/*
* Mesa 3-D graphics library
- * Version: 6.5.3
*
- * Copyright (C) 2005-2007 Brian Paul All Rights Reserved.
+ * Copyright (C) 2005-2008 Brian Paul All Rights Reserved.
+ * Copyright (C) 2008 VMware, Inc. All Rights Reserved.
*
* Permission is hereby granted, free of charge, to any person obtaining a
* copy of this software and associated documentation files (the "Software"),
***/
-#include "imports.h"
-#include "context.h"
-#include "macros.h"
-#include "program.h"
-#include "prog_instruction.h"
-#include "prog_parameter.h"
-#include "prog_print.h"
+#include "main/imports.h"
+#include "main/context.h"
+#include "main/macros.h"
+#include "shader/program.h"
+#include "shader/prog_instruction.h"
+#include "shader/prog_parameter.h"
+#include "shader/prog_print.h"
#include "slang_builtin.h"
#include "slang_emit.h"
+#include "slang_mem.h"
#define PEEPHOLE_OPTIMIZATIONS 1
#define ANNOTATE 0
-/* XXX temporarily here */
-
-
typedef struct
{
slang_info_log *log;
slang_var_table *vt;
struct gl_program *prog;
+ struct gl_program **Subroutines;
+ GLuint NumSubroutines;
+
+ GLuint MaxInstructions; /**< size of prog->Instructions[] buffer */
+
/* code-gen options */
GLboolean EmitHighLevelInstructions;
+ GLboolean EmitCondCodes;
GLboolean EmitComments;
+ GLboolean EmitBeginEndSub; /* XXX TEMPORARY */
} slang_emit_info;
+
+static struct gl_program *
+new_subroutine(slang_emit_info *emitInfo, GLuint *id)
+{
+ GET_CURRENT_CONTEXT(ctx);
+ const GLuint n = emitInfo->NumSubroutines;
+
+ emitInfo->Subroutines = (struct gl_program **)
+ _mesa_realloc(emitInfo->Subroutines,
+ n * sizeof(struct gl_program),
+ (n + 1) * sizeof(struct gl_program));
+ emitInfo->Subroutines[n] = ctx->Driver.NewProgram(ctx, emitInfo->prog->Target, 0);
+ emitInfo->Subroutines[n]->Parameters = emitInfo->prog->Parameters;
+ emitInfo->NumSubroutines++;
+ *id = n;
+ return emitInfo->Subroutines[n];
+}
+
+
/**
- * Assembly and IR info
+ * Convert a writemask to a swizzle. Used for testing cond codes because
+ * we only want to test the cond code component(s) that was set by the
+ * previous instruction.
*/
-typedef struct
-{
- slang_ir_opcode IrOpcode;
- const char *IrName;
- gl_inst_opcode InstOpcode;
- GLuint ResultSize, NumParams;
-} slang_ir_info;
-
-
-
-static const slang_ir_info IrInfo[] = {
- /* binary ops */
- { IR_ADD, "IR_ADD", OPCODE_ADD, 4, 2 },
- { IR_SUB, "IR_SUB", OPCODE_SUB, 4, 2 },
- { IR_MUL, "IR_MUL", OPCODE_MUL, 4, 2 },
- { IR_DIV, "IR_DIV", OPCODE_NOP, 0, 2 }, /* XXX broke */
- { IR_DOT4, "IR_DOT_4", OPCODE_DP4, 1, 2 },
- { IR_DOT3, "IR_DOT_3", OPCODE_DP3, 1, 2 },
- { IR_CROSS, "IR_CROSS", OPCODE_XPD, 3, 2 },
- { IR_LRP, "IR_LRP", OPCODE_LRP, 4, 3 },
- { IR_MIN, "IR_MIN", OPCODE_MIN, 4, 2 },
- { IR_MAX, "IR_MAX", OPCODE_MAX, 4, 2 },
- { IR_CLAMP, "IR_CLAMP", OPCODE_NOP, 4, 3 }, /* special case: emit_clamp() */
- { IR_SEQUAL, "IR_SEQUAL", OPCODE_SEQ, 4, 2 },
- { IR_SNEQUAL, "IR_SNEQUAL", OPCODE_SNE, 4, 2 },
- { IR_SGE, "IR_SGE", OPCODE_SGE, 4, 2 },
- { IR_SGT, "IR_SGT", OPCODE_SGT, 4, 2 },
- { IR_SLE, "IR_SLE", OPCODE_SLE, 4, 2 },
- { IR_SLT, "IR_SLT", OPCODE_SLT, 4, 2 },
- { IR_POW, "IR_POW", OPCODE_POW, 1, 2 },
- /* unary ops */
- { IR_I_TO_F, "IR_I_TO_F", OPCODE_NOP, 1, 1 },
- { IR_F_TO_I, "IR_F_TO_I", OPCODE_INT, 4, 1 }, /* 4 floats to 4 ints */
- { IR_EXP, "IR_EXP", OPCODE_EXP, 1, 1 },
- { IR_EXP2, "IR_EXP2", OPCODE_EX2, 1, 1 },
- { IR_LOG2, "IR_LOG2", OPCODE_LG2, 1, 1 },
- { IR_RSQ, "IR_RSQ", OPCODE_RSQ, 1, 1 },
- { IR_RCP, "IR_RCP", OPCODE_RCP, 1, 1 },
- { IR_FLOOR, "IR_FLOOR", OPCODE_FLR, 4, 1 },
- { IR_FRAC, "IR_FRAC", OPCODE_FRC, 4, 1 },
- { IR_ABS, "IR_ABS", OPCODE_ABS, 4, 1 },
- { IR_NEG, "IR_NEG", OPCODE_NOP, 4, 1 }, /* special case: emit_negation() */
- { IR_DDX, "IR_DDX", OPCODE_DDX, 4, 1 },
- { IR_DDX, "IR_DDY", OPCODE_DDX, 4, 1 },
- { IR_SIN, "IR_SIN", OPCODE_SIN, 1, 1 },
- { IR_COS, "IR_COS", OPCODE_COS, 1, 1 },
- { IR_NOISE1, "IR_NOISE1", OPCODE_NOISE1, 1, 1 },
- { IR_NOISE2, "IR_NOISE2", OPCODE_NOISE2, 1, 1 },
- { IR_NOISE3, "IR_NOISE3", OPCODE_NOISE3, 1, 1 },
- { IR_NOISE4, "IR_NOISE4", OPCODE_NOISE4, 1, 1 },
-
- /* other */
- { IR_SEQ, "IR_SEQ", OPCODE_NOP, 0, 0 },
- { IR_SCOPE, "IR_SCOPE", OPCODE_NOP, 0, 0 },
- { IR_LABEL, "IR_LABEL", OPCODE_NOP, 0, 0 },
- { IR_JUMP, "IR_JUMP", OPCODE_NOP, 0, 0 },
- { IR_IF, "IR_IF", OPCODE_NOP, 0, 0 },
- { IR_KILL, "IR_KILL", OPCODE_NOP, 0, 0 },
- { IR_COND, "IR_COND", OPCODE_NOP, 0, 0 },
- { IR_CALL, "IR_CALL", OPCODE_NOP, 0, 0 },
- { IR_MOVE, "IR_MOVE", OPCODE_NOP, 0, 1 },
- { IR_NOT, "IR_NOT", OPCODE_NOP, 1, 1 },
- { IR_VAR, "IR_VAR", OPCODE_NOP, 0, 0 },
- { IR_VAR_DECL, "IR_VAR_DECL", OPCODE_NOP, 0, 0 },
- { IR_TEX, "IR_TEX", OPCODE_TEX, 4, 1 },
- { IR_TEXB, "IR_TEXB", OPCODE_TXB, 4, 1 },
- { IR_TEXP, "IR_TEXP", OPCODE_TXP, 4, 1 },
- { IR_FLOAT, "IR_FLOAT", OPCODE_NOP, 0, 0 }, /* float literal */
- { IR_FIELD, "IR_FIELD", OPCODE_NOP, 0, 0 },
- { IR_ELEMENT, "IR_ELEMENT", OPCODE_NOP, 0, 0 },
- { IR_SWIZZLE, "IR_SWIZZLE", OPCODE_NOP, 0, 0 },
- { IR_NOP, NULL, OPCODE_NOP, 0, 0 }
-};
-
-
-static const slang_ir_info *
-slang_find_ir_info(slang_ir_opcode opcode)
+static GLuint
+writemask_to_swizzle(GLuint writemask)
{
- GLuint i;
- for (i = 0; IrInfo[i].IrName; i++) {
- if (IrInfo[i].IrOpcode == opcode) {
- return IrInfo + i;
- }
- }
- return NULL;
+ if (writemask == WRITEMASK_X)
+ return SWIZZLE_XXXX;
+ if (writemask == WRITEMASK_Y)
+ return SWIZZLE_YYYY;
+ if (writemask == WRITEMASK_Z)
+ return SWIZZLE_ZZZZ;
+ if (writemask == WRITEMASK_W)
+ return SWIZZLE_WWWW;
+ return SWIZZLE_XYZW; /* shouldn't be hit */
}
-static const char *
-slang_ir_name(slang_ir_opcode opcode)
+
+/**
+ * Convert a swizzle mask to a writemask.
+ * Note that the slang_ir_storage->Swizzle field can represent either a
+ * swizzle mask or a writemask, depending on how it's used. For example,
+ * when we parse "direction.yz" alone, we don't know whether .yz is a
+ * writemask or a swizzle. In this case, we encode ".yz" in store->Swizzle
+ * as a swizzle mask (.yz?? actually). Later, if direction.yz is used as
+ * an R-value, we use store->Swizzle as-is. Otherwise, if direction.yz is
+ * used as an L-value, we convert it to a writemask.
+ */
+static GLuint
+swizzle_to_writemask(GLuint swizzle)
{
- return slang_find_ir_info(opcode)->IrName;
+ GLuint i, writemask = 0x0;
+ for (i = 0; i < 4; i++) {
+ GLuint swz = GET_SWZ(swizzle, i);
+ if (swz <= SWIZZLE_W) {
+ writemask |= (1 << swz);
+ }
+ }
+ return writemask;
}
/**
- * Swizzle a swizzle. That is, return swz2(swz1)
+ * Swizzle a swizzle (function composition).
+ * That is, return swz2(swz1), or said another way: swz1.szw2
+ * Example: swizzle_swizzle(".zwxx", ".xxyw") yields ".zzwx"
*/
-static GLuint
-swizzle_swizzle(GLuint swz1, GLuint swz2)
+GLuint
+_slang_swizzle_swizzle(GLuint swz1, GLuint swz2)
{
GLuint i, swz, s[4];
for (i = 0; i < 4; i++) {
GLuint c = GET_SWZ(swz2, i);
- s[i] = GET_SWZ(swz1, c);
+ if (c <= SWIZZLE_W)
+ s[i] = GET_SWZ(swz1, c);
+ else
+ s[i] = c;
}
swz = MAKE_SWIZZLE4(s[0], s[1], s[2], s[3]);
return swz;
}
-slang_ir_storage *
-_slang_new_ir_storage(enum register_file file, GLint index, GLint size)
+/**
+ * Return the default swizzle mask for accessing a variable of the
+ * given size (in floats). If size = 1, comp is used to identify
+ * which component [0..3] of the register holds the variable.
+ */
+GLuint
+_slang_var_swizzle(GLint size, GLint comp)
{
- slang_ir_storage *st;
- st = (slang_ir_storage *) _mesa_calloc(sizeof(slang_ir_storage));
- if (st) {
- st->File = file;
- st->Index = index;
- st->Size = size;
- st->Swizzle = SWIZZLE_NOOP;
- }
- return st;
+ switch (size) {
+ case 1:
+ return MAKE_SWIZZLE4(comp, comp, comp, comp);
+ case 2:
+ return MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_Y, SWIZZLE_NIL, SWIZZLE_NIL);
+ case 3:
+ return MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_Y, SWIZZLE_Z, SWIZZLE_NIL);
+ default:
+ return SWIZZLE_XYZW;
+ }
}
-static const char *
-swizzle_string(GLuint swizzle)
-{
- static char s[6];
- GLuint i;
- s[0] = '.';
- for (i = 1; i < 5; i++) {
- s[i] = "xyzw"[GET_SWZ(swizzle, i-1)];
- }
- s[i] = 0;
- return s;
-}
-static const char *
-writemask_string(GLuint writemask)
+/**
+ * Allocate storage for the given node (if it hasn't already been allocated).
+ *
+ * Typically this is temporary storage for an intermediate result (such as
+ * for a multiply or add, etc).
+ *
+ * If n->Store does not exist it will be created and will be of the size
+ * specified by defaultSize.
+ */
+static GLboolean
+alloc_node_storage(slang_emit_info *emitInfo, slang_ir_node *n,
+ GLint defaultSize)
{
- static char s[6];
- GLuint i, j = 0;
- s[j++] = '.';
- for (i = 0; i < 4; i++) {
- if (writemask & (1 << i))
- s[j++] = "xyzw"[i];
+ assert(!n->Var);
+ if (!n->Store) {
+ assert(defaultSize > 0);
+ n->Store = _slang_new_ir_storage(PROGRAM_TEMPORARY, -1, defaultSize);
}
- s[j] = 0;
- return s;
-}
-static const char *
-storage_string(const slang_ir_storage *st)
-{
- static const char *files[] = {
- "TEMP",
- "LOCAL_PARAM",
- "ENV_PARAM",
- "STATE",
- "INPUT",
- "OUTPUT",
- "NAMED_PARAM",
- "CONSTANT",
- "UNIFORM",
- "WRITE_ONLY",
- "ADDRESS",
- "SAMPLER",
- "UNDEFINED"
- };
- static char s[100];
-#if 0
- if (st->Size == 1)
- sprintf(s, "%s[%d]", files[st->File], st->Index);
- else
- sprintf(s, "%s[%d..%d]", files[st->File], st->Index,
- st->Index + st->Size - 1);
-#endif
- assert(st->File < (GLint) (sizeof(files) / sizeof(files[0])));
- sprintf(s, "%s[%d]", files[st->File], st->Index);
- return s;
+ /* now allocate actual register(s). I.e. set n->Store->Index >= 0 */
+ if (n->Store->Index < 0) {
+ if (!_slang_alloc_temp(emitInfo->vt, n->Store)) {
+ slang_info_log_error(emitInfo->log,
+ "Ran out of registers, too many temporaries");
+ _slang_free(n->Store);
+ n->Store = NULL;
+ return GL_FALSE;
+ }
+ }
+ return GL_TRUE;
}
+/**
+ * Free temporary storage, if n->Store is, in fact, temp storage.
+ * Otherwise, no-op.
+ */
static void
-spaces(int n)
+free_node_storage(slang_var_table *vt, slang_ir_node *n)
{
- while (n-- > 0) {
- printf(" ");
- }
-}
-
-#define IND 0
-void
-slang_print_ir(const slang_ir_node *n, int indent)
-{
- if (!n)
- return;
-#if !IND
- if (n->Opcode != IR_SEQ)
-#else
- printf("%3d:", indent);
-#endif
- spaces(indent);
-
- switch (n->Opcode) {
- case IR_SEQ:
-#if IND
- printf("SEQ at %p\n", (void*) n);
-#endif
- assert(n->Children[0]);
- assert(n->Children[1]);
- slang_print_ir(n->Children[0], indent + IND);
- slang_print_ir(n->Children[1], indent + IND);
- break;
- case IR_SCOPE:
- printf("NEW SCOPE\n");
- assert(!n->Children[1]);
- slang_print_ir(n->Children[0], indent + 3);
- break;
- case IR_MOVE:
- printf("MOVE (writemask = %s)\n", writemask_string(n->Writemask));
- slang_print_ir(n->Children[0], indent+3);
- slang_print_ir(n->Children[1], indent+3);
- break;
- case IR_LABEL:
- printf("LABEL: %s\n", n->Label->Name);
- break;
- case IR_COND:
- printf("COND\n");
- slang_print_ir(n->Children[0], indent + 3);
- break;
- case IR_JUMP:
- printf("JUMP %s\n", n->Label->Name);
- break;
-
- case IR_IF:
- printf("IF \n");
- slang_print_ir(n->Children[0], indent+3);
- spaces(indent);
- printf("THEN\n");
- slang_print_ir(n->Children[1], indent+3);
- if (n->Children[2]) {
- spaces(indent);
- printf("ELSE\n");
- slang_print_ir(n->Children[2], indent+3);
+ if (n->Store->File == PROGRAM_TEMPORARY &&
+ n->Store->Index >= 0 &&
+ n->Opcode != IR_SWIZZLE) {
+ if (_slang_is_temp(vt, n->Store)) {
+ _slang_free_temp(vt, n->Store);
+ n->Store->Index = -1;
+ n->Store = NULL; /* XXX this may not be needed */
}
- printf("ENDIF\n");
- break;
-
- case IR_BEGIN_SUB:
- printf("BEGIN_SUB\n");
- break;
- case IR_END_SUB:
- printf("END_SUB\n");
- break;
- case IR_RETURN:
- printf("RETURN\n");
- break;
- case IR_CALL:
- printf("CALL\n");
- break;
-
- case IR_LOOP:
- printf("LOOP\n");
- slang_print_ir(n->Children[0], indent+3);
- spaces(indent);
- printf("ENDLOOP\n");
- break;
- case IR_CONT:
- printf("CONT\n");
- break;
- case IR_BREAK:
- printf("BREAK\n");
- break;
- case IR_BREAK_IF_FALSE:
- printf("BREAK_IF_FALSE\n");
- slang_print_ir(n->Children[0], indent+3);
- break;
- case IR_BREAK_IF_TRUE:
- printf("BREAK_IF_TRUE\n");
- slang_print_ir(n->Children[0], indent+3);
- break;
- case IR_CONT_IF_FALSE:
- printf("CONT_IF_FALSE\n");
- slang_print_ir(n->Children[0], indent+3);
- break;
- case IR_CONT_IF_TRUE:
- printf("CONT_IF_TRUE\n");
- slang_print_ir(n->Children[0], indent+3);
- break;
-
- case IR_VAR:
- printf("VAR %s%s at %s store %p\n",
- (n->Var ? (char *) n->Var->a_name : "TEMP"),
- swizzle_string(n->Store->Swizzle),
- storage_string(n->Store), (void*) n->Store);
- break;
- case IR_VAR_DECL:
- printf("VAR_DECL %s (%p) at %s store %p\n",
- (n->Var ? (char *) n->Var->a_name : "TEMP"),
- (void*) n->Var, storage_string(n->Store),
- (void*) n->Store);
- break;
- case IR_FIELD:
- printf("FIELD %s of\n", n->Field);
- slang_print_ir(n->Children[0], indent+3);
- break;
- case IR_FLOAT:
- printf("FLOAT %g %g %g %g\n",
- n->Value[0], n->Value[1], n->Value[2], n->Value[3]);
- break;
- case IR_I_TO_F:
- printf("INT_TO_FLOAT\n");
- slang_print_ir(n->Children[0], indent+3);
- break;
- case IR_F_TO_I:
- printf("FLOAT_TO_INT\n");
- slang_print_ir(n->Children[0], indent+3);
- break;
- case IR_SWIZZLE:
- printf("SWIZZLE %s of (store %p) \n",
- swizzle_string(n->Store->Swizzle), (void*) n->Store);
- slang_print_ir(n->Children[0], indent + 3);
- break;
- default:
- printf("%s (%p, %p) (store %p)\n", slang_ir_name(n->Opcode),
- (void*) n->Children[0], (void*) n->Children[1], (void*) n->Store);
- slang_print_ir(n->Children[0], indent+3);
- slang_print_ir(n->Children[1], indent+3);
}
}
/**
- * Allocate temporary storage for an intermediate result (such as for
- * a multiply or add, etc.
+ * Helper function to allocate a short-term temporary.
+ * Free it with _slang_free_temp().
*/
static GLboolean
-alloc_temp_storage(slang_emit_info *emitInfo, slang_ir_node *n, GLint size)
+alloc_local_temp(slang_emit_info *emitInfo, slang_ir_storage *temp, GLint size)
{
- assert(!n->Var);
- assert(!n->Store);
- assert(size > 0);
- n->Store = _slang_new_ir_storage(PROGRAM_TEMPORARY, -1, size);
- if (!_slang_alloc_temp(emitInfo->vt, n->Store)) {
- slang_info_log_error(emitInfo->log,
- "Ran out of registers, too many temporaries");
- return GL_FALSE;
- }
- return GL_TRUE;
+ assert(size >= 1);
+ assert(size <= 4);
+ _mesa_bzero(temp, sizeof(*temp));
+ temp->Size = size;
+ temp->File = PROGRAM_TEMPORARY;
+ temp->Index = -1;
+ return _slang_alloc_temp(emitInfo->vt, temp);
}
/**
- * Free temporary storage, if n->Store is, in fact, temp storage.
- * Otherwise, no-op.
+ * Remove any SWIZZLE_NIL terms from given swizzle mask.
+ * For a swizzle like .z??? generate .zzzz (replicate single component).
+ * Else, for .wx?? generate .wxzw (insert default component for the position).
*/
-static void
-free_temp_storage(slang_var_table *vt, slang_ir_node *n)
+static GLuint
+fix_swizzle(GLuint swizzle)
{
- if (n->Store->File == PROGRAM_TEMPORARY && n->Store->Index >= 0) {
- if (_slang_is_temp(vt, n->Store)) {
- _slang_free_temp(vt, n->Store);
- n->Store->Index = -1;
- n->Store->Size = -1;
- }
+ GLuint c0 = GET_SWZ(swizzle, 0),
+ c1 = GET_SWZ(swizzle, 1),
+ c2 = GET_SWZ(swizzle, 2),
+ c3 = GET_SWZ(swizzle, 3);
+ if (c1 == SWIZZLE_NIL && c2 == SWIZZLE_NIL && c3 == SWIZZLE_NIL) {
+ /* smear first component across all positions */
+ c1 = c2 = c3 = c0;
}
+ else {
+ /* insert default swizzle components */
+ if (c0 == SWIZZLE_NIL)
+ c0 = SWIZZLE_X;
+ if (c1 == SWIZZLE_NIL)
+ c1 = SWIZZLE_Y;
+ if (c2 == SWIZZLE_NIL)
+ c2 = SWIZZLE_Z;
+ if (c3 == SWIZZLE_NIL)
+ c3 = SWIZZLE_W;
+ }
+ return MAKE_SWIZZLE4(c0, c1, c2, c3);
}
+
/**
* Convert IR storage to an instruction dst register.
*/
static void
-storage_to_dst_reg(struct prog_dst_register *dst, const slang_ir_storage *st,
- GLuint writemask)
+storage_to_dst_reg(struct prog_dst_register *dst, const slang_ir_storage *st)
{
- static const GLuint defaultWritemask[4] = {
- WRITEMASK_X,
- WRITEMASK_X | WRITEMASK_Y,
- WRITEMASK_X | WRITEMASK_Y | WRITEMASK_Z,
- WRITEMASK_X | WRITEMASK_Y | WRITEMASK_Z | WRITEMASK_W
- };
- assert(st->Index >= 0);
- dst->File = st->File;
- dst->Index = st->Index;
+ 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);
+ }
+
assert(st->File != PROGRAM_UNDEFINED);
- assert(st->Size >= 1);
- assert(st->Size <= 4);
- if (st->Size == 1) {
- GLuint comp = GET_SWZ(st->Swizzle, 0);
- assert(comp < 4);
- assert(writemask & WRITEMASK_X);
- dst->WriteMask = WRITEMASK_X << comp;
+ dst->File = st->File;
+
+ assert(index >= 0);
+ dst->Index = index;
+
+ assert(size >= 1);
+ assert(size <= 4);
+
+ if (swizzle != SWIZZLE_XYZW) {
+ dst->WriteMask = swizzle_to_writemask(swizzle);
}
else {
- dst->WriteMask = defaultWritemask[st->Size - 1] & writemask;
+ GLuint writemask;
+ switch (size) {
+ case 1:
+ writemask = WRITEMASK_X << GET_SWZ(st->Swizzle, 0);
+ break;
+ case 2:
+ writemask = WRITEMASK_XY;
+ break;
+ case 3:
+ writemask = WRITEMASK_XYZ;
+ break;
+ case 4:
+ writemask = WRITEMASK_XYZW;
+ break;
+ default:
+ ; /* error would have been caught above */
+ }
+ dst->WriteMask = writemask;
}
+
+ dst->RelAddr = relAddr;
}
static void
storage_to_src_reg(struct prog_src_register *src, const slang_ir_storage *st)
{
- static const GLuint defaultSwizzle[4] = {
- MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_X, SWIZZLE_X, SWIZZLE_X),
- MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_Y, SWIZZLE_Z, SWIZZLE_W),
- MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_Y, SWIZZLE_Z, SWIZZLE_W),
- MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_Y, SWIZZLE_Z, SWIZZLE_W)
- };
+ const GLboolean relAddr = st->RelAddr;
+ GLint index = st->Index;
+ 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);
+ }
+
assert(st->File >= 0);
+#if 1 /* XXX temporary */
+ if (st->File == PROGRAM_UNDEFINED) {
+ slang_ir_storage *st0 = (slang_ir_storage *) st;
+ st0->File = PROGRAM_TEMPORARY;
+ }
+#endif
assert(st->File < PROGRAM_UNDEFINED);
- assert(st->Size >= 1);
- assert(st->Size <= 4);
src->File = st->File;
- src->Index = st->Index;
- if (st->Swizzle != SWIZZLE_NOOP)
- src->Swizzle = st->Swizzle;
- else
- src->Swizzle = defaultSwizzle[st->Size - 1]; /*XXX really need this?*/
- assert(GET_SWZ(src->Swizzle, 0) != SWIZZLE_NIL);
- assert(GET_SWZ(src->Swizzle, 1) != SWIZZLE_NIL);
- assert(GET_SWZ(src->Swizzle, 2) != SWIZZLE_NIL);
- assert(GET_SWZ(src->Swizzle, 3) != SWIZZLE_NIL);
+ assert(index >= 0);
+ src->Index = index;
+
+ swizzle = fix_swizzle(swizzle);
+ assert(GET_SWZ(swizzle, 0) <= SWIZZLE_W);
+ assert(GET_SWZ(swizzle, 1) <= SWIZZLE_W);
+ assert(GET_SWZ(swizzle, 2) <= SWIZZLE_W);
+ assert(GET_SWZ(swizzle, 3) <= SWIZZLE_W);
+ src->Swizzle = swizzle;
+
+ src->RelAddr = relAddr;
}
+/*
+ * Setup storage pointing to a scalar constant/literal.
+ */
+static void
+constant_to_storage(slang_emit_info *emitInfo,
+ GLfloat val,
+ slang_ir_storage *store)
+{
+ GLuint swizzle;
+ GLint reg;
+ GLfloat value[4];
+
+ value[0] = val;
+ reg = _mesa_add_unnamed_constant(emitInfo->prog->Parameters,
+ value, 1, &swizzle);
+
+ memset(store, 0, sizeof(*store));
+ store->File = PROGRAM_CONSTANT;
+ store->Index = reg;
+ store->Swizzle = swizzle;
+}
+
/**
* Add new instruction at end of given program.
{
struct gl_program *prog = emitInfo->prog;
struct prog_instruction *inst;
- prog->Instructions = _mesa_realloc_instructions(prog->Instructions,
- prog->NumInstructions,
- prog->NumInstructions + 1);
+
+#if 0
+ /* print prev inst */
+ if (prog->NumInstructions > 0) {
+ _mesa_print_instruction(prog->Instructions + prog->NumInstructions - 1);
+ }
+#endif
+ 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);
}
-#if 0
+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)
+{
+ 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];
+ }
+ }
+ }
+ }
+
+ /* 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;
+ }
+
+ /* 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);
+
+ for (i = 0; i < 3; i++) {
+ if (src[i])
+ storage_to_src_reg(&inst->SrcReg[i], src[i]);
+ }
+
+ /* Free any temp registers that we allocated above */
+ for (i = 0; i < 3; i++) {
+ if (isTemp[i])
+ _slang_free_temp(emitInfo->vt, &newSrc[i]);
+ }
+
+ return inst;
+}
+
+
+
+/**
+ * Put a comment on the given instruction.
+ */
+static void
+inst_comment(struct prog_instruction *inst, const char *comment)
+{
+ if (inst)
+ inst->Comment = _mesa_strdup(comment);
+}
+
+
+
/**
* Return pointer to last instruction in program.
*/
static struct prog_instruction *
-prev_instruction(struct gl_program *prog)
+prev_instruction(slang_emit_info *emitInfo)
{
+ struct gl_program *prog = emitInfo->prog;
if (prog->NumInstructions == 0)
return NULL;
else
return prog->Instructions + prog->NumInstructions - 1;
}
-#endif
static struct prog_instruction *
case OPCODE_MUL:
operator = "*";
break;
+ case OPCODE_DP2:
+ operator = "DP2";
+ break;
case OPCODE_DP3:
operator = "DP3";
break;
}
+/**
+ * Emit an instruction that's just a comment.
+ */
+static struct prog_instruction *
+emit_comment(slang_emit_info *emitInfo, const char *comment)
+{
+ struct prog_instruction *inst = new_instruction(emitInfo, OPCODE_NOP);
+ inst_comment(inst, comment);
+ return inst;
+}
+
/**
* Generate code for a simple arithmetic instruction.
static struct prog_instruction *
emit_arith(slang_emit_info *emitInfo, slang_ir_node *n)
{
+ const slang_ir_info *info = _slang_ir_info(n->Opcode);
struct prog_instruction *inst;
- const slang_ir_info *info = slang_find_ir_info(n->Opcode);
- char *srcAnnot[3], *dstAnnot;
GLuint i;
assert(info);
assert(info->InstOpcode != OPCODE_NOP);
- srcAnnot[0] = srcAnnot[1] = srcAnnot[2] = dstAnnot = NULL;
-
#if PEEPHOLE_OPTIMIZATIONS
/* Look for MAD opportunity */
if (info->NumParams == 2 &&
emit(emitInfo, n->Children[0]->Children[0]); /* A */
emit(emitInfo, n->Children[0]->Children[1]); /* B */
emit(emitInfo, n->Children[1]); /* C */
- /* generate MAD instruction */
- inst = new_instruction(emitInfo, OPCODE_MAD);
- /* operands: A, B, C: */
- storage_to_src_reg(&inst->SrcReg[0], n->Children[0]->Children[0]->Store);
- storage_to_src_reg(&inst->SrcReg[1], n->Children[0]->Children[1]->Store);
- storage_to_src_reg(&inst->SrcReg[2], n->Children[1]->Store);
- free_temp_storage(emitInfo->vt, n->Children[0]->Children[0]);
- free_temp_storage(emitInfo->vt, n->Children[0]->Children[1]);
- free_temp_storage(emitInfo->vt, n->Children[1]);
- }
- else if (info->NumParams == 2 &&
- n->Opcode == IR_ADD && n->Children[1]->Opcode == IR_MUL) {
+ alloc_node_storage(emitInfo, n, -1); /* dest */
+
+ inst = emit_instruction(emitInfo,
+ OPCODE_MAD,
+ n->Store,
+ n->Children[0]->Children[0]->Store,
+ n->Children[0]->Children[1]->Store,
+ n->Children[1]->Store);
+
+ free_node_storage(emitInfo->vt, n->Children[0]->Children[0]);
+ free_node_storage(emitInfo->vt, n->Children[0]->Children[1]);
+ free_node_storage(emitInfo->vt, n->Children[1]);
+ return inst;
+ }
+
+ if (info->NumParams == 2 &&
+ n->Opcode == IR_ADD && n->Children[1]->Opcode == IR_MUL) {
/* found pattern IR_ADD(A, IR_MUL(B, C)) */
emit(emitInfo, n->Children[0]); /* A */
emit(emitInfo, n->Children[1]->Children[0]); /* B */
emit(emitInfo, n->Children[1]->Children[1]); /* C */
- /* generate MAD instruction */
- inst = new_instruction(emitInfo, OPCODE_MAD);
- /* operands: B, C, A */
- storage_to_src_reg(&inst->SrcReg[0], n->Children[1]->Children[0]->Store);
- storage_to_src_reg(&inst->SrcReg[1], n->Children[1]->Children[1]->Store);
- storage_to_src_reg(&inst->SrcReg[2], n->Children[0]->Store);
- free_temp_storage(emitInfo->vt, n->Children[1]->Children[0]);
- free_temp_storage(emitInfo->vt, n->Children[1]->Children[1]);
- free_temp_storage(emitInfo->vt, n->Children[0]);
+ alloc_node_storage(emitInfo, n, -1); /* dest */
+
+ inst = emit_instruction(emitInfo,
+ OPCODE_MAD,
+ n->Store,
+ n->Children[1]->Children[0]->Store,
+ n->Children[1]->Children[1]->Store,
+ n->Children[0]->Store);
+
+ free_node_storage(emitInfo->vt, n->Children[1]->Children[0]);
+ free_node_storage(emitInfo->vt, n->Children[1]->Children[1]);
+ free_node_storage(emitInfo->vt, n->Children[0]);
+ return inst;
}
- else
#endif
- {
- /* normal case */
- /* gen code for children */
- for (i = 0; i < info->NumParams; i++)
- emit(emitInfo, n->Children[i]);
+ /* gen code for children, may involve temp allocation */
+ for (i = 0; i < info->NumParams; i++) {
+ emit(emitInfo, n->Children[i]);
+ if (!n->Children[i] || !n->Children[i]->Store) {
+ /* error recovery */
+ return NULL;
+ }
+ }
+
+ /* result storage */
+ alloc_node_storage(emitInfo, n, -1);
+
+ inst = emit_instruction(emitInfo,
+ info->InstOpcode,
+ n->Store, /* dest */
+ (info->NumParams > 0 ? n->Children[0]->Store : NULL),
+ (info->NumParams > 1 ? n->Children[1]->Store : NULL),
+ (info->NumParams > 2 ? n->Children[2]->Store : NULL)
+ );
+
+ /* free temps */
+ for (i = 0; i < info->NumParams; i++)
+ free_node_storage(emitInfo->vt, n->Children[i]);
+
+ return inst;
+}
+
+
+/**
+ * Emit code for == and != operators. These could normally be handled
+ * by emit_arith() except we need to be able to handle structure comparisons.
+ */
+static struct prog_instruction *
+emit_compare(slang_emit_info *emitInfo, slang_ir_node *n)
+{
+ struct prog_instruction *inst = NULL;
+ GLint size;
+
+ assert(n->Opcode == IR_EQUAL || n->Opcode == IR_NOTEQUAL);
+
+ /* gen code for children */
+ emit(emitInfo, n->Children[0]);
+ emit(emitInfo, n->Children[1]);
+
+ if (n->Children[0]->Store->Size != n->Children[1]->Store->Size) {
+ slang_info_log_error(emitInfo->log, "invalid operands to == or !=");
+ return NULL;
+ }
- /* gen this instruction and src registers */
- inst = new_instruction(emitInfo, info->InstOpcode);
- for (i = 0; i < info->NumParams; i++)
- storage_to_src_reg(&inst->SrcReg[i], n->Children[i]->Store);
+ /* final result is 1 bool */
+ if (!alloc_node_storage(emitInfo, n, 1))
+ return NULL;
- /* annotation */
- for (i = 0; i < info->NumParams; i++)
- srcAnnot[i] = storage_annotation(n->Children[i], emitInfo->prog);
+ size = n->Children[0]->Store->Size;
- /* free temps */
- for (i = 0; i < info->NumParams; i++)
- free_temp_storage(emitInfo->vt, n->Children[i]);
+ if (size == 1) {
+ gl_inst_opcode opcode = n->Opcode == IR_EQUAL ? OPCODE_SEQ : OPCODE_SNE;
+ inst = emit_instruction(emitInfo,
+ opcode,
+ n->Store, /* dest */
+ n->Children[0]->Store,
+ n->Children[1]->Store,
+ NULL);
}
+ else if (size <= 4) {
+ /* compare two vectors.
+ * Unfortunately, there's no instruction to compare vectors and
+ * return a scalar result. Do it with some compare and dot product
+ * instructions...
+ */
+ GLuint swizzle;
+ gl_inst_opcode dotOp;
+ slang_ir_storage tempStore;
- /* result storage */
- if (!n->Store) {
- if (!alloc_temp_storage(emitInfo, n, info->ResultSize))
+ if (!alloc_local_temp(emitInfo, &tempStore, 4)) {
return NULL;
+ /* out of temps */
+ }
+
+ if (size == 4) {
+ dotOp = OPCODE_DP4;
+ swizzle = SWIZZLE_XYZW;
+ }
+ else if (size == 3) {
+ dotOp = OPCODE_DP3;
+ swizzle = SWIZZLE_XYZW;
+ }
+ else {
+ assert(size == 2);
+ dotOp = OPCODE_DP3; /* XXX use OPCODE_DP2 eventually */
+ swizzle = MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_Y, SWIZZLE_Y, SWIZZLE_Y);
+ }
+
+ /* Compute inequality (temp = (A != B)) */
+ inst = emit_instruction(emitInfo,
+ OPCODE_SNE,
+ &tempStore,
+ n->Children[0]->Store,
+ n->Children[1]->Store,
+ NULL);
+ inst_comment(inst, "Compare values");
+
+ /* Compute val = DOT(temp, temp) (reduction) */
+ inst = emit_instruction(emitInfo,
+ dotOp,
+ n->Store,
+ &tempStore,
+ &tempStore,
+ NULL);
+ inst->SrcReg[0].Swizzle = inst->SrcReg[1].Swizzle = swizzle; /*override*/
+ inst_comment(inst, "Reduce vec to bool");
+
+ _slang_free_temp(emitInfo->vt, &tempStore); /* free temp */
+
+ if (n->Opcode == IR_EQUAL) {
+ /* compute val = !val.x with SEQ val, val, 0; */
+ slang_ir_storage zero;
+ constant_to_storage(emitInfo, 0.0, &zero);
+ inst = emit_instruction(emitInfo,
+ OPCODE_SEQ,
+ n->Store, /* dest */
+ n->Store,
+ &zero,
+ NULL);
+ inst_comment(inst, "Invert true/false");
+ }
}
- storage_to_dst_reg(&inst->DstReg, n->Store, n->Writemask);
+ else {
+ /* size > 4, struct or array compare.
+ * XXX this won't work reliably for structs with padding!!
+ */
+ GLint i, num = (n->Children[0]->Store->Size + 3) / 4;
+ slang_ir_storage accTemp, sneTemp;
+
+ if (!alloc_local_temp(emitInfo, &accTemp, 4))
+ return NULL;
- dstAnnot = storage_annotation(n, emitInfo->prog);
+ if (!alloc_local_temp(emitInfo, &sneTemp, 4))
+ return NULL;
+
+ for (i = 0; i < num; i++) {
+ slang_ir_storage srcStore0 = *n->Children[0]->Store;
+ slang_ir_storage srcStore1 = *n->Children[1]->Store;
+ srcStore0.Index += i;
+ srcStore1.Index += i;
+
+ if (i == 0) {
+ /* SNE accTemp, left[i], right[i] */
+ inst = emit_instruction(emitInfo, OPCODE_SNE,
+ &accTemp, /* dest */
+ &srcStore0,
+ &srcStore1,
+ NULL);
+ inst_comment(inst, "Begin struct/array comparison");
+ }
+ else {
+ /* SNE sneTemp, left[i], right[i] */
+ inst = emit_instruction(emitInfo, OPCODE_SNE,
+ &sneTemp, /* dest */
+ &srcStore0,
+ &srcStore1,
+ NULL);
+ /* ADD accTemp, accTemp, sneTemp; # like logical-OR */
+ inst = emit_instruction(emitInfo, OPCODE_ADD,
+ &accTemp, /* dest */
+ &accTemp,
+ &sneTemp,
+ NULL);
+ }
+ }
- inst->Comment = instruction_annotation(inst->Opcode, dstAnnot, srcAnnot[0],
- srcAnnot[1], srcAnnot[2]);
+ /* compute accTemp.x || accTemp.y || accTemp.z || accTemp.w with DOT4 */
+ inst = emit_instruction(emitInfo, OPCODE_DP4,
+ n->Store,
+ &accTemp,
+ &accTemp,
+ NULL);
+ inst_comment(inst, "End struct/array comparison");
+
+ if (n->Opcode == IR_EQUAL) {
+ /* compute tmp.x = !tmp.x via tmp.x = (tmp.x == 0) */
+ slang_ir_storage zero;
+ constant_to_storage(emitInfo, 0.0, &zero);
+ inst = emit_instruction(emitInfo, OPCODE_SEQ,
+ n->Store, /* dest */
+ n->Store,
+ &zero,
+ NULL);
+ inst_comment(inst, "Invert true/false");
+ }
+
+ _slang_free_temp(emitInfo->vt, &accTemp);
+ _slang_free_temp(emitInfo->vt, &sneTemp);
+ }
+
+ /* free temps */
+ free_node_storage(emitInfo->vt, n->Children[0]);
+ free_node_storage(emitInfo->vt, n->Children[1]);
- /*_mesa_print_instruction(inst);*/
return inst;
}
+
/**
* Generate code for an IR_CLAMP instruction.
*/
emit_clamp(slang_emit_info *emitInfo, slang_ir_node *n)
{
struct prog_instruction *inst;
+ slang_ir_node tmpNode;
assert(n->Opcode == IR_CLAMP);
/* ch[0] = value
}
#endif
- if (!n->Store)
- if (!alloc_temp_storage(emitInfo, n, n->Children[0]->Store->Size))
- return NULL;
+ if (!alloc_node_storage(emitInfo, n, n->Children[0]->Store->Size))
+ return NULL;
emit(emitInfo, n->Children[1]);
emit(emitInfo, n->Children[2]);
+ /* Some GPUs don't allow reading from output registers. So if the
+ * dest for this clamp() is an output reg, we can't use that reg for
+ * the intermediate result. Use a temp register instead.
+ */
+ _mesa_bzero(&tmpNode, sizeof(tmpNode));
+ alloc_node_storage(emitInfo, &tmpNode, n->Store->Size);
+
/* tmp = max(ch[0], ch[1]) */
- inst = new_instruction(emitInfo, OPCODE_MAX);
- storage_to_dst_reg(&inst->DstReg, n->Store, n->Writemask);
- storage_to_src_reg(&inst->SrcReg[0], n->Children[0]->Store);
- storage_to_src_reg(&inst->SrcReg[1], n->Children[1]->Store);
+ inst = emit_instruction(emitInfo, OPCODE_MAX,
+ tmpNode.Store, /* dest */
+ n->Children[0]->Store,
+ n->Children[1]->Store,
+ NULL);
- /* tmp = min(tmp, ch[2]) */
- inst = new_instruction(emitInfo, OPCODE_MIN);
- storage_to_dst_reg(&inst->DstReg, n->Store, n->Writemask);
- storage_to_src_reg(&inst->SrcReg[0], n->Store);
- storage_to_src_reg(&inst->SrcReg[1], n->Children[2]->Store);
+ /* n->dest = min(tmp, ch[2]) */
+ inst = emit_instruction(emitInfo, OPCODE_MIN,
+ n->Store, /* dest */
+ tmpNode.Store,
+ n->Children[2]->Store,
+ NULL);
+
+ free_node_storage(emitInfo->vt, &tmpNode);
return inst;
}
emit(emitInfo, n->Children[0]);
- if (!n->Store)
- if (!alloc_temp_storage(emitInfo, n, n->Children[0]->Store->Size))
- return NULL;
+ if (!alloc_node_storage(emitInfo, n, n->Children[0]->Store->Size))
+ return NULL;
- inst = new_instruction(emitInfo, OPCODE_MOV);
- storage_to_dst_reg(&inst->DstReg, n->Store, n->Writemask);
- storage_to_src_reg(&inst->SrcReg[0], n->Children[0]->Store);
+ inst = emit_instruction(emitInfo,
+ OPCODE_MOV,
+ n->Store, /* dest */
+ n->Children[0]->Store,
+ NULL,
+ NULL);
inst->SrcReg[0].NegateBase = NEGATE_XYZW;
return inst;
}
emit_label(slang_emit_info *emitInfo, const slang_ir_node *n)
{
assert(n->Label);
+#if 0
+ /* XXX this fails in loop tail code - investigate someday */
assert(_slang_label_get_location(n->Label) < 0);
_slang_label_set_location(n->Label, emitInfo->prog->NumInstructions,
emitInfo->prog);
+#else
+ if (_slang_label_get_location(n->Label) < 0)
+ _slang_label_set_location(n->Label, emitInfo->prog->NumInstructions,
+ emitInfo->prog);
+#endif
return NULL;
}
+/**
+ * Emit code for a function call.
+ * Note that for each time a function is called, we emit the function's
+ * body code again because the set of available registers may be different.
+ */
static struct prog_instruction *
-emit_jump(slang_emit_info *emitInfo, slang_ir_node *n)
+emit_fcall(slang_emit_info *emitInfo, slang_ir_node *n)
{
+ struct gl_program *progSave;
struct prog_instruction *inst;
- assert(n);
+ GLuint subroutineId;
+ GLuint maxInstSave;
+
+ assert(n->Opcode == IR_CALL);
assert(n->Label);
- inst = new_instruction(emitInfo, OPCODE_BRA);
- inst->DstReg.CondMask = COND_TR; /* always branch */
- inst->BranchTarget = _slang_label_get_location(n->Label);
- if (inst->BranchTarget < 0) {
- _slang_label_add_reference(n->Label, emitInfo->prog->NumInstructions - 1);
+
+ /* 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);
+
+ if (emitInfo->EmitBeginEndSub) {
+ /* BGNSUB isn't a real instruction.
+ * We require a label (i.e. "foobar:") though, if we're going to
+ * print the program in the NV format. The BNGSUB instruction is
+ * really just a NOP to attach the label to.
+ */
+ inst = new_instruction(emitInfo, OPCODE_BGNSUB);
+ inst_comment(inst, n->Label->Name);
}
+
+ /* body of function: */
+ emit(emitInfo, n->Children[0]);
+ n->Store = n->Children[0]->Store;
+
+ /* add RET instruction now, if needed */
+ inst = prev_instruction(emitInfo);
+ if (inst && inst->Opcode != OPCODE_RET) {
+ inst = new_instruction(emitInfo, OPCODE_RET);
+ }
+
+ if (emitInfo->EmitBeginEndSub) {
+ inst = new_instruction(emitInfo, OPCODE_ENDSUB);
+ inst_comment(inst, n->Label->Name);
+ }
+
+ /* pop/restore cur program */
+ emitInfo->prog = progSave;
+ emitInfo->MaxInstructions = maxInstSave;
+
+ /* emit the function call */
+ inst = new_instruction(emitInfo, OPCODE_CAL);
+ /* The branch target is just the subroutine number (changed later) */
+ inst->BranchTarget = subroutineId;
+ inst_comment(inst, n->Label->Name);
+ assert(inst->BranchTarget >= 0);
+
+ return inst;
+}
+
+
+/**
+ * Emit code for a 'return' statement.
+ */
+static struct prog_instruction *
+emit_return(slang_emit_info *emitInfo, slang_ir_node *n)
+{
+ struct prog_instruction *inst;
+ assert(n);
+ assert(n->Opcode == IR_RETURN);
+ assert(n->Label);
+ inst = new_instruction(emitInfo, OPCODE_RET);
+ inst->DstReg.CondMask = COND_TR; /* always return */
return inst;
}
static struct prog_instruction *
emit_kill(slang_emit_info *emitInfo)
{
+ struct gl_fragment_program *fp;
struct prog_instruction *inst;
/* NV-KILL - discard fragment depending on condition code.
* Note that ARB-KILL depends on sign of vector operand.
*/
inst = new_instruction(emitInfo, OPCODE_KIL_NV);
- inst->DstReg.CondMask = COND_TR; /* always branch */
+ inst->DstReg.CondMask = COND_TR; /* always kill */
+
+ assert(emitInfo->prog->Target == GL_FRAGMENT_PROGRAM_ARB);
+ fp = (struct gl_fragment_program *) emitInfo->prog;
+ fp->UsesKill = GL_TRUE;
+
return inst;
}
emit_tex(slang_emit_info *emitInfo, slang_ir_node *n)
{
struct prog_instruction *inst;
-
- (void) emit(emitInfo, n->Children[1]);
+ gl_inst_opcode opcode;
if (n->Opcode == IR_TEX) {
- inst = new_instruction(emitInfo, OPCODE_TEX);
+ opcode = OPCODE_TEX;
}
else if (n->Opcode == IR_TEXB) {
- inst = new_instruction(emitInfo, OPCODE_TXB);
+ opcode = OPCODE_TXB;
}
else {
assert(n->Opcode == IR_TEXP);
- inst = new_instruction(emitInfo, OPCODE_TXP);
+ opcode = OPCODE_TXP;
}
- if (!n->Store)
- if (!alloc_temp_storage(emitInfo, n, 4))
- return NULL;
+ /* emit code for the texcoord operand */
+ (void) emit(emitInfo, n->Children[1]);
- storage_to_dst_reg(&inst->DstReg, n->Store, n->Writemask);
+ /* alloc storage for result of texture fetch */
+ if (!alloc_node_storage(emitInfo, n, 4))
+ return NULL;
- /* Child[1] is the coord */
- assert(n->Children[1]->Store->File != PROGRAM_UNDEFINED);
- assert(n->Children[1]->Store->Index >= 0);
- storage_to_src_reg(&inst->SrcReg[0], n->Children[1]->Store);
+ /* emit TEX instruction; Child[1] is the texcoord */
+ inst = emit_instruction(emitInfo,
+ opcode,
+ n->Store,
+ n->Children[1]->Store,
+ NULL,
+ NULL);
/* Child[0] is the sampler (a uniform which'll indicate the texture unit) */
assert(n->Children[0]->Store);
+ /* Store->Index is the sampler index */
+ assert(n->Children[0]->Store->Index >= 0);
+ /* Store->Size is the texture target */
assert(n->Children[0]->Store->Size >= TEXTURE_1D_INDEX);
+ assert(n->Children[0]->Store->Size <= TEXTURE_RECT_INDEX);
- inst->Sampler = n->Children[0]->Store->Index; /* i.e. uniform's index */
inst->TexSrcTarget = n->Children[0]->Store->Size;
- inst->TexSrcUnit = 27; /* Dummy value; the TexSrcUnit will be computed at
- * link time, using the sampler uniform's value.
- */
+ inst->TexSrcUnit = n->Children[0]->Store->Index; /* i.e. uniform's index */
+
return inst;
}
+/**
+ * Assignment/copy
+ */
static struct prog_instruction *
-emit_move(slang_emit_info *emitInfo, slang_ir_node *n)
+emit_copy(slang_emit_info *emitInfo, slang_ir_node *n)
{
struct prog_instruction *inst;
+ assert(n->Opcode == IR_COPY);
+
/* lhs */
emit(emitInfo, n->Children[0]);
+ if (!n->Children[0]->Store || n->Children[0]->Store->Index < 0) {
+ /* an error should have been already recorded */
+ return NULL;
+ }
/* rhs */
assert(n->Children[1]);
inst = emit(emitInfo, n->Children[1]);
+ if (!n->Children[1]->Store || n->Children[1]->Store->Index < 0) {
+ if (!emitInfo->log->text) {
+ slang_info_log_error(emitInfo->log, "invalid assignment");
+ }
+ return NULL;
+ }
+
assert(n->Children[1]->Store->Index >= 0);
- assert(!n->Store);
+ /*assert(n->Children[0]->Store->Size == n->Children[1]->Store->Size);*/
+
n->Store = n->Children[0]->Store;
+ if (n->Store->File == PROGRAM_SAMPLER) {
+ /* no code generated for sampler assignments,
+ * just copy the sampler index at compile time.
+ */
+ n->Store->Index = n->Children[1]->Store->Index;
+ return NULL;
+ }
+
#if PEEPHOLE_OPTIMIZATIONS
- if (inst && _slang_is_temp(emitInfo->vt, n->Children[1]->Store)) {
+ 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) &&
+ !n->Children[0]->Store->IsIndirect &&
+ n->Children[0]->Store->Size <= 4) {
/* Peephole optimization:
- * Just modify the RHS to put its result into the dest of this
- * MOVE operation. Then, this MOVE is a no-op.
+ * The Right-Hand-Side has its results in a temporary place.
+ * Modify the RHS (and the prev instruction) to store its results
+ * in the destination specified by n->Children[0].
+ * Then, this MOVE is a no-op.
+ * Ex:
+ * MUL tmp, x, y;
+ * MOV a, tmp;
+ * becomes:
+ * MUL a, x, y;
*/
- _slang_free_temp(emitInfo->vt, n->Children[1]->Store);
+ if (n->Children[1]->Opcode != IR_SWIZZLE)
+ _slang_free_temp(emitInfo->vt, n->Children[1]->Store);
*n->Children[1]->Store = *n->Children[0]->Store;
- /* fixup the prev (RHS) instruction */
+
+ /* fixup the previous instruction (which stored the RHS result) */
assert(n->Children[0]->Store->Index >= 0);
- assert(n->Children[0]->Store->Index < 16);
- storage_to_dst_reg(&inst->DstReg, n->Children[0]->Store, n->Writemask);
+
+ storage_to_dst_reg(&inst->DstReg, n->Children[0]->Store);
return inst;
}
else
slang_ir_storage dstStore = *n->Children[0]->Store;
slang_ir_storage srcStore = *n->Children[1]->Store;
GLint size = srcStore.Size;
- ASSERT(n->Children[0]->Writemask == WRITEMASK_XYZW);
ASSERT(n->Children[1]->Store->Swizzle == SWIZZLE_NOOP);
dstStore.Size = 4;
srcStore.Size = 4;
while (size >= 4) {
- inst = new_instruction(emitInfo, OPCODE_MOV);
- inst->Comment = _mesa_strdup("IR_MOVE block");
- storage_to_dst_reg(&inst->DstReg, &dstStore, n->Writemask);
- storage_to_src_reg(&inst->SrcReg[0], &srcStore);
+ inst = emit_instruction(emitInfo, OPCODE_MOV,
+ &dstStore,
+ &srcStore,
+ NULL,
+ NULL);
+ inst_comment(inst, "IR_COPY block");
srcStore.Index++;
dstStore.Index++;
size -= 4;
else {
/* single register move */
char *srcAnnot, *dstAnnot;
- inst = new_instruction(emitInfo, OPCODE_MOV);
assert(n->Children[0]->Store->Index >= 0);
- storage_to_dst_reg(&inst->DstReg, n->Children[0]->Store, n->Writemask);
- storage_to_src_reg(&inst->SrcReg[0], n->Children[1]->Store);
+ inst = emit_instruction(emitInfo, OPCODE_MOV,
+ n->Children[0]->Store, /* dest */
+ n->Children[1]->Store,
+ NULL,
+ NULL);
dstAnnot = storage_annotation(n->Children[0], emitInfo->prog);
srcAnnot = storage_annotation(n->Children[1], emitInfo->prog);
inst->Comment = instruction_annotation(inst->Opcode, dstAnnot,
srcAnnot, NULL, NULL);
}
- free_temp_storage(emitInfo->vt, n->Children[1]);
+ free_node_storage(emitInfo->vt, n->Children[1]);
return inst;
}
}
+/**
+ * An IR_COND node wraps a boolean expression which is used by an
+ * IF or WHILE test. This is where we'll set condition codes, if needed.
+ */
static struct prog_instruction *
emit_cond(slang_emit_info *emitInfo, slang_ir_node *n)
{
- /* Conditional expression (in if/while/for stmts).
- * Need to update condition code register.
- * Next instruction is typically an IR_IF.
- */
struct prog_instruction *inst;
+ assert(n->Opcode == IR_COND);
+
if (!n->Children[0])
return NULL;
+ /* emit code for the expression */
inst = emit(emitInfo, n->Children[0]);
- if (inst) {
- /* set inst's CondUpdate flag */
- inst->CondUpdate = GL_TRUE;
- return inst; /* XXX or null? */
+
+ if (!n->Children[0]->Store) {
+ /* error recovery */
+ return NULL;
+ }
+
+ assert(n->Children[0]->Store);
+ /*assert(n->Children[0]->Store->Size == 1);*/
+
+ if (emitInfo->EmitCondCodes) {
+ if (inst &&
+ n->Children[0]->Store &&
+ inst->DstReg.File == n->Children[0]->Store->File &&
+ inst->DstReg.Index == n->Children[0]->Store->Index) {
+ /* The previous instruction wrote to the register who's value
+ * we're testing. Just fix that instruction so that the
+ * condition codes are computed.
+ */
+ inst->CondUpdate = GL_TRUE;
+ n->Store = n->Children[0]->Store;
+ return inst;
+ }
+ else {
+ /* This'll happen for things like "if (i) ..." where no code
+ * is normally generated for the expression "i".
+ * Generate a move instruction just to set condition codes.
+ */
+ if (!alloc_node_storage(emitInfo, n, 1))
+ return NULL;
+ inst = emit_instruction(emitInfo, OPCODE_MOV,
+ n->Store, /* dest */
+ n->Children[0]->Store,
+ NULL,
+ NULL);
+ inst->CondUpdate = GL_TRUE;
+ inst_comment(inst, "COND expr");
+ _slang_free_temp(emitInfo->vt, n->Store);
+ return inst;
+ }
}
else {
- /* This'll happen for things like "if (i) ..." where no code
- * is normally generated for the expression "i".
- * Generate a move instruction just to set condition codes.
- * Note: must use full 4-component vector since all four
- * condition codes must be set identically.
- */
- if (!alloc_temp_storage(emitInfo, n, 4))
- return NULL;
- inst = new_instruction(emitInfo, OPCODE_MOV);
- inst->CondUpdate = GL_TRUE;
- storage_to_dst_reg(&inst->DstReg, n->Store, n->Writemask);
- storage_to_src_reg(&inst->SrcReg[0], n->Children[0]->Store);
- _slang_free_temp(emitInfo->vt, n->Store);
- inst->Comment = _mesa_strdup("COND expr");
- return inst; /* XXX or null? */
+ /* No-op: the boolean result of the expression is in a regular reg */
+ n->Store = n->Children[0]->Store;
+ return inst;
}
}
static struct prog_instruction *
emit_not(slang_emit_info *emitInfo, slang_ir_node *n)
{
- GLfloat zero = 0.0;
- slang_ir_storage st;
+ static const struct {
+ gl_inst_opcode op, opNot;
+ } operators[] = {
+ { OPCODE_SLT, OPCODE_SGE },
+ { OPCODE_SLE, OPCODE_SGT },
+ { OPCODE_SGT, OPCODE_SLE },
+ { OPCODE_SGE, OPCODE_SLT },
+ { OPCODE_SEQ, OPCODE_SNE },
+ { OPCODE_SNE, OPCODE_SEQ },
+ { 0, 0 }
+ };
struct prog_instruction *inst;
-
- /* need zero constant */
- st.File = PROGRAM_CONSTANT;
- st.Size = 1;
- st.Index = _mesa_add_unnamed_constant(emitInfo->prog->Parameters, &zero,
- 1, &st.Swizzle);
+ slang_ir_storage zero;
+ GLuint i;
/* child expr */
- (void) emit(emitInfo, n->Children[0]);
- /* XXXX if child instr is SGT convert to SLE, if SEQ, SNE, etc */
+ inst = emit(emitInfo, n->Children[0]);
- if (!n->Store)
- if (!alloc_temp_storage(emitInfo, n, n->Children[0]->Store->Size))
- return NULL;
+#if PEEPHOLE_OPTIMIZATIONS
+ if (inst) {
+ /* if the prev instruction was a comparison instruction, invert it */
+ for (i = 0; operators[i].op; i++) {
+ if (inst->Opcode == operators[i].op) {
+ inst->Opcode = operators[i].opNot;
+ n->Store = n->Children[0]->Store;
+ return inst;
+ }
+ }
+ }
+#endif
- inst = new_instruction(emitInfo, OPCODE_SEQ);
- storage_to_dst_reg(&inst->DstReg, n->Store, n->Writemask);
- storage_to_src_reg(&inst->SrcReg[0], n->Children[0]->Store);
- storage_to_src_reg(&inst->SrcReg[1], &st);
+ /* else, invert using SEQ (v = v == 0) */
+ if (!alloc_node_storage(emitInfo, n, n->Children[0]->Store->Size))
+ return NULL;
+
+ constant_to_storage(emitInfo, 0.0, &zero);
+ inst = emit_instruction(emitInfo,
+ OPCODE_SEQ,
+ n->Store,
+ n->Children[0]->Store,
+ &zero,
+ NULL);
+ inst_comment(inst, "NOT");
- free_temp_storage(emitInfo->vt, n->Children[0]);
+ free_node_storage(emitInfo->vt, n->Children[0]);
- inst->Comment = _mesa_strdup("NOT");
return inst;
}
emit_if(slang_emit_info *emitInfo, slang_ir_node *n)
{
struct gl_program *prog = emitInfo->prog;
- struct prog_instruction *ifInst;
GLuint ifInstLoc, elseInstLoc = 0;
+ GLuint condWritemask = 0;
+
+ /* emit condition expression code */
+ {
+ struct prog_instruction *inst;
+ inst = emit(emitInfo, n->Children[0]);
+ if (emitInfo->EmitCondCodes) {
+ if (!inst) {
+ /* error recovery */
+ return NULL;
+ }
+ condWritemask = inst->DstReg.WriteMask;
+ }
+ }
+
+ if (!n->Children[0]->Store)
+ return NULL;
+
+#if 0
+ assert(n->Children[0]->Store->Size == 1); /* a bool! */
+#endif
- emit(emitInfo, n->Children[0]); /* the condition */
ifInstLoc = prog->NumInstructions;
if (emitInfo->EmitHighLevelInstructions) {
- ifInst = new_instruction(emitInfo, OPCODE_IF);
- ifInst->DstReg.CondMask = COND_NE; /* if cond is non-zero */
- ifInst->DstReg.CondSwizzle = SWIZZLE_X;
+ if (emitInfo->EmitCondCodes) {
+ /* IF condcode THEN ... */
+ struct prog_instruction *ifInst;
+ ifInst = new_instruction(emitInfo, OPCODE_IF);
+ ifInst->DstReg.CondMask = COND_NE; /* if cond is non-zero */
+ /* only test the cond code (1 of 4) that was updated by the
+ * previous instruction.
+ */
+ ifInst->DstReg.CondSwizzle = writemask_to_swizzle(condWritemask);
+ }
+ else {
+ /* IF src[0] THEN ... */
+ emit_instruction(emitInfo, OPCODE_IF,
+ NULL, /* dst */
+ n->Children[0]->Store, /* op0 */
+ NULL,
+ NULL);
+ }
}
else {
/* conditional jump to else, or endif */
- ifInst = new_instruction(emitInfo, OPCODE_BRA);
+ struct prog_instruction *ifInst = new_instruction(emitInfo, OPCODE_BRA);
ifInst->DstReg.CondMask = COND_EQ; /* BRA if cond is zero */
- ifInst->DstReg.CondSwizzle = SWIZZLE_X;
- ifInst->Comment = _mesa_strdup("if zero");
+ inst_comment(ifInst, "if zero");
+ ifInst->DstReg.CondSwizzle = writemask_to_swizzle(condWritemask);
}
/* if body */
/* jump to endif instruction */
struct prog_instruction *inst;
inst = new_instruction(emitInfo, OPCODE_BRA);
- inst->Comment = _mesa_strdup("else");
+ inst_comment(inst, "else");
inst->DstReg.CondMask = COND_TR; /* always branch */
}
- ifInst = prog->Instructions + ifInstLoc;
- ifInst->BranchTarget = prog->NumInstructions;
-
+ prog->Instructions[ifInstLoc].BranchTarget = prog->NumInstructions;
emit(emitInfo, n->Children[2]);
}
else {
/* no else body */
- ifInst = prog->Instructions + ifInstLoc;
- ifInst->BranchTarget = prog->NumInstructions + 1;
+ prog->Instructions[ifInstLoc].BranchTarget = prog->NumInstructions;
}
if (emitInfo->EmitHighLevelInstructions) {
}
if (n->Children[2]) {
- struct prog_instruction *elseInst;
- elseInst = prog->Instructions + elseInstLoc;
- elseInst->BranchTarget = prog->NumInstructions;
+ prog->Instructions[elseInstLoc].BranchTarget = prog->NumInstructions;
}
return NULL;
}
emit_loop(slang_emit_info *emitInfo, slang_ir_node *n)
{
struct gl_program *prog = emitInfo->prog;
- struct prog_instruction *beginInst, *endInst;
- GLuint beginInstLoc, endInstLoc;
+ struct prog_instruction *endInst;
+ GLuint beginInstLoc, tailInstLoc, endInstLoc;
slang_ir_node *ir;
/* emit OPCODE_BGNLOOP */
/* body */
emit(emitInfo, n->Children[0]);
+ /* tail */
+ tailInstLoc = prog->NumInstructions;
+ if (n->Children[1]) {
+ if (emitInfo->EmitComments)
+ emit_comment(emitInfo, "Loop tail code:");
+ emit(emitInfo, n->Children[1]);
+ }
+
endInstLoc = prog->NumInstructions;
if (emitInfo->EmitHighLevelInstructions) {
/* emit OPCODE_ENDLOOP */
endInst = new_instruction(emitInfo, OPCODE_BRA);
endInst->DstReg.CondMask = COND_TR; /* always true */
}
- /* end instruction's BranchTarget points to top of loop */
+ /* ENDLOOP's BranchTarget points to the BGNLOOP inst */
endInst->BranchTarget = beginInstLoc;
if (emitInfo->EmitHighLevelInstructions) {
/* BGNLOOP's BranchTarget points to the ENDLOOP inst */
- beginInst = prog->Instructions + beginInstLoc;
- beginInst->BranchTarget = prog->NumInstructions - 1;
+ prog->Instructions[beginInstLoc].BranchTarget = prog->NumInstructions -1;
}
/* Done emitting loop code. Now walk over the loop's linked list of
* BREAK and CONT nodes, filling in their BranchTarget fields (which
* will point to the ENDLOOP+1 or BGNLOOP instructions, respectively).
*/
- for (ir = n->BranchNode; ir; ir = ir->BranchNode) {
+ for (ir = n->List; ir; ir = ir->List) {
struct prog_instruction *inst = prog->Instructions + ir->InstLocation;
assert(inst->BranchTarget < 0);
if (ir->Opcode == IR_BREAK ||
- ir->Opcode == IR_BREAK_IF_FALSE ||
ir->Opcode == IR_BREAK_IF_TRUE) {
assert(inst->Opcode == OPCODE_BRK ||
inst->Opcode == OPCODE_BRA);
}
else {
assert(ir->Opcode == IR_CONT ||
- ir->Opcode == IR_CONT_IF_FALSE ||
ir->Opcode == IR_CONT_IF_TRUE);
assert(inst->Opcode == OPCODE_CONT ||
inst->Opcode == OPCODE_BRA);
- /* to go instruction at top of loop */
- inst->BranchTarget = beginInstLoc;
+ /* go to instruction at tail of loop */
+ inst->BranchTarget = endInstLoc;
}
}
return NULL;
/**
- * "Continue" or "break" statement.
+ * Unconditional "continue" or "break" statement.
* Either OPCODE_CONT, OPCODE_BRK or OPCODE_BRA will be emitted.
*/
static struct prog_instruction *
{
gl_inst_opcode opcode;
struct prog_instruction *inst;
- n->InstLocation = emitInfo->prog->NumInstructions;
+
+ if (n->Opcode == IR_CONT) {
+ /* we need to execute the loop's tail code before doing CONT */
+ assert(n->Parent);
+ assert(n->Parent->Opcode == IR_LOOP);
+ if (n->Parent->Children[1]) {
+ /* emit tail code */
+ if (emitInfo->EmitComments) {
+ emit_comment(emitInfo, "continue - tail code:");
+ }
+ emit(emitInfo, n->Parent->Children[1]);
+ }
+ }
+
+ /* opcode selection */
if (emitInfo->EmitHighLevelInstructions) {
opcode = (n->Opcode == IR_CONT) ? OPCODE_CONT : OPCODE_BRK;
}
else {
opcode = OPCODE_BRA;
}
+ n->InstLocation = emitInfo->prog->NumInstructions;
inst = new_instruction(emitInfo, opcode);
inst->DstReg.CondMask = COND_TR; /* always true */
return inst;
* Either OPCODE_CONT, OPCODE_BRK or OPCODE_BRA will be emitted.
*/
static struct prog_instruction *
-emit_cont_break_if(slang_emit_info *emitInfo, slang_ir_node *n,
- GLboolean breakTrue)
+emit_cont_break_if_true(slang_emit_info *emitInfo, slang_ir_node *n)
{
- gl_inst_opcode opcode;
struct prog_instruction *inst;
+ assert(n->Opcode == IR_CONT_IF_TRUE ||
+ n->Opcode == IR_BREAK_IF_TRUE);
+
/* evaluate condition expr, setting cond codes */
inst = emit(emitInfo, n->Children[0]);
- assert(inst);
- inst->CondUpdate = GL_TRUE;
+ if (emitInfo->EmitCondCodes) {
+ assert(inst);
+ inst->CondUpdate = GL_TRUE;
+ }
n->InstLocation = emitInfo->prog->NumInstructions;
+
+ /* opcode selection */
if (emitInfo->EmitHighLevelInstructions) {
- if (n->Opcode == IR_CONT_IF_TRUE ||
- n->Opcode == IR_CONT_IF_FALSE)
- opcode = OPCODE_CONT;
- else
- opcode = OPCODE_BRK;
+ const gl_inst_opcode opcode
+ = (n->Opcode == IR_CONT_IF_TRUE) ? OPCODE_CONT : OPCODE_BRK;
+ if (emitInfo->EmitCondCodes) {
+ /* Get the writemask from the previous instruction which set
+ * the condcodes. Use that writemask as the CondSwizzle.
+ */
+ const GLuint condWritemask = inst->DstReg.WriteMask;
+ inst = new_instruction(emitInfo, opcode);
+ inst->DstReg.CondMask = COND_NE;
+ inst->DstReg.CondSwizzle = writemask_to_swizzle(condWritemask);
+ return inst;
+ }
+ else {
+ /* IF reg
+ * BRK/CONT;
+ * ENDIF
+ */
+ GLint ifInstLoc;
+ ifInstLoc = emitInfo->prog->NumInstructions;
+ inst = emit_instruction(emitInfo, OPCODE_IF,
+ NULL, /* dest */
+ n->Children[0]->Store,
+ NULL,
+ NULL);
+ n->InstLocation = emitInfo->prog->NumInstructions;
+
+ inst = new_instruction(emitInfo, opcode);
+ inst = new_instruction(emitInfo, OPCODE_ENDIF);
+
+ emitInfo->prog->Instructions[ifInstLoc].BranchTarget
+ = emitInfo->prog->NumInstructions;
+ return inst;
+ }
}
else {
- opcode = OPCODE_BRA;
+ const GLuint condWritemask = inst->DstReg.WriteMask;
+ assert(emitInfo->EmitCondCodes);
+ inst = new_instruction(emitInfo, OPCODE_BRA);
+ inst->DstReg.CondMask = COND_NE;
+ inst->DstReg.CondSwizzle = writemask_to_swizzle(condWritemask);
+ return inst;
}
- inst = new_instruction(emitInfo, opcode);
- inst->DstReg.CondMask = breakTrue ? COND_NE : COND_EQ;
- return inst;
}
+static struct prog_instruction *
+emit_swizzle(slang_emit_info *emitInfo, slang_ir_node *n)
+{
+ struct prog_instruction *inst;
+
+ inst = emit(emitInfo, n->Children[0]);
+
+#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;
+}
+
/**
- * Remove any SWIZZLE_NIL terms from given swizzle mask (smear prev term).
- * Ex: fix_swizzle("zyNN") -> "zyyy"
+ * 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 GLuint
-fix_swizzle(GLuint swizzle)
+static struct prog_instruction *
+emit_array_element(slang_emit_info *emitInfo, slang_ir_node *n)
{
- GLuint swz[4], i;
- for (i = 0; i < 4; i++) {
- swz[i] = GET_SWZ(swizzle, i);
- if (swz[i] == SWIZZLE_NIL) {
- swz[i] = swz[i - 1];
+ 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;
+
+ assert(n->Opcode == IR_ELEMENT);
+ 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;
+
+ if (root->File == PROGRAM_STATE_VAR) {
+ GLint index = _slang_alloc_statevar(n, emitInfo->prog->Parameters);
+ assert(n->Store->Index == index);
+ return NULL;
+ }
+ }
+
+ /* 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 */
+ const GLint element = (GLint) n->Children[1]->Value[0];
+
+ /* this element's storage is the array's storage, plus constant offset */
+ n->Store->Index += elemSizeVec * element;
+ }
+ else {
+ /* Variable array index */
+
+ /* 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.
+ */
+
+
+ /*indexStore = tempstore();*/
+ }
+
+
+ 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 */
+ 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);
+
+ /* multiply array index by element size */
+ inst = emit_instruction(emitInfo,
+ OPCODE_MUL,
+ indexTemp, /* dest */
+ indexStore, /* the index */
+ &elemSizeStore,
+ NULL);
+
+ indexStore = indexTemp;
+ }
+
+ 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;
}
+
+ /* 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;
}
- return MAKE_SWIZZLE4(swz[0], swz[1], swz[2], swz[3]);
+
+ n->Store->Size = elemSize;
+ n->Store->Swizzle = _slang_var_swizzle(elemSize, 0);
+
+ return NULL; /* no instruction */
}
+/**
+ * Resolve storage for accessing a structure field.
+ */
static struct prog_instruction *
-emit_swizzle(slang_emit_info *emitInfo, slang_ir_node *n)
+emit_struct_field(slang_emit_info *emitInfo, slang_ir_node *n)
{
- GLuint swizzle;
+ slang_ir_storage *root = n->Store;
+ GLint fieldOffset, fieldSize;
- /* swizzled storage access */
- (void) emit(emitInfo, n->Children[0]);
+ assert(n->Opcode == IR_FIELD);
- /* "pull-up" the child's storage info, applying our swizzle info */
- n->Store->File = n->Children[0]->Store->File;
- n->Store->Index = n->Children[0]->Store->Index;
- n->Store->Size = n->Children[0]->Store->Size;
- /*n->Var = n->Children[0]->Var; XXX for debug */
- assert(n->Store->Index >= 0);
+ assert(!root->Parent);
+ while (root->Parent)
+ root = root->Parent;
- swizzle = fix_swizzle(n->Store->Swizzle);
-#ifdef DEBUG
- {
- GLuint s = n->Children[0]->Store->Swizzle;
- assert(GET_SWZ(s, 0) != SWIZZLE_NIL);
- assert(GET_SWZ(s, 1) != SWIZZLE_NIL);
- assert(GET_SWZ(s, 2) != SWIZZLE_NIL);
- assert(GET_SWZ(s, 3) != SWIZZLE_NIL);
+ /* If this is the field of a state var, allocate constant/uniform
+ * storage for it now if we haven't already.
+ * Note that we allocate storage (uniform/constant slots) for state
+ * variables here rather than at declaration time so we only allocate
+ * space for the ones that we actually use!
+ */
+ if (root->File == PROGRAM_STATE_VAR) {
+ root->Index = _slang_alloc_statevar(n, emitInfo->prog->Parameters);
+ if (root->Index < 0) {
+ 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);
}
-#endif
- /* apply this swizzle to child's swizzle to get composed swizzle */
- n->Store->Swizzle = swizzle_swizzle(n->Children[0]->Store->Swizzle,
- swizzle);
- return NULL;
+ 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 */
}
/**
- * Dereference array element. Just resolve storage for the array
- * element represented by this node.
+ * Emit code for a variable declaration.
+ * This usually doesn't result in any code generation, but just
+ * memory allocation.
*/
static struct prog_instruction *
-emit_array_element(slang_emit_info *emitInfo, slang_ir_node *n)
+emit_var_decl(slang_emit_info *emitInfo, slang_ir_node *n)
{
assert(n->Store);
assert(n->Store->File != PROGRAM_UNDEFINED);
assert(n->Store->Size > 0);
-
- if (n->Store->File == PROGRAM_STATE_VAR) {
- n->Store->Index = _slang_alloc_statevar(n, emitInfo->prog->Parameters);
- return NULL;
- }
-
-
- if (n->Children[1]->Opcode == IR_FLOAT) {
- /* Constant index */
- const GLint arrayAddr = n->Children[0]->Store->Index;
- const GLint index = (GLint) n->Children[1]->Value[0];
- n->Store->Index = arrayAddr + index;
+ /*assert(n->Store->Index < 0);*/
+
+ if (!n->Var || n->Var->isTemp) {
+ /* a nameless/temporary variable, will be freed after first use */
+ /*NEW*/
+ if (n->Store->Index < 0 && !_slang_alloc_temp(emitInfo->vt, n->Store)) {
+ slang_info_log_error(emitInfo->log,
+ "Ran out of registers, too many temporaries");
+ return NULL;
+ }
}
else {
- /* Variable index - PROBLEM */
- const GLint arrayAddr = n->Children[0]->Store->Index;
- const GLint index = 0;
- _mesa_problem(NULL, "variable array indexes not supported yet!");
- n->Store->Index = arrayAddr + index;
+ /* a regular variable */
+ _slang_add_variable(emitInfo->vt, n->Var);
+ if (!_slang_alloc_var(emitInfo->vt, n->Store)) {
+ slang_info_log_error(emitInfo->log,
+ "Ran out of registers, too many variables");
+ return NULL;
+ }
+ /*
+ printf("IR_VAR_DECL %s %d store %p\n",
+ (char*) n->Var->a_name, n->Store->Index, (void*) n->Store);
+ */
+ assert(n->Var->store == n->Store);
}
- return NULL; /* no instruction */
+ if (emitInfo->EmitComments) {
+ /* emit NOP with comment describing the variable's storage location */
+ char s[1000];
+ sprintf(s, "TEMP[%d]%s = variable %s (size %d)",
+ n->Store->Index,
+ _mesa_swizzle_string(n->Store->Swizzle, 0, GL_FALSE),
+ (n->Var ? (char *) n->Var->a_name : "anonymous"),
+ n->Store->Size);
+ emit_comment(emitInfo, s);
+ }
+ return NULL;
}
/**
- * Resolve storage for accessing a structure field.
+ * Emit code for a reference to a variable.
+ * 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 *
-emit_struct_field(slang_emit_info *emitInfo, slang_ir_node *n)
+emit_var_ref(slang_emit_info *emitInfo, slang_ir_node *n)
{
- if (n->Store->File == PROGRAM_STATE_VAR) {
+ assert(n->Store);
+ assert(n->Store->File != PROGRAM_UNDEFINED);
+
+ if (n->Store->File == PROGRAM_STATE_VAR && n->Store->Index < 0) {
n->Store->Index = _slang_alloc_statevar(n, emitInfo->prog->Parameters);
}
- else {
- _mesa_problem(NULL, "structs/fields not supported yet");
+ else if (n->Store->File == PROGRAM_UNIFORM) {
+ /* mark var as used */
+ _mesa_use_uniform(emitInfo->prog->Parameters, (char *) n->Var->a_name);
}
- return NULL; /* no instruction */
+
+ if (n->Store->Index < 0) {
+ /* probably ran out of registers */
+ return NULL;
+ }
+ assert(n->Store->Size > 0);
+
+ return NULL;
}
if (!n)
return NULL;
+ if (emitInfo->log->error_flag) {
+ return NULL;
+ }
+
switch (n->Opcode) {
case IR_SEQ:
/* sequence of two sub-trees */
assert(n->Children[0]);
assert(n->Children[1]);
emit(emitInfo, n->Children[0]);
+ if (emitInfo->log->error_flag)
+ return NULL;
inst = emit(emitInfo, n->Children[1]);
+#if 0
assert(!n->Store);
+#endif
n->Store = n->Children[1]->Store;
return inst;
case IR_VAR_DECL:
/* Variable declaration - allocate a register for it */
- assert(n->Store);
- assert(n->Store->File != PROGRAM_UNDEFINED);
- assert(n->Store->Size > 0);
- assert(n->Store->Index < 0);
- if (!n->Var || n->Var->isTemp) {
- /* a nameless/temporary variable, will be freed after first use */
- if (!_slang_alloc_temp(emitInfo->vt, n->Store)) {
- slang_info_log_error(emitInfo->log,
- "Ran out of registers, too many temporaries");
- return NULL;
- }
- }
- else {
- /* a regular variable */
- _slang_add_variable(emitInfo->vt, n->Var);
- if (!_slang_alloc_var(emitInfo->vt, n->Store)) {
- slang_info_log_error(emitInfo->log,
- "Ran out of registers, too many variables");
- return NULL;
- }
- /*
- printf("IR_VAR_DECL %s %d store %p\n",
- (char*) n->Var->a_name, n->Store->Index, (void*) n->Store);
- */
- assert(n->Var->aux == n->Store);
- }
- if (emitInfo->EmitComments) {
- /* emit NOP with comment describing the variable's storage location */
- char s[1000];
- sprintf(s, "TEMP[%d]%s = %s (size %d)",
- n->Store->Index,
- _mesa_swizzle_string(n->Store->Swizzle, 0, GL_FALSE),
- (char *) n->Var->a_name,
- n->Store->Size);
- inst = new_instruction(emitInfo, OPCODE_NOP);
- inst->Comment = _mesa_strdup(s);
- return inst;
- }
- return NULL;
+ inst = emit_var_decl(emitInfo, n);
+ return inst;
case IR_VAR:
/* Reference to a variable
* Storage should have already been resolved/allocated.
*/
- assert(n->Store);
- assert(n->Store->File != PROGRAM_UNDEFINED);
-
- if (n->Store->File == PROGRAM_STATE_VAR &&
- n->Store->Index < 0) {
- n->Store->Index = _slang_alloc_statevar(n, emitInfo->prog->Parameters);
- }
-
- if (n->Store->Index < 0) {
- printf("#### VAR %s not allocated!\n", (char*)n->Var->a_name);
- }
- assert(n->Store->Index >= 0);
- assert(n->Store->Size > 0);
- break;
+ return emit_var_ref(emitInfo, n);
case IR_ELEMENT:
return emit_array_element(emitInfo, n);
case IR_SWIZZLE:
return emit_swizzle(emitInfo, n);
- case IR_I_TO_F:
- /* just move */
- emit(emitInfo, n->Children[0]);
- inst = new_instruction(emitInfo, OPCODE_MOV);
- if (!n->Store) {
- if (!alloc_temp_storage(emitInfo, n, 1))
- return NULL;
- }
- storage_to_dst_reg(&inst->DstReg, n->Store, n->Writemask);
- storage_to_src_reg(&inst->SrcReg[0], n->Children[0]->Store);
- if (emitInfo->EmitComments)
- inst->Comment = _mesa_strdup("int to float");
- return NULL;
-
/* Simple arithmetic */
/* unary */
+ case IR_MOVE:
case IR_RSQ:
case IR_RCP:
case IR_FLOOR:
case IR_FRAC:
case IR_F_TO_I:
+ case IR_I_TO_F:
case IR_ABS:
case IR_SIN:
case IR_COS:
case IR_DDX:
case IR_DDY:
+ case IR_EXP:
+ case IR_EXP2:
+ case IR_LOG2:
case IR_NOISE1:
case IR_NOISE2:
case IR_NOISE3:
case IR_NOISE4:
+ case IR_NRM4:
+ case IR_NRM3:
/* binary */
case IR_ADD:
case IR_SUB:
case IR_MUL:
case IR_DOT4:
case IR_DOT3:
+ case IR_DOT2:
case IR_CROSS:
case IR_MIN:
case IR_MAX:
case IR_SLE:
case IR_SLT:
case IR_POW:
- case IR_EXP:
- case IR_EXP2:
/* trinary operators */
case IR_LRP:
return emit_arith(emitInfo, n);
+
+ case IR_EQUAL:
+ case IR_NOTEQUAL:
+ return emit_compare(emitInfo, n);
+
case IR_CLAMP:
return emit_clamp(emitInfo, n);
case IR_TEX:
return emit_negation(emitInfo, n);
case IR_FLOAT:
/* find storage location for this float constant */
- n->Store->Index = _mesa_add_unnamed_constant(emitInfo->prog->Parameters, n->Value,
+ n->Store->Index = _mesa_add_unnamed_constant(emitInfo->prog->Parameters,
+ n->Value,
n->Store->Size,
&n->Store->Swizzle);
if (n->Store->Index < 0) {
}
return NULL;
- case IR_MOVE:
- return emit_move(emitInfo, n);
+ case IR_COPY:
+ return emit_copy(emitInfo, n);
case IR_COND:
return emit_cond(emitInfo, n);
case IR_LABEL:
return emit_label(emitInfo, n);
- case IR_JUMP:
- assert(n);
- assert(n->Label);
- return emit_jump(emitInfo, n);
+
case IR_KILL:
return emit_kill(emitInfo);
+ case IR_CALL:
+ /* new variable scope for subroutines/function calls */
+ _slang_push_var_table(emitInfo->vt);
+ inst = emit_fcall(emitInfo, n);
+ _slang_pop_var_table(emitInfo->vt);
+ return inst;
+
case IR_IF:
return emit_if(emitInfo, n);
case IR_LOOP:
return emit_loop(emitInfo, n);
- case IR_BREAK_IF_FALSE:
- case IR_CONT_IF_FALSE:
- return emit_cont_break_if(emitInfo, n, GL_FALSE);
case IR_BREAK_IF_TRUE:
case IR_CONT_IF_TRUE:
- return emit_cont_break_if(emitInfo, n, GL_TRUE);
+ return emit_cont_break_if_true(emitInfo, n);
case IR_BREAK:
/* fall-through */
case IR_CONT:
case IR_END_SUB:
return new_instruction(emitInfo, OPCODE_ENDSUB);
case IR_RETURN:
- return new_instruction(emitInfo, OPCODE_RET);
+ return emit_return(emitInfo, n);
case IR_NOP:
return NULL;
default:
_mesa_problem(NULL, "Unexpected IR opcode in emit()\n");
- abort();
}
return NULL;
}
+/**
+ * After code generation, any subroutines will be in separate program
+ * objects. This function appends all the subroutines onto the main
+ * program and resolves the linking of all the branch/call instructions.
+ * XXX this logic should really be part of the linking process...
+ */
+static void
+_slang_resolve_subroutines(slang_emit_info *emitInfo)
+{
+ GET_CURRENT_CONTEXT(ctx);
+ struct gl_program *mainP = emitInfo->prog;
+ GLuint *subroutineLoc, i, total;
+
+ subroutineLoc
+ = (GLuint *) _mesa_malloc(emitInfo->NumSubroutines * sizeof(GLuint));
+
+ /* total number of instructions */
+ total = mainP->NumInstructions;
+ for (i = 0; i < emitInfo->NumSubroutines; i++) {
+ subroutineLoc[i] = total;
+ total += emitInfo->Subroutines[i]->NumInstructions;
+ }
+
+ /* adjust BranchTargets within the functions */
+ for (i = 0; i < emitInfo->NumSubroutines; i++) {
+ struct gl_program *sub = emitInfo->Subroutines[i];
+ GLuint j;
+ for (j = 0; j < sub->NumInstructions; j++) {
+ struct prog_instruction *inst = sub->Instructions + j;
+ if (inst->Opcode != OPCODE_CAL && inst->BranchTarget >= 0) {
+ inst->BranchTarget += subroutineLoc[i];
+ }
+ }
+ }
+
+ /* append subroutines' instructions after main's instructions */
+ mainP->Instructions = _mesa_realloc_instructions(mainP->Instructions,
+ mainP->NumInstructions,
+ total);
+ mainP->NumInstructions = total;
+ for (i = 0; i < emitInfo->NumSubroutines; i++) {
+ struct gl_program *sub = emitInfo->Subroutines[i];
+ _mesa_copy_instructions(mainP->Instructions + subroutineLoc[i],
+ sub->Instructions,
+ sub->NumInstructions);
+ /* delete subroutine code */
+ sub->Parameters = NULL; /* prevent double-free */
+ _mesa_reference_program(ctx, &emitInfo->Subroutines[i], NULL);
+ }
+
+ /* free subroutine list */
+ if (emitInfo->Subroutines) {
+ _mesa_free(emitInfo->Subroutines);
+ emitInfo->Subroutines = NULL;
+ }
+ emitInfo->NumSubroutines = 0;
+
+ /* Examine CAL instructions.
+ * At this point, the BranchTarget field of the CAL instruction is
+ * the number/id of the subroutine to call (an index into the
+ * emitInfo->Subroutines list).
+ * Translate that into an actual instruction location now.
+ */
+ for (i = 0; i < mainP->NumInstructions; i++) {
+ struct prog_instruction *inst = mainP->Instructions + i;
+ if (inst->Opcode == OPCODE_CAL) {
+ const GLuint f = inst->BranchTarget;
+ inst->BranchTarget = subroutineLoc[f];
+ }
+ }
+
+ _mesa_free(subroutineLoc);
+}
+
+
+
+
GLboolean
_slang_emit_code(slang_ir_node *n, slang_var_table *vt,
struct gl_program *prog, GLboolean withEnd,
GET_CURRENT_CONTEXT(ctx);
GLboolean success;
slang_emit_info emitInfo;
+ GLuint maxUniforms;
emitInfo.log = log;
emitInfo.vt = vt;
emitInfo.prog = prog;
+ emitInfo.Subroutines = NULL;
+ emitInfo.NumSubroutines = 0;
+ emitInfo.MaxInstructions = prog->NumInstructions;
emitInfo.EmitHighLevelInstructions = ctx->Shader.EmitHighLevelInstructions;
+ emitInfo.EmitCondCodes = ctx->Shader.EmitCondCodes;
emitInfo.EmitComments = ctx->Shader.EmitComments;
+ emitInfo.EmitBeginEndSub = GL_TRUE;
+
+ if (!emitInfo.EmitCondCodes) {
+ emitInfo.EmitHighLevelInstructions = GL_TRUE;
+ }
+
+ /* Check uniform/constant limits */
+ if (prog->Target == GL_FRAGMENT_PROGRAM_ARB) {
+ maxUniforms = ctx->Const.FragmentProgram.MaxUniformComponents / 4;
+ }
+ else {
+ assert(prog->Target == GL_VERTEX_PROGRAM_ARB);
+ maxUniforms = ctx->Const.VertexProgram.MaxUniformComponents / 4;
+ }
+ if (prog->Parameters->NumParameters > maxUniforms) {
+ slang_info_log_error(log, "Constant/uniform register limit exceeded");
+ return GL_FALSE;
+ }
(void) emit(&emitInfo, n);
struct prog_instruction *inst;
inst = new_instruction(&emitInfo, OPCODE_END);
}
+
+ _slang_resolve_subroutines(&emitInfo);
+
success = GL_TRUE;
- printf("*********** End generate code (%u inst):\n", prog->NumInstructions);
#if 0
+ printf("*********** End emit code (%u inst):\n", prog->NumInstructions);
_mesa_print_program(prog);
_mesa_print_program_parameters(ctx,prog);
#endif