i965/fs: Lower 32x32 bit multiplication on BXT.

[mesa.git] / src / mesa / drivers / dri / i965 / brw_fs_channel_expressions.cpp

diff --git a/src/mesa/drivers/dri/i965/brw_fs_channel_expressions.cpp b/src/mesa/drivers/dri/i965/brw_fs_channel_expressions.cpp
index 0f3d4abdd2678bb691c5b6fa165b21c20b5c42ae..a8883a35ef296f38ff11c9e9ac2f332b81de9a5b 100644 (file)
--- a/src/mesa/drivers/dri/i965/brw_fs_channel_expressions.cpp
+++ b/src/mesa/drivers/dri/i965/brw_fs_channel_expressions.cpp
@@ -41,10 +41,8 @@
  * we do retain the vector types in that case.
  */
 
-extern "C" {
 #include "main/core.h"
 #include "brw_wm.h"
-}
 #include "glsl/ir.h"
 #include "glsl/ir_expression_flattening.h"
 #include "glsl/glsl_types.h"
@@ -75,6 +73,18 @@ channel_expressions_predicate(ir_instruction *ir)
    if (!expr)
       return false;
 
+   switch (expr->operation) {
+      /* these opcodes need to act on the whole vector,
+       * just like texturing.
+       */
+      case ir_unop_interpolate_at_centroid:
+      case ir_binop_interpolate_at_offset:
+      case ir_binop_interpolate_at_sample:
+         return false;
+      default:
+         break;
+   }
+
    for (i = 0; i < expr->get_num_operands(); i++) {
       if (expr->operands[i]->type->is_vector())
         return true;
@@ -153,6 +163,16 @@ ir_channel_expressions_visitor::visit_leave(ir_assignment *ir)
    if (!found_vector)
       return visit_continue;
 
+   switch (expr->operation) {
+      case ir_unop_interpolate_at_centroid:
+      case ir_binop_interpolate_at_offset:
+      case ir_binop_interpolate_at_sample:
+         return visit_continue;
+
+      default:
+         break;
+   }
+
    /* Store the expression operands in temps so we can use them
     * multiple times.
     */
@@ -212,14 +232,18 @@ ir_channel_expressions_visitor::visit_leave(ir_assignment *ir)
    case ir_unop_round_even:
    case ir_unop_sin:
    case ir_unop_cos:
-   case ir_unop_sin_reduced:
-   case ir_unop_cos_reduced:
    case ir_unop_dFdx:
+   case ir_unop_dFdx_coarse:
+   case ir_unop_dFdx_fine:
    case ir_unop_dFdy:
+   case ir_unop_dFdy_coarse:
+   case ir_unop_dFdy_fine:
    case ir_unop_bitfield_reverse:
    case ir_unop_bit_count:
    case ir_unop_find_msb:
    case ir_unop_find_lsb:
+   case ir_unop_saturate:
+   case ir_unop_subroutine_to_int:
       for (i = 0; i < vector_elements; i++) {
         ir_rvalue *op0 = get_element(op_var[0], i);
 
@@ -233,7 +257,10 @@ ir_channel_expressions_visitor::visit_leave(ir_assignment *ir)
    case ir_binop_add:
    case ir_binop_sub:
    case ir_binop_mul:
+   case ir_binop_imul_high:
    case ir_binop_div:
+   case ir_binop_carry:
+   case ir_binop_borrow:
    case ir_binop_mod:
    case ir_binop_min:
    case ir_binop_max:
@@ -243,6 +270,9 @@ ir_channel_expressions_visitor::visit_leave(ir_assignment *ir)
    case ir_binop_bit_and:
    case ir_binop_bit_xor:
    case ir_binop_bit_or:
+   case ir_binop_logic_and:
+   case ir_binop_logic_xor:
+   case ir_binop_logic_or:
    case ir_binop_less:
    case ir_binop_greater:
    case ir_binop_lequal:
@@ -301,13 +331,6 @@ ir_channel_expressions_visitor::visit_leave(ir_assignment *ir)
       break;
    }
 
-   case ir_binop_logic_and:
-   case ir_binop_logic_xor:
-   case ir_binop_logic_or:
-      ir->print();
-      printf("\n");
-      assert(!"not reached: expression operates on scalars only");
-      break;
    case ir_binop_all_equal:
    case ir_binop_any_nequal: {
       ir_expression *last = NULL;
@@ -339,8 +362,7 @@ ir_channel_expressions_visitor::visit_leave(ir_assignment *ir)
       break;
    }
    case ir_unop_noise:
-      assert(!"noise should have been broken down to function call");
-      break;
+      unreachable("noise should have been broken down to function call");
 
    case ir_binop_bfm: {
       /* Does not need to be scalarized, since its result will be identical
@@ -357,10 +379,11 @@ ir_channel_expressions_visitor::visit_leave(ir_assignment *ir)
    }
 
    case ir_binop_ubo_load:
-      assert(!"not yet supported");
-      break;
+      unreachable("not yet supported");
 
+   case ir_triop_fma:
    case ir_triop_lrp:
+   case ir_triop_csel:
    case ir_triop_bitfield_extract:
       for (i = 0; i < vector_elements; i++) {
         ir_rvalue *op0 = get_element(op_var[0], i);
@@ -402,16 +425,33 @@ ir_channel_expressions_visitor::visit_leave(ir_assignment *ir)
    case ir_unop_unpack_unorm_2x16:
    case ir_unop_unpack_unorm_4x8:
    case ir_unop_unpack_half_2x16:
+   case ir_binop_ldexp:
+   case ir_binop_vector_extract:
+   case ir_triop_vector_insert:
    case ir_quadop_bitfield_insert:
    case ir_quadop_vector:
-      assert(!"should have been lowered");
-      break;
+      unreachable("should have been lowered");
 
    case ir_unop_unpack_half_2x16_split_x:
    case ir_unop_unpack_half_2x16_split_y:
    case ir_binop_pack_half_2x16_split:
-      assert("!not reached: expression operates on scalars only");
-      break;
+   case ir_unop_interpolate_at_centroid:
+   case ir_binop_interpolate_at_offset:
+   case ir_binop_interpolate_at_sample:
+      unreachable("not reached: expression operates on scalars only");
+
+   case ir_unop_pack_double_2x32:
+   case ir_unop_unpack_double_2x32:
+   case ir_unop_frexp_sig:
+   case ir_unop_frexp_exp:
+   case ir_unop_d2f:
+   case ir_unop_f2d:
+   case ir_unop_d2i:
+   case ir_unop_i2d:
+   case ir_unop_d2u:
+   case ir_unop_u2d:
+   case ir_unop_d2b:
+      unreachable("no fp64 support yet");
    }
 
    ir->remove();