Merge ../mesa into vulkan
[mesa.git] / src / gallium / auxiliary / tgsi / tgsi_scan.c
index 05b711148ce70a9fc6f915d3e73e2b83fcfe86b9..7a02e27e01e866c0b1ded20af07cdf3a5a4bb39d 100644 (file)
@@ -1,6 +1,6 @@
 /**************************************************************************
  * 
- * Copyright 2008 Tungsten Graphics, Inc., Cedar Park, Texas.
+ * Copyright 2008 VMware, Inc.
  * All Rights Reserved.
  * Copyright 2008 VMware, Inc.  All rights Reserved.
  *
@@ -19,7 +19,7 @@
  * 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 NON-INFRINGEMENT.
- * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR
+ * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS 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.
@@ -56,10 +56,14 @@ tgsi_scan_shader(const struct tgsi_token *tokens,
 {
    uint procType, i;
    struct tgsi_parse_context parse;
+   unsigned current_depth = 0;
 
    memset(info, 0, sizeof(*info));
    for (i = 0; i < TGSI_FILE_COUNT; i++)
       info->file_max[i] = -1;
+   for (i = 0; i < Elements(info->const_file_max); i++)
+      info->const_file_max[i] = -1;
+   info->properties[TGSI_PROPERTY_GS_INVOCATIONS] = 1;
 
    /**
     ** Setup to begin parsing input shader
@@ -72,6 +76,8 @@ tgsi_scan_shader(const struct tgsi_token *tokens,
    assert(procType == TGSI_PROCESSOR_FRAGMENT ||
           procType == TGSI_PROCESSOR_VERTEX ||
           procType == TGSI_PROCESSOR_GEOMETRY ||
+          procType == TGSI_PROCESSOR_TESS_CTRL ||
+          procType == TGSI_PROCESSOR_TESS_EVAL ||
           procType == TGSI_PROCESSOR_COMPUTE);
    info->processor = procType;
 
@@ -95,6 +101,72 @@ tgsi_scan_shader(const struct tgsi_token *tokens,
             assert(fullinst->Instruction.Opcode < TGSI_OPCODE_LAST);
             info->opcode_count[fullinst->Instruction.Opcode]++;
 
+            switch (fullinst->Instruction.Opcode) {
+            case TGSI_OPCODE_IF:
+            case TGSI_OPCODE_UIF:
+            case TGSI_OPCODE_BGNLOOP:
+               current_depth++;
+               info->max_depth = MAX2(info->max_depth, current_depth);
+               break;
+            case TGSI_OPCODE_ENDIF:
+            case TGSI_OPCODE_ENDLOOP:
+               current_depth--;
+               break;
+            default:
+               break;
+            }
+
+            if (fullinst->Instruction.Opcode == TGSI_OPCODE_INTERP_CENTROID ||
+                fullinst->Instruction.Opcode == TGSI_OPCODE_INTERP_OFFSET ||
+                fullinst->Instruction.Opcode == TGSI_OPCODE_INTERP_SAMPLE) {
+               const struct tgsi_full_src_register *src0 = &fullinst->Src[0];
+               unsigned input;
+
+               if (src0->Register.Indirect && src0->Indirect.ArrayID)
+                  input = info->input_array_first[src0->Indirect.ArrayID];
+               else
+                  input = src0->Register.Index;
+
+               /* For the INTERP opcodes, the interpolation is always
+                * PERSPECTIVE unless LINEAR is specified.
+                */
+               switch (info->input_interpolate[input]) {
+               case TGSI_INTERPOLATE_COLOR:
+               case TGSI_INTERPOLATE_CONSTANT:
+               case TGSI_INTERPOLATE_PERSPECTIVE:
+                  switch (fullinst->Instruction.Opcode) {
+                  case TGSI_OPCODE_INTERP_CENTROID:
+                     info->uses_persp_opcode_interp_centroid = true;
+                     break;
+                  case TGSI_OPCODE_INTERP_OFFSET:
+                     info->uses_persp_opcode_interp_offset = true;
+                     break;
+                  case TGSI_OPCODE_INTERP_SAMPLE:
+                     info->uses_persp_opcode_interp_sample = true;
+                     break;
+                  }
+                  break;
+
+               case TGSI_INTERPOLATE_LINEAR:
+                  switch (fullinst->Instruction.Opcode) {
+                  case TGSI_OPCODE_INTERP_CENTROID:
+                     info->uses_linear_opcode_interp_centroid = true;
+                     break;
+                  case TGSI_OPCODE_INTERP_OFFSET:
+                     info->uses_linear_opcode_interp_offset = true;
+                     break;
+                  case TGSI_OPCODE_INTERP_SAMPLE:
+                     info->uses_linear_opcode_interp_sample = true;
+                     break;
+                  }
+                  break;
+               }
+            }
+
+            if (fullinst->Instruction.Opcode >= TGSI_OPCODE_F2D &&
+                fullinst->Instruction.Opcode <= TGSI_OPCODE_DSSG)
+               info->uses_doubles = true;
+
             for (i = 0; i < fullinst->Instruction.NumSrcRegs; i++) {
                const struct tgsi_full_src_register *src =
                   &fullinst->Src[i];
@@ -115,19 +187,35 @@ tgsi_scan_shader(const struct tgsi_token *tokens,
                   }
 
                   if (procType == TGSI_PROCESSOR_FRAGMENT &&
-                      info->reads_position &&
-                      src->Register.Index == 0 &&
-                      (src->Register.SwizzleX == TGSI_SWIZZLE_Z ||
-                       src->Register.SwizzleY == TGSI_SWIZZLE_Z ||
-                       src->Register.SwizzleZ == TGSI_SWIZZLE_Z ||
-                       src->Register.SwizzleW == TGSI_SWIZZLE_Z)) {
-                     info->reads_z = TRUE;
+                      !src->Register.Indirect) {
+                     unsigned name =
+                        info->input_semantic_name[src->Register.Index];
+                     unsigned index =
+                        info->input_semantic_index[src->Register.Index];
+
+                     if (name == TGSI_SEMANTIC_POSITION &&
+                         (src->Register.SwizzleX == TGSI_SWIZZLE_Z ||
+                          src->Register.SwizzleY == TGSI_SWIZZLE_Z ||
+                          src->Register.SwizzleZ == TGSI_SWIZZLE_Z ||
+                          src->Register.SwizzleW == TGSI_SWIZZLE_Z))
+                        info->reads_z = TRUE;
+
+                     if (name == TGSI_SEMANTIC_COLOR) {
+                        unsigned mask =
+                              (1 << src->Register.SwizzleX) |
+                              (1 << src->Register.SwizzleY) |
+                              (1 << src->Register.SwizzleZ) |
+                              (1 << src->Register.SwizzleW);
+
+                        info->colors_read |= mask << (index * 4);
+                     }
                   }
                }
 
                /* check for indirect register reads */
                if (src->Register.Indirect) {
                   info->indirect_files |= (1 << src->Register.File);
+                  info->indirect_files_read |= (1 << src->Register.File);
                }
 
                /* MSAA samplers */
@@ -148,6 +236,7 @@ tgsi_scan_shader(const struct tgsi_token *tokens,
                const struct tgsi_full_dst_register *dst = &fullinst->Dst[i];
                if (dst->Register.Indirect) {
                   info->indirect_files |= (1 << dst->Register.File);
+                  info->indirect_files_written |= (1 << dst->Register.File);
                }
             }
 
@@ -161,30 +250,102 @@ tgsi_scan_shader(const struct tgsi_token *tokens,
                = &parse.FullToken.FullDeclaration;
             const uint file = fulldecl->Declaration.File;
             uint reg;
+
+            if (fulldecl->Declaration.Array) {
+               unsigned array_id = fulldecl->Array.ArrayID;
+
+               switch (file) {
+               case TGSI_FILE_INPUT:
+                  assert(array_id < ARRAY_SIZE(info->input_array_first));
+                  info->input_array_first[array_id] = fulldecl->Range.First;
+                  info->input_array_last[array_id] = fulldecl->Range.Last;
+                  break;
+               case TGSI_FILE_OUTPUT:
+                  assert(array_id < ARRAY_SIZE(info->output_array_first));
+                  info->output_array_first[array_id] = fulldecl->Range.First;
+                  info->output_array_last[array_id] = fulldecl->Range.Last;
+                  break;
+               }
+               info->array_max[file] = MAX2(info->array_max[file], array_id);
+            }
+
             for (reg = fulldecl->Range.First;
                  reg <= fulldecl->Range.Last;
                  reg++) {
                unsigned semName = fulldecl->Semantic.Name;
-               unsigned semIndex = fulldecl->Semantic.Index;
+               unsigned semIndex =
+                  fulldecl->Semantic.Index + (reg - fulldecl->Range.First);
 
                /* only first 32 regs will appear in this bitfield */
                info->file_mask[file] |= (1 << reg);
                info->file_count[file]++;
                info->file_max[file] = MAX2(info->file_max[file], (int)reg);
 
-               if (file == TGSI_FILE_INPUT) {
+               if (file == TGSI_FILE_CONSTANT) {
+                  int buffer = 0;
+
+                  if (fulldecl->Declaration.Dimension)
+                     buffer = fulldecl->Dim.Index2D;
+
+                  info->const_file_max[buffer] =
+                        MAX2(info->const_file_max[buffer], (int)reg);
+               }
+               else if (file == TGSI_FILE_INPUT) {
                   info->input_semantic_name[reg] = (ubyte) semName;
                   info->input_semantic_index[reg] = (ubyte) semIndex;
                   info->input_interpolate[reg] = (ubyte)fulldecl->Interp.Interpolate;
-                  info->input_centroid[reg] = (ubyte)fulldecl->Interp.Centroid;
+                  info->input_interpolate_loc[reg] = (ubyte)fulldecl->Interp.Location;
                   info->input_cylindrical_wrap[reg] = (ubyte)fulldecl->Interp.CylindricalWrap;
                   info->num_inputs++;
 
-                  if (procType == TGSI_PROCESSOR_FRAGMENT) {
+                  /* Only interpolated varyings. Don't include POSITION.
+                   * Don't include integer varyings, because they are not
+                   * interpolated.
+                   */
+                  if (semName == TGSI_SEMANTIC_GENERIC ||
+                      semName == TGSI_SEMANTIC_TEXCOORD ||
+                      semName == TGSI_SEMANTIC_COLOR ||
+                      semName == TGSI_SEMANTIC_BCOLOR ||
+                      semName == TGSI_SEMANTIC_FOG ||
+                      semName == TGSI_SEMANTIC_CLIPDIST ||
+                      semName == TGSI_SEMANTIC_CULLDIST) {
+                     switch (fulldecl->Interp.Interpolate) {
+                     case TGSI_INTERPOLATE_COLOR:
+                     case TGSI_INTERPOLATE_PERSPECTIVE:
+                        switch (fulldecl->Interp.Location) {
+                        case TGSI_INTERPOLATE_LOC_CENTER:
+                           info->uses_persp_center = true;
+                           break;
+                        case TGSI_INTERPOLATE_LOC_CENTROID:
+                           info->uses_persp_centroid = true;
+                           break;
+                        case TGSI_INTERPOLATE_LOC_SAMPLE:
+                           info->uses_persp_sample = true;
+                           break;
+                        }
+                        break;
+                     case TGSI_INTERPOLATE_LINEAR:
+                        switch (fulldecl->Interp.Location) {
+                        case TGSI_INTERPOLATE_LOC_CENTER:
+                           info->uses_linear_center = true;
+                           break;
+                        case TGSI_INTERPOLATE_LOC_CENTROID:
+                           info->uses_linear_centroid = true;
+                           break;
+                        case TGSI_INTERPOLATE_LOC_SAMPLE:
+                           info->uses_linear_sample = true;
+                           break;
+                        }
+                        break;
+                     /* TGSI_INTERPOLATE_CONSTANT doesn't do any interpolation. */
+                     }
+                  }
+
+                  if (semName == TGSI_SEMANTIC_PRIMID)
+                     info->uses_primid = TRUE;
+                  else if (procType == TGSI_PROCESSOR_FRAGMENT) {
                      if (semName == TGSI_SEMANTIC_POSITION)
                         info->reads_position = TRUE;
-                     else if (semName == TGSI_SEMANTIC_PRIMID)
-                        info->uses_primid = TRUE;
                      else if (semName == TGSI_SEMANTIC_FACE)
                         info->uses_frontface = TRUE;
                   }
@@ -202,24 +363,44 @@ tgsi_scan_shader(const struct tgsi_token *tokens,
                   else if (semName == TGSI_SEMANTIC_VERTEXID) {
                      info->uses_vertexid = TRUE;
                   }
+                  else if (semName == TGSI_SEMANTIC_VERTEXID_NOBASE) {
+                     info->uses_vertexid_nobase = TRUE;
+                  }
+                  else if (semName == TGSI_SEMANTIC_BASEVERTEX) {
+                     info->uses_basevertex = TRUE;
+                  }
                   else if (semName == TGSI_SEMANTIC_PRIMID) {
                      info->uses_primid = TRUE;
-                  }
+                  } else if (semName == TGSI_SEMANTIC_INVOCATIONID) {
+                     info->uses_invocationid = TRUE;
+                  } else if (semName == TGSI_SEMANTIC_POSITION)
+                     info->reads_position = TRUE;
+                  else if (semName == TGSI_SEMANTIC_FACE)
+                     info->uses_frontface = TRUE;
                }
                else if (file == TGSI_FILE_OUTPUT) {
                   info->output_semantic_name[reg] = (ubyte) semName;
                   info->output_semantic_index[reg] = (ubyte) semIndex;
                   info->num_outputs++;
 
+                  if (semName == TGSI_SEMANTIC_COLOR)
+                     info->colors_written |= 1 << semIndex;
+
                   if (procType == TGSI_PROCESSOR_VERTEX ||
-                      procType == TGSI_PROCESSOR_GEOMETRY) {
-                     if (semName == TGSI_SEMANTIC_CLIPDIST) {
-                        info->num_written_clipdistance +=
-                           util_bitcount(fulldecl->Declaration.UsageMask);
+                      procType == TGSI_PROCESSOR_GEOMETRY ||
+                      procType == TGSI_PROCESSOR_TESS_CTRL ||
+                      procType == TGSI_PROCESSOR_TESS_EVAL) {
+                     if (semName == TGSI_SEMANTIC_VIEWPORT_INDEX) {
+                        info->writes_viewport_index = TRUE;
                      }
-                     else if (semName == TGSI_SEMANTIC_CULLDIST) {
-                        info->num_written_culldistance +=
-                           util_bitcount(fulldecl->Declaration.UsageMask);
+                     else if (semName == TGSI_SEMANTIC_LAYER) {
+                        info->writes_layer = TRUE;
+                     }
+                     else if (semName == TGSI_SEMANTIC_PSIZE) {
+                        info->writes_psize = TRUE;
+                     }
+                     else if (semName == TGSI_SEMANTIC_CLIPVERTEX) {
+                        info->writes_clipvertex = TRUE;
                      }
                   }
 
@@ -229,6 +410,8 @@ tgsi_scan_shader(const struct tgsi_token *tokens,
                      }
                      else if (semName == TGSI_SEMANTIC_STENCIL) {
                         info->writes_stencil = TRUE;
+                     } else if (semName == TGSI_SEMANTIC_SAMPLEMASK) {
+                        info->writes_samplemask = TRUE;
                      }
                   }
 
@@ -237,15 +420,8 @@ tgsi_scan_shader(const struct tgsi_token *tokens,
                         info->writes_edgeflag = TRUE;
                      }
                   }
-
-                  if (procType == TGSI_PROCESSOR_GEOMETRY) {
-                     if (semName == TGSI_SEMANTIC_VIEWPORT_INDEX) {
-                        info->writes_viewport_index = TRUE;
-                     }
-                     else if (semName == TGSI_SEMANTIC_LAYER) {
-                        info->writes_layer = TRUE;
-                     }
-                  }
+               } else if (file == TGSI_FILE_SAMPLER) {
+                  info->samplers_declared |= 1 << reg;
                }
             }
          }
@@ -266,13 +442,22 @@ tgsi_scan_shader(const struct tgsi_token *tokens,
          {
             const struct tgsi_full_property *fullprop
                = &parse.FullToken.FullProperty;
-
-            info->properties[info->num_properties].name =
-               fullprop->Property.PropertyName;
-            memcpy(info->properties[info->num_properties].data,
-                   fullprop->u, 8 * sizeof(unsigned));;
-
-            ++info->num_properties;
+            unsigned name = fullprop->Property.PropertyName;
+            unsigned value = fullprop->u[0].Data;
+
+            assert(name < Elements(info->properties));
+            info->properties[name] = value;
+
+            switch (name) {
+            case TGSI_PROPERTY_NUM_CLIPDIST_ENABLED:
+               info->num_written_clipdistance = value;
+               info->clipdist_writemask |= (1 << value) - 1;
+               break;
+            case TGSI_PROPERTY_NUM_CULLDIST_ENABLED:
+               info->num_written_culldistance = value;
+               info->culldist_writemask |= (1 << value) - 1;
+               break;
+            }
          }
          break;
 
@@ -284,36 +469,19 @@ tgsi_scan_shader(const struct tgsi_token *tokens,
    info->uses_kill = (info->opcode_count[TGSI_OPCODE_KILL_IF] ||
                       info->opcode_count[TGSI_OPCODE_KILL]);
 
-   /* extract simple properties */
-   for (i = 0; i < info->num_properties; ++i) {
-      switch (info->properties[i].name) {
-      case TGSI_PROPERTY_FS_COORD_ORIGIN:
-         info->origin_lower_left = info->properties[i].data[0];
-         break;
-      case TGSI_PROPERTY_FS_COORD_PIXEL_CENTER:
-         info->pixel_center_integer = info->properties[i].data[0];
-         break;
-      case TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS:
-         info->color0_writes_all_cbufs = info->properties[i].data[0];
-         break;
-      case TGSI_PROPERTY_GS_INPUT_PRIM:
-         /* The dimensions of the IN decleration in geometry shader have
-          * to be deduced from the type of the input primitive.
-          */
-         if (procType == TGSI_PROCESSOR_GEOMETRY) {
-            unsigned input_primitive = info->properties[i].data[0];
-            int num_verts = u_vertices_per_prim(input_primitive);
-            int j;
-            info->file_count[TGSI_FILE_INPUT] = num_verts;
-            info->file_max[TGSI_FILE_INPUT] =
-               MAX2(info->file_max[TGSI_FILE_INPUT], num_verts - 1);
-            for (j = 0; j < num_verts; ++j) {
-               info->file_mask[TGSI_FILE_INPUT] |= (1 << j);
-            }
-         }
-         break;
-      default:
-         ;
+   /* The dimensions of the IN decleration in geometry shader have
+    * to be deduced from the type of the input primitive.
+    */
+   if (procType == TGSI_PROCESSOR_GEOMETRY) {
+      unsigned input_primitive =
+            info->properties[TGSI_PROPERTY_GS_INPUT_PRIM];
+      int num_verts = u_vertices_per_prim(input_primitive);
+      int j;
+      info->file_count[TGSI_FILE_INPUT] = num_verts;
+      info->file_max[TGSI_FILE_INPUT] =
+            MAX2(info->file_max[TGSI_FILE_INPUT], num_verts - 1);
+      for (j = 0; j < num_verts; ++j) {
+         info->file_mask[TGSI_FILE_INPUT] |= (1 << j);
       }
    }