Merge commit 'origin/master' into gallium-0.2

13 files changed, 687 insertions, 355 deletions

diff --git a/src/mesa/drivers/dri/i915/intel_tris.c b/src/mesa/drivers/dri/i915/intel_tris.c
index 06210035af5..797d6c58587 100644
--- a/src/mesa/drivers/dri/i915/intel_tris.c
+++ b/src/mesa/drivers/dri/i915/intel_tris.c
@@ -186,7 +186,7 @@ void intel_flush_prim(struct intel_context *intel)
       OUT_RELOC(vb_bo, I915_GEM_DOMAIN_VERTEX, 0,
 		offset | (intel->vertex_size << S0_VB_PITCH_SHIFT_830) |
 		S0_VB_ENABLE_830);
-      /* S1
+      /* S2
        * This is somewhat unfortunate -- VB width is tied up with
        * vertex format data that we've already uploaded through
        * _3DSTATE_VFT[01]_CMD.  We may want to replace emits of VFT state with
diff --git a/src/mesa/drivers/dri/intel/intel_decode.c b/src/mesa/drivers/dri/intel/intel_decode.c
index 9c105013c05..0e72ca08b22 100644
--- a/src/mesa/drivers/dri/intel/intel_decode.c
+++ b/src/mesa/drivers/dri/intel/intel_decode.c
@@ -308,6 +308,15 @@ decode_3d_1c(uint32_t *data, int count, uint32_t hw_offset, int *failures)
     case 0x10:
 	instr_out(data, hw_offset, 0, "3DSTATE_SCISSOR_ENABLE\n");
 	return 1;
+    case 0x01:
+	instr_out(data, hw_offset, 0, "3DSTATE_MAP_COORD_SET_I830\n");
+	return 1;
+    case 0x0a:
+	instr_out(data, hw_offset, 0, "3DSTATE_MAP_CUBE_I830\n");
+	return 1;
+    case 0x05:
+	instr_out(data, hw_offset, 0, "3DSTATE_MAP_TEX_STREAM_I830\n");
+	return 1;
     }
 
     instr_out(data, hw_offset, 0, "3D UNKNOWN\n");
@@ -316,33 +325,39 @@ decode_3d_1c(uint32_t *data, int count, uint32_t hw_offset, int *failures)
 }
 
 static int
-decode_3d_1d(uint32_t *data, int count, uint32_t hw_offset, int *failures)
+decode_3d_1d(uint32_t *data, int count, uint32_t hw_offset, int *failures, int i830)
 {
     unsigned int len, i, c, opcode, word, map, sampler, instr;
 
     struct {
 	uint32_t opcode;
+	int i830_only;
 	int min_len;
 	int max_len;
 	char *name;
     } opcodes_3d_1d[] = {
-	{ 0x8e, 3, 3, "3DSTATE_BUFFER_INFO" },
-	{ 0x86, 4, 4, "3DSTATE_CHROMA_KEY" },
-	{ 0x9c, 1, 1, "3DSTATE_CLEAR_PARAMETERS" },
-	{ 0x88, 2, 2, "3DSTATE_CONSTANT_BLEND_COLOR" },
-	{ 0x99, 2, 2, "3DSTATE_DEFAULT_DIFFUSE" },
-	{ 0x9a, 2, 2, "3DSTATE_DEFAULT_SPECULAR" },
-	{ 0x98, 2, 2, "3DSTATE_DEFAULT_Z" },
-	{ 0x97, 2, 2, "3DSTATE_DEPTH_OFFSET_SCALE" },
-	{ 0x85, 2, 2, "3DSTATE_DEST_BUFFER_VARIABLES" },
-	{ 0x80, 5, 5, "3DSTATE_DRAWING_RECTANGLE" },
-	{ 0x8e, 3, 3, "3DSTATE_BUFFER_INFO" },
-	{ 0x9d, 65, 65, "3DSTATE_FILTER_COEFFICIENTS_4X4" },
-	{ 0x9e, 4, 4, "3DSTATE_MONO_FILTER" },
-	{ 0x89, 4, 4, "3DSTATE_FOG_MODE" },
-	{ 0x8f, 2, 16, "3DSTATE_MAP_PALLETE_LOAD_32" },
-	{ 0x81, 3, 3, "3DSTATE_SCISSOR_RECTANGLE" },
-	{ 0x83, 2, 2, "3DSTATE_SPAN_STIPPLE" },
+	{ 0x8e, 0, 3, 3, "3DSTATE_BUFFER_INFO" },
+	{ 0x86, 0, 4, 4, "3DSTATE_CHROMA_KEY" },
+	{ 0x9c, 0, 1, 1, "3DSTATE_CLEAR_PARAMETERS" },
+	{ 0x88, 0, 2, 2, "3DSTATE_CONSTANT_BLEND_COLOR" },
+	{ 0x99, 0, 2, 2, "3DSTATE_DEFAULT_DIFFUSE" },
+	{ 0x9a, 0, 2, 2, "3DSTATE_DEFAULT_SPECULAR" },
+	{ 0x98, 0, 2, 2, "3DSTATE_DEFAULT_Z" },
+	{ 0x97, 0, 2, 2, "3DSTATE_DEPTH_OFFSET_SCALE" },
+	{ 0x85, 0, 2, 2, "3DSTATE_DEST_BUFFER_VARIABLES" },
+	{ 0x80, 0, 5, 5, "3DSTATE_DRAWING_RECTANGLE" },
+	{ 0x8e, 0, 3, 3, "3DSTATE_BUFFER_INFO" },
+	{ 0x9d, 0, 65, 65, "3DSTATE_FILTER_COEFFICIENTS_4X4" },
+	{ 0x9e, 0, 4, 4, "3DSTATE_MONO_FILTER" },
+	{ 0x89, 0, 4, 4, "3DSTATE_FOG_MODE" },
+	{ 0x8f, 0, 2, 16, "3DSTATE_MAP_PALLETE_LOAD_32" },
+	{ 0x81, 0, 3, 3, "3DSTATE_SCISSOR_RECTANGLE" },
+	{ 0x83, 0, 2, 2, "3DSTATE_SPAN_STIPPLE" },
+	{ 0x8c, 1, 2, 2, "3DSTATE_MAP_COORD_TRANSFORM_I830" },
+	{ 0x8b, 1, 2, 2, "3DSTATE_MAP_VERTEX_TRANSFORM_I830" },
+	{ 0x8d, 1, 3, 3, "3DSTATE_W_STATE_I830" },
+	{ 0x01, 1, 2, 2, "3DSTATE_COLOR_FACTOR_I830" },
+	{ 0x02, 1, 2, 2, "3DSTATE_MAP_COORD_SETBIND_I830" },
     };
 
     switch ((data[0] & 0x00ff0000) >> 16) {
@@ -488,6 +503,8 @@ decode_3d_1d(uint32_t *data, int count, uint32_t hw_offset, int *failures)
 	}
 	return len;
     case 0x01:
+	if (i830)
+	    break;
 	instr_out(data, hw_offset, 0, "3DSTATE_SAMPLER_STATE\n");
 	len = (data[0] & 0x0000003f) + 2;
 	i = 1;
@@ -513,6 +530,9 @@ decode_3d_1d(uint32_t *data, int count, uint32_t hw_offset, int *failures)
     for (opcode = 0; opcode < sizeof(opcodes_3d_1d) / sizeof(opcodes_3d_1d[0]);
 	 opcode++)
     {
+	if (opcodes_3d_1d[opcode].i830_only && !i830)
+	    continue;
+
 	if (((data[0] & 0x00ff0000) >> 16) == opcodes_3d_1d[opcode].opcode) {
 	    len = 1;
 
@@ -756,7 +776,7 @@ decode_3d(uint32_t *data, int count, uint32_t hw_offset, int *failures)
     case 0x1f:
 	return decode_3d_primitive(data, count, hw_offset, failures);
     case 0x1d:
-	return decode_3d_1d(data, count, hw_offset, failures);
+	return decode_3d_1d(data, count, hw_offset, failures, 0);
     case 0x1c:
 	return decode_3d_1c(data, count, hw_offset, failures);
     }
@@ -999,6 +1019,73 @@ decode_3d_965(uint32_t *data, int count, uint32_t hw_offset, int *failures)
     return 1;
 }
 
+static int
+decode_3d_i830(uint32_t *data, int count, uint32_t hw_offset, int *failures)
+{
+    unsigned int opcode;
+
+    struct {
+	uint32_t opcode;
+	int min_len;
+	int max_len;
+	char *name;
+    } opcodes_3d[] = {
+	{ 0x02, 1, 1, "3DSTATE_MODES_3" },
+	{ 0x03, 1, 1, "3DSTATE_ENABLES_1"},
+	{ 0x04, 1, 1, "3DSTATE_ENABLES_2"},
+	{ 0x05, 1, 1, "3DSTATE_VFT0"},
+	{ 0x06, 1, 1, "3DSTATE_AA"},
+	{ 0x07, 1, 1, "3DSTATE_RASTERIZATION_RULES" },
+	{ 0x08, 1, 1, "3DSTATE_MODES_1" },
+	{ 0x09, 1, 1, "3DSTATE_STENCIL_TEST" },
+	{ 0x0a, 1, 1, "3DSTATE_VFT1"},
+	{ 0x0b, 1, 1, "3DSTATE_INDPT_ALPHA_BLEND" },
+	{ 0x0c, 1, 1, "3DSTATE_MODES_5" },
+	{ 0x0d, 1, 1, "3DSTATE_MAP_BLEND_OP" },
+	{ 0x0e, 1, 1, "3DSTATE_MAP_BLEND_ARG" },
+	{ 0x0f, 1, 1, "3DSTATE_MODES_2" },
+	{ 0x15, 1, 1, "3DSTATE_FOG_COLOR" },
+	{ 0x16, 1, 1, "3DSTATE_MODES_4" },
+    };
+
+    switch ((data[0] & 0x1f000000) >> 24) {
+    case 0x1f:
+	return decode_3d_primitive(data, count, hw_offset, failures);
+    case 0x1d:
+	return decode_3d_1d(data, count, hw_offset, failures, 1);
+    case 0x1c:
+	return decode_3d_1c(data, count, hw_offset, failures);
+    }
+
+    for (opcode = 0; opcode < sizeof(opcodes_3d) / sizeof(opcodes_3d[0]);
+	 opcode++) {
+	if ((data[0] & 0x1f000000) >> 24 == opcodes_3d[opcode].opcode) {
+	    unsigned int len = 1, i;
+
+	    instr_out(data, hw_offset, 0, "%s\n", opcodes_3d[opcode].name);
+	    if (opcodes_3d[opcode].max_len > 1) {
+		len = (data[0] & 0xff) + 2;
+		if (len < opcodes_3d[opcode].min_len ||
+		    len > opcodes_3d[opcode].max_len)
+		{
+		    fprintf(out, "Bad count in %s\n", opcodes_3d[opcode].name);
+		}
+	    }
+
+	    for (i = 1; i < len; i++) {
+		if (i >= count)
+		    BUFFER_FAIL(count, len, opcodes_3d[opcode].name);
+		instr_out(data, hw_offset, i, "dword %d\n", i);
+	    }
+	    return len;
+	}
+    }
+
+    instr_out(data, hw_offset, 0, "3D UNKNOWN\n");
+    (*failures)++;
+    return 1;
+}
+
 /**
  * Decodes an i830-i915 batch buffer, writing the output to stdout.
  *
@@ -1028,9 +1115,12 @@ intel_decode(uint32_t *data, int count, uint32_t hw_offset, uint32_t devid)
 	    if (IS_965(devid)) {
 		index += decode_3d_965(data + index, count - index,
 				       hw_offset + index * 4, &failures);
-	    } else {
+	    } else if (IS_9XX(devid)) {
 		index += decode_3d(data + index, count - index,
 				   hw_offset + index * 4, &failures);
+	    } else {
+		index += decode_3d_i830(data + index, count - index,
+					hw_offset + index * 4, &failures);
 	    }
 	    break;
 	default:
diff --git a/src/mesa/drivers/dri/intel/intel_reg.h b/src/mesa/drivers/dri/intel/intel_reg.h
index 81a7386e429..68d8a05b836 100644
--- a/src/mesa/drivers/dri/intel/intel_reg.h
+++ b/src/mesa/drivers/dri/intel/intel_reg.h
@@ -80,7 +80,7 @@
  */
 #define S0_VB_OFFSET_MASK_830		0xffffff8
 #define S0_VB_PITCH_SHIFT_830		1
-#define S0_VB_ENABLE_830		0
+#define S0_VB_ENABLE_830		(1<<0)
 /** @} */
 
 #define S1_VERTEX_WIDTH_SHIFT          24
@@ -100,8 +100,8 @@
 #define S2_TEXCOORD_FMT(unit, type)    ((type)<<(unit*4))
 #define S2_TEXCOORD_NONE               (~0)
 #define S2_TEX_COUNT_SHIFT_830		12
-#define S2_VERTEX_0_WIDTH_SHIFT_830	0
-#define S2_VERTEX_1_WIDTH_SHIFT_830	6
+#define S2_VERTEX_1_WIDTH_SHIFT_830	0
+#define S2_VERTEX_0_WIDTH_SHIFT_830	6
 /* S3 not interesting */
 
 #define S4_POINT_WIDTH_SHIFT           23
diff --git a/src/mesa/main/image.c b/src/mesa/main/image.c
index bd38e0c3d8c..d3282d42e99 100644
--- a/src/mesa/main/image.c
+++ b/src/mesa/main/image.c
@@ -1689,7 +1689,7 @@ _mesa_pack_rgba_span_float(GLcontext *ctx, GLuint n, GLfloat rgba[][4],
 
    if (dstFormat == GL_LUMINANCE || dstFormat == GL_LUMINANCE_ALPHA) {
       /* compute luminance values */
-      if (dstType != GL_FLOAT || ctx->Color.ClampReadColor == GL_TRUE) {
+      if (transferOps & IMAGE_CLAMP_BIT) {
          for (i = 0; i < n; i++) {
             GLfloat sum = rgba[i][RCOMP] + rgba[i][GCOMP] + rgba[i][BCOMP];
             luminance[i] = CLAMP(sum, 0.0F, 1.0F);
diff --git a/src/mesa/main/texenvprogram.c b/src/mesa/main/texenvprogram.c
index 237ddaa6838..3cb103f51f8 100644
--- a/src/mesa/main/texenvprogram.c
+++ b/src/mesa/main/texenvprogram.c
@@ -297,10 +297,13 @@ static void make_state_key( GLcontext *ctx,  struct state_key *key )
 
    for (i=0;i<MAX_TEXTURE_UNITS;i++) {
       const struct gl_texture_unit *texUnit = &ctx->Texture.Unit[i];
+      GLenum format;
 
       if (!texUnit->_ReallyEnabled || !texUnit->Enabled)
          continue;
 
+      format = texUnit->_Current->Image[0][texUnit->_Current->BaseLevel]->_BaseFormat;
+
       key->unit[i].enabled = 1;
       key->enabled_units |= (1<<i);
       key->nr_enabled_units = i+1;
@@ -308,7 +311,9 @@ static void make_state_key( GLcontext *ctx,  struct state_key *key )
 
       key->unit[i].source_index = 
 	 translate_tex_src_bit(texUnit->_ReallyEnabled);		
-      key->unit[i].shadow = texUnit->_Current->CompareMode == GL_COMPARE_R_TO_TEXTURE;
+      key->unit[i].shadow = ((texUnit->_Current->CompareMode == GL_COMPARE_R_TO_TEXTURE) && 
+                             ((format == GL_DEPTH_COMPONENT) || 
+                              (format == GL_DEPTH_STENCIL_EXT)));
 
       key->unit[i].NumArgsRGB = texUnit->_CurrentCombine->_NumArgsRGB;
       key->unit[i].NumArgsA = texUnit->_CurrentCombine->_NumArgsA;
diff --git a/src/mesa/main/texstore.c b/src/mesa/main/texstore.c
index c1fafbae30b..7278180a5cc 100644
--- a/src/mesa/main/texstore.c
+++ b/src/mesa/main/texstore.c
@@ -3719,6 +3719,14 @@ _mesa_get_teximage(GLcontext *ctx, GLenum target, GLint level,
                /* general case:  convert row to RGBA format */
                GLfloat rgba[MAX_WIDTH][4];
                GLint col;
+               GLbitfield transferOps = 0x0;
+
+               if (type == GL_FLOAT && 
+                   ((ctx->Color.ClampReadColor == GL_TRUE) ||
+                    (ctx->Color.ClampReadColor == GL_FIXED_ONLY_ARB &&
+                     texImage->TexFormat->DataType != GL_FLOAT)))
+                  transferOps |= IMAGE_CLAMP_BIT;
+
                for (col = 0; col < width; col++) {
                   (*texImage->FetchTexelf)(texImage, col, row, img, rgba[col]);
                   if (texImage->TexFormat->BaseFormat == GL_ALPHA) {
@@ -3743,7 +3751,7 @@ _mesa_get_teximage(GLcontext *ctx, GLenum target, GLint level,
                }
                _mesa_pack_rgba_span_float(ctx, width, (GLfloat (*)[4]) rgba,
                                           format, type, dest,
-                                          &ctx->Pack, 0x0 /*image xfer ops*/);
+                                          &ctx->Pack, transferOps /*image xfer ops*/);
             } /* format */
          } /* row */
       } /* img */
diff --git a/src/mesa/shader/slang/library/slang_shader.syn b/src/mesa/shader/slang/library/slang_shader.syn
index cbfbda651af..47245e0d878 100644
--- a/src/mesa/shader/slang/library/slang_shader.syn
+++ b/src/mesa/shader/slang/library/slang_shader.syn
@@ -837,10 +837,8 @@ parameter_type_specifier_3
  * <init_declarator_list> ::= <single_declaration>
  *                          | <init_declarator_list> "," <identifier>
  *                          | <init_declarator_list> "," <identifier> "[" "]"
- *                          | <init_declarator_list> "," <identifier> "["
- *                            <constant_expression> "]"
- *                          | <init_declarator_list> "," <identifier> "="
- *                            <initializer>
+ *                          | <init_declarator_list> "," <identifier> "[" <constant_expression> "]"
+ *                          | <init_declarator_list> "," <identifier> "=" <initializer>
  */
 init_declarator_list
     single_declaration .and .loop init_declarator_list_1 .emit DECLARATOR_NEXT .and
@@ -860,10 +858,9 @@ init_declarator_list_5
  * <single_declaration> ::= <fully_specified_type>
  *                        | <fully_specified_type> <identifier>
  *                        | <fully_specified_type> <identifier> "[" "]"
- *                        | <fully_specified_type> <identifier> "["
- *                          <constant_expression> "]"
+ *                        | <fully_specified_type> <identifier> "[" <constant_expression> "]"
  *                        | <fully_specified_type> <identifier> "=" <initializer>
-*/
+ */
 single_declaration
     single_declaration_nospace .or single_declaration_space;
 single_declaration_space
@@ -924,12 +921,12 @@ centroid_qualifier
 
 
 /*
- *  <type_qualifier> ::= "const"
- *                     | "attribute" // Vertex only.
- *                     | "varying"
- *                     | "uniform"
- *                     | "__fixed_output"
- *                     | "__fixed_input"
+ * <type_qualifier> ::= "const"
+ *                    | "attribute" // Vertex only.
+ *                    | "varying"
+ *                    | "uniform"
+ *                    | "__fixed_output"
+ *                    | "__fixed_input"
  *
  * note: this is an extension to the standard language specification,
  * normally slang disallows __fixed_output and __fixed_input type qualifiers
@@ -1013,7 +1010,7 @@ type_specifier_nospace
 
 /*
  * <struct_specifier> ::= "struct" <identifier> "{" <struct_declaration_list> "}"
- *                       | "struct" "{" <struct_declaration_list> "}"
+ *                      | "struct" "{" <struct_declaration_list> "}"
  */
 struct_specifier
     "struct" .and struct_specifier_1 .and optional_space .and lbrace .error LBRACE_EXPECTED .and
@@ -1025,7 +1022,7 @@ struct_specifier_2
 
 /*
  * <struct_declaration_list> ::= <struct_declaration>
- *                              | <struct_declaration_list> <struct_declaration>
+ *                             | <struct_declaration_list> <struct_declaration>
  */
 struct_declaration_list
     struct_declaration .and .loop struct_declaration .emit FIELD_NEXT;
@@ -1042,7 +1039,7 @@ struct_declaration_nospace
 
 /*
  * <struct_declarator_list> ::= <struct_declarator>
- *                             | <struct_declarator_list> "," <struct_declarator>
+ *                            | <struct_declarator_list> "," <struct_declarator>
  */
 struct_declarator_list
     struct_declarator .and .loop struct_declarator_list_1 .emit FIELD_NEXT;
diff --git a/src/mesa/shader/slang/slang_codegen.c b/src/mesa/shader/slang/slang_codegen.c
index d23ae4d8cb5..851335d9c58 100644
--- a/src/mesa/shader/slang/slang_codegen.c
+++ b/src/mesa/shader/slang/slang_codegen.c
@@ -1477,6 +1477,8 @@ _slang_simple_writemask(GLuint writemask, GLuint swizzle)
  * Convert the given swizzle into a writemask.  In some cases this
  * is trivial, in other cases, we'll need to also swizzle the right
  * hand side to put components in the right places.
+ * See comment above for more info.
+ * XXX this function could be simplified and should probably be renamed.
  * \param swizzle  the incoming swizzle
  * \param writemaskOut  returns the writemask
  * \param swizzleOut  swizzle to apply to the right-hand-side
@@ -3114,7 +3116,7 @@ _slang_gen_struct_field(slang_assemble_ctx * A, slang_operation *oper)
       /* oper->a_id is the field name */
       slang_ir_node *base, *n;
       slang_typeinfo field_ti;
-      GLint fieldSize, fieldOffset = -1, swz;
+      GLint fieldSize, fieldOffset = -1;
 
       /* type of field */
       slang_typeinfo_construct(&field_ti);
@@ -3147,22 +3149,12 @@ _slang_gen_struct_field(slang_assemble_ctx * A, slang_operation *oper)
       if (!n)
          return NULL;
 
-
-      /* setup the storage info for this node */
-      swz = fieldOffset % 4;
-
       n->Field = (char *) oper->a_id;
-      n->Store = _slang_new_ir_storage_relative(fieldOffset / 4,
-                                                fieldSize,
-                                                base->Store);
-      if (fieldSize == 1)
-         n->Store->Swizzle = MAKE_SWIZZLE4(swz, swz, swz, swz);
-      else if (fieldSize == 2)
-         n->Store->Swizzle = MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_Y,
-                                           SWIZZLE_NIL, SWIZZLE_NIL);
-      else if (fieldSize == 3)
-         n->Store->Swizzle = MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_Y,
-                                           SWIZZLE_Z, SWIZZLE_NIL);
+
+      /* Store the field's offset in storage->Index */
+      n->Store = _slang_new_ir_storage(base->Store->File,
+                                       fieldOffset,
+                                       fieldSize);
 
       return n;
    }
@@ -3255,12 +3247,12 @@ _slang_gen_array_element(slang_assemble_ctx * A, slang_operation *oper)
          }
 
          elem = new_node2(IR_ELEMENT, array, index);
-         elem->Store = _slang_new_ir_storage_relative(constIndex,
-                                                      elemSize,
-                                                      array->Store);
 
-         assert(elem->Store->Parent);
-         /* XXX try to do some array bounds checking here */
+         /* The storage info here will be updated during code emit */
+         elem->Store = _slang_new_ir_storage(array->Store->File,
+                                             array->Store->Index,
+                                             elemSize);
+
          return elem;
       }
       else {
diff --git a/src/mesa/shader/slang/slang_compile.c b/src/mesa/shader/slang/slang_compile.c
index 72b83c04224..07c40eaa324 100644
--- a/src/mesa/shader/slang/slang_compile.c
+++ b/src/mesa/shader/slang/slang_compile.c
@@ -167,6 +167,16 @@ typedef struct slang_output_ctx_
 
 /* _slang_compile() */
 
+
+/* Debugging aid, print file/line where parsing error is detected */
+#define RETURN0 \
+   do { \
+      if (0) \
+         printf("slang error at %s:%d\n", __FILE__, __LINE__); \
+      return 0; \
+   } while (0)
+
+
 static void
 parse_identifier_str(slang_parse_ctx * C, char **id)
 {
@@ -223,7 +233,7 @@ parse_float(slang_parse_ctx * C, float *number)
                                   _mesa_strlen(exponent) + 3) * sizeof(char));
    if (whole == NULL) {
       slang_info_log_memory(C->L);
-      return 0;
+      RETURN0;
    }
 
    slang_string_copy(whole, integral);
@@ -247,7 +257,7 @@ check_revision(slang_parse_ctx * C)
 {
    if (*C->I != REVISION) {
       slang_info_log_error(C->L, "Internal compiler error.");
-      return 0;
+      RETURN0;
    }
    C->I++;
    return 1;
@@ -387,23 +397,23 @@ parse_struct_field(slang_parse_ctx * C, slang_output_ctx * O,
 
    o.structs = st->structs;
    if (!parse_type_specifier(C, &o, sp))
-      return 0;
+      RETURN0;
 
    do {
       slang_atom a_name;
       slang_variable *var = slang_variable_scope_grow(st->fields);
       if (!var) {
          slang_info_log_memory(C->L);
-         return 0;
+         RETURN0;
       }
       a_name = parse_identifier(C);
       if (_slang_locate_variable(st->fields, a_name, GL_FALSE)) {
          slang_info_log_error(C->L, "duplicate field '%s'", (char *) a_name);
-         return 0;
+         RETURN0;
       }
 
       if (!parse_struct_field_var(C, &o, var, a_name, sp))
-         return 0;
+         RETURN0;
    }
    while (*C->I++ != FIELD_NONE);
 
@@ -419,26 +429,26 @@ parse_struct(slang_parse_ctx * C, slang_output_ctx * O, slang_struct ** st)
    /* parse struct name (if any) and make sure it is unique in current scope */
    a_name = parse_identifier(C);
    if (a_name == SLANG_ATOM_NULL)
-      return 0;
+      RETURN0;
 
    name = slang_atom_pool_id(C->atoms, a_name);
    if (name[0] != '\0'
        && slang_struct_scope_find(O->structs, a_name, 0) != NULL) {
       slang_info_log_error(C->L, "%s: duplicate type name.", name);
-      return 0;
+      RETURN0;
    }
 
    /* set-up a new struct */
    *st = (slang_struct *) _slang_alloc(sizeof(slang_struct));
    if (*st == NULL) {
       slang_info_log_memory(C->L);
-      return 0;
+      RETURN0;
    }
    if (!slang_struct_construct(*st)) {
       _slang_free(*st);
       *st = NULL;
       slang_info_log_memory(C->L);
-      return 0;
+      RETURN0;
    }
    (**st).a_name = a_name;
    (**st).structs->outer_scope = O->structs;
@@ -450,7 +460,7 @@ parse_struct(slang_parse_ctx * C, slang_output_ctx * O, slang_struct ** st)
       slang_type_specifier_ctr(&sp);
       if (!parse_struct_field(C, O, *st, &sp)) {
          slang_type_specifier_dtr(&sp);
-         return 0;
+         RETURN0;
       }
       slang_type_specifier_dtr(&sp);
    }
@@ -468,14 +478,14 @@ parse_struct(slang_parse_ctx * C, slang_output_ctx * O, slang_struct ** st)
                                          * sizeof(slang_struct));
       if (O->structs->structs == NULL) {
          slang_info_log_memory(C->L);
-         return 0;
+         RETURN0;
       }
       s = &O->structs->structs[O->structs->num_structs];
       if (!slang_struct_construct(s))
-         return 0;
+         RETURN0;
       O->structs->num_structs++;
       if (!slang_struct_copy(s, *st))
-         return 0;
+         RETURN0;
    }
 
    return 1;
@@ -498,7 +508,7 @@ parse_type_variant(slang_parse_ctx * C, slang_type_variant *variant)
       *variant = SLANG_INVARIANT;
       return 1;
    default:
-      return 0;
+      RETURN0;
    }
 }
 
@@ -519,7 +529,7 @@ parse_type_centroid(slang_parse_ctx * C, slang_type_centroid *centroid)
       *centroid = SLANG_CENTROID;
       return 1;
    default:
-      return 0;
+      RETURN0;
    }
 }
 
@@ -560,7 +570,7 @@ parse_type_qualifier(slang_parse_ctx * C, slang_type_qualifier * qual)
       *qual = SLANG_QUAL_FIXEDINPUT;
       break;
    default:
-      return 0;
+      RETURN0;
    }
    return 1;
 }
@@ -698,7 +708,7 @@ parse_type_specifier(slang_parse_ctx * C, slang_output_ctx * O,
    case TYPE_SPECIFIER_STRUCT:
       spec->type = SLANG_SPEC_STRUCT;
       if (!parse_struct(C, O, &spec->_struct))
-         return 0;
+         RETURN0;
       break;
    case TYPE_SPECIFIER_TYPENAME:
       spec->type = SLANG_SPEC_STRUCT;
@@ -708,31 +718,31 @@ parse_type_specifier(slang_parse_ctx * C, slang_output_ctx * O,
 
          a_name = parse_identifier(C);
          if (a_name == NULL)
-            return 0;
+            RETURN0;
 
          stru = slang_struct_scope_find(O->structs, a_name, 1);
          if (stru == NULL) {
             slang_info_log_error(C->L, "undeclared type name '%s'",
                                  slang_atom_pool_id(C->atoms, a_name));
-            return 0;
+            RETURN0;
          }
 
          spec->_struct = (slang_struct *) _slang_alloc(sizeof(slang_struct));
          if (spec->_struct == NULL) {
             slang_info_log_memory(C->L);
-            return 0;
+            RETURN0;
          }
          if (!slang_struct_construct(spec->_struct)) {
             _slang_free(spec->_struct);
             spec->_struct = NULL;
-            return 0;
+            RETURN0;
          }
          if (!slang_struct_copy(spec->_struct, stru))
-            return 0;
+            RETURN0;
       }
       break;
    default:
-      return 0;
+      RETURN0;
    }
    return 1;
 }
@@ -762,7 +772,7 @@ parse_type_precision(slang_parse_ctx *C,
       *precision = SLANG_PREC_HIGH;
       return 1;
    default:
-      return 0;
+      RETURN0;
    }
 }
 
@@ -771,36 +781,36 @@ parse_fully_specified_type(slang_parse_ctx * C, slang_output_ctx * O,
                            slang_fully_specified_type * type)
 {
    if (!parse_type_variant(C, &type->variant))
-      return 0;
+      RETURN0;
   
    if (!parse_type_centroid(C, &type->centroid))
-      return 0;
+      RETURN0;
 
    if (!parse_type_qualifier(C, &type->qualifier))
-      return 0;
+      RETURN0;
 
    if (!parse_type_precision(C, &type->precision))
-      return 0;
+      RETURN0;
 
    if (!parse_type_specifier(C, O, &type->specifier))
-      return 0;
+      RETURN0;
 
    if (!O->allow_invariant && type->variant == SLANG_INVARIANT) {
       slang_info_log_error(C->L,
          "'invariant' keyword not allowed (perhaps set #version 120)");
-      return 0;
+      RETURN0;
    }
 
    if (!O->allow_centroid && type->centroid == SLANG_CENTROID) {
       slang_info_log_error(C->L,
          "'centroid' keyword not allowed (perhaps set #version 120)");
-      return 0;
+      RETURN0;
    }
    else if (type->centroid == SLANG_CENTROID &&
             type->qualifier != SLANG_QUAL_VARYING) {
       slang_info_log_error(C->L,
          "'centroid' keyword only allowed for varying vars");
-      return 0;
+      RETURN0;
    }
 
 
@@ -809,7 +819,7 @@ parse_fully_specified_type(slang_parse_ctx * C, slang_output_ctx * O,
        type->variant == SLANG_INVARIANT) {
       slang_info_log_error(C->L,
                            "invariant qualifer only allowed for varying vars");
-      return 0;
+      RETURN0;
    }
    */
 
@@ -824,7 +834,7 @@ parse_fully_specified_type(slang_parse_ctx * C, slang_output_ctx * O,
       /* only default is allowed */
       if (type->precision != SLANG_PREC_DEFAULT) {
          slang_info_log_error(C->L, "precision qualifiers not allowed");
-         return 0;
+         RETURN0;
       }
    }
 
@@ -935,7 +945,7 @@ parse_statement(slang_parse_ctx * C, slang_output_ctx * O,
       oper->type = SLANG_OPER_BLOCK_NO_NEW_SCOPE;
       while (*C->I != OP_END)
          if (!parse_child_operation(C, O, oper, 1))
-            return 0;
+            RETURN0;
       C->I++;
       break;
    case OP_BLOCK_BEGIN_NEW_SCOPE:
@@ -947,7 +957,7 @@ parse_statement(slang_parse_ctx * C, slang_output_ctx * O,
          o.vars = oper->locals;
          while (*C->I != OP_END)
             if (!parse_child_operation(C, &o, oper, 1))
-               return 0;
+               RETURN0;
          C->I++;
       }
       break;
@@ -963,7 +973,7 @@ parse_statement(slang_parse_ctx * C, slang_output_ctx * O,
           * than one declarators
           */
          if (!parse_declaration(C, O))
-            return 0;
+            RETURN0;
          if (first_var < O->vars->num_variables) {
             const unsigned int num_vars = O->vars->num_variables - first_var;
             unsigned int i;
@@ -972,7 +982,7 @@ parse_statement(slang_parse_ctx * C, slang_output_ctx * O,
             oper->children = slang_operation_new(num_vars);
             if (oper->children == NULL) {
                slang_info_log_memory(C->L);
-               return 0;
+               RETURN0;
             }
             for (i = first_var; i < O->vars->num_variables; i++) {
                slang_operation *o = &oper->children[i - first_var];
@@ -983,7 +993,7 @@ parse_statement(slang_parse_ctx * C, slang_output_ctx * O,
                if (!legal_identifier(o->a_id)) {
                   slang_info_log_error(C->L, "illegal variable name '%s'",
                                        (char *) o->a_id);
-                  return 0;
+                  RETURN0;
                }
             }
          }
@@ -996,10 +1006,10 @@ parse_statement(slang_parse_ctx * C, slang_output_ctx * O,
       oper->type = SLANG_OPER_ASM;
       oper->a_id = parse_identifier(C);
       if (oper->a_id == SLANG_ATOM_NULL)
-         return 0;
+         RETURN0;
       while (*C->I != OP_END) {
          if (!parse_child_operation(C, O, oper, 0))
-            return 0;
+            RETURN0;
       }
       C->I++;
       break;
@@ -1015,21 +1025,21 @@ parse_statement(slang_parse_ctx * C, slang_output_ctx * O,
    case OP_RETURN:
       oper->type = SLANG_OPER_RETURN;
       if (!parse_child_operation(C, O, oper, 0))
-         return 0;
+         RETURN0;
       break;
    case OP_EXPRESSION:
       oper->type = SLANG_OPER_EXPRESSION;
       if (!parse_child_operation(C, O, oper, 0))
-         return 0;
+         RETURN0;
       break;
    case OP_IF:
       oper->type = SLANG_OPER_IF;
       if (!parse_child_operation(C, O, oper, 0))
-         return 0;
+         RETURN0;
       if (!parse_child_operation(C, O, oper, 1))
-         return 0;
+         RETURN0;
       if (!parse_child_operation(C, O, oper, 1))
-         return 0;
+         RETURN0;
       break;
    case OP_WHILE:
       {
@@ -1038,17 +1048,17 @@ parse_statement(slang_parse_ctx * C, slang_output_ctx * O,
          oper->type = SLANG_OPER_WHILE;
          o.vars = oper->locals;
          if (!parse_child_operation(C, &o, oper, 1))
-            return 0;
+            RETURN0;
          if (!parse_child_operation(C, &o, oper, 1))
-            return 0;
+            RETURN0;
       }
       break;
    case OP_DO:
       oper->type = SLANG_OPER_DO;
       if (!parse_child_operation(C, O, oper, 1))
-         return 0;
+         RETURN0;
       if (!parse_child_operation(C, O, oper, 0))
-         return 0;
+         RETURN0;
       break;
    case OP_FOR:
       {
@@ -1057,13 +1067,13 @@ parse_statement(slang_parse_ctx * C, slang_output_ctx * O,
          oper->type = SLANG_OPER_FOR;
          o.vars = oper->locals;
          if (!parse_child_operation(C, &o, oper, 1))
-            return 0;
+            RETURN0;
          if (!parse_child_operation(C, &o, oper, 1))
-            return 0;
+            RETURN0;
          if (!parse_child_operation(C, &o, oper, 0))
-            return 0;
+            RETURN0;
          if (!parse_child_operation(C, &o, oper, 1))
-            return 0;
+            RETURN0;
       }
       break;
    case OP_PRECISION:
@@ -1077,7 +1087,7 @@ parse_statement(slang_parse_ctx * C, slang_output_ctx * O,
       }
       break;
    default:
-      return 0;
+      RETURN0;
    }
    return 1;
 }
@@ -1092,7 +1102,7 @@ handle_nary_expression(slang_parse_ctx * C, slang_operation * op,
    op->children = slang_operation_new(n);
    if (op->children == NULL) {
       slang_info_log_memory(C->L);
-      return 0;
+      RETURN0;
    }
    op->num_children = n;
 
@@ -1110,7 +1120,7 @@ handle_nary_expression(slang_parse_ctx * C, slang_operation * op,
                      *total_ops * sizeof(slang_operation));
    if (*ops == NULL) {
       slang_info_log_memory(C->L);
-      return 0;
+      RETURN0;
    }
    return 1;
 }
@@ -1143,12 +1153,12 @@ parse_expression(slang_parse_ctx * C, slang_output_ctx * O,
                         (num_ops + 1) * sizeof(slang_operation));
       if (ops == NULL) {
          slang_info_log_memory(C->L);
-         return 0;
+         RETURN0;
       }
       op = &ops[num_ops];
       if (!slang_operation_construct(op)) {
          slang_info_log_memory(C->L);
-         return 0;
+         RETURN0;
       }
       num_ops++;
       op->locals->outer_scope = O->vars;
@@ -1160,7 +1170,7 @@ parse_expression(slang_parse_ctx * C, slang_output_ctx * O,
       case OP_PUSH_BOOL:
          op->type = SLANG_OPER_LITERAL_BOOL;
          if (!parse_number(C, &number))
-            return 0;
+            RETURN0;
          op->literal[0] =
          op->literal[1] =
          op->literal[2] =
@@ -1170,7 +1180,7 @@ parse_expression(slang_parse_ctx * C, slang_output_ctx * O,
       case OP_PUSH_INT:
          op->type = SLANG_OPER_LITERAL_INT;
          if (!parse_number(C, &number))
-            return 0;
+            RETURN0;
          op->literal[0] =
          op->literal[1] =
          op->literal[2] =
@@ -1180,7 +1190,7 @@ parse_expression(slang_parse_ctx * C, slang_output_ctx * O,
       case OP_PUSH_FLOAT:
          op->type = SLANG_OPER_LITERAL_FLOAT;
          if (!parse_float(C, &op->literal[0]))
-            return 0;
+            RETURN0;
          op->literal[1] =
          op->literal[2] =
          op->literal[3] = op->literal[0];
@@ -1190,37 +1200,37 @@ parse_expression(slang_parse_ctx * C, slang_output_ctx * O,
          op->type = SLANG_OPER_IDENTIFIER;
          op->a_id = parse_identifier(C);
          if (op->a_id == SLANG_ATOM_NULL)
-            return 0;
+            RETURN0;
          break;
       case OP_SEQUENCE:
          op->type = SLANG_OPER_SEQUENCE;
          if (!handle_nary_expression(C, op, &ops, &num_ops, 2))
-            return 0;
+            RETURN0;
          break;
       case OP_ASSIGN:
          op->type = SLANG_OPER_ASSIGN;
          if (!handle_nary_expression(C, op, &ops, &num_ops, 2))
-            return 0;
+            RETURN0;
          break;
       case OP_ADDASSIGN:
          op->type = SLANG_OPER_ADDASSIGN;
          if (!handle_nary_expression(C, op, &ops, &num_ops, 2))
-            return 0;
+            RETURN0;
          break;
       case OP_SUBASSIGN:
          op->type = SLANG_OPER_SUBASSIGN;
          if (!handle_nary_expression(C, op, &ops, &num_ops, 2))
-            return 0;
+            RETURN0;
          break;
       case OP_MULASSIGN:
          op->type = SLANG_OPER_MULASSIGN;
          if (!handle_nary_expression(C, op, &ops, &num_ops, 2))
-            return 0;
+            RETURN0;
          break;
       case OP_DIVASSIGN:
          op->type = SLANG_OPER_DIVASSIGN;
          if (!handle_nary_expression(C, op, &ops, &num_ops, 2))
-            return 0;
+            RETURN0;
          break;
          /*case OP_MODASSIGN: */
          /*case OP_LSHASSIGN: */
@@ -1231,22 +1241,22 @@ parse_expression(slang_parse_ctx * C, slang_output_ctx * O,
       case OP_SELECT:
          op->type = SLANG_OPER_SELECT;
          if (!handle_nary_expression(C, op, &ops, &num_ops, 3))
-            return 0;
+            RETURN0;
          break;
       case OP_LOGICALOR:
          op->type = SLANG_OPER_LOGICALOR;
          if (!handle_nary_expression(C, op, &ops, &num_ops, 2))
-            return 0;
+            RETURN0;
          break;
       case OP_LOGICALXOR:
          op->type = SLANG_OPER_LOGICALXOR;
          if (!handle_nary_expression(C, op, &ops, &num_ops, 2))
-            return 0;
+            RETURN0;
          break;
       case OP_LOGICALAND:
          op->type = SLANG_OPER_LOGICALAND;
          if (!handle_nary_expression(C, op, &ops, &num_ops, 2))
-            return 0;
+            RETURN0;
          break;
          /*case OP_BITOR: */
          /*case OP_BITXOR: */
@@ -1254,95 +1264,95 @@ parse_expression(slang_parse_ctx * C, slang_output_ctx * O,
       case OP_EQUAL:
          op->type = SLANG_OPER_EQUAL;
          if (!handle_nary_expression(C, op, &ops, &num_ops, 2))
-            return 0;
+            RETURN0;
          break;
       case OP_NOTEQUAL:
          op->type = SLANG_OPER_NOTEQUAL;
          if (!handle_nary_expression(C, op, &ops, &num_ops, 2))
-            return 0;
+            RETURN0;
          break;
       case OP_LESS:
          op->type = SLANG_OPER_LESS;
          if (!handle_nary_expression(C, op, &ops, &num_ops, 2))
-            return 0;
+            RETURN0;
          break;
       case OP_GREATER:
          op->type = SLANG_OPER_GREATER;
          if (!handle_nary_expression(C, op, &ops, &num_ops, 2))
-            return 0;
+            RETURN0;
          break;
       case OP_LESSEQUAL:
          op->type = SLANG_OPER_LESSEQUAL;
          if (!handle_nary_expression(C, op, &ops, &num_ops, 2))
-            return 0;
+            RETURN0;
          break;
       case OP_GREATEREQUAL:
          op->type = SLANG_OPER_GREATEREQUAL;
          if (!handle_nary_expression(C, op, &ops, &num_ops, 2))
-            return 0;
+            RETURN0;
          break;
          /*case OP_LSHIFT: */
          /*case OP_RSHIFT: */
       case OP_ADD:
          op->type = SLANG_OPER_ADD;
          if (!handle_nary_expression(C, op, &ops, &num_ops, 2))
-            return 0;
+            RETURN0;
          break;
       case OP_SUBTRACT:
          op->type = SLANG_OPER_SUBTRACT;
          if (!handle_nary_expression(C, op, &ops, &num_ops, 2))
-            return 0;
+            RETURN0;
          break;
       case OP_MULTIPLY:
          op->type = SLANG_OPER_MULTIPLY;
          if (!handle_nary_expression(C, op, &ops, &num_ops, 2))
-            return 0;
+            RETURN0;
          break;
       case OP_DIVIDE:
          op->type = SLANG_OPER_DIVIDE;
          if (!handle_nary_expression(C, op, &ops, &num_ops, 2))
-            return 0;
+            RETURN0;
          break;
          /*case OP_MODULUS: */
       case OP_PREINCREMENT:
          op->type = SLANG_OPER_PREINCREMENT;
          if (!handle_nary_expression(C, op, &ops, &num_ops, 1))
-            return 0;
+            RETURN0;
          break;
       case OP_PREDECREMENT:
          op->type = SLANG_OPER_PREDECREMENT;
          if (!handle_nary_expression(C, op, &ops, &num_ops, 1))
-            return 0;
+            RETURN0;
          break;
       case OP_PLUS:
          op->type = SLANG_OPER_PLUS;
          if (!handle_nary_expression(C, op, &ops, &num_ops, 1))
-            return 0;
+            RETURN0;
          break;
       case OP_MINUS:
          op->type = SLANG_OPER_MINUS;
          if (!handle_nary_expression(C, op, &ops, &num_ops, 1))
-            return 0;
+            RETURN0;
          break;
       case OP_NOT:
          op->type = SLANG_OPER_NOT;
          if (!handle_nary_expression(C, op, &ops, &num_ops, 1))
-            return 0;
+            RETURN0;
          break;
          /*case OP_COMPLEMENT: */
       case OP_SUBSCRIPT:
          op->type = SLANG_OPER_SUBSCRIPT;
          if (!handle_nary_expression(C, op, &ops, &num_ops, 2))
-            return 0;
+            RETURN0;
          break;
       case OP_CALL:
          op->type = SLANG_OPER_CALL;
          op->a_id = parse_identifier(C);
          if (op->a_id == SLANG_ATOM_NULL)
-            return 0;
+            RETURN0;
          while (*C->I != OP_END)
             if (!parse_child_operation(C, O, op, 0))
-               return 0;
+               RETURN0;
          C->I++;
 
          if (!C->parsing_builtin
@@ -1352,7 +1362,7 @@ parse_expression(slang_parse_ctx * C, slang_output_ctx * O,
             id = slang_atom_pool_id(C->atoms, op->a_id);
             if (!is_constructor_name(id, op->a_id, O->structs)) {
                slang_info_log_error(C->L, "%s: undeclared function name.", id);
-               return 0;
+               RETURN0;
             }
          }
          break;
@@ -1360,22 +1370,22 @@ parse_expression(slang_parse_ctx * C, slang_output_ctx * O,
          op->type = SLANG_OPER_FIELD;
          op->a_id = parse_identifier(C);
          if (op->a_id == SLANG_ATOM_NULL)
-            return 0;
+            RETURN0;
          if (!handle_nary_expression(C, op, &ops, &num_ops, 1))
-            return 0;
+            RETURN0;
          break;
       case OP_POSTINCREMENT:
          op->type = SLANG_OPER_POSTINCREMENT;
          if (!handle_nary_expression(C, op, &ops, &num_ops, 1))
-            return 0;
+            RETURN0;
          break;
       case OP_POSTDECREMENT:
          op->type = SLANG_OPER_POSTDECREMENT;
          if (!handle_nary_expression(C, op, &ops, &num_ops, 1))
-            return 0;
+            RETURN0;
          break;
       default:
-         return 0;
+         RETURN0;
       }
    }
    C->I++;
@@ -1406,7 +1416,7 @@ parse_parameter_declaration(slang_parse_ctx * C, slang_output_ctx * O,
     * two at most) because not all combinations are valid
     */
    if (!parse_type_qualifier(C, &param->type.qualifier))
-      return 0;
+      RETURN0;
 
    param_qual = *C->I++;
    switch (param_qual) {
@@ -1414,7 +1424,7 @@ parse_parameter_declaration(slang_parse_ctx * C, slang_output_ctx * O,
       if (param->type.qualifier != SLANG_QUAL_CONST
           && param->type.qualifier != SLANG_QUAL_NONE) {
          slang_info_log_error(C->L, "Invalid type qualifier.");
-         return 0;
+         RETURN0;
       }
       break;
    case PARAM_QUALIFIER_OUT:
@@ -1422,7 +1432,7 @@ parse_parameter_declaration(slang_parse_ctx * C, slang_output_ctx * O,
          param->type.qualifier = SLANG_QUAL_OUT;
       else {
          slang_info_log_error(C->L, "Invalid type qualifier.");
-         return 0;
+         RETURN0;
       }
       break;
    case PARAM_QUALIFIER_INOUT:
@@ -1430,11 +1440,11 @@ parse_parameter_declaration(slang_parse_ctx * C, slang_output_ctx * O,
          param->type.qualifier = SLANG_QUAL_INOUT;
       else {
          slang_info_log_error(C->L, "Invalid type qualifier.");
-         return 0;
+         RETURN0;
       }
       break;
    default:
-      return 0;
+      RETURN0;
    }
 
    /* parse precision qualifier (lowp, mediump, highp */
@@ -1444,10 +1454,10 @@ parse_parameter_declaration(slang_parse_ctx * C, slang_output_ctx * O,
 
    /* parse parameter's type specifier and name */
    if (!parse_type_specifier(C, O, &param->type.specifier))
-      return 0;
+      RETURN0;
    param->a_name = parse_identifier(C);
    if (param->a_name == SLANG_ATOM_NULL)
-      return 0;
+      RETURN0;
 
    /* if the parameter is an array, parse its size (the size must be
     * explicitly defined
@@ -1564,13 +1574,13 @@ parse_operator_name(slang_parse_ctx * C)
             slang_atom_pool_atom(C->atoms, operator_names[i].o_name);
          if (atom == SLANG_ATOM_NULL) {
             slang_info_log_memory(C->L);
-            return 0;
+            RETURN0;
          }
          C->I++;
          return atom;
       }
    }
-   return 0;
+   RETURN0;
 }
 
 
@@ -1581,7 +1591,7 @@ parse_function_prototype(slang_parse_ctx * C, slang_output_ctx * O,
    GLuint functype;
    /* parse function type and name */
    if (!parse_fully_specified_type(C, O, &func->header.type))
-      return 0;
+      RETURN0;
 
    functype = *C->I++;
    switch (functype) {
@@ -1589,35 +1599,35 @@ parse_function_prototype(slang_parse_ctx * C, slang_output_ctx * O,
       func->kind = SLANG_FUNC_ORDINARY;
       func->header.a_name = parse_identifier(C);
       if (func->header.a_name == SLANG_ATOM_NULL)
-         return 0;
+         RETURN0;
       break;
    case FUNCTION_CONSTRUCTOR:
       func->kind = SLANG_FUNC_CONSTRUCTOR;
       if (func->header.type.specifier.type == SLANG_SPEC_STRUCT)
-         return 0;
+         RETURN0;
       func->header.a_name =
          slang_atom_pool_atom(C->atoms,
                               slang_type_specifier_type_to_string
                               (func->header.type.specifier.type));
       if (func->header.a_name == SLANG_ATOM_NULL) {
          slang_info_log_memory(C->L);
-         return 0;
+         RETURN0;
       }
       break;
    case FUNCTION_OPERATOR:
       func->kind = SLANG_FUNC_OPERATOR;
       func->header.a_name = parse_operator_name(C);
       if (func->header.a_name == SLANG_ATOM_NULL)
-         return 0;
+         RETURN0;
       break;
    default:
-      return 0;
+      RETURN0;
    }
 
    if (!legal_identifier(func->header.a_name)) {
       slang_info_log_error(C->L, "illegal function name '%s'",
                            (char *) func->header.a_name);
-      return 0;
+      RETURN0;
    }
 
    /* parse function parameters */
@@ -1625,10 +1635,10 @@ parse_function_prototype(slang_parse_ctx * C, slang_output_ctx * O,
       slang_variable *p = slang_variable_scope_grow(func->parameters);
       if (!p) {
          slang_info_log_memory(C->L);
-         return 0;
+         RETURN0;
       }
       if (!parse_parameter_declaration(C, O, p))
-         return 0;
+         RETURN0;
    }
 
    /* if the function returns a value, append a hidden __retVal 'out'
@@ -1662,19 +1672,19 @@ parse_function_definition(slang_parse_ctx * C, slang_output_ctx * O,
    slang_output_ctx o = *O;
 
    if (!parse_function_prototype(C, O, func))
-      return 0;
+      RETURN0;
 
    /* create function's body operation */
    func->body = (slang_operation *) _slang_alloc(sizeof(slang_operation));
    if (func->body == NULL) {
       slang_info_log_memory(C->L);
-      return 0;
+      RETURN0;
    }
    if (!slang_operation_construct(func->body)) {
       _slang_free(func->body);
       func->body = NULL;
       slang_info_log_memory(C->L);
-      return 0;
+      RETURN0;
    }
 
    /* to parse the body the parse context is modified in order to
@@ -1683,7 +1693,7 @@ parse_function_definition(slang_parse_ctx * C, slang_output_ctx * O,
    C->global_scope = GL_FALSE;
    o.vars = func->parameters;
    if (!parse_statement(C, &o, func->body))
-      return 0;
+      RETURN0;
 
    C->global_scope = GL_TRUE;
    return 1;
@@ -1779,46 +1789,46 @@ parse_init_declarator(slang_parse_ctx * C, slang_output_ctx * O,
       slang_info_log_error(C->L,
                    "declaration of '%s' conflicts with previous declaration",
                    (char *) a_name);
-      return 0;
+      RETURN0;
    }
 
    /* make room for the new variable and initialize it */
    var = slang_variable_scope_grow(O->vars);
    if (!var) {
       slang_info_log_memory(C->L);
-      return 0;
+      RETURN0;
    }
 
    /* copy the declarator qualifier type, parse the identifier */
    var->type.qualifier = type->qualifier;
    var->a_name = a_name;
    if (var->a_name == SLANG_ATOM_NULL)
-      return 0;
+      RETURN0;
 
    switch (*C->I++) {
    case VARIABLE_NONE:
       /* simple variable declarator - just copy the specifier */
       if (!slang_type_specifier_copy(&var->type.specifier, &type->specifier))
-         return 0;
+         RETURN0;
       break;
    case VARIABLE_INITIALIZER:
       /* initialized variable - copy the specifier and parse the expression */
       if (!slang_type_specifier_copy(&var->type.specifier, &type->specifier))
-         return 0;
+         RETURN0;
       var->initializer =
          (slang_operation *) _slang_alloc(sizeof(slang_operation));
       if (var->initializer == NULL) {
          slang_info_log_memory(C->L);
-         return 0;
+         RETURN0;
       }
       if (!slang_operation_construct(var->initializer)) {
          _slang_free(var->initializer);
          var->initializer = NULL;
          slang_info_log_memory(C->L);
-         return 0;
+         RETURN0;
       }
       if (!parse_expression(C, O, var->initializer))
-         return 0;
+         RETURN0;
       break;
    case VARIABLE_ARRAY_UNKNOWN:
       /* unsized array - mark it as array and copy the specifier to
@@ -1834,7 +1844,7 @@ parse_init_declarator(slang_parse_ctx * C, slang_output_ctx * O,
          return GL_FALSE;
       break;
    default:
-      return 0;
+      RETURN0;
    }
 
    /* allocate global address space for a variable with a known size */
@@ -1858,7 +1868,7 @@ parse_init_declarator(slang_parse_ctx * C, slang_output_ctx * O,
       A.log = C->L;
       A.curFuncEndLabel = NULL;
       if (!_slang_codegen_global_variable(&A, var, C->type))
-         return 0;
+         RETURN0;
    }
 
    /* initialize global variable */
@@ -1871,7 +1881,7 @@ parse_init_declarator(slang_parse_ctx * C, slang_output_ctx * O,
          A.space.structs = O->structs;
          A.space.vars = O->vars;
          if (!initialize_global(&A, var))
-            return 0;
+            RETURN0;
       }
    }
    return 1;
@@ -1888,17 +1898,17 @@ parse_init_declarator_list(slang_parse_ctx * C, slang_output_ctx * O)
 
    /* parse the fully specified type, common to all declarators */
    if (!slang_fully_specified_type_construct(&type))
-      return 0;
+      RETURN0;
    if (!parse_fully_specified_type(C, O, &type)) {
       slang_fully_specified_type_destruct(&type);
-      return 0;
+      RETURN0;
    }
 
    /* parse declarators, pass-in the parsed type */
    do {
       if (!parse_init_declarator(C, O, &type)) {
          slang_fully_specified_type_destruct(&type);
-         return 0;
+         RETURN0;
       }
    }
    while (*C->I++ == DECLARATOR_NEXT);
@@ -2005,18 +2015,18 @@ parse_declaration(slang_parse_ctx * C, slang_output_ctx * O)
    switch (*C->I++) {
    case DECLARATION_INIT_DECLARATOR_LIST:
       if (!parse_init_declarator_list(C, O))
-         return 0;
+         RETURN0;
       break;
    case DECLARATION_FUNCTION_PROTOTYPE:
       {
          slang_function *dummy_func;
 
          if (!parse_function(C, O, 0, &dummy_func))
-            return 0;
+            RETURN0;
       }
       break;
    default:
-      return 0;
+      RETURN0;
    }
    return 1;
 }
@@ -2028,7 +2038,7 @@ parse_default_precision(slang_parse_ctx * C, slang_output_ctx * O)
 
    if (!O->allow_precision) {
       slang_info_log_error(C->L, "syntax error at \"precision\"");
-      return 0;
+      RETURN0;
    }
 
    precision = *C->I++;
@@ -2041,7 +2051,7 @@ parse_default_precision(slang_parse_ctx * C, slang_output_ctx * O)
    default:
       _mesa_problem(NULL, "unexpected precision %d at %s:%d\n",
                     precision, __FILE__, __LINE__);
-      return 0;
+      RETURN0;
    }
 
    type = *C->I++;
@@ -2061,7 +2071,7 @@ parse_default_precision(slang_parse_ctx * C, slang_output_ctx * O)
    default:
       _mesa_problem(NULL, "unexpected type %d at %s:%d\n",
                     type, __FILE__, __LINE__);
-      return 0;
+      RETURN0;
    }
 
    assert(type < TYPE_SPECIFIER_COUNT);
@@ -2108,7 +2118,7 @@ parse_invariant(slang_parse_ctx * C, slang_output_ctx * O)
    }
    else {
       slang_info_log_error(C->L, "syntax error at \"invariant\"");
-      return 0;
+      RETURN0;
    }
 }
       
diff --git a/src/mesa/shader/slang/slang_emit.c b/src/mesa/shader/slang/slang_emit.c
index b67cea7617d..0a3ab39eb10 100644
--- a/src/mesa/shader/slang/slang_emit.c
+++ b/src/mesa/shader/slang/slang_emit.c
@@ -60,6 +60,8 @@ typedef struct
    struct gl_program **Subroutines;
    GLuint NumSubroutines;
 
+   GLuint MaxInstructions;  /**< size of prog->Instructions[] buffer */
+
    /* code-gen options */
    GLboolean EmitHighLevelInstructions;
    GLboolean EmitCondCodes;
@@ -259,13 +261,16 @@ fix_swizzle(GLuint swizzle)
 static void
 storage_to_dst_reg(struct prog_dst_register *dst, const slang_ir_storage *st)
 {
+   const GLboolean relAddr = st->RelAddr;
    const GLint size = st->Size;
    GLint index = st->Index;
    GLuint swizzle = st->Swizzle;
 
+   assert(index >= 0);
    /* if this is storage relative to some parent storage, walk up the tree */
    while (st->Parent) {
       st = st->Parent;
+      assert(st->Index >= 0);
       index += st->Index;
       swizzle = _slang_swizzle_swizzle(st->Swizzle, swizzle);
    }
@@ -302,6 +307,8 @@ storage_to_dst_reg(struct prog_dst_register *dst, const slang_ir_storage *st)
       }
       dst->WriteMask = writemask;
    }
+
+   dst->RelAddr = relAddr;
 }
 
 
@@ -316,8 +323,10 @@ storage_to_src_reg(struct prog_src_register *src, const slang_ir_storage *st)
    GLuint swizzle = st->Swizzle;
 
    /* if this is storage relative to some parent storage, walk up the tree */
+   assert(index >= 0);
    while (st->Parent) {
       st = st->Parent;
+      assert(st->Index >= 0);
       index += st->Index;
       swizzle = _slang_swizzle_swizzle(fix_swizzle(st->Swizzle), swizzle);
    }
@@ -387,9 +396,17 @@ new_instruction(slang_emit_info *emitInfo, gl_inst_opcode opcode)
       _mesa_print_instruction(prog->Instructions + prog->NumInstructions - 1);
    }
 #endif
-   prog->Instructions = _mesa_realloc_instructions(prog->Instructions,
-                                                   prog->NumInstructions,
-                                                   prog->NumInstructions + 1);
+   assert(prog->NumInstructions <= emitInfo->MaxInstructions);
+
+   if (prog->NumInstructions == emitInfo->MaxInstructions) {
+      /* grow the instruction buffer */
+      emitInfo->MaxInstructions += 20;
+      prog->Instructions =
+         _mesa_realloc_instructions(prog->Instructions,
+                                    prog->NumInstructions,
+                                    emitInfo->MaxInstructions);
+   }
+
    inst = prog->Instructions + prog->NumInstructions;
    prog->NumInstructions++;
    _mesa_init_instructions(inst, 1);
@@ -403,60 +420,149 @@ new_instruction(slang_emit_info *emitInfo, gl_inst_opcode opcode)
 }
 
 
+static struct prog_instruction *
+emit_arl_load(slang_emit_info *emitInfo,
+              enum register_file file, GLint index, GLuint swizzle)
+{
+   struct prog_instruction *inst = new_instruction(emitInfo, OPCODE_ARL);
+   inst->SrcReg[0].File = file;
+   inst->SrcReg[0].Index = index;
+   inst->SrcReg[0].Swizzle = swizzle;
+   inst->DstReg.File = PROGRAM_ADDRESS;
+   inst->DstReg.Index = 0;
+   inst->DstReg.WriteMask = WRITEMASK_X;
+   return inst;
+}
+
+
 /**
  * Emit a new instruction with given opcode, operands.
+ * At this point the instruction may have multiple indirect register
+ * loads/stores.  We convert those into ARL loads and address-relative
+ * operands.  See comments inside.
+ * At some point in the future we could directly emit indirectly addressed
+ * registers in Mesa GPU instructions.
  */
 static struct prog_instruction *
 emit_instruction(slang_emit_info *emitInfo,
                  gl_inst_opcode opcode,
                  const slang_ir_storage *dst,
+                 const slang_ir_storage *src0,
                  const slang_ir_storage *src1,
-                 const slang_ir_storage *src2,
-                 const slang_ir_storage *src3)
+                 const slang_ir_storage *src2)
 {
-   struct gl_program *prog = emitInfo->prog;
    struct prog_instruction *inst;
+   GLuint numIndirect = 0;
+   const slang_ir_storage *src[3];
+   slang_ir_storage newSrc[3], newDst;
+   GLuint i;
+   GLboolean isTemp[3];
+
+   isTemp[0] = isTemp[1] = isTemp[2] = GL_FALSE;
+
+   src[0] = src0;
+   src[1] = src1;
+   src[2] = src2;
+
+   /* count up how many operands are indirect loads */
+   for (i = 0; i < 3; i++) {
+      if (src[i] && src[i]->IsIndirect)
+         numIndirect++;
+   }
+   if (dst && dst->IsIndirect)
+      numIndirect++;
+
+   /* Take special steps for indirect register loads.
+    * If we had multiple address registers this would be simpler.
+    * For example, this GLSL code:
+    *    x[i] = y[j] + z[k];
+    * would translate into something like:
+    *    ARL ADDR.x, i;
+    *    ARL ADDR.y, j;
+    *    ARL ADDR.z, k;
+    *    ADD TEMP[ADDR.x+5], TEMP[ADDR.y+9], TEMP[ADDR.z+4];
+    * But since we currently only have one address register we have to do this:
+    *    ARL ADDR.x, i;
+    *    MOV t1, TEMP[ADDR.x+9];
+    *    ARL ADDR.x, j;
+    *    MOV t2, TEMP[ADDR.x+4];
+    *    ARL ADDR.x, k;
+    *    ADD TEMP[ADDR.x+5], t1, t2;
+    * The code here figures this out...
+    */
+   if (numIndirect > 0) {
+      for (i = 0; i < 3; i++) {
+         if (src[i] && src[i]->IsIndirect) {
+            /* load the ARL register with the indirect register */
+            emit_arl_load(emitInfo,
+                          src[i]->IndirectFile,
+                          src[i]->IndirectIndex,
+                          src[i]->IndirectSwizzle);
+
+            if (numIndirect > 1) {
+               /* Need to load src[i] into a temporary register */
+               slang_ir_storage srcRelAddr;
+               alloc_local_temp(emitInfo, &newSrc[i], src[i]->Size);
+               isTemp[i] = GL_TRUE;
+
+               /* set RelAddr flag on src register */
+               srcRelAddr = *src[i];
+               srcRelAddr.RelAddr = GL_TRUE;
+               srcRelAddr.IsIndirect = GL_FALSE; /* not really needed */
+
+               /* MOV newSrc, srcRelAddr; */
+               inst = emit_instruction(emitInfo,
+                                       OPCODE_MOV,
+                                       &newSrc[i],
+                                       &srcRelAddr,
+                                       NULL,
+                                       NULL);
+
+               src[i] = &newSrc[i];
+            }
+            else {
+               /* just rewrite the src[i] storage to be ARL-relative */
+               newSrc[i] = *src[i];
+               newSrc[i].RelAddr = GL_TRUE;
+               newSrc[i].IsIndirect = GL_FALSE; /* not really needed */
+               src[i] = &newSrc[i];
+            }
+         }
+      }
+   }
 
-   prog->Instructions = _mesa_realloc_instructions(prog->Instructions,
-                                                   prog->NumInstructions,
-                                                   prog->NumInstructions + 1);
-   inst = prog->Instructions + prog->NumInstructions;
-   prog->NumInstructions++;
+   /* Take special steps for indirect dest register write */
+   if (dst && dst->IsIndirect) {
+      /* load the ARL register with the indirect register */
+      emit_arl_load(emitInfo,
+                    dst->IndirectFile,
+                    dst->IndirectIndex,
+                    dst->IndirectSwizzle);
+      newDst = *dst;
+      newDst.RelAddr = GL_TRUE;
+      newDst.IsIndirect = GL_FALSE;
+      dst = &newDst;
+   }
 
-   _mesa_init_instructions(inst, 1);
-   inst->Opcode = opcode;
-   inst->BranchTarget = -1; /* invalid */
+   /* OK, emit the instruction and its dst, src regs */
+   inst = new_instruction(emitInfo, opcode);
+   if (!inst)
+      return NULL;
 
    if (dst)
       storage_to_dst_reg(&inst->DstReg, dst);
 
-   if (src1)
-      storage_to_src_reg(&inst->SrcReg[0], src1);
-   if (src2)
-      storage_to_src_reg(&inst->SrcReg[1], src2);
-   if (src3)
-      storage_to_src_reg(&inst->SrcReg[2], src3);   
-
-   return inst;
-}
-
+   for (i = 0; i < 3; i++) {
+      if (src[i])
+         storage_to_src_reg(&inst->SrcReg[i], src[i]);
+   }
 
-/**
- * Emit an ARL instruction.
- */
-static struct prog_instruction *
-emit_arl_instruction(slang_emit_info *emitInfo,
-                     GLint addrReg,
-                     const slang_ir_storage *src)
-{
-   struct prog_instruction *inst;
+   /* Free any temp registers that we allocated above */
+   for (i = 0; i < 3; i++) {
+      if (isTemp[i])
+         _slang_free_temp(emitInfo->vt, &newSrc[i]);
+   }
 
-   assert(addrReg == 0); /* only one addr reg at this time */
-   inst = new_instruction(emitInfo, OPCODE_ARL);
-   storage_to_src_reg(&inst->SrcReg[0], src);
-   inst->DstReg.File = PROGRAM_ADDRESS;
-   inst->DstReg.Index = addrReg;
-   inst->DstReg.WriteMask = WRITEMASK_X;
    return inst;
 }
 
@@ -1034,13 +1140,17 @@ emit_fcall(slang_emit_info *emitInfo, slang_ir_node *n)
    struct gl_program *progSave;
    struct prog_instruction *inst;
    GLuint subroutineId;
+   GLuint maxInstSave;
 
    assert(n->Opcode == IR_CALL);
    assert(n->Label);
 
    /* save/push cur program */
+   maxInstSave = emitInfo->MaxInstructions;
    progSave = emitInfo->prog;
+
    emitInfo->prog = new_subroutine(emitInfo, &subroutineId);
+   emitInfo->MaxInstructions = emitInfo->prog->NumInstructions;
 
    _slang_label_set_location(n->Label, emitInfo->prog->NumInstructions,
                              emitInfo->prog);
@@ -1072,6 +1182,7 @@ emit_fcall(slang_emit_info *emitInfo, slang_ir_node *n)
 
    /* pop/restore cur program */
    emitInfo->prog = progSave;
+   emitInfo->MaxInstructions = maxInstSave;
 
    /* emit the function call */
    inst = new_instruction(emitInfo, OPCODE_CAL);
@@ -1212,7 +1323,8 @@ emit_copy(slang_emit_info *emitInfo, slang_ir_node *n)
    if (inst &&
        _slang_is_temp(emitInfo->vt, n->Children[1]->Store) &&
        (inst->DstReg.File == n->Children[1]->Store->File) &&
-       (inst->DstReg.Index == n->Children[1]->Store->Index)) {
+       (inst->DstReg.Index == n->Children[1]->Store->Index) &&
+       !n->Children[0]->Store->IsIndirect) {
       /* Peephole optimization:
        * The Right-Hand-Side has its results in a temporary place.
        * Modify the RHS (and the prev instruction) to store its results
@@ -1670,80 +1782,38 @@ emit_swizzle(slang_emit_info *emitInfo, slang_ir_node *n)
 
    inst = emit(emitInfo, n->Children[0]);
 
-   /* setup storage info, if needed */
-   if (!n->Store->Parent)
-      n->Store->Parent = n->Children[0]->Store;
-
+#if 0
    assert(n->Store->Parent);
-
+   /* Apply this node's swizzle to parent's storage */
+   GLuint swizzle = n->Store->Swizzle;
+   _slang_copy_ir_storage(n->Store, n->Store->Parent);
+   n->Store->Swizzle = _slang_swizzle_swizzle(n->Store->Swizzle, swizzle);
+   assert(!n->Store->Parent);
+#endif
    return inst;
 }
 
 
 /**
- * Move a block registers from src to dst (or move a single register).
- * \param size  size of block, in floats (<=4 means one register)
+ * Dereference array element:  element == array[index]
+ * This basically involves emitting code for computing the array index
+ * and updating the node/element's storage info.
  */
 static struct prog_instruction *
-move_block(slang_emit_info *emitInfo,
-           GLuint size, GLboolean relAddr,
-           const slang_ir_storage *dst,
-           const slang_ir_storage *src)
+emit_array_element(slang_emit_info *emitInfo, slang_ir_node *n)
 {
+   slang_ir_storage *arrayStore, *indexStore;
+   const int elemSize = n->Store->Size;           /* number of floats */
+   const GLint elemSizeVec = (elemSize + 3) / 4;  /* number of vec4 */
    struct prog_instruction *inst;
 
-   if (size > 4) {
-      /* move matrix/struct etc (block of registers) */
-      slang_ir_storage dstStore = *dst;
-      slang_ir_storage srcStore = *src;
-
-      dstStore.Size = 4;
-      srcStore.Size = 4;
-      while (size >= 4) {
-         inst = emit_instruction(emitInfo, OPCODE_MOV,
-                                 &dstStore,
-                                 &srcStore,
-                                 NULL,
-                                 NULL);
-         inst->SrcReg[0].RelAddr = relAddr;
-         inst_comment(inst, "IR_COPY block");
-         srcStore.Index++;
-         dstStore.Index++;
-         size -= 4;
-      }
-   }
-   else {
-      /* single register move */
-      inst = emit_instruction(emitInfo,
-                              OPCODE_MOV,
-                              dst,
-                              src,
-                              NULL,
-                              NULL);
-      inst->SrcReg[0].RelAddr = relAddr;
-   }
-   return inst;
-}
-
-
-
-/**
- * Dereference array element.  Just resolve storage for the array
- * element represented by this node.
- * This is typically where Indirect addressing comes into play.
- * See comments on struct slang_ir_storage.
- */
-static struct prog_instruction *
-emit_array_element(slang_emit_info *emitInfo, slang_ir_node *n)
-{
    assert(n->Opcode == IR_ELEMENT);
-   assert(n->Store);
-   assert(n->Store->File == PROGRAM_UNDEFINED);
-   assert(n->Store->Parent);
-   assert(n->Store->Size > 0);
+   assert(elemSize > 0);
 
+   /* special case for built-in state variables, like light state */
    {
       slang_ir_storage *root = n->Store;
+      assert(!root->Parent);
       while (root->Parent)
          root = root->Parent;
 
@@ -1754,69 +1824,98 @@ emit_array_element(slang_emit_info *emitInfo, slang_ir_node *n)
       }
    }
 
-   /* do codegen for array */
+   /* do codegen for array itself */
    emit(emitInfo, n->Children[0]);
+   arrayStore = n->Children[0]->Store;
+
+   /* The initial array element storage is the array's storage,
+    * then modified below.
+    */
+   _slang_copy_ir_storage(n->Store, arrayStore);
+
 
    if (n->Children[1]->Opcode == IR_FLOAT) {
-      /* Constant array index.
-       * Set Store's index to be the offset of the array element in
-       * the register file.
-       */
+      /* Constant array index */
       const GLint element = (GLint) n->Children[1]->Value[0];
-      const GLint sz = (n->Store->Size + 3) / 4; /* size in slots/registers */
 
-      n->Store->Index = sz * element;
-      assert(n->Store->Parent);
+      /* this element's storage is the array's storage, plus constant offset */
+      n->Store->Index += elemSizeVec * element;
    }
    else {
       /* Variable array index */
-      struct prog_instruction *inst;
 
       /* do codegen for array index expression */
       emit(emitInfo, n->Children[1]);
+      indexStore = n->Children[1]->Store;
+
+      if (indexStore->IsIndirect) {
+         /* need to put the array index into a temporary since we can't
+          * directly support a[b[i]] constructs.
+          */
+
 
-      /* allocate temp storage for the array element */
-      assert(n->Store->Index < 0);
-      n->Store->File = PROGRAM_TEMPORARY;
-      n->Store->Parent = NULL;
-      alloc_node_storage(emitInfo, n, -1);
+         /*indexStore = tempstore();*/
+      }
 
-      if (n->Store->Size > 4) {
-         /* need to multiply the index by the element size */
-         const GLint elemSize = (n->Store->Size + 3) / 4;
-         slang_ir_storage indexTemp, elemSizeStore;
 
-         /* constant containing the element size */
-         constant_to_storage(emitInfo, (float) elemSize, &elemSizeStore);
+      if (elemSize > 4) {
+         /* need to multiply array index by array element size */
+         struct prog_instruction *inst;
+         slang_ir_storage *indexTemp;
+         slang_ir_storage elemSizeStore;
 
          /* allocate 1 float indexTemp */
-         alloc_local_temp(emitInfo, &indexTemp, 1);
+         indexTemp = _slang_new_ir_storage(PROGRAM_TEMPORARY, -1, 1);
+         _slang_alloc_temp(emitInfo->vt, indexTemp);
+
+         /* allocate a constant containing the element size */
+         constant_to_storage(emitInfo, (float) elemSizeVec, &elemSizeStore);
 
-         /* MUL temp, index, elemSize */
-         inst = emit_instruction(emitInfo, OPCODE_MUL,
-                                 &indexTemp, /* dest */
-                                 n->Children[1]->Store, /* the index */
+         /* multiply array index by element size */
+         inst = emit_instruction(emitInfo,
+                                 OPCODE_MUL,
+                                 indexTemp, /* dest */
+                                 indexStore, /* the index */
                                  &elemSizeStore,
                                  NULL);
 
-         /* load ADDR[0].X = temp */
-         inst = emit_arl_instruction(emitInfo, 0, &indexTemp);
-
-         _slang_free_temp(emitInfo->vt, &indexTemp);
+         indexStore = indexTemp;
       }
-      else {
-         /* simply load address reg w/ array index */
-         inst = emit_arl_instruction(emitInfo, 0, n->Children[1]->Store);
+
+      if (arrayStore->IsIndirect) {
+         /* ex: in a[i][j], a[i] (the arrayStore) is indirect */
+         /* Need to add indexStore to arrayStore->Indirect store */
+         slang_ir_storage indirectArray;
+         slang_ir_storage *indexTemp;
+
+         _slang_init_ir_storage(&indirectArray,
+                                arrayStore->IndirectFile,
+                                arrayStore->IndirectIndex,
+                                1,
+                                arrayStore->IndirectSwizzle);
+
+         /* allocate 1 float indexTemp */
+         indexTemp = _slang_new_ir_storage(PROGRAM_TEMPORARY, -1, 1);
+         _slang_alloc_temp(emitInfo->vt, indexTemp);
+
+         inst = emit_instruction(emitInfo,
+                                 OPCODE_ADD,
+                                 indexTemp,      /* dest */
+                                 indexStore,     /* the index */
+                                 &indirectArray, /* indirect array base */
+                                 NULL);
+
+         indexStore = indexTemp;
       }
 
-      /* copy from array element to temp storage */
-      move_block(emitInfo, n->Store->Size, GL_TRUE,
-                 n->Store, n->Children[0]->Store);
+      /* update the array element storage info */
+      n->Store->IsIndirect = GL_TRUE;
+      n->Store->IndirectFile = indexStore->File;
+      n->Store->IndirectIndex = indexStore->Index;
+      n->Store->IndirectSwizzle = indexStore->Swizzle;
    }
 
-   /* if array element size is one, make sure we only access X */
-   if (n->Store->Size == 1)
-      n->Store->Swizzle = SWIZZLE_XXXX;
+   n->Store->Size = elemSize;
 
    return NULL; /* no instruction */
 }
@@ -1829,9 +1928,11 @@ static struct prog_instruction *
 emit_struct_field(slang_emit_info *emitInfo, slang_ir_node *n)
 {
    slang_ir_storage *root = n->Store;
+   GLint fieldOffset, fieldSize;
 
    assert(n->Opcode == IR_FIELD);
 
+   assert(!root->Parent);
    while (root->Parent)
       root = root->Parent;
 
@@ -1847,12 +1948,45 @@ emit_struct_field(slang_emit_info *emitInfo, slang_ir_node *n)
          slang_info_log_error(emitInfo->log, "Error parsing state variable");
          return NULL;
       }
+      return NULL;
    }
    else {
       /* do codegen for struct */
       emit(emitInfo, n->Children[0]);
+      assert(n->Children[0]->Store->Index >= 0);
    }
 
+   fieldOffset = n->Store->Index;
+   fieldSize = n->Store->Size;
+
+   _slang_copy_ir_storage(n->Store, n->Children[0]->Store);
+
+   n->Store->Index = n->Children[0]->Store->Index + fieldOffset / 4;
+   /* XXX test this:
+   n->Store->Index += fieldOffset / 4;
+   */
+
+   switch (fieldSize) {
+   case 1:
+      {
+         GLint swz = fieldOffset % 4;
+         n->Store->Swizzle = MAKE_SWIZZLE4(swz, swz, swz, swz);
+      }
+      break;
+   case 2:
+      n->Store->Swizzle = MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_Y,
+                                        SWIZZLE_NIL, SWIZZLE_NIL);
+      break;
+   case 3:
+      n->Store->Swizzle = MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_Y,
+                                        SWIZZLE_Z, SWIZZLE_NIL);
+      break;
+   default:
+      n->Store->Swizzle = SWIZZLE_XYZW;
+   }
+
+   assert(n->Store->Index >= 0);
+
    return NULL; /* no instruction */
 }
 
@@ -1909,7 +2043,7 @@ emit_var_decl(slang_emit_info *emitInfo, slang_ir_node *n)
 
 /**
  * Emit code for a reference to a variable.
- * Actually, no code is generated but we may do some memory alloation.
+ * Actually, no code is generated but we may do some memory allocation.
  * In particular, state vars (uniforms) are allocated on an as-needed basis.
  */
 static struct prog_instruction *
@@ -2130,7 +2264,7 @@ _slang_resolve_subroutines(slang_emit_info *emitInfo)
       total += emitInfo->Subroutines[i]->NumInstructions;
    }
 
-   /* adjust BrancTargets within the functions */
+   /* adjust BranchTargets within the functions */
    for (i = 0; i < emitInfo->NumSubroutines; i++) {
       struct gl_program *sub = emitInfo->Subroutines[i];
       GLuint j;
@@ -2199,6 +2333,7 @@ _slang_emit_code(slang_ir_node *n, slang_var_table *vt,
    emitInfo.prog = prog;
    emitInfo.Subroutines = NULL;
    emitInfo.NumSubroutines = 0;
+   emitInfo.MaxInstructions = prog->NumInstructions;
 
    emitInfo.EmitHighLevelInstructions = ctx->Shader.EmitHighLevelInstructions;
    emitInfo.EmitCondCodes = ctx->Shader.EmitCondCodes;
diff --git a/src/mesa/shader/slang/slang_ir.c b/src/mesa/shader/slang/slang_ir.c
index 4fe3a2b85d5..c33c80da999 100644
--- a/src/mesa/shader/slang/slang_ir.c
+++ b/src/mesa/shader/slang/slang_ir.c
@@ -116,6 +116,20 @@ _slang_ir_info(slang_ir_opcode opcode)
 }
 
 
+void
+_slang_init_ir_storage(slang_ir_storage *st,
+                       enum register_file file, GLint index, GLint size,
+                       GLuint swizzle)
+{
+   st->File = file;
+   st->Index = index;
+   st->Size = size;
+   st->Swizzle = swizzle;
+   st->Parent = NULL;
+   st->IsIndirect = GL_FALSE;
+}
+
+
 /**
  * Return a new slang_ir_storage object.
  */
@@ -130,6 +144,7 @@ _slang_new_ir_storage(enum register_file file, GLint index, GLint size)
       st->Size = size;
       st->Swizzle = SWIZZLE_NOOP;
       st->Parent = NULL;
+      st->IsIndirect = GL_FALSE;
    }
    return st;
 }
@@ -150,6 +165,7 @@ _slang_new_ir_storage_swz(enum register_file file, GLint index, GLint size,
       st->Size = size;
       st->Swizzle = swizzle;
       st->Parent = NULL;
+      st->IsIndirect = GL_FALSE;
    }
    return st;
 }
@@ -170,11 +186,45 @@ _slang_new_ir_storage_relative(GLint index, GLint size,
       st->Size = size;
       st->Swizzle = SWIZZLE_NOOP;
       st->Parent = parent;
+      st->IsIndirect = GL_FALSE;
+   }
+   return st;
+}
+
+
+slang_ir_storage *
+_slang_new_ir_storage_indirect(enum register_file file,
+                               GLint index,
+                               GLint size,
+                               enum register_file indirectFile,
+                               GLint indirectIndex,
+                               GLuint indirectSwizzle)
+{
+   slang_ir_storage *st;
+   st = (slang_ir_storage *) _slang_alloc(sizeof(slang_ir_storage));
+   if (st) {
+      st->File = file;
+      st->Index = index;
+      st->Size = size;
+      st->Swizzle = SWIZZLE_NOOP;
+      st->IsIndirect = GL_TRUE;
+      st->IndirectFile = indirectFile;
+      st->IndirectIndex = indirectIndex;
+      st->IndirectSwizzle = indirectSwizzle;
    }
    return st;
 }
 
 
+/* XXX temporary function */
+void
+_slang_copy_ir_storage(slang_ir_storage *dst, const slang_ir_storage *src)
+{
+   *dst = *src;
+   dst->Parent = NULL;
+}
+
+
 
 static const char *
 _slang_ir_name(slang_ir_opcode opcode)
diff --git a/src/mesa/shader/slang/slang_ir.h b/src/mesa/shader/slang/slang_ir.h
index a4552ae1442..a258e92e06f 100644
--- a/src/mesa/shader/slang/slang_ir.h
+++ b/src/mesa/shader/slang/slang_ir.h
@@ -140,24 +140,53 @@ typedef enum
 
 
 /**
- * Describes where data storage is allocated.
+ * Describes where data/variables are stored in the various register files.
+ *
+ * In the simple case, the File, Index and Size fields indicate where
+ * a variable is stored.  For example, a vec3 variable may be stored
+ * as (File=PROGRAM_TEMPORARY, Index=6, Size=3).  Or, File[Index].
+ * Or, a program input like color may be stored as
+ * (File=PROGRAM_INPUT,Index=3,Size=4);
+ *
+ * For single-float values, the Swizzle field indicates which component
+ * of the vector contains the float.
+ *
+ * If IsIndirect is set, the storage is accessed through an indirect
+ * register lookup.  The value in question will be located at:
+ *   File[Index + IndirectFile[IndirectIndex]]
+ *
+ * This is primary used for indexing arrays.  For example, consider this
+ * GLSL code:
+ *   uniform int i;
+ *   float a[10];
+ *   float x = a[i];
+ *
+ * here, storage for a[i] would be described by (File=PROGRAM_TEMPORAY,
+ * Index=aPos, IndirectFile=PROGRAM_UNIFORM, IndirectIndex=iPos), which
+ * would mean TEMP[aPos + UNIFORM[iPos]]
  */
-struct _slang_ir_storage
+struct slang_ir_storage_
 {
    enum register_file File;  /**< PROGRAM_TEMPORARY, PROGRAM_INPUT, etc */
    GLint Index;  /**< -1 means unallocated */
    GLint Size;  /**< number of floats */
-   GLuint Swizzle;
+   GLuint Swizzle; /**< Swizzle AND writemask info */
    GLint RefCount; /**< Used during IR tree delete */
-   GLboolean RelAddr;
+
+   GLboolean RelAddr; /* we'll remove this eventually */
+
+   GLboolean IsIndirect;
+   enum register_file IndirectFile;
+   GLint IndirectIndex;
+   GLuint IndirectSwizzle;
 
    /** If Parent is non-null, Index is relative to parent.
     * The other fields are ignored.
     */
-   struct _slang_ir_storage *Parent;
+   struct slang_ir_storage_ *Parent;
 };
 
-typedef struct _slang_ir_storage slang_ir_storage;
+typedef struct slang_ir_storage_ slang_ir_storage;
 
 
 /**
@@ -199,6 +228,11 @@ extern const slang_ir_info *
 _slang_ir_info(slang_ir_opcode opcode);
 
 
+extern void
+_slang_init_ir_storage(slang_ir_storage *st,
+                       enum register_file file, GLint index, GLint size,
+                       GLuint swizzle);
+
 extern slang_ir_storage *
 _slang_new_ir_storage(enum register_file file, GLint index, GLint size);
 
@@ -212,6 +246,17 @@ _slang_new_ir_storage_relative(GLint index, GLint size,
                                slang_ir_storage *parent);
 
 
+extern slang_ir_storage *
+_slang_new_ir_storage_indirect(enum register_file file,
+                               GLint index,
+                               GLint size,
+                               enum register_file indirectFile,
+                               GLint indirectIndex,
+                               GLuint indirectSwizzle);
+
+extern void
+_slang_copy_ir_storage(slang_ir_storage *dst, const slang_ir_storage *src);
+
 
 extern void
 _slang_free_ir_tree(slang_ir_node *n);
diff --git a/src/mesa/shader/slang/slang_vartable.h b/src/mesa/shader/slang/slang_vartable.h
index 8a3b992c969..94bcd63f45a 100644
--- a/src/mesa/shader/slang/slang_vartable.h
+++ b/src/mesa/shader/slang/slang_vartable.h
@@ -2,7 +2,7 @@
 #ifndef SLANG_VARTABLE_H
 #define SLANG_VARTABLE_H
 
-struct _slang_ir_storage;
+struct slang_ir_storage_;
 
 typedef struct slang_var_table_ slang_var_table;
 
@@ -27,16 +27,16 @@ extern struct slang_variable_ *
 _slang_find_variable(const slang_var_table *t, slang_atom name);
 
 extern GLboolean
-_slang_alloc_var(slang_var_table *t, struct _slang_ir_storage *store);
+_slang_alloc_var(slang_var_table *t, struct slang_ir_storage_ *store);
 
 extern GLboolean
-_slang_alloc_temp(slang_var_table *t, struct _slang_ir_storage *store);
+_slang_alloc_temp(slang_var_table *t, struct slang_ir_storage_ *store);
 
 extern void
-_slang_free_temp(slang_var_table *t, struct _slang_ir_storage *store);
+_slang_free_temp(slang_var_table *t, struct slang_ir_storage_ *store);
 
 extern GLboolean
-_slang_is_temp(const slang_var_table *t, const struct _slang_ir_storage *store);
+_slang_is_temp(const slang_var_table *t, const struct slang_ir_storage_ *store);
 
 
 #endif /* SLANG_VARTABLE_H */