
-       * m + 1: r1
-       * m + 2: g0
-       * m + 3: g1
-       * m + 4: b0
-       * m + 5: b1
-       * m + 6: a0
-       * m + 7: a1
-       */
-      emit(BRW_OPCODE_MOV, fs_reg(MRF, first_color_mrf + index * reg_width),
-          color);
-   } else {
-      /* pre-gen6 SIMD16 single source DP write looks like:
-       * m + 0: r0
-       * m + 1: g0
-       * m + 2: b0
-       * m + 3: a0
-       * m + 4: r1
-       * m + 5: g1
-       * m + 6: b1
-       * m + 7: a1
-       */
-      if (brw->has_compr4) {
-        /* By setting the high bit of the MRF register number, we
-         * indicate that we want COMPR4 mode - instead of doing the
-         * usual destination + 1 for the second half we get
-         * destination + 4.
-         */
-        emit(BRW_OPCODE_MOV,
-             fs_reg(MRF, BRW_MRF_COMPR4 + first_color_mrf + index), color);
-      } else {
-        push_force_uncompressed();
-        emit(BRW_OPCODE_MOV, fs_reg(MRF, first_color_mrf + index), color);
-        pop_force_uncompressed();
-
-        push_force_sechalf();
-        color.sechalf = true;
-        emit(BRW_OPCODE_MOV, fs_reg(MRF, first_color_mrf + index + 4), color);
-        pop_force_sechalf();
-        color.sechalf = false;
-      }
-   }
-}
-
-void
-fs_visitor::emit_fb_writes()
-{
-   this->current_annotation = "FB write header";
-   GLboolean header_present = GL_TRUE;
-   int nr = 0;
-   int reg_width = c->dispatch_width / 8;
-
-   if (intel->gen >= 6 &&
-       !this->kill_emitted &&
-       c->key.nr_color_regions == 1) {
-      header_present = false;
-   }
-
-   if (header_present) {
-      /* m0, m1 header */
-      nr += 2;
-   }
-
-   if (c->aa_dest_stencil_reg) {
-      push_force_uncompressed();
-      emit(BRW_OPCODE_MOV, fs_reg(MRF, nr++),
-          fs_reg(brw_vec8_grf(c->aa_dest_stencil_reg, 0)));
-      pop_force_uncompressed();
-   }
-
-   /* Reserve space for color. It'll be filled in per MRT below. */
-   int color_mrf = nr;
-   nr += 4 * reg_width;
-
-   if (c->source_depth_to_render_target) {
-      if (intel->gen == 6 && c->dispatch_width == 16) {
-        /* For outputting oDepth on gen6, SIMD8 writes have to be
-         * used.  This would require 8-wide moves of each half to
-         * message regs, kind of like pre-gen5 SIMD16 FB writes.
-         * Just bail on doing so for now.
-         */
-        fail("Missing support for simd16 depth writes on gen6\n");
-      }
-
-      if (c->computes_depth) {
-        /* Hand over gl_FragDepth. */
-        assert(this->frag_depth);
-        fs_reg depth = *(variable_storage(this->frag_depth));
-
-        emit(BRW_OPCODE_MOV, fs_reg(MRF, nr), depth);
-      } else {
-        /* Pass through the payload depth. */
-        emit(BRW_OPCODE_MOV, fs_reg(MRF, nr),
-             fs_reg(brw_vec8_grf(c->source_depth_reg, 0)));
-      }
-      nr += reg_width;
-   }
-
-   if (c->dest_depth_reg) {
-      emit(BRW_OPCODE_MOV, fs_reg(MRF, nr),
-          fs_reg(brw_vec8_grf(c->dest_depth_reg, 0)));
-      nr += reg_width;
-   }
-
-   fs_reg color = reg_undef;
-   if (this->frag_color)
-      color = *(variable_storage(this->frag_color));
-   else if (this->frag_data) {
-      color = *(variable_storage(this->frag_data));
-      color.type = BRW_REGISTER_TYPE_F;
-   }
-
-   for (int target = 0; target < c->key.nr_color_regions; target++) {
-      this->current_annotation = ralloc_asprintf(this->mem_ctx,
-                                                "FB write target %d",
-                                                target);
-      if (this->frag_color || this->frag_data) {
-        for (int i = 0; i < 4; i++) {
-           emit_color_write(i, color_mrf, color);
-           color.reg_offset++;
-        }
-      }
-
-      if (this->frag_color)
-        color.reg_offset -= 4;
-
-      fs_inst *inst = emit(FS_OPCODE_FB_WRITE);
-      inst->target = target;
-      inst->base_mrf = 0;
-      inst->mlen = nr;
-      if (target == c->key.nr_color_regions - 1)
-        inst->eot = true;
-      inst->header_present = header_present;
-   }
-
-   if (c->key.nr_color_regions == 0) {
-      if (c->key.alpha_test && (this->frag_color || this->frag_data)) {
-        /* If the alpha test is enabled but there's no color buffer,
-         * we still need to send alpha out the pipeline to our null
-         * renderbuffer.
-         */
-        color.reg_offset += 3;
-        emit_color_write(3, color_mrf, color);
-      }
-
-      fs_inst *inst = emit(FS_OPCODE_FB_WRITE);
-      inst->base_mrf = 0;
-      inst->mlen = nr;
-      inst->eot = true;
-      inst->header_present = header_present;
-   }
-
-   this->current_annotation = NULL;
-}
-
 /**
  * To be called after the last _mesa_add_state_reference() call, to
  * set up prog_data.param[] for assign_curb_setup() and
index 41e5004f37bd3d5d1da163c2dad994607c192f9a..d1f70ee2117723932b868672893e34ddd02052f9 100644 (file)
@@ -466,6 +466,8 @@ public:
       return emit(fs_inst(opcode, dst, src0, src1, src2));
    }
 
+   int type_size(const struct glsl_type *type);
+
    bool run();
    void setup_paramvalues_refs();
    void assign_curb_setup();
diff --git a/src/mesa/drivers/dri/i965/brw_fs_visitor.cpp b/src/mesa/drivers/dri/i965/brw_fs_visitor.cpp
new file mode 100644 (file)
index 0000000..1445582
--- /dev/null
@@ -0,0 +1,1724 @@
+/*
+ * Copyright Â© 2010 Intel Corporation
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+/** @file brw_fs_visitor.cpp
+ *
+ * This file supports generating the FS LIR from the GLSL IR.  The LIR
+ * makes it easier to do backend-specific optimizations than doing so
+ * in the GLSL IR or in the native code.
+ */
+extern "C" {
+
+#include <sys/types.h>
+
+#include "main/macros.h"
+#include "main/shaderobj.h"
+#include "main/uniforms.h"
+#include "program/prog_parameter.h"
+#include "program/prog_print.h"
+#include "program/prog_optimize.h"
+#include "program/register_allocate.h"
+#include "program/sampler.h"
+#include "program/hash_table.h"
+#include "brw_context.h"
+#include "brw_eu.h"
+#include "brw_wm.h"
+}
+#include "brw_shader.h"
+#include "brw_fs.h"
+#include "../glsl/glsl_types.h"
+#include "../glsl/ir_optimization.h"
+#include "../glsl/ir_print_visitor.h"
+
+void
+fs_visitor::visit(ir_variable *ir)
+{
+   fs_reg *reg = NULL;
+
+   if (variable_storage(ir))
+      return;
+
+   if (strcmp(ir->name, "gl_FragColor") == 0) {
+      this->frag_color = ir;
+   } else if (strcmp(ir->name, "gl_FragData") == 0) {
+      this->frag_data = ir;
+   } else if (strcmp(ir->name, "gl_FragDepth") == 0) {
+      this->frag_depth = ir;
+   }
+
+   if (ir->mode == ir_var_in) {
+      if (!strcmp(ir->name, "gl_FragCoord")) {
+        reg = emit_fragcoord_interpolation(ir);
+      } else if (!strcmp(ir->name, "gl_FrontFacing")) {
+        reg = emit_frontfacing_interpolation(ir);
+      } else {
+        reg = emit_general_interpolation(ir);
+      }
+      assert(reg);
+      hash_table_insert(this->variable_ht, reg, ir);
+      return;
+   }
+
+   if (ir->mode == ir_var_uniform) {
+      int param_index = c->prog_data.nr_params;
+
+      if (c->dispatch_width == 16) {
+        if (!variable_storage(ir)) {
+           fail("Failed to find uniform '%s' in 16-wide\n", ir->name);
+        }
+        return;
+      }
+
+      if (!strncmp(ir->name, "gl_", 3)) {
+        setup_builtin_uniform_values(ir);
+      } else {
+        setup_uniform_values(ir->location, ir->type);
+      }
+
+      reg = new(this->mem_ctx) fs_reg(UNIFORM, param_index);
+      reg->type = brw_type_for_base_type(ir->type);
+   }
+
+   if (!reg)
+      reg = new(this->mem_ctx) fs_reg(this, ir->type);
+
+   hash_table_insert(this->variable_ht, reg, ir);
+}
+
+void
+fs_visitor::visit(ir_dereference_variable *ir)
+{
+   fs_reg *reg = variable_storage(ir->var);
+   this->result = *reg;
+}
+
+void
+fs_visitor::visit(ir_dereference_record *ir)
+{
+   const glsl_type *struct_type = ir->record->type;
+
+   ir->record->accept(this);
+
+   unsigned int offset = 0;
+   for (unsigned int i = 0; i < struct_type->length; i++) {
+      if (strcmp(struct_type->fields.structure[i].name, ir->field) == 0)
+        break;
+      offset += type_size(struct_type->fields.structure[i].type);
+   }
+   this->result.reg_offset += offset;
+   this->result.type = brw_type_for_base_type(ir->type);
+}
+
+void
+fs_visitor::visit(ir_dereference_array *ir)
+{
+   ir_constant *index;
+   int element_size;
+
+   ir->array->accept(this);
+   index = ir->array_index->as_constant();
+
+   element_size = type_size(ir->type);
+   this->result.type = brw_type_for_base_type(ir->type);
+
+   if (index) {
+      assert(this->result.file == UNIFORM ||
+            (this->result.file == GRF &&
+             this->result.reg != 0));
+      this->result.reg_offset += index->value.i[0] * element_size;
+   } else {
+      assert(!"FINISHME: non-constant array element");
+   }
+}
+
+/* Instruction selection: Produce a MOV.sat instead of
+ * MIN(MAX(val, 0), 1) when possible.
+ */
+bool
+fs_visitor::try_emit_saturate(ir_expression *ir)
+{
+   ir_rvalue *sat_val = ir->as_rvalue_to_saturate();
+
+   if (!sat_val)
+      return false;
+
+   this->result = reg_undef;
+   sat_val->accept(this);
+   fs_reg src = this->result;
+
+   this->result = fs_reg(this, ir->type);
+   fs_inst *inst = emit(BRW_OPCODE_MOV, this->result, src);
+   inst->saturate = true;
+
+   return true;
+}
+
+void
+fs_visitor::visit(ir_expression *ir)
+{
+   unsigned int operand;
+   fs_reg op[2], temp;
+   fs_inst *inst;
+
+   assert(ir->get_num_operands() <= 2);
+
+   if (try_emit_saturate(ir))
+      return;
+
+   /* This is where our caller would like us to put the result, if possible. */
+   fs_reg saved_result_storage = this->result;
+
+   for (operand = 0; operand < ir->get_num_operands(); operand++) {
+      this->result = reg_undef;
+      ir->operands[operand]->accept(this);
+      if (this->result.file == BAD_FILE) {
+        ir_print_visitor v;
+        fail("Failed to get tree for expression operand:\n");
+        ir->operands[operand]->accept(&v);
+      }
+      op[operand] = this->result;
+
+      /* Matrix expression operands should have been broken down to vector
+       * operations already.
+       */
+      assert(!ir->operands[operand]->type->is_matrix());
+      /* And then those vector operands should have been broken down to scalar.
+       */
+      assert(!ir->operands[operand]->type->is_vector());
+   }
+
+   /* Inherit storage from our parent if possible, and otherwise we
+    * alloc a temporary.
+    */
+   if (saved_result_storage.file == BAD_FILE) {
+      this->result = fs_reg(this, ir->type);
+   } else {
+      this->result = saved_result_storage;
+   }
+
+   switch (ir->operation) {
+   case ir_unop_logic_not:
+      /* Note that BRW_OPCODE_NOT is not appropriate here, since it is
+       * ones complement of the whole register, not just bit 0.
+       */
+      emit(BRW_OPCODE_XOR, this->result, op[0], fs_reg(1));
+      break;
+   case ir_unop_neg:
+      op[0].negate = !op[0].negate;
+      this->result = op[0];
+      break;
+   case ir_unop_abs:
+      op[0].abs = true;
+      op[0].negate = false;
+      this->result = op[0];
+      break;
+   case ir_unop_sign:
+      temp = fs_reg(this, ir->type);
+
+      /* Unalias the destination.  (imagine a = sign(a)) */
+      this->result = fs_reg(this, ir->type);
+
+      emit(BRW_OPCODE_MOV, this->result, fs_reg(0.0f));
+
+      inst = emit(BRW_OPCODE_CMP, reg_null_f, op[0], fs_reg(0.0f));
+      inst->conditional_mod = BRW_CONDITIONAL_G;
+      inst = emit(BRW_OPCODE_MOV, this->result, fs_reg(1.0f));
+      inst->predicated = true;
+
+      inst = emit(BRW_OPCODE_CMP, reg_null_f, op[0], fs_reg(0.0f));
+      inst->conditional_mod = BRW_CONDITIONAL_L;
+      inst = emit(BRW_OPCODE_MOV, this->result, fs_reg(-1.0f));
+      inst->predicated = true;
+
+      break;
+   case ir_unop_rcp:
+      emit_math(FS_OPCODE_RCP, this->result, op[0]);
+      break;
+
+   case ir_unop_exp2:
+      emit_math(FS_OPCODE_EXP2, this->result, op[0]);
+      break;
+   case ir_unop_log2:
+      emit_math(FS_OPCODE_LOG2, this->result, op[0]);
+      break;
+   case ir_unop_exp:
+   case ir_unop_log:
+      assert(!"not reached: should be handled by ir_explog_to_explog2");
+      break;
+   case ir_unop_sin:
+   case ir_unop_sin_reduced:
+      emit_math(FS_OPCODE_SIN, this->result, op[0]);
+      break;
+   case ir_unop_cos:
+   case ir_unop_cos_reduced:
+      emit_math(FS_OPCODE_COS, this->result, op[0]);
+      break;
+
+   case ir_unop_dFdx:
+      emit(FS_OPCODE_DDX, this->result, op[0]);
+      break;
+   case ir_unop_dFdy:
+      emit(FS_OPCODE_DDY, this->result, op[0]);
+      break;
+
+   case ir_binop_add:
+      emit(BRW_OPCODE_ADD, this->result, op[0], op[1]);
+      break;
+   case ir_binop_sub:
+      assert(!"not reached: should be handled by ir_sub_to_add_neg");
+      break;
+
+   case ir_binop_mul:
+      emit(BRW_OPCODE_MUL, this->result, op[0], op[1]);
+      break;
+   case ir_binop_div:
+      assert(!"not reached: should be handled by ir_div_to_mul_rcp");
+      break;
+   case ir_binop_mod:
+      assert(!"ir_binop_mod should have been converted to b * fract(a/b)");
+      break;
+
+   case ir_binop_less:
+   case ir_binop_greater:
+   case ir_binop_lequal:
+   case ir_binop_gequal:
+   case ir_binop_equal:
+   case ir_binop_all_equal:
+   case ir_binop_nequal:
+   case ir_binop_any_nequal:
+      temp = this->result;
+      /* original gen4 does implicit conversion before comparison. */
+      if (intel->gen < 5)
+        temp.type = op[0].type;
+
+      inst = emit(BRW_OPCODE_CMP, temp, op[0], op[1]);
+      inst->conditional_mod = brw_conditional_for_comparison(ir->operation);
+      emit(BRW_OPCODE_AND, this->result, this->result, fs_reg(0x1));
+      break;
+
+   case ir_binop_logic_xor:
+      emit(BRW_OPCODE_XOR, this->result, op[0], op[1]);
+      break;
+
+   case ir_binop_logic_or:
+      emit(BRW_OPCODE_OR, this->result, op[0], op[1]);
+      break;
+
+   case ir_binop_logic_and:
+      emit(BRW_OPCODE_AND, this->result, op[0], op[1]);
+      break;
+
+   case ir_binop_dot:
+   case ir_unop_any:
+      assert(!"not reached: should be handled by brw_fs_channel_expressions");
+      break;
+
+   case ir_unop_noise:
+      assert(!"not reached: should be handled by lower_noise");
+      break;
+
+   case ir_quadop_vector:
+      assert(!"not reached: should be handled by lower_quadop_vector");
+      break;
+
+   case ir_unop_sqrt:
+      emit_math(FS_OPCODE_SQRT, this->result, op[0]);
+      break;
+
+   case ir_unop_rsq:
+      emit_math(FS_OPCODE_RSQ, this->result, op[0]);
+      break;
+
+   case ir_unop_i2f:
+   case ir_unop_b2f:
+   case ir_unop_b2i:
+   case ir_unop_f2i:
+      emit(BRW_OPCODE_MOV, this->result, op[0]);
+      break;
+   case ir_unop_f2b:
+   case ir_unop_i2b:
+      temp = this->result;
+      /* original gen4 does implicit conversion before comparison. */
+      if (intel->gen < 5)
+        temp.type = op[0].type;
+
+      inst = emit(BRW_OPCODE_CMP, temp, op[0], fs_reg(0.0f));
+      inst->conditional_mod = BRW_CONDITIONAL_NZ;
+      inst = emit(BRW_OPCODE_AND, this->result, this->result, fs_reg(1));
+      break;
+
+   case ir_unop_trunc:
+      emit(BRW_OPCODE_RNDZ, this->result, op[0]);
+      break;
+   case ir_unop_ceil:
+      op[0].negate = !op[0].negate;
+      inst = emit(BRW_OPCODE_RNDD, this->result, op[0]);
+      this->result.negate = true;
+      break;
+   case ir_unop_floor:
+      inst = emit(BRW_OPCODE_RNDD, this->result, op[0]);
+      break;
+   case ir_unop_fract:
+      inst = emit(BRW_OPCODE_FRC, this->result, op[0]);
+      break;
+   case ir_unop_round_even:
+      emit(BRW_OPCODE_RNDE, this->result, op[0]);
+      break;
+
+   case ir_binop_min:
+      /* Unalias the destination */
+      this->result = fs_reg(this, ir->type);
+
+      inst = emit(BRW_OPCODE_CMP, this->result, op[0], op[1]);
+      inst->conditional_mod = BRW_CONDITIONAL_L;
+
+      inst = emit(BRW_OPCODE_SEL, this->result, op[0], op[1]);
+      inst->predicated = true;
+      break;
+   case ir_binop_max:
+      /* Unalias the destination */
+      this->result = fs_reg(this, ir->type);
+
+      inst = emit(BRW_OPCODE_CMP, this->result, op[0], op[1]);
+      inst->conditional_mod = BRW_CONDITIONAL_G;
+
+      inst = emit(BRW_OPCODE_SEL, this->result, op[0], op[1]);
+      inst->predicated = true;
+      break;
+
+   case ir_binop_pow:
+      emit_math(FS_OPCODE_POW, this->result, op[0], op[1]);
+      break;
+
+   case ir_unop_bit_not:
+      inst = emit(BRW_OPCODE_NOT, this->result, op[0]);
+      break;
+   case ir_binop_bit_and:
+      inst = emit(BRW_OPCODE_AND, this->result, op[0], op[1]);
+      break;
+   case ir_binop_bit_xor:
+      inst = emit(BRW_OPCODE_XOR, this->result, op[0], op[1]);
+      break;
+   case ir_binop_bit_or:
+      inst = emit(BRW_OPCODE_OR, this->result, op[0], op[1]);
+      break;
+
+   case ir_unop_u2f:
+   case ir_binop_lshift:
+   case ir_binop_rshift:
+      assert(!"GLSL 1.30 features unsupported");
+      break;
+   }
+}
+
+void
+fs_visitor::emit_assignment_writes(fs_reg &l, fs_reg &r,
+                                  const glsl_type *type, bool predicated)
+{
+   switch (type->base_type) {
+   case GLSL_TYPE_FLOAT:
+   case GLSL_TYPE_UINT:
+   case GLSL_TYPE_INT:
+   case GLSL_TYPE_BOOL:
+      for (unsigned int i = 0; i < type->components(); i++) {
+        l.type = brw_type_for_base_type(type);
+        r.type = brw_type_for_base_type(type);
+
+        if (predicated || !l.equals(&r)) {
+           fs_inst *inst = emit(BRW_OPCODE_MOV, l, r);
+           inst->predicated = predicated;
+        }
+
+        l.reg_offset++;
+        r.reg_offset++;
+      }
+      break;
+   case GLSL_TYPE_ARRAY:
+      for (unsigned int i = 0; i < type->length; i++) {
+        emit_assignment_writes(l, r, type->fields.array, predicated);
+      }
+      break;
+
+   case GLSL_TYPE_STRUCT:
+      for (unsigned int i = 0; i < type->length; i++) {
+        emit_assignment_writes(l, r, type->fields.structure[i].type,
+                               predicated);
+      }
+      break;
+
+   case GLSL_TYPE_SAMPLER:
+      break;
+
+   default:
+      assert(!"not reached");
+      break;
+   }
+}
+
+void
+fs_visitor::visit(ir_assignment *ir)
+{
+   struct fs_reg l, r;
+   fs_inst *inst;
+
+   /* FINISHME: arrays on the lhs */
+   this->result = reg_undef;
+   ir->lhs->accept(this);
+   l = this->result;
+
+   /* If we're doing a direct assignment, an RHS expression could
+    * drop its result right into our destination.  Otherwise, tell it
+    * not to.
+    */
+   if (ir->condition ||
+       !(ir->lhs->type->is_scalar() ||
+        (ir->lhs->type->is_vector() &&
+         ir->write_mask == (1 << ir->lhs->type->vector_elements) - 1))) {
+      this->result = reg_undef;
+   }
+
+   ir->rhs->accept(this);
+   r = this->result;
+
+   assert(l.file != BAD_FILE);
+   assert(r.file != BAD_FILE);
+
+   if (ir->condition) {
+      emit_bool_to_cond_code(ir->condition);
+   }
+
+   if (ir->lhs->type->is_scalar() ||
+       ir->lhs->type->is_vector()) {
+      for (int i = 0; i < ir->lhs->type->vector_elements; i++) {
+        if (ir->write_mask & (1 << i)) {
+           if (ir->condition) {
+              inst = emit(BRW_OPCODE_MOV, l, r);
+              inst->predicated = true;
+           } else if (!l.equals(&r)) {
+              inst = emit(BRW_OPCODE_MOV, l, r);
+           }
+
+           r.reg_offset++;
+        }
+        l.reg_offset++;
+      }
+   } else {
+      emit_assignment_writes(l, r, ir->lhs->type, ir->condition != NULL);
+   }
+}
+
+fs_inst *
+fs_visitor::emit_texture_gen4(ir_texture *ir, fs_reg dst, fs_reg coordinate,
+                             int sampler)
+{
+   int mlen;
+   int base_mrf = 1;
+   bool simd16 = false;
+   fs_reg orig_dst;
+
+   /* g0 header. */
+   mlen = 1;
+
+   if (ir->shadow_comparitor) {
+      for (int i = 0; i < ir->coordinate->type->vector_elements; i++) {
+        fs_inst *inst = emit(BRW_OPCODE_MOV,
+                             fs_reg(MRF, base_mrf + mlen + i), coordinate);
+        if (i < 3 && c->key.gl_clamp_mask[i] & (1 << sampler))
+           inst->saturate = true;
+
+        coordinate.reg_offset++;
+      }
+      /* gen4's SIMD8 sampler always has the slots for u,v,r present. */
+      mlen += 3;
+
+      if (ir->op == ir_tex) {
+        /* There's no plain shadow compare message, so we use shadow
+         * compare with a bias of 0.0.
+         */
+        emit(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen), fs_reg(0.0f));
+        mlen++;
+      } else if (ir->op == ir_txb) {
+        this->result = reg_undef;
+        ir->lod_info.bias->accept(this);
+        emit(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen), this->result);
+        mlen++;
+      } else {
+        assert(ir->op == ir_txl);
+        this->result = reg_undef;
+        ir->lod_info.lod->accept(this);
+        emit(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen), this->result);
+        mlen++;
+      }
+
+      this->result = reg_undef;
+      ir->shadow_comparitor->accept(this);
+      emit(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen), this->result);
+      mlen++;
+   } else if (ir->op == ir_tex) {
+      for (int i = 0; i < ir->coordinate->type->vector_elements; i++) {
+        fs_inst *inst = emit(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen + i),
+                             coordinate);
+        if (i < 3 && c->key.gl_clamp_mask[i] & (1 << sampler))
+           inst->saturate = true;
+        coordinate.reg_offset++;
+      }
+      /* gen4's SIMD8 sampler always has the slots for u,v,r present. */
+      mlen += 3;
+   } else if (ir->op == ir_txd) {
+      assert(!"TXD isn't supported on gen4 yet.");
+   } else {
+      /* Oh joy.  gen4 doesn't have SIMD8 non-shadow-compare bias/lod
+       * instructions.  We'll need to do SIMD16 here.
+       */
+      assert(ir->op == ir_txb || ir->op == ir_txl);
+
+      for (int i = 0; i < ir->coordinate->type->vector_elements; i++) {
+        fs_inst *inst = emit(BRW_OPCODE_MOV, fs_reg(MRF,
+                                                    base_mrf + mlen + i * 2),
+                             coordinate);
+        if (i < 3 && c->key.gl_clamp_mask[i] & (1 << sampler))
+           inst->saturate = true;
+        coordinate.reg_offset++;
+      }
+
+      /* lod/bias appears after u/v/r. */
+      mlen += 6;
+
+      if (ir->op == ir_txb) {
+        this->result = reg_undef;
+        ir->lod_info.bias->accept(this);
+        emit(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen), this->result);
+        mlen++;
+      } else {
+        this->result = reg_undef;
+        ir->lod_info.lod->accept(this);
+        emit(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen), this->result);
+        mlen++;
+      }
+
+      /* The unused upper half. */
+      mlen++;
+
+      /* Now, since we're doing simd16, the return is 2 interleaved
+       * vec4s where the odd-indexed ones are junk. We'll need to move
+       * this weirdness around to the expected layout.
+       */
+      simd16 = true;
+      orig_dst = dst;
+      dst = fs_reg(this, glsl_type::get_array_instance(glsl_type::vec4_type,
+                                                      2));
+      dst.type = BRW_REGISTER_TYPE_F;
+   }
+
+   fs_inst *inst = NULL;
+   switch (ir->op) {
+   case ir_tex:
+      inst = emit(FS_OPCODE_TEX, dst);
+      break;
+   case ir_txb:
+      inst = emit(FS_OPCODE_TXB, dst);
+      break;
+   case ir_txl:
+      inst = emit(FS_OPCODE_TXL, dst);
+      break;
+   case ir_txd:
+      inst = emit(FS_OPCODE_TXD, dst);
+      break;
+   case ir_txf:
+      assert(!"GLSL 1.30 features unsupported");
+      break;
+   }
+   inst->base_mrf = base_mrf;
+   inst->mlen = mlen;
+   inst->header_present = true;
+
+   if (simd16) {
+      for (int i = 0; i < 4; i++) {
+        emit(BRW_OPCODE_MOV, orig_dst, dst);
+        orig_dst.reg_offset++;
+        dst.reg_offset += 2;
+      }
+   }
+
+   return inst;
+}
+
+/* gen5's sampler has slots for u, v, r, array index, then optional
+ * parameters like shadow comparitor or LOD bias.  If optional
+ * parameters aren't present, those base slots are optional and don't
+ * need to be included in the message.
+ *
+ * We don't fill in the unnecessary slots regardless, which may look
+ * surprising in the disassembly.
+ */
+fs_inst *
+fs_visitor::emit_texture_gen5(ir_texture *ir, fs_reg dst, fs_reg coordinate,
+                             int sampler)
+{
+   int mlen = 0;
+   int base_mrf = 2;
+   int reg_width = c->dispatch_width / 8;
+   bool header_present = false;
+
+   if (ir->offset) {
+      /* The offsets set up by the ir_texture visitor are in the
+       * m1 header, so we can't go headerless.
+       */
+      header_present = true;
+      mlen++;
+      base_mrf--;
+   }
+
+   for (int i = 0; i < ir->coordinate->type->vector_elements; i++) {
+      fs_inst *inst = emit(BRW_OPCODE_MOV,
+                          fs_reg(MRF, base_mrf + mlen + i * reg_width),
+                          coordinate);
+      if (i < 3 && c->key.gl_clamp_mask[i] & (1 << sampler))
+        inst->saturate = true;
+      coordinate.reg_offset++;
+   }
+   mlen += ir->coordinate->type->vector_elements * reg_width;
+
+   if (ir->shadow_comparitor) {
+      mlen = MAX2(mlen, header_present + 4 * reg_width);
+
+      this->result = reg_undef;
+      ir->shadow_comparitor->accept(this);
+      emit(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen), this->result);
+      mlen += reg_width;
+   }
+
+   fs_inst *inst = NULL;
+   switch (ir->op) {
+   case ir_tex:
+      inst = emit(FS_OPCODE_TEX, dst);
+      break;
+   case ir_txb:
+      this->result = reg_undef;
+      ir->lod_info.bias->accept(this);
+      mlen = MAX2(mlen, header_present + 4 * reg_width);
+      emit(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen), this->result);
+      mlen += reg_width;
+
+      inst = emit(FS_OPCODE_TXB, dst);
+
+      break;
+   case ir_txl:
+      this->result = reg_undef;
+      ir->lod_info.lod->accept(this);
+      mlen = MAX2(mlen, header_present + 4 * reg_width);
+      emit(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen), this->result);
+      mlen += reg_width;
+
+      inst = emit(FS_OPCODE_TXL, dst);
+      break;
+   case ir_txd:
+   case ir_txf:
+      assert(!"GLSL 1.30 features unsupported");
+      break;
+   }
+   inst->base_mrf = base_mrf;
+   inst->mlen = mlen;
+   inst->header_present = header_present;
+
+   if (mlen > 11) {
+      fail("Message length >11 disallowed by hardware\n");
+   }
+
+   return inst;
+}
+
+fs_inst *
+fs_visitor::emit_texture_gen7(ir_texture *ir, fs_reg dst, fs_reg coordinate,
+                             int sampler)
+{
+   int mlen = 0;
+   int base_mrf = 2;
+   int reg_width = c->dispatch_width / 8;
+   bool header_present = false;
+
+   if (ir->offset) {
+      /* The offsets set up by the ir_texture visitor are in the
+       * m1 header, so we can't go headerless.
+       */
+      header_present = true;
+      mlen++;
+      base_mrf--;
+   }
+
+   if (ir->shadow_comparitor) {
+      ir->shadow_comparitor->accept(this);
+      emit(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen), this->result);
+      mlen += reg_width;
+   }
+
+   /* Set up the LOD info */
+   switch (ir->op) {
+   case ir_tex:
+      break;
+   case ir_txb:
+      ir->lod_info.bias->accept(this);
+      emit(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen), this->result);
+      mlen += reg_width;
+      break;
+   case ir_txl:
+      ir->lod_info.lod->accept(this);
+      emit(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen), this->result);
+      mlen += reg_width;
+      break;
+   case ir_txd:
+   case ir_txf:
+      assert(!"GLSL 1.30 features unsupported");
+      break;
+   }
+
+   /* Set up the coordinate */
+   for (int i = 0; i < ir->coordinate->type->vector_elements; i++) {
+      fs_inst *inst = emit(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen),
+                          coordinate);
+      if (i < 3 && c->key.gl_clamp_mask[i] & (1 << sampler))
+        inst->saturate = true;
+      coordinate.reg_offset++;
+      mlen += reg_width;
+   }
+
+   /* Generate the SEND */
+   fs_inst *inst = NULL;
+   switch (ir->op) {
+   case ir_tex: inst = emit(FS_OPCODE_TEX, dst); break;
+   case ir_txb: inst = emit(FS_OPCODE_TXB, dst); break;
+   case ir_txl: inst = emit(FS_OPCODE_TXL, dst); break;
+   case ir_txd: inst = emit(FS_OPCODE_TXD, dst); break;
+   case ir_txf: assert(!"TXF unsupported.");
+   }
+   inst->base_mrf = base_mrf;
+   inst->mlen = mlen;
+   inst->header_present = header_present;
+
+   if (mlen > 11) {
+      fail("Message length >11 disallowed by hardware\n");
+   }
+
+   return inst;
+}
+
+void
+fs_visitor::visit(ir_texture *ir)
+{
+   int sampler;
+   fs_inst *inst = NULL;
+
+   this->result = reg_undef;
+   ir->coordinate->accept(this);
+   fs_reg coordinate = this->result;
+
+   if (ir->offset != NULL) {
+      ir_constant *offset = ir->offset->as_constant();
+      assert(offset != NULL);
+
+      signed char offsets[3];
+      for (unsigned i = 0; i < ir->offset->type->vector_elements; i++)
+        offsets[i] = (signed char) offset->value.i[i];
+
+      /* Combine all three offsets into a single unsigned dword:
+       *
+       *    bits 11:8 - U Offset (X component)
+       *    bits  7:4 - V Offset (Y component)
+       *    bits  3:0 - R Offset (Z component)
+       */
+      unsigned offset_bits = 0;
+      for (unsigned i = 0; i < ir->offset->type->vector_elements; i++) {
+        const unsigned shift = 4 * (2 - i);
+        offset_bits |= (offsets[i] << shift) & (0xF << shift);
+      }
+
+      /* Explicitly set up the message header by copying g0 to msg reg m1. */
+      emit(BRW_OPCODE_MOV, fs_reg(MRF, 1, BRW_REGISTER_TYPE_UD),
+          fs_reg(GRF, 0, BRW_REGISTER_TYPE_UD));
+
+      /* Then set the offset bits in DWord 2 of the message header. */
+      emit(BRW_OPCODE_MOV,
+          fs_reg(retype(brw_vec1_reg(BRW_MESSAGE_REGISTER_FILE, 1, 2),
+                        BRW_REGISTER_TYPE_UD)),
+          fs_reg(brw_imm_uw(offset_bits)));
+   }
+
+   /* Should be lowered by do_lower_texture_projection */
+   assert(!ir->projector);
+
+   sampler = _mesa_get_sampler_uniform_value(ir->sampler,
+                                            ctx->Shader.CurrentFragmentProgram,
+                                            &brw->fragment_program->Base);
+   sampler = c->fp->program.Base.SamplerUnits[sampler];
+
+   /* The 965 requires the EU to do the normalization of GL rectangle
+    * texture coordinates.  We use the program parameter state
+    * tracking to get the scaling factor.
+    */
+   if (ir->sampler->type->sampler_dimensionality == GLSL_SAMPLER_DIM_RECT) {
+      struct gl_program_parameter_list *params = c->fp->program.Base.Parameters;
+      int tokens[STATE_LENGTH] = {
+        STATE_INTERNAL,
+        STATE_TEXRECT_SCALE,
+        sampler,
+        0,
+        0
+      };
+
+      if (c->dispatch_width == 16) {
+        fail("rectangle scale uniform setup not supported on 16-wide\n");
+        this->result = fs_reg(this, ir->type);
+        return;
+      }
+
+      c->prog_data.param_convert[c->prog_data.nr_params] =
+        PARAM_NO_CONVERT;
+      c->prog_data.param_convert[c->prog_data.nr_params + 1] =
+        PARAM_NO_CONVERT;
+
+      fs_reg scale_x = fs_reg(UNIFORM, c->prog_data.nr_params);
+      fs_reg scale_y = fs_reg(UNIFORM, c->prog_data.nr_params + 1);
+      GLuint index = _mesa_add_state_reference(params,
+                                              (gl_state_index *)tokens);
+
+      this->param_index[c->prog_data.nr_params] = index;
+      this->param_offset[c->prog_data.nr_params] = 0;
+      c->prog_data.nr_params++;
+      this->param_index[c->prog_data.nr_params] = index;
+      this->param_offset[c->prog_data.nr_params] = 1;
+      c->prog_data.nr_params++;
+
+      fs_reg dst = fs_reg(this, ir->coordinate->type);
+      fs_reg src = coordinate;
+      coordinate = dst;
+
+      emit(BRW_OPCODE_MUL, dst, src, scale_x);
+      dst.reg_offset++;
+      src.reg_offset++;
+      emit(BRW_OPCODE_MUL, dst, src, scale_y);
+   }
+
+   /* Writemasking doesn't eliminate channels on SIMD8 texture
+    * samples, so don't worry about them.
+    */
+   fs_reg dst = fs_reg(this, glsl_type::vec4_type);
+
+   if (intel->gen >= 7) {
+      inst = emit_texture_gen7(ir, dst, coordinate, sampler);
+   } else if (intel->gen >= 5) {
+      inst = emit_texture_gen5(ir, dst, coordinate, sampler);
+   } else {
+      inst = emit_texture_gen4(ir, dst, coordinate, sampler);
+   }
+
+   /* If there's an offset, we already set up m1.  To avoid the implied move,
+    * use the null register.  Otherwise, we want an implied move from g0.
+    */
+   if (ir->offset != NULL || !inst->header_present)
+      inst->src[0] = reg_undef;
+   else
+      inst->src[0] = fs_reg(retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UW));
+
+   inst->sampler = sampler;
+
+   this->result = dst;
+
+   if (ir->shadow_comparitor)
+      inst->shadow_compare = true;
+
+   if (ir->type == glsl_type::float_type) {
+      /* Ignore DEPTH_TEXTURE_MODE swizzling. */
+      assert(ir->sampler->type->sampler_shadow);
+   } else if (c->key.tex_swizzles[inst->sampler] != SWIZZLE_NOOP) {
+      fs_reg swizzle_dst = fs_reg(this, glsl_type::vec4_type);
+
+      for (int i = 0; i < 4; i++) {
+        int swiz = GET_SWZ(c->key.tex_swizzles[inst->sampler], i);
+        fs_reg l = swizzle_dst;
+        l.reg_offset += i;
+
+        if (swiz == SWIZZLE_ZERO) {
+           emit(BRW_OPCODE_MOV, l, fs_reg(0.0f));
+        } else if (swiz == SWIZZLE_ONE) {
+           emit(BRW_OPCODE_MOV, l, fs_reg(1.0f));
+        } else {
+           fs_reg r = dst;
+           r.reg_offset += GET_SWZ(c->key.tex_swizzles[inst->sampler], i);
+           emit(BRW_OPCODE_MOV, l, r);
+        }
+      }
+      this->result = swizzle_dst;
+   }
+}
+
+void
+fs_visitor::visit(ir_swizzle *ir)
+{
+   this->result = reg_undef;
+   ir->val->accept(this);
+   fs_reg val = this->result;
+
+   if (ir->type->vector_elements == 1) {
+      this->result.reg_offset += ir->mask.x;
+      return;
+   }
+
+   fs_reg result = fs_reg(this, ir->type);
+   this->result = result;
+
+   for (unsigned int i = 0; i < ir->type->vector_elements; i++) {
+      fs_reg channel = val;
+      int swiz = 0;
+
+      switch (i) {
+      case 0:
+        swiz = ir->mask.x;
+        break;
+      case 1:
+        swiz = ir->mask.y;
+        break;
+      case 2:
+        swiz = ir->mask.z;
+        break;
+      case 3:
+        swiz = ir->mask.w;
+        break;
+      }
+
+      channel.reg_offset += swiz;
+      emit(BRW_OPCODE_MOV, result, channel);
+      result.reg_offset++;
+   }
+}
+
+void
+fs_visitor::visit(ir_discard *ir)
+{
+   assert(ir->condition == NULL); /* FINISHME */
+
+   emit(FS_OPCODE_DISCARD);
+   kill_emitted = true;
+}
+
+void
+fs_visitor::visit(ir_constant *ir)
+{
+   /* Set this->result to reg at the bottom of the function because some code
+    * paths will cause this visitor to be applied to other fields.  This will
+    * cause the value stored in this->result to be modified.
+    *
+    * Make reg constant so that it doesn't get accidentally modified along the
+    * way.  Yes, I actually had this problem. :(
+    */
+   const fs_reg reg(this, ir->type);
+   fs_reg dst_reg = reg;
+
+   if (ir->type->is_array()) {
+      const unsigned size = type_size(ir->type->fields.array);
+
+      for (unsigned i = 0; i < ir->type->length; i++) {
+        this->result = reg_undef;
+        ir->array_elements[i]->accept(this);
+        fs_reg src_reg = this->result;
+
+        dst_reg.type = src_reg.type;
+        for (unsigned j = 0; j < size; j++) {
+           emit(BRW_OPCODE_MOV, dst_reg, src_reg);
+           src_reg.reg_offset++;
+           dst_reg.reg_offset++;
+        }
+      }
+   } else if (ir->type->is_record()) {
+      foreach_list(node, &ir->components) {
+        ir_instruction *const field = (ir_instruction *) node;
+        const unsigned size = type_size(field->type);
+
+        this->result = reg_undef;
+        field->accept(this);
+        fs_reg src_reg = this->result;
+
+        dst_reg.type = src_reg.type;
+        for (unsigned j = 0; j < size; j++) {
+           emit(BRW_OPCODE_MOV, dst_reg, src_reg);
+           src_reg.reg_offset++;
+           dst_reg.reg_offset++;
+        }
+      }
+   } else {
+      const unsigned size = type_size(ir->type);
+
+      for (unsigned i = 0; i < size; i++) {
+        switch (ir->type->base_type) {
+        case GLSL_TYPE_FLOAT:
+           emit(BRW_OPCODE_MOV, dst_reg, fs_reg(ir->value.f[i]));
+           break;
+        case GLSL_TYPE_UINT:
+           emit(BRW_OPCODE_MOV, dst_reg, fs_reg(ir->value.u[i]));
+           break;
+        case GLSL_TYPE_INT:
+           emit(BRW_OPCODE_MOV, dst_reg, fs_reg(ir->value.i[i]));
+           break;
+        case GLSL_TYPE_BOOL:
+           emit(BRW_OPCODE_MOV, dst_reg, fs_reg((int)ir->value.b[i]));
+           break;
+        default:
+           assert(!"Non-float/uint/int/bool constant");
+        }
+        dst_reg.reg_offset++;
+      }
+   }
+
+   this->result = reg;
+}
+
+void
+fs_visitor::emit_bool_to_cond_code(ir_rvalue *ir)
+{
+   ir_expression *expr = ir->as_expression();
+
+   if (expr) {
+      fs_reg op[2];
+      fs_inst *inst;
+
+      assert(expr->get_num_operands() <= 2);
+      for (unsigned int i = 0; i < expr->get_num_operands(); i++) {
+        assert(expr->operands[i]->type->is_scalar());
+
+        this->result = reg_undef;
+        expr->operands[i]->accept(this);
+        op[i] = this->result;
+      }
+
+      switch (expr->operation) {
+      case ir_unop_logic_not:
+        inst = emit(BRW_OPCODE_AND, reg_null_d, op[0], fs_reg(1));
+        inst->conditional_mod = BRW_CONDITIONAL_Z;
+        break;
+
+      case ir_binop_logic_xor:
+        inst = emit(BRW_OPCODE_XOR, reg_null_d, op[0], op[1]);
+        inst->conditional_mod = BRW_CONDITIONAL_NZ;
+        break;
+
+      case ir_binop_logic_or:
+        inst = emit(BRW_OPCODE_OR, reg_null_d, op[0], op[1]);
+        inst->conditional_mod = BRW_CONDITIONAL_NZ;
+        break;
+
+      case ir_binop_logic_and:
+        inst = emit(BRW_OPCODE_AND, reg_null_d, op[0], op[1]);
+        inst->conditional_mod = BRW_CONDITIONAL_NZ;
+        break;
+
+      case ir_unop_f2b:
+        if (intel->gen >= 6) {
+           inst = emit(BRW_OPCODE_CMP, reg_null_d, op[0], fs_reg(0.0f));
+        } else {
+           inst = emit(BRW_OPCODE_MOV, reg_null_f, op[0]);
+        }
+        inst->conditional_mod = BRW_CONDITIONAL_NZ;
+        break;
+
+      case ir_unop_i2b:
+        if (intel->gen >= 6) {
+           inst = emit(BRW_OPCODE_CMP, reg_null_d, op[0], fs_reg(0));
+        } else {
+           inst = emit(BRW_OPCODE_MOV, reg_null_d, op[0]);
+        }
+        inst->conditional_mod = BRW_CONDITIONAL_NZ;
+        break;
+
+      case ir_binop_greater:
+      case ir_binop_gequal:
+      case ir_binop_less:
+      case ir_binop_lequal:
+      case ir_binop_equal:
+      case ir_binop_all_equal:
+      case ir_binop_nequal:
+      case ir_binop_any_nequal:
+        inst = emit(BRW_OPCODE_CMP, reg_null_cmp, op[0], op[1]);
+        inst->conditional_mod =
+           brw_conditional_for_comparison(expr->operation);
+        break;
+
+      default:
+        assert(!"not reached");
+        fail("bad cond code\n");
+        break;
+      }
+      return;
+   }
+
+   this->result = reg_undef;
+   ir->accept(this);
+
+   if (intel->gen >= 6) {
+      fs_inst *inst = emit(BRW_OPCODE_AND, reg_null_d, this->result, fs_reg(1));
+      inst->conditional_mod = BRW_CONDITIONAL_NZ;
+   } else {
+      fs_inst *inst = emit(BRW_OPCODE_MOV, reg_null_d, this->result);
+      inst->conditional_mod = BRW_CONDITIONAL_NZ;
+   }
+}
+
+/**
+ * Emit a gen6 IF statement with the comparison folded into the IF
+ * instruction.
+ */
+void
+fs_visitor::emit_if_gen6(ir_if *ir)
+{
+   ir_expression *expr = ir->condition->as_expression();
+
+   if (expr) {
+      fs_reg op[2];
+      fs_inst *inst;
+      fs_reg temp;
+
+      assert(expr->get_num_operands() <= 2);
+      for (unsigned int i = 0; i < expr->get_num_operands(); i++) {
+        assert(expr->operands[i]->type->is_scalar());
+
+        this->result = reg_undef;
+        expr->operands[i]->accept(this);
+        op[i] = this->result;
+      }
+
+      switch (expr->operation) {
+      case ir_unop_logic_not:
+        inst = emit(BRW_OPCODE_IF, temp, op[0], fs_reg(0));
+        inst->conditional_mod = BRW_CONDITIONAL_Z;
+        return;
+
+      case ir_binop_logic_xor:
+        inst = emit(BRW_OPCODE_IF, reg_null_d, op[0], op[1]);
+        inst->conditional_mod = BRW_CONDITIONAL_NZ;
+        return;
+
+      case ir_binop_logic_or:
+        temp = fs_reg(this, glsl_type::bool_type);
+        emit(BRW_OPCODE_OR, temp, op[0], op[1]);
+        inst = emit(BRW_OPCODE_IF, reg_null_d, temp, fs_reg(0));
+        inst->conditional_mod = BRW_CONDITIONAL_NZ;
+        return;
+
+      case ir_binop_logic_and:
+        temp = fs_reg(this, glsl_type::bool_type);
+        emit(BRW_OPCODE_AND, temp, op[0], op[1]);
+        inst = emit(BRW_OPCODE_IF, reg_null_d, temp, fs_reg(0));
+        inst->conditional_mod = BRW_CONDITIONAL_NZ;
+        return;
+
+      case ir_unop_f2b:
+        inst = emit(BRW_OPCODE_IF, reg_null_f, op[0], fs_reg(0));
+        inst->conditional_mod = BRW_CONDITIONAL_NZ;
+        return;
+
+      case ir_unop_i2b:
+        inst = emit(BRW_OPCODE_IF, reg_null_d, op[0], fs_reg(0));
+        inst->conditional_mod = BRW_CONDITIONAL_NZ;
+        return;
+
+      case ir_binop_greater:
+      case ir_binop_gequal:
+      case ir_binop_less:
+      case ir_binop_lequal:
+      case ir_binop_equal:
+      case ir_binop_all_equal:
+      case ir_binop_nequal:
+      case ir_binop_any_nequal:
+        inst = emit(BRW_OPCODE_IF, reg_null_d, op[0], op[1]);
+        inst->conditional_mod =
+           brw_conditional_for_comparison(expr->operation);
+        return;
+      default:
+        assert(!"not reached");
+        inst = emit(BRW_OPCODE_IF, reg_null_d, op[0], fs_reg(0));
+        inst->conditional_mod = BRW_CONDITIONAL_NZ;
+        fail("bad condition\n");
+        return;
+      }
+      return;
+   }
+
+   this->result = reg_undef;
+   ir->condition->accept(this);
+
+   fs_inst *inst = emit(BRW_OPCODE_IF, reg_null_d, this->result, fs_reg(0));
+   inst->conditional_mod = BRW_CONDITIONAL_NZ;
+}
+
+void
+fs_visitor::visit(ir_if *ir)
+{
+   fs_inst *inst;
+
+   if (intel->gen != 6 && c->dispatch_width == 16) {
+      fail("Can't support (non-uniform) control flow on 16-wide\n");
+   }
+
+   /* Don't point the annotation at the if statement, because then it plus
+    * the then and else blocks get printed.
+    */
+   this->base_ir = ir->condition;
+
+   if (intel->gen == 6) {
+      emit_if_gen6(ir);
+   } else {
+      emit_bool_to_cond_code(ir->condition);
+
+      inst = emit(BRW_OPCODE_IF);
+      inst->predicated = true;
+   }
+
+   foreach_iter(exec_list_iterator, iter, ir->then_instructions) {
+      ir_instruction *ir = (ir_instruction *)iter.get();
+      this->base_ir = ir;
+      this->result = reg_undef;
+      ir->accept(this);
+   }
+
+   if (!ir->else_instructions.is_empty()) {
+      emit(BRW_OPCODE_ELSE);
+
+      foreach_iter(exec_list_iterator, iter, ir->else_instructions) {
+        ir_instruction *ir = (ir_instruction *)iter.get();
+        this->base_ir = ir;
+        this->result = reg_undef;
+        ir->accept(this);
+      }
+   }
+
+   emit(BRW_OPCODE_ENDIF);
+}
+
+void
+fs_visitor::visit(ir_loop *ir)
+{
+   fs_reg counter = reg_undef;
+
+   if (c->dispatch_width == 16) {
+      fail("Can't support (non-uniform) control flow on 16-wide\n");
+   }
+
+   if (ir->counter) {
+      this->base_ir = ir->counter;
+      ir->counter->accept(this);
+      counter = *(variable_storage(ir->counter));
+
+      if (ir->from) {
+        this->result = counter;
+
+        this->base_ir = ir->from;
+        this->result = counter;
+        ir->from->accept(this);
+
+        if (!this->result.equals(&counter))
+           emit(BRW_OPCODE_MOV, counter, this->result);
+      }
+   }
+
+   emit(BRW_OPCODE_DO);
+
+   if (ir->to) {
+      this->base_ir = ir->to;
+      this->result = reg_undef;
+      ir->to->accept(this);
+
+      fs_inst *inst = emit(BRW_OPCODE_CMP, reg_null_cmp, counter, this->result);
+      inst->conditional_mod = brw_conditional_for_comparison(ir->cmp);
+
+      inst = emit(BRW_OPCODE_BREAK);
+      inst->predicated = true;
+   }
+
+   foreach_iter(exec_list_iterator, iter, ir->body_instructions) {
+      ir_instruction *ir = (ir_instruction *)iter.get();
+
+      this->base_ir = ir;
+      this->result = reg_undef;
+      ir->accept(this);
+   }
+
+   if (ir->increment) {
+      this->base_ir = ir->increment;
+      this->result = reg_undef;
+      ir->increment->accept(this);
+      emit(BRW_OPCODE_ADD, counter, counter, this->result);
+   }
+
+   emit(BRW_OPCODE_WHILE);
+}
+
+void
+fs_visitor::visit(ir_loop_jump *ir)
+{
+   switch (ir->mode) {
+   case ir_loop_jump::jump_break:
+      emit(BRW_OPCODE_BREAK);
+      break;
+   case ir_loop_jump::jump_continue:
+      emit(BRW_OPCODE_CONTINUE);
+      break;
+   }
+}
+
+void
+fs_visitor::visit(ir_call *ir)
+{
+   assert(!"FINISHME");
+}
+
+void
+fs_visitor::visit(ir_return *ir)
+{
+   assert(!"FINISHME");
+}
+
+void
+fs_visitor::visit(ir_function *ir)
+{
+   /* Ignore function bodies other than main() -- we shouldn't see calls to
+    * them since they should all be inlined before we get to ir_to_mesa.
+    */
+   if (strcmp(ir->name, "main") == 0) {
+      const ir_function_signature *sig;
+      exec_list empty;
+
+      sig = ir->matching_signature(&empty);
+
+      assert(sig);
+
+      foreach_iter(exec_list_iterator, iter, sig->body) {
+        ir_instruction *ir = (ir_instruction *)iter.get();
+        this->base_ir = ir;
+        this->result = reg_undef;
+        ir->accept(this);
+      }
+   }
+}
+
+void
+fs_visitor::visit(ir_function_signature *ir)
+{
+   assert(!"not reached");
+   (void)ir;
+}
+
+fs_inst *
+fs_visitor::emit(fs_inst inst)
+{
+   fs_inst *list_inst = new(mem_ctx) fs_inst;
+   *list_inst = inst;
+
+   if (force_uncompressed_stack > 0)
+      list_inst->force_uncompressed = true;
+   else if (force_sechalf_stack > 0)
+      list_inst->force_sechalf = true;
+
+   list_inst->annotation = this->current_annotation;
+   list_inst->ir = this->base_ir;
+
+   this->instructions.push_tail(list_inst);
+
+   return list_inst;
+}
+
+/** Emits a dummy fragment shader consisting of magenta for bringup purposes. */
+void
+fs_visitor::emit_dummy_fs()
+{
+   /* Everyone's favorite color. */
+   emit(BRW_OPCODE_MOV, fs_reg(MRF, 2), fs_reg(1.0f));
+   emit(BRW_OPCODE_MOV, fs_reg(MRF, 3), fs_reg(0.0f));
+   emit(BRW_OPCODE_MOV, fs_reg(MRF, 4), fs_reg(1.0f));
+   emit(BRW_OPCODE_MOV, fs_reg(MRF, 5), fs_reg(0.0f));
+
+   fs_inst *write;
+   write = emit(FS_OPCODE_FB_WRITE, fs_reg(0), fs_reg(0));
+   write->base_mrf = 0;
+}
+
+/* The register location here is relative to the start of the URB
+ * data.  It will get adjusted to be a real location before
+ * generate_code() time.
+ */
+struct brw_reg
+fs_visitor::interp_reg(int location, int channel)
+{
+   int regnr = urb_setup[location] * 2 + channel / 2;
+   int stride = (channel & 1) * 4;
+
+   assert(urb_setup[location] != -1);
+
+   return brw_vec1_grf(regnr, stride);
+}
+
+/** Emits the interpolation for the varying inputs. */
+void
+fs_visitor::emit_interpolation_setup_gen4()
+{
+   this->current_annotation = "compute pixel centers";
+   this->pixel_x = fs_reg(this, glsl_type::uint_type);
+   this->pixel_y = fs_reg(this, glsl_type::uint_type);
+   this->pixel_x.type = BRW_REGISTER_TYPE_UW;
+   this->pixel_y.type = BRW_REGISTER_TYPE_UW;
+
+   emit(FS_OPCODE_PIXEL_X, this->pixel_x);
+   emit(FS_OPCODE_PIXEL_Y, this->pixel_y);
+
+   this->current_annotation = "compute pixel deltas from v0";
+   if (brw->has_pln) {
+      this->delta_x = fs_reg(this, glsl_type::vec2_type);
+      this->delta_y = this->delta_x;
+      this->delta_y.reg_offset++;
+   } else {
+      this->delta_x = fs_reg(this, glsl_type::float_type);
+      this->delta_y = fs_reg(this, glsl_type::float_type);
+   }
+   emit(BRW_OPCODE_ADD, this->delta_x,
+       this->pixel_x, fs_reg(negate(brw_vec1_grf(1, 0))));
+   emit(BRW_OPCODE_ADD, this->delta_y,
+       this->pixel_y, fs_reg(negate(brw_vec1_grf(1, 1))));
+
+   this->current_annotation = "compute pos.w and 1/pos.w";
+   /* Compute wpos.w.  It's always in our setup, since it's needed to
+    * interpolate the other attributes.
+    */
+   this->wpos_w = fs_reg(this, glsl_type::float_type);
+   emit(FS_OPCODE_LINTERP, wpos_w, this->delta_x, this->delta_y,
+       interp_reg(FRAG_ATTRIB_WPOS, 3));
+   /* Compute the pixel 1/W value from wpos.w. */
+   this->pixel_w = fs_reg(this, glsl_type::float_type);
+   emit_math(FS_OPCODE_RCP, this->pixel_w, wpos_w);
+   this->current_annotation = NULL;
+}
+
+/** Emits the interpolation for the varying inputs. */
+void
+fs_visitor::emit_interpolation_setup_gen6()
+{
+   struct brw_reg g1_uw = retype(brw_vec1_grf(1, 0), BRW_REGISTER_TYPE_UW);
+
+   /* If the pixel centers end up used, the setup is the same as for gen4. */
+   this->current_annotation = "compute pixel centers";
+   fs_reg int_pixel_x = fs_reg(this, glsl_type::uint_type);
+   fs_reg int_pixel_y = fs_reg(this, glsl_type::uint_type);
+   int_pixel_x.type = BRW_REGISTER_TYPE_UW;
+   int_pixel_y.type = BRW_REGISTER_TYPE_UW;
+   emit(BRW_OPCODE_ADD,
+       int_pixel_x,
+       fs_reg(stride(suboffset(g1_uw, 4), 2, 4, 0)),
+       fs_reg(brw_imm_v(0x10101010)));
+   emit(BRW_OPCODE_ADD,
+       int_pixel_y,
+       fs_reg(stride(suboffset(g1_uw, 5), 2, 4, 0)),
+       fs_reg(brw_imm_v(0x11001100)));
+
+   /* As of gen6, we can no longer mix float and int sources.  We have
+    * to turn the integer pixel centers into floats for their actual
+    * use.
+    */
+   this->pixel_x = fs_reg(this, glsl_type::float_type);
+   this->pixel_y = fs_reg(this, glsl_type::float_type);
+   emit(BRW_OPCODE_MOV, this->pixel_x, int_pixel_x);
+   emit(BRW_OPCODE_MOV, this->pixel_y, int_pixel_y);
+
+   this->current_annotation = "compute pos.w";
+   this->pixel_w = fs_reg(brw_vec8_grf(c->source_w_reg, 0));
+   this->wpos_w = fs_reg(this, glsl_type::float_type);
+   emit_math(FS_OPCODE_RCP, this->wpos_w, this->pixel_w);
+
+   this->delta_x = fs_reg(brw_vec8_grf(2, 0));
+   this->delta_y = fs_reg(brw_vec8_grf(3, 0));
+
+   this->current_annotation = NULL;
+}
+
+void
+fs_visitor::emit_color_write(int index, int first_color_mrf, fs_reg color)
+{
+   int reg_width = c->dispatch_width / 8;
+
+   if (c->dispatch_width == 8 || intel->gen == 6) {
+      /* SIMD8 write looks like:
+       * m + 0: r0
+       * m + 1: r1
+       * m + 2: g0
+       * m + 3: g1
+       *
+       * gen6 SIMD16 DP write looks like:
+       * m + 0: r0
+       * m + 1: r1
+       * m + 2: g0
+       * m + 3: g1
+       * m + 4: b0
+       * m + 5: b1
+       * m + 6: a0
+       * m + 7: a1
+       */
+      emit(BRW_OPCODE_MOV, fs_reg(MRF, first_color_mrf + index * reg_width),
+          color);
+   } else {
+      /* pre-gen6 SIMD16 single source DP write looks like:
+       * m + 0: r0
+       * m + 1: g0
+       * m + 2: b0
+       * m + 3: a0
+       * m + 4: r1
+       * m + 5: g1
+       * m + 6: b1
+       * m + 7: a1
+       */
+      if (brw->has_compr4) {
+        /* By setting the high bit of the MRF register number, we
+         * indicate that we want COMPR4 mode - instead of doing the
+         * usual destination + 1 for the second half we get
+         * destination + 4.
+         */
+        emit(BRW_OPCODE_MOV,
+             fs_reg(MRF, BRW_MRF_COMPR4 + first_color_mrf + index), color);
+      } else {
+        push_force_uncompressed();
+        emit(BRW_OPCODE_MOV, fs_reg(MRF, first_color_mrf + index), color);
+        pop_force_uncompressed();
+
+        push_force_sechalf();
+        color.sechalf = true;
+        emit(BRW_OPCODE_MOV, fs_reg(MRF, first_color_mrf + index + 4), color);
+        pop_force_sechalf();
+        color.sechalf = false;
+      }
+   }
+}
+
+void
+fs_visitor::emit_fb_writes()
+{
+   this->current_annotation = "FB write header";
+   GLboolean header_present = GL_TRUE;
+   int nr = 0;
+   int reg_width = c->dispatch_width / 8;
+
+   if (intel->gen >= 6 &&
+       !this->kill_emitted &&
+       c->key.nr_color_regions == 1) {
+      header_present = false;
+   }
+
+   if (header_present) {
+      /* m0, m1 header */
+      nr += 2;
+   }
+
+   if (c->aa_dest_stencil_reg) {
+      push_force_uncompressed();
+      emit(BRW_OPCODE_MOV, fs_reg(MRF, nr++),
+          fs_reg(brw_vec8_grf(c->aa_dest_stencil_reg, 0)));
+      pop_force_uncompressed();
+   }
+
+   /* Reserve space for color. It'll be filled in per MRT below. */
+   int color_mrf = nr;
+   nr += 4 * reg_width;
+
+   if (c->source_depth_to_render_target) {
+      if (intel->gen == 6 && c->dispatch_width == 16) {
+        /* For outputting oDepth on gen6, SIMD8 writes have to be
+         * used.  This would require 8-wide moves of each half to
+         * message regs, kind of like pre-gen5 SIMD16 FB writes.
+         * Just bail on doing so for now.
+         */
+        fail("Missing support for simd16 depth writes on gen6\n");
+      }
+
+      if (c->computes_depth) {
+        /* Hand over gl_FragDepth. */
+        assert(this->frag_depth);
+        fs_reg depth = *(variable_storage(this->frag_depth));
+
+        emit(BRW_OPCODE_MOV, fs_reg(MRF, nr), depth);
+      } else {
+        /* Pass through the payload depth. */
+        emit(BRW_OPCODE_MOV, fs_reg(MRF, nr),
+             fs_reg(brw_vec8_grf(c->source_depth_reg, 0)));
+      }
+      nr += reg_width;
+   }
+
+   if (c->dest_depth_reg) {
+      emit(BRW_OPCODE_MOV, fs_reg(MRF, nr),
+          fs_reg(brw_vec8_grf(c->dest_depth_reg, 0)));
+      nr += reg_width;
+   }
+
+   fs_reg color = reg_undef;
+   if (this->frag_color)
+      color = *(variable_storage(this->frag_color));
+   else if (this->frag_data) {
+      color = *(variable_storage(this->frag_data));
+      color.type = BRW_REGISTER_TYPE_F;
+   }
+
+   for (int target = 0; target < c->key.nr_color_regions; target++) {
+      this->current_annotation = ralloc_asprintf(this->mem_ctx,
+                                                "FB write target %d",
+                                                target);
+      if (this->frag_color || this->frag_data) {
+        for (int i = 0; i < 4; i++) {
+           emit_color_write(i, color_mrf, color);
+           color.reg_offset++;
+        }
+      }
+
+      if (this->frag_color)
+        color.reg_offset -= 4;
+
+      fs_inst *inst = emit(FS_OPCODE_FB_WRITE);
+      inst->target = target;
+      inst->base_mrf = 0;
+      inst->mlen = nr;
+      if (target == c->key.nr_color_regions - 1)
+        inst->eot = true;
+      inst->header_present = header_present;
+   }
+
+   if (c->key.nr_color_regions == 0) {
+      if (c->key.alpha_test && (this->frag_color || this->frag_data)) {
+        /* If the alpha test is enabled but there's no color buffer,
+         * we still need to send alpha out the pipeline to our null
+         * renderbuffer.
+         */
+        color.reg_offset += 3;
+        emit_color_write(3, color_mrf, color);
+      }
+
+      fs_inst *inst = emit(FS_OPCODE_FB_WRITE);
+      inst->base_mrf = 0;
+      inst->mlen = nr;
+      inst->eot = true;
+      inst->header_present = header_present;
+   }
+
+   this->current_annotation = NULL;
+}