Enable vec4 extension for x86 back-end.

5 files changed, 220 insertions, 37 deletions

diff --git a/src/mesa/shader/slang/slang_assemble.c b/src/mesa/shader/slang/slang_assemble.c
index 4e0395226d..9e5851383c 100644
--- a/src/mesa/shader/slang/slang_assemble.c
+++ b/src/mesa/shader/slang/slang_assemble.c
@@ -652,10 +652,19 @@ static const struct
 	{ "vec4_texcube",   slang_asm_vec4_texcube,   slang_asm_none },
 	{ "vec4_shad1d",    slang_asm_vec4_shad1d,    slang_asm_none },
 	{ "vec4_shad2d",    slang_asm_vec4_shad2d,    slang_asm_none },
-	/* mesa-specific extensions */
+   /* GL_MESA_shader_debug */
 	{ "float_print",    slang_asm_float_deref,    slang_asm_float_print },
 	{ "int_print",      slang_asm_int_deref,      slang_asm_int_print },
 	{ "bool_print",     slang_asm_bool_deref,     slang_asm_bool_print },
+   /* vec4 */
+   { "float_to_vec4",  slang_asm_float_to_vec4,  slang_asm_none },
+   { "vec4_add",       slang_asm_vec4_add,       slang_asm_none },
+   { "vec4_subtract",  slang_asm_vec4_subtract,  slang_asm_none },
+   { "vec4_multiply",  slang_asm_vec4_multiply,  slang_asm_none },
+   { "vec4_divide",    slang_asm_vec4_divide,    slang_asm_none },
+   { "vec4_negate",    slang_asm_vec4_negate,    slang_asm_none },
+   { "vec4_dot",       slang_asm_vec4_dot,       slang_asm_none },
+
 	{ NULL,             slang_asm_none,           slang_asm_none }
 };
 
diff --git a/src/mesa/shader/slang/slang_assemble.h b/src/mesa/shader/slang/slang_assemble.h
index 5c3afcaa5d..fbf88bd6d1 100644
--- a/src/mesa/shader/slang/slang_assemble.h
+++ b/src/mesa/shader/slang/slang_assemble.h
@@ -93,10 +93,19 @@ typedef enum slang_assembly_type_
 	slang_asm_return,
 	slang_asm_discard,
 	slang_asm_exit,
-	/* mesa-specific extensions */
+   /* GL_MESA_shader_debug */
 	slang_asm_float_print,
 	slang_asm_int_print,
 	slang_asm_bool_print,
+   /* vec4 */
+   slang_asm_float_to_vec4,
+   slang_asm_vec4_add,
+   slang_asm_vec4_subtract,
+   slang_asm_vec4_multiply,
+   slang_asm_vec4_divide,
+   slang_asm_vec4_negate,
+   slang_asm_vec4_dot,
+   /* not a real assembly instruction */
 	slang_asm__last
 } slang_assembly_type;
 
diff --git a/src/mesa/shader/slang/slang_compile.c b/src/mesa/shader/slang/slang_compile.c
index 28c4120111..bcbb756182 100644
--- a/src/mesa/shader/slang/slang_compile.c
+++ b/src/mesa/shader/slang/slang_compile.c
@@ -80,9 +80,10 @@ _slang_code_unit_dtr (slang_code_unit *self)
 GLvoid
 _slang_code_object_ctr (slang_code_object *self)
 {
-   _slang_code_unit_ctr (&self->builtin[0], self);
-   _slang_code_unit_ctr (&self->builtin[1], self);
-   _slang_code_unit_ctr (&self->builtin[2], self);
+   GLuint i;
+
+   for (i = 0; i < SLANG_BUILTIN_TOTAL; i++)
+      _slang_code_unit_ctr (&self->builtin[i], self);
    _slang_code_unit_ctr (&self->unit, self);
    _slang_assembly_file_ctr (&self->assembly);
    slang_machine_ctr (&self->machine);
@@ -97,9 +98,10 @@ _slang_code_object_ctr (slang_code_object *self)
 GLvoid
 _slang_code_object_dtr (slang_code_object *self)
 {
-   _slang_code_unit_dtr (&self->builtin[0]);
-   _slang_code_unit_dtr (&self->builtin[1]);
-   _slang_code_unit_dtr (&self->builtin[2]);
+   GLuint i;
+
+   for (i = 0; i < SLANG_BUILTIN_TOTAL; i++)
+      _slang_code_unit_dtr (&self->builtin[i]);
    _slang_code_unit_dtr (&self->unit);
    slang_assembly_file_destruct (&self->assembly);
    slang_machine_dtr (&self->machine);
@@ -1675,18 +1677,22 @@ static int parse_init_declarator (slang_parse_ctx *C, slang_output_ctx *O,
 	}
 
 	/* initialize global variable */
-	if (C->global_scope && var->initializer != NULL)
-	{
-		slang_assemble_ctx A;
-
-		A.file = O->assembly;
-		A.mach = O->machine;
-		A.atoms = C->atoms;
-		A.space.funcs = O->funs;
-		A.space.structs = O->structs;
-		A.space.vars = O->vars;
-		if (!initialize_global (&A, var))
-			return 0;
+   if (C->global_scope) {
+      if (var->initializer != NULL) {
+         slang_assemble_ctx A;
+
+         A.file = O->assembly;
+         A.mach = O->machine;
+         A.atoms = C->atoms;
+         A.space.funcs = O->funs;
+         A.space.structs = O->structs;
+         A.space.vars = O->vars;
+         if (!initialize_global (&A, var))
+            return 0;
+      }
+      else {
+         _mesa_memset ((GLubyte *) (O->machine->mem) + var->address, 0, var->size);
+      }
 	}
 	return 1;
 }
@@ -1880,11 +1886,6 @@ parse_code_unit (slang_parse_ctx *C, slang_code_unit *unit)
 	return 1;
 }
 
-#define BUILTIN_CORE 0
-#define BUILTIN_COMMON 1
-#define BUILTIN_TARGET 2
-#define BUILTIN_TOTAL 3
-
 static GLboolean
 compile_binary (const byte *prod, slang_code_unit *unit, slang_unit_type type,
                 slang_info_log *infolog, slang_code_unit *builtin, slang_code_unit *downlink)
@@ -1935,7 +1936,7 @@ compile_with_grammar (grammar id, const char *source, slang_code_unit *unit, sla
 	}
 
 	/* syntax is okay - translate it to internal representation */
-   if (!compile_binary (prod, unit, type, infolog, builtin, &builtin[BUILTIN_TARGET])) {
+   if (!compile_binary (prod, unit, type, infolog, builtin, &builtin[SLANG_BUILTIN_TOTAL - 1])) {
       grammar_alloc_free (prod);
       return GL_FALSE;
    }
@@ -1964,6 +1965,12 @@ static const byte slang_vertex_builtin_gc[] = {
 #include "library/slang_vertex_builtin_gc.h"
 };
 
+#if defined(USE_X86_ASM) || defined(SLANG_X86)
+static const byte slang_builtin_vec4_gc[] = {
+#include "library/slang_builtin_vec4_gc.h"
+};
+#endif
+
 static GLboolean
 compile_object (grammar *id, const char *source, slang_code_object *object, slang_unit_type type,
                 slang_info_log *infolog)
@@ -1995,32 +2002,40 @@ compile_object (grammar *id, const char *source, slang_code_object *object, slan
 	if (type == slang_unit_fragment_shader || type == slang_unit_vertex_shader)
 	{
 		/* compile core functionality first */
-      if (!compile_binary (slang_core_gc, &object->builtin[BUILTIN_CORE],
+      if (!compile_binary (slang_core_gc, &object->builtin[SLANG_BUILTIN_CORE],
                            slang_unit_fragment_builtin, infolog, NULL, NULL))
          return GL_FALSE;
 
 		/* compile common functions and variables, link to core */
-      if (!compile_binary (slang_common_builtin_gc, &object->builtin[BUILTIN_COMMON],
+      if (!compile_binary (slang_common_builtin_gc, &object->builtin[SLANG_BUILTIN_COMMON],
                            slang_unit_fragment_builtin, infolog, NULL,
-                           &object->builtin[BUILTIN_CORE]))
+                           &object->builtin[SLANG_BUILTIN_CORE]))
          return GL_FALSE;
 
 		/* compile target-specific functions and variables, link to common */
 		if (type == slang_unit_fragment_shader)
 		{
-         if (!compile_binary (slang_fragment_builtin_gc, &object->builtin[BUILTIN_TARGET],
+         if (!compile_binary (slang_fragment_builtin_gc, &object->builtin[SLANG_BUILTIN_TARGET],
                               slang_unit_fragment_builtin, infolog, NULL,
-                              &object->builtin[BUILTIN_COMMON]))
+                              &object->builtin[SLANG_BUILTIN_COMMON]))
             return GL_FALSE;
 		}
 		else if (type == slang_unit_vertex_shader)
 		{
-         if (!compile_binary (slang_vertex_builtin_gc, &object->builtin[BUILTIN_TARGET],
+         if (!compile_binary (slang_vertex_builtin_gc, &object->builtin[SLANG_BUILTIN_TARGET],
                               slang_unit_vertex_builtin, infolog, NULL,
-                              &object->builtin[BUILTIN_COMMON]))
+                              &object->builtin[SLANG_BUILTIN_COMMON]))
             return GL_FALSE;
 		}
 
+#if defined(USE_X86_ASM) || defined(SLANG_X86)
+      /* compile x86 4-component vector overrides, link to target */
+      if (!compile_binary (slang_builtin_vec4_gc, &object->builtin[SLANG_BUILTIN_VEC4],
+                           slang_unit_fragment_builtin, infolog, NULL,
+                           &object->builtin[SLANG_BUILTIN_TARGET]))
+         return GL_FALSE;
+#endif
+
 		/* disable language extensions */
 		grammar_set_reg8 (*id, (const byte *) "parsing_builtin", 0);
       builtins = object->builtin;
diff --git a/src/mesa/shader/slang/slang_compile.h b/src/mesa/shader/slang/slang_compile.h
index 89c3dcfdba..02987f4e1b 100644
--- a/src/mesa/shader/slang/slang_compile.h
+++ b/src/mesa/shader/slang/slang_compile.h
@@ -64,9 +64,20 @@ _slang_code_unit_ctr (slang_code_unit *, struct slang_code_object_ *);
 extern GLvoid
 _slang_code_unit_dtr (slang_code_unit *);
 
+#define SLANG_BUILTIN_CORE   0
+#define SLANG_BUILTIN_COMMON 1
+#define SLANG_BUILTIN_TARGET 2
+
+#if defined(USE_X86_ASM) || defined(SLANG_X86)
+#define SLANG_BUILTIN_VEC4   3
+#define SLANG_BUILTIN_TOTAL  4
+#else
+#define SLANG_BUILTIN_TOTAL  3
+#endif
+
 typedef struct slang_code_object_
 {
-   slang_code_unit builtin[3];
+   slang_code_unit builtin[SLANG_BUILTIN_TOTAL];
    slang_code_unit unit;
    slang_assembly_file assembly;
    slang_machine machine;
diff --git a/src/mesa/shader/slang/slang_execute_x86.c b/src/mesa/shader/slang/slang_execute_x86.c
index 9d2967f2b7..59c3aadcaa 100644
--- a/src/mesa/shader/slang/slang_execute_x86.c
+++ b/src/mesa/shader/slang/slang_execute_x86.c
@@ -56,6 +56,7 @@ typedef struct
 	struct x86_reg r_st1;
 	struct x86_reg r_st2;
 	struct x86_reg r_st3;
+   struct x86_reg r_st4;
 	fixup *fixups;
 	GLuint fixup_count;
 	GLubyte **labels;
@@ -481,7 +482,7 @@ static GLvoid codegen_assem (codegen_ctx *G, slang_assembly *a, slang_info_log *
 	case slang_asm_exit:
 		x86_jmp (&G->f, G->l_exit);
 		break;
-	/* mesa-specific extensions */
+   /* GL_MESA_shader_debug */
    case slang_asm_float_print:
       /* TODO: use push imm32 */
       x86_mov_reg_imm (&G->f, G->r_eax, (GLint) (infolog));
@@ -503,6 +504,138 @@ static GLvoid codegen_assem (codegen_ctx *G, slang_assembly *a, slang_info_log *
 		x86_call (&G->f, (GLubyte *) do_print_bool);
       x86_lea (&G->f, G->r_esp, x86_make_disp (G->r_esp, 4));
 		break;
+   /* vec4 */
+   case slang_asm_float_to_vec4:
+      /* [vec4] | float */
+      x87_fld (&G->f, x86_deref (G->r_esp));
+      x86_lea (&G->f, G->r_esp, x86_make_disp (G->r_esp, 4));
+      x86_mov (&G->f, G->r_eax, x86_deref (G->r_esp));
+      x87_fst (&G->f, x86_make_disp (G->r_eax, 12));
+      x87_fst (&G->f, x86_make_disp (G->r_eax, 8));
+      x87_fst (&G->f, x86_make_disp (G->r_eax, 4));
+      x87_fstp (&G->f, x86_deref (G->r_eax));
+      break;
+   case slang_asm_vec4_add:
+      /* [vec4] | vec4 */
+      x87_fld (&G->f, x86_deref (G->r_esp));
+      x87_fld (&G->f, x86_make_disp (G->r_esp, 4));
+      x87_fld (&G->f, x86_make_disp (G->r_esp, 8));
+      x87_fld (&G->f, x86_make_disp (G->r_esp, 12));
+      x86_lea (&G->f, G->r_esp, x86_make_disp (G->r_esp, 16));
+      x86_mov (&G->f, G->r_eax, x86_deref (G->r_esp));
+      x87_fld (&G->f, x86_deref (G->r_eax));
+      x87_fld (&G->f, x86_make_disp (G->r_eax, 4));
+      x87_fld (&G->f, x86_make_disp (G->r_eax, 8));
+      x87_fld (&G->f, x86_make_disp (G->r_eax, 12));
+      x87_faddp (&G->f, G->r_st4);
+      x87_faddp (&G->f, G->r_st4);
+      x87_faddp (&G->f, G->r_st4);
+      x87_faddp (&G->f, G->r_st4);
+      x87_fstp (&G->f, x86_make_disp (G->r_eax, 12));
+      x87_fstp (&G->f, x86_make_disp (G->r_eax, 8));
+      x87_fstp (&G->f, x86_make_disp (G->r_eax, 4));
+      x87_fstp (&G->f, x86_deref (G->r_eax));
+      break;
+   case slang_asm_vec4_subtract:
+      /* [vec4] | vec4 */
+      x86_mov (&G->f, G->r_eax, x86_make_disp (G->r_esp, 16));
+      x87_fld (&G->f, x86_deref (G->r_eax));
+      x87_fld (&G->f, x86_make_disp (G->r_eax, 4));
+      x87_fld (&G->f, x86_make_disp (G->r_eax, 8));
+      x87_fld (&G->f, x86_make_disp (G->r_eax, 12));
+      x87_fld (&G->f, x86_deref (G->r_esp));
+      x87_fld (&G->f, x86_make_disp (G->r_esp, 4));
+      x87_fld (&G->f, x86_make_disp (G->r_esp, 8));
+      x87_fld (&G->f, x86_make_disp (G->r_esp, 12));
+      x86_lea (&G->f, G->r_esp, x86_make_disp (G->r_esp, 16));
+      x87_fsubp (&G->f, G->r_st4);
+      x87_fsubp (&G->f, G->r_st4);
+      x87_fsubp (&G->f, G->r_st4);
+      x87_fsubp (&G->f, G->r_st4);
+      x87_fstp (&G->f, x86_make_disp (G->r_eax, 12));
+      x87_fstp (&G->f, x86_make_disp (G->r_eax, 8));
+      x87_fstp (&G->f, x86_make_disp (G->r_eax, 4));
+      x87_fstp (&G->f, x86_deref (G->r_eax));
+      break;
+   case slang_asm_vec4_multiply:
+      /* [vec4] | vec4 */
+      x87_fld (&G->f, x86_deref (G->r_esp));
+      x87_fld (&G->f, x86_make_disp (G->r_esp, 4));
+      x87_fld (&G->f, x86_make_disp (G->r_esp, 8));
+      x87_fld (&G->f, x86_make_disp (G->r_esp, 12));
+      x86_lea (&G->f, G->r_esp, x86_make_disp (G->r_esp, 16));
+      x86_mov (&G->f, G->r_eax, x86_deref (G->r_esp));
+      x87_fld (&G->f, x86_deref (G->r_eax));
+      x87_fld (&G->f, x86_make_disp (G->r_eax, 4));
+      x87_fld (&G->f, x86_make_disp (G->r_eax, 8));
+      x87_fld (&G->f, x86_make_disp (G->r_eax, 12));
+      x87_fmulp (&G->f, G->r_st4);
+      x87_fmulp (&G->f, G->r_st4);
+      x87_fmulp (&G->f, G->r_st4);
+      x87_fmulp (&G->f, G->r_st4);
+      x87_fstp (&G->f, x86_make_disp (G->r_eax, 12));
+      x87_fstp (&G->f, x86_make_disp (G->r_eax, 8));
+      x87_fstp (&G->f, x86_make_disp (G->r_eax, 4));
+      x87_fstp (&G->f, x86_deref (G->r_eax));
+      break;
+   case slang_asm_vec4_divide:
+      /* [vec4] | vec4 */
+      x86_mov (&G->f, G->r_eax, x86_make_disp (G->r_esp, 16));
+      x87_fld (&G->f, x86_deref (G->r_eax));
+      x87_fld (&G->f, x86_make_disp (G->r_eax, 4));
+      x87_fld (&G->f, x86_make_disp (G->r_eax, 8));
+      x87_fld (&G->f, x86_make_disp (G->r_eax, 12));
+      x87_fld (&G->f, x86_deref (G->r_esp));
+      x87_fld (&G->f, x86_make_disp (G->r_esp, 4));
+      x87_fld (&G->f, x86_make_disp (G->r_esp, 8));
+      x87_fld (&G->f, x86_make_disp (G->r_esp, 12));
+      x86_lea (&G->f, G->r_esp, x86_make_disp (G->r_esp, 16));
+      x87_fdivp (&G->f, G->r_st4);
+      x87_fdivp (&G->f, G->r_st4);
+      x87_fdivp (&G->f, G->r_st4);
+      x87_fdivp (&G->f, G->r_st4);
+      x87_fstp (&G->f, x86_make_disp (G->r_eax, 12));
+      x87_fstp (&G->f, x86_make_disp (G->r_eax, 8));
+      x87_fstp (&G->f, x86_make_disp (G->r_eax, 4));
+      x87_fstp (&G->f, x86_deref (G->r_eax));
+      break;
+   case slang_asm_vec4_negate:
+      /* [vec4] */
+      x86_mov (&G->f, G->r_eax, x86_deref (G->r_esp));
+      x87_fld (&G->f, x86_deref (G->r_eax));
+      x87_fld (&G->f, x86_make_disp (G->r_eax, 4));
+      x87_fld (&G->f, x86_make_disp (G->r_eax, 8));
+      x87_fld (&G->f, x86_make_disp (G->r_eax, 12));
+      x87_fchs (&G->f);
+      x87_fstp (&G->f, x86_make_disp (G->r_eax, 12));
+      x87_fchs (&G->f);
+      x87_fstp (&G->f, x86_make_disp (G->r_eax, 8));
+      x87_fchs (&G->f);
+      x87_fstp (&G->f, x86_make_disp (G->r_eax, 4));
+      x87_fchs (&G->f);
+      x87_fstp (&G->f, x86_deref (G->r_eax));
+      break;
+   case slang_asm_vec4_dot:
+      /* [vec4] | vec4 */
+      x87_fld (&G->f, x86_deref (G->r_esp));
+      x87_fld (&G->f, x86_make_disp (G->r_esp, 4));
+      x87_fld (&G->f, x86_make_disp (G->r_esp, 8));
+      x87_fld (&G->f, x86_make_disp (G->r_esp, 12));
+      x86_lea (&G->f, G->r_esp, x86_make_disp (G->r_esp, 16));
+      x86_mov (&G->f, G->r_eax, x86_deref (G->r_esp));
+      x87_fld (&G->f, x86_deref (G->r_eax));
+      x87_fld (&G->f, x86_make_disp (G->r_eax, 4));
+      x87_fld (&G->f, x86_make_disp (G->r_eax, 8));
+      x87_fld (&G->f, x86_make_disp (G->r_eax, 12));
+      x87_fmulp (&G->f, G->r_st4);
+      x87_fmulp (&G->f, G->r_st4);
+      x87_fmulp (&G->f, G->r_st4);
+      x87_fmulp (&G->f, G->r_st4);
+      x87_faddp (&G->f, G->r_st1);
+      x87_faddp (&G->f, G->r_st1);
+      x87_faddp (&G->f, G->r_st1);
+      x87_fstp (&G->f, x86_deref (G->r_eax));
+      break;
 	default:
 		assert (0);
 	}
@@ -514,6 +647,13 @@ GLboolean _slang_x86_codegen (slang_machine *mach, slang_assembly_file *file, GL
 	GLubyte *j_body, *j_exit;
 	GLuint i;
 
+   /* Free the old code - if any.
+    */
+   if (mach->x86.compiled_func != NULL) {
+      _mesa_exec_free (mach->x86.compiled_func);
+      mach->x86.compiled_func = NULL;
+   }
+
 	/*
 	 * We need as much as 1M because *all* assembly, including built-in library, is
 	 * being translated to x86.
@@ -530,6 +670,7 @@ GLboolean _slang_x86_codegen (slang_machine *mach, slang_assembly_file *file, GL
 	G.r_st1 = x86_make_reg (file_x87, 1);
 	G.r_st2 = x86_make_reg (file_x87, 2);
 	G.r_st3 = x86_make_reg (file_x87, 3);
+   G.r_st4 = x86_make_reg (file_x87, 4);
 	G.fixups = NULL;
 	G.fixup_count = 0;
 	G.labels = (GLubyte **) slang_alloc_malloc (file->count * sizeof (GLubyte *));
@@ -591,8 +732,6 @@ GLboolean _slang_x86_codegen (slang_machine *mach, slang_assembly_file *file, GL
 	slang_alloc_free (G.labels);
 
 	/* install new code */
-	if (mach->x86.compiled_func != NULL)
-		_mesa_exec_free (mach->x86.compiled_func);
 	mach->x86.compiled_func = (GLvoid (*) (slang_machine *)) x86_get_func (&G.f);
 
 	return GL_TRUE;