1 files changed, 343 insertions, 45 deletions

diff --git a/src/mesa/shader/slang/slang_codegen.c b/src/mesa/shader/slang/slang_codegen.c
index 712f5975dc..5e2ce0ce3a 100644
--- a/src/mesa/shader/slang/slang_codegen.c
+++ b/src/mesa/shader/slang/slang_codegen.c
@@ -57,6 +57,21 @@
 #include "slang_print.h"
 
 
+/** Max iterations to unroll */
+const GLuint MAX_FOR_LOOP_UNROLL_ITERATIONS = 32;
+
+/** Max for-loop body size (in slang operations) to unroll */
+const GLuint MAX_FOR_LOOP_UNROLL_BODY_SIZE = 50;
+
+/** Max for-loop body complexity to unroll.
+ * We'll compute complexity as the product of the number of iterations
+ * and the size of the body.  So long-ish loops with very simple bodies
+ * can be unrolled, as well as short loops with larger bodies.
+ */
+const GLuint MAX_FOR_LOOP_UNROLL_COMPLEXITY = 256;
+
+
+
 static slang_ir_node *
 _slang_gen_operation(slang_assemble_ctx * A, slang_operation *oper);
 
@@ -398,7 +413,7 @@ _slang_input_index(const char *name, GLenum target, GLuint *swizzleOut)
    const struct input_info *inputs
       = (target == GL_VERTEX_PROGRAM_ARB) ? vertInputs : fragInputs;
 
-   ASSERT(MAX_TEXTURE_UNITS == 8); /* if this fails, fix vertInputs above */
+   ASSERT(MAX_TEXTURE_COORD_UNITS == 8); /* if this fails, fix vertInputs above */
 
    for (i = 0; inputs[i].Name; i++) {
       if (strcmp(inputs[i].Name, name) == 0) {
@@ -476,6 +491,9 @@ static slang_asm_info AsmInfo[] = {
    { "vec4_multiply", IR_MUL, 1, 2 },
    { "vec4_dot", IR_DOT4, 1, 2 },
    { "vec3_dot", IR_DOT3, 1, 2 },
+   { "vec2_dot", IR_DOT2, 1, 2 },
+   { "vec3_nrm", IR_NRM3, 1, 1 },
+   { "vec4_nrm", IR_NRM4, 1, 1 },
    { "vec3_cross", IR_CROSS, 1, 2 },
    { "vec4_lrp", IR_LRP, 1, 3 },
    { "vec4_min", IR_MIN, 1, 2 },
@@ -498,18 +516,28 @@ static slang_asm_info AsmInfo[] = {
    /* float binary op */
    { "float_power", IR_POW, 1, 2 },
    /* texture / sampler */
-   { "vec4_tex1d", IR_TEX, 1, 2 },
-   { "vec4_texb1d", IR_TEXB, 1, 2 },  /* 1d w/ bias */
-   { "vec4_texp1d", IR_TEXP, 1, 2 },  /* 1d w/ projection */
-   { "vec4_tex2d", IR_TEX, 1, 2 },
-   { "vec4_texb2d", IR_TEXB, 1, 2 },  /* 2d w/ bias */
-   { "vec4_texp2d", IR_TEXP, 1, 2 },  /* 2d w/ projection */
-   { "vec4_tex3d", IR_TEX, 1, 2 },
-   { "vec4_texb3d", IR_TEXB, 1, 2 },  /* 3d w/ bias */
-   { "vec4_texp3d", IR_TEXP, 1, 2 },  /* 3d w/ projection */
-   { "vec4_texcube", IR_TEX, 1, 2 },  /* cubemap */
-   { "vec4_tex_rect", IR_TEX, 1, 2 }, /* rectangle */
-   { "vec4_texp_rect", IR_TEX, 1, 2 },/* rectangle w/ projection */
+   { "vec4_tex_1d", IR_TEX, 1, 2 },
+   { "vec4_tex_1d_bias", IR_TEXB, 1, 2 },  /* 1d w/ bias */
+   { "vec4_tex_1d_proj", IR_TEXP, 1, 2 },  /* 1d w/ projection */
+   { "vec4_tex_2d", IR_TEX, 1, 2 },
+   { "vec4_tex_2d_bias", IR_TEXB, 1, 2 },  /* 2d w/ bias */
+   { "vec4_tex_2d_proj", IR_TEXP, 1, 2 },  /* 2d w/ projection */
+   { "vec4_tex_3d", IR_TEX, 1, 2 },
+   { "vec4_tex_3d_bias", IR_TEXB, 1, 2 },  /* 3d w/ bias */
+   { "vec4_tex_3d_proj", IR_TEXP, 1, 2 },  /* 3d w/ projection */
+   { "vec4_tex_cube", IR_TEX, 1, 2 },      /* cubemap */
+   { "vec4_tex_rect", IR_TEX, 1, 2 },      /* rectangle */
+   { "vec4_tex_rect_bias", IR_TEX, 1, 2 }, /* rectangle w/ projection */
+
+   /* texture / sampler but with shadow comparison */
+   { "vec4_tex_1d_shadow", IR_TEX_SH, 1, 2 },
+   { "vec4_tex_1d_bias_shadow", IR_TEXB_SH, 1, 2 },
+   { "vec4_tex_1d_proj_shadow", IR_TEXP_SH, 1, 2 },
+   { "vec4_tex_2d_shadow", IR_TEX_SH, 1, 2 },
+   { "vec4_tex_2d_bias_shadow", IR_TEXB_SH, 1, 2 },
+   { "vec4_tex_2d_proj_shadow", IR_TEXP_SH, 1, 2 },
+   { "vec4_tex_rect_shadow", IR_TEX_SH, 1, 2 },
+   { "vec4_tex_rect_proj_shadow", IR_TEXP_SH, 1, 2 },
 
    /* unary op */
    { "ivec4_to_vec4", IR_I_TO_F, 1, 1 }, /* int[4] to float[4] */
@@ -1574,6 +1602,7 @@ swizzle_to_writemask(slang_assemble_ctx *A, GLuint swizzle,
 }
 
 
+#if 0 /* not used, but don't remove just yet */
 /**
  * Recursively traverse 'oper' to produce a swizzle mask in the event
  * of any vector subscripts and swizzle suffixes.
@@ -1627,8 +1656,10 @@ resolve_swizzle(const slang_operation *oper)
       return SWIZZLE_XYZW;
    }
 }
+#endif
 
 
+#if 0
 /**
  * Recursively descend through swizzle nodes to find the node's storage info.
  */
@@ -1640,7 +1671,7 @@ get_store(const slang_ir_node *n)
    }
    return n->Store;
 }
-
+#endif
 
 
 /**
@@ -1713,6 +1744,7 @@ _slang_gen_asm(slang_assemble_ctx *A, slang_operation *oper,
 }
 
 
+#if 0
 static void
 print_funcs(struct slang_function_scope_ *scope, const char *name)
 {
@@ -1726,6 +1758,7 @@ print_funcs(struct slang_function_scope_ *scope, const char *name)
    if (scope->outer_scope)
       print_funcs(scope->outer_scope, name);
 }
+#endif
 
 
 /**
@@ -2443,40 +2476,259 @@ _slang_gen_do(slang_assemble_ctx * A, const slang_operation *oper)
 
 
 /**
- * Generate for-loop using high-level IR_LOOP instruction.
+ * Recursively count the number of operations rooted at 'oper'.
+ * This gives some kind of indication of the size/complexity of an operation.
+ */
+static GLuint
+sizeof_operation(const slang_operation *oper)
+{
+   if (oper) {
+      GLuint count = 1; /* me */
+      GLuint i;
+      for (i = 0; i < oper->num_children; i++) {
+         count += sizeof_operation(&oper->children[i]);
+      }
+      return count;
+   }
+   else {
+      return 0;
+   }
+}
+
+
+/**
+ * Determine if a for-loop can be unrolled.
+ * At this time, only a rather narrow class of for loops can be unrolled.
+ * See code for details.
+ * When a loop can't be unrolled because it's too large we'll emit a
+ * message to the log.
+ */
+static GLboolean
+_slang_can_unroll_for_loop(slang_assemble_ctx * A, const slang_operation *oper)
+{
+   GLuint bodySize;
+   GLint start, end;
+   const char *varName;
+   slang_atom varId;
+
+   assert(oper->type == SLANG_OPER_FOR);
+   assert(oper->num_children == 4);
+
+   /* children[0] must be either "int i=constant" or "i=constant" */
+   if (oper->children[0].type == SLANG_OPER_BLOCK_NO_NEW_SCOPE) {
+      slang_variable *var;
+
+      if (oper->children[0].children[0].type != SLANG_OPER_VARIABLE_DECL)
+         return GL_FALSE;
+
+      varId = oper->children[0].children[0].a_id;
+
+      var = _slang_variable_locate(oper->children[0].children[0].locals,
+                                   varId, GL_TRUE);
+      if (!var)
+         return GL_FALSE;
+      if (!var->initializer)
+         return GL_FALSE;
+      if (var->initializer->type != SLANG_OPER_LITERAL_INT)
+         return GL_FALSE;
+      start = (GLint) var->initializer->literal[0];
+   }
+   else if (oper->children[0].type == SLANG_OPER_EXPRESSION) {
+      if (oper->children[0].children[0].type != SLANG_OPER_ASSIGN)
+         return GL_FALSE;
+      if (oper->children[0].children[0].children[0].type != SLANG_OPER_IDENTIFIER)
+         return GL_FALSE;
+      if (oper->children[0].children[0].children[1].type != SLANG_OPER_LITERAL_INT)
+         return GL_FALSE;
+
+      varId = oper->children[0].children[0].children[0].a_id;
+
+      start = (GLint) oper->children[0].children[0].children[1].literal[0];
+   }
+   else {
+      return GL_FALSE;
+   }
+
+   /* children[1] must be "i<constant" */
+   if (oper->children[1].type != SLANG_OPER_EXPRESSION)
+      return GL_FALSE;
+   if (oper->children[1].children[0].type != SLANG_OPER_LESS)
+      return GL_FALSE;
+   if (oper->children[1].children[0].children[0].type != SLANG_OPER_IDENTIFIER)
+      return GL_FALSE;
+   if (oper->children[1].children[0].children[1].type != SLANG_OPER_LITERAL_INT)
+      return GL_FALSE;
+
+   end = (GLint) oper->children[1].children[0].children[1].literal[0];
+
+   /* children[2] must be "i++" or "++i" */
+   if (oper->children[2].type != SLANG_OPER_POSTINCREMENT &&
+       oper->children[2].type != SLANG_OPER_PREINCREMENT)
+      return GL_FALSE;
+   if (oper->children[2].children[0].type != SLANG_OPER_IDENTIFIER)
+      return GL_FALSE;
+
+   /* make sure the same variable name is used in all places */
+   if ((oper->children[1].children[0].children[0].a_id != varId) ||
+       (oper->children[2].children[0].a_id != varId))
+      return GL_FALSE;
+
+   varName = (const char *) varId;
+
+   /* children[3], the loop body, can't be too large */
+   bodySize = sizeof_operation(&oper->children[3]);
+   if (bodySize > MAX_FOR_LOOP_UNROLL_BODY_SIZE) {
+      slang_info_log_print(A->log,
+                           "Note: 'for (%s ... )' body is too large/complex"
+                           " to unroll",
+                           varName);
+      return GL_FALSE;
+   }
+
+   if (start >= end)
+      return GL_FALSE; /* degenerate case */
+
+   if (end - start > MAX_FOR_LOOP_UNROLL_ITERATIONS) {
+      slang_info_log_print(A->log,
+                           "Note: 'for (%s=%d; %s<%d; ++%s)' is too"
+                           " many iterations to unroll",
+                           varName, start, varName, end, varName);
+      return GL_FALSE;
+   }
+
+   if ((end - start) * bodySize > MAX_FOR_LOOP_UNROLL_COMPLEXITY) {
+      slang_info_log_print(A->log,
+                           "Note: 'for (%s=%d; %s<%d; ++%s)' will generate"
+                           " too much code to unroll",
+                           varName, start, varName, end, varName);
+      return GL_FALSE;
+   }
+
+   return GL_TRUE; /* we can unroll the loop */
+}
+
+
+/**
+ * Unroll a for-loop.
+ * First we determine the number of iterations to unroll.
+ * Then for each iteration:
+ *   make a copy of the loop body
+ *   replace instances of the loop variable with the current iteration value
+ *   generate IR code for the body
+ * \return pointer to generated IR code or NULL if error, out of memory, etc.
+ */
+static slang_ir_node *
+_slang_unroll_for_loop(slang_assemble_ctx * A, const slang_operation *oper)
+{
+   GLint start, end, iter;
+   slang_ir_node *n, *root = NULL;
+   slang_atom varId;
+
+   if (oper->children[0].type == SLANG_OPER_BLOCK_NO_NEW_SCOPE) {
+      /* for (int i=0; ... */
+      slang_variable *var;
+
+      varId = oper->children[0].children[0].a_id;
+      var = _slang_variable_locate(oper->children[0].children[0].locals,
+                                   varId, GL_TRUE);
+      start = (GLint) var->initializer->literal[0];
+   }
+   else {
+      /* for (i=0; ... */
+      varId = oper->children[0].children[0].children[0].a_id;
+      start = (GLint) oper->children[0].children[0].children[1].literal[0];
+   }
+
+   end = (GLint) oper->children[1].children[0].children[1].literal[0];
+
+   for (iter = start; iter < end; iter++) {
+      slang_operation *body;
+
+      /* make a copy of the loop body */
+      body = slang_operation_new(1);
+      if (!body)
+         return NULL;
+
+      if (!slang_operation_copy(body, &oper->children[3]))
+         return NULL;
+
+      /* in body, replace instances of 'varId' with literal 'iter' */
+      {
+         slang_variable *oldVar;
+         slang_operation *newOper;
+
+         oldVar = _slang_variable_locate(oper->locals, varId, GL_TRUE);
+         if (!oldVar) {
+            /* undeclared loop variable */
+            slang_operation_delete(body);
+            return NULL;
+         }
+
+         newOper = slang_operation_new(1);
+         newOper->type = SLANG_OPER_LITERAL_INT;
+         newOper->literal_size = 1;
+         newOper->literal[0] = iter;
+
+         /* replace instances of the loop variable with newOper */
+         slang_substitute(A, body, 1, &oldVar, &newOper, GL_FALSE);
+      }
+
+      /* do IR codegen for body */
+      n = _slang_gen_operation(A, body);
+      root = new_seq(root, n);
+
+      slang_operation_delete(body);
+   }
+
+   return root;
+}
+
+
+/**
+ * Generate IR for a for-loop.  Unrolling will be done when possible.
  */
 static slang_ir_node *
 _slang_gen_for(slang_assemble_ctx * A, const slang_operation *oper)
 {
-   /*
-    * init code (child[0])
-    * LOOP:
-    *    BREAK if !expr (child[1])
-    *    body code (child[3])
-    *    tail code:
-    *       incr code (child[2])   // XXX continue here
-    */
-   slang_ir_node *prevLoop, *loop, *cond, *breakIf, *body, *init, *incr;
+   GLboolean unroll = _slang_can_unroll_for_loop(A, oper);
 
-   init = _slang_gen_operation(A, &oper->children[0]);
-   loop = new_loop(NULL);
+   if (unroll) {
+      slang_ir_node *code = _slang_unroll_for_loop(A, oper);
+      if (code)
+         return code;
+   }
 
-   /* save old, push new loop */
-   prevLoop = A->CurLoop;
-   A->CurLoop = loop;
+   /* conventional for-loop code generation */
+   {
+      /*
+       * init code (child[0])
+       * LOOP:
+       *    BREAK if !expr (child[1])
+       *    body code (child[3])
+       *    tail code:
+       *       incr code (child[2])   // XXX continue here
+       */
+      slang_ir_node *prevLoop, *loop, *cond, *breakIf, *body, *init, *incr;
+      init = _slang_gen_operation(A, &oper->children[0]);
+      loop = new_loop(NULL);
 
-   cond = new_cond(new_not(_slang_gen_operation(A, &oper->children[1])));
-   breakIf = new_break_if_true(A->CurLoop, cond);
-   body = _slang_gen_operation(A, &oper->children[3]);
-   incr = _slang_gen_operation(A, &oper->children[2]);
+      /* save old, push new loop */
+      prevLoop = A->CurLoop;
+      A->CurLoop = loop;
 
-   loop->Children[0] = new_seq(breakIf, body);
-   loop->Children[1] = incr;  /* tail code */
+      cond = new_cond(new_not(_slang_gen_operation(A, &oper->children[1])));
+      breakIf = new_break_if_true(A->CurLoop, cond);
+      body = _slang_gen_operation(A, &oper->children[3]);
+      incr = _slang_gen_operation(A, &oper->children[2]);
 
-   /* pop loop, restore prev */
-   A->CurLoop = prevLoop;
+      loop->Children[0] = new_seq(breakIf, body);
+      loop->Children[1] = incr;  /* tail code */
+
+      /* pop loop, restore prev */
+      A->CurLoop = prevLoop;
 
-   return new_seq(init, loop);
+      return new_seq(init, loop);
+   }
 }
 
 
@@ -3093,6 +3345,7 @@ _slang_gen_return(slang_assemble_ctx * A, slang_operation *oper)
 }
 
 
+#if 0
 /**
  * Determine if the given operation/expression is const-valued.
  */
@@ -3116,6 +3369,7 @@ _slang_is_constant_expr(const slang_operation *oper)
       return GL_TRUE;
    }
 }
+#endif
 
 
 /**
@@ -3418,7 +3672,7 @@ _slang_gen_assignment(slang_assemble_ctx * A, slang_operation *oper)
       if (lhs && rhs) {
          /* convert lhs swizzle into writemask */
          const GLuint swizzle = root_swizzle(lhs->Store);
-         GLuint writemask, newSwizzle;
+         GLuint writemask, newSwizzle = 0x0;
          if (!swizzle_to_writemask(A, swizzle, &writemask, &newSwizzle)) {
             /* Non-simple writemask, need to swizzle right hand side in
              * order to put components into the right place.
@@ -3560,8 +3814,16 @@ _slang_gen_array_element(slang_assemble_ctx * A, slang_operation *oper)
       index = (GLint) oper->children[1].literal[0];
       if (oper->children[1].type != SLANG_OPER_LITERAL_INT ||
           index >= (GLint) max) {
+#if 0
          slang_info_log_error(A->log, "Invalid array index for vector type");
+         printf("type = %d\n", oper->children[1].type);
+         printf("index = %d, max = %d\n", index, max);
+         printf("array = %s\n", (char*)oper->children[0].a_id);
+         printf("index = %s\n", (char*)oper->children[1].a_id);
          return NULL;
+#else
+         index = 0;
+#endif
       }
 
       n = _slang_gen_operation(A, &oper->children[0]);
@@ -3987,6 +4249,21 @@ _slang_gen_operation(slang_assemble_ctx * A, slang_operation *oper)
 
 
 /**
+ * Check if the given type specifier is a rectangular texture sampler.
+ */
+static GLboolean
+is_rect_sampler_spec(const slang_type_specifier *spec)
+{
+   while (spec->_array) {
+      spec = spec->_array;
+   }
+   return spec->type == SLANG_SPEC_SAMPLER2DRECT ||
+          spec->type == SLANG_SPEC_SAMPLER2DRECTSHADOW;
+}
+
+
+
+/**
  * Called by compiler when a global variable has been parsed/compiled.
  * Here we examine the variable's type to determine what kind of register
  * storage will be used.
@@ -4009,10 +4286,14 @@ _slang_codegen_global_variable(slang_assemble_ctx *A, slang_variable *var,
    slang_ir_storage *store = NULL;
    int dbg = 0;
    const GLenum datatype = _slang_gltype_from_specifier(&var->type.specifier);
-   const GLint texIndex = sampler_to_texture_index(var->type.specifier.type);
    const GLint size = _slang_sizeof_type_specifier(&var->type.specifier);
    const GLint arrayLen = _slang_array_length(var);
    const GLint totalSize = _slang_array_size(size, arrayLen);
+   GLint texIndex = sampler_to_texture_index(var->type.specifier.type);
+
+   /* check for sampler2D arrays */
+   if (texIndex == -1 && var->type.specifier._array)
+      texIndex = sampler_to_texture_index(var->type.specifier._array->type);
 
    if (texIndex != -1) {
       /* This is a texture sampler variable...
@@ -4026,15 +4307,32 @@ _slang_codegen_global_variable(slang_assemble_ctx *A, slang_variable *var,
       }
 #if FEATURE_es2_glsl /* XXX should use FEATURE_texture_rect */
       /* disallow rect samplers */
-      if (var->type.specifier.type == SLANG_SPEC_SAMPLER2DRECT ||
-          var->type.specifier.type == SLANG_SPEC_SAMPLER2DRECTSHADOW) {
+      if (is_rect_sampler_spec(&var->type.specifier)) {
          slang_info_log_error(A->log, "invalid sampler type for '%s'", varName);
          return GL_FALSE;
       }
+#else
+      (void) is_rect_sampler_spec; /* silence warning */
 #endif
       {
          GLint sampNum = _mesa_add_sampler(prog->Parameters, varName, datatype);
-         store = _slang_new_ir_storage(PROGRAM_SAMPLER, sampNum, texIndex);
+         store = _slang_new_ir_storage_sampler(sampNum, texIndex, totalSize);
+
+         /* If we have a sampler array, then we need to allocate the 
+	  * additional samplers to ensure we don't allocate them elsewhere.
+	  * We can't directly use _mesa_add_sampler() as that checks the
+	  * varName and gets a match, so we call _mesa_add_parameter()
+	  * directly and use the last sampler number from the call above.
+	  */
+	 if (arrayLen > 0) {
+	    GLint a = arrayLen - 1;
+	    GLint i;
+	    for (i = 0; i < a; i++) {
+               GLfloat value = (GLfloat)(i + sampNum + 1);
+               (void) _mesa_add_parameter(prog->Parameters, PROGRAM_SAMPLER,
+                                 varName, 1, datatype, &value, NULL, 0x0);
+	    }
+	 }
       }
       if (dbg) printf("SAMPLER ");
    }
@@ -4218,7 +4516,7 @@ _slang_codegen_global_variable(slang_assemble_ctx *A, slang_variable *var,
       n = _slang_gen_var_decl(A, var, var->initializer);
 
       /* emit GPU instructions */
-      success = _slang_emit_code(n, A->vartable, A->program, GL_FALSE, A->log);
+      success = _slang_emit_code(n, A->vartable, A->program, A->pragmas, GL_FALSE, A->log);
 
       _slang_free_ir_tree(n);
    }
@@ -4328,7 +4626,7 @@ _slang_codegen_function(slang_assemble_ctx * A, slang_function * fun)
 #endif
 
    /* Emit program instructions */
-   success = _slang_emit_code(n, A->vartable, A->program, GL_TRUE, A->log);
+   success = _slang_emit_code(n, A->vartable, A->program, A->pragmas, GL_TRUE, A->log);
    _slang_free_ir_tree(n);
 
    /* free codegen context */