/*
* Mesa 3-D graphics library
- * Version: 7.5
*
* Copyright (C) 2009 VMware, Inc. All Rights Reserved.
*
#define MAX_LOOP_NESTING 50
-
+/* MAX_PROGRAM_TEMPS is a low number (256), and we want to be able to
+ * register allocate many temporary values into that small number of
+ * temps. So allow large temporary indices coming into the register
+ * allocator.
+ */
+#define REG_ALLOCATE_MAX_PROGRAM_TEMPS ((1 << INST_INDEX_BITS) - 1)
static GLboolean dbg = GL_FALSE;
case OPCODE_MAD:
case OPCODE_MUL:
case OPCODE_SUB:
+ case OPCODE_CMP:
+ case OPCODE_FLR:
+ case OPCODE_FRC:
+ case OPCODE_LRP:
+ case OPCODE_SEQ:
+ case OPCODE_SGE:
+ case OPCODE_SGT:
+ case OPCODE_SLE:
+ case OPCODE_SLT:
+ case OPCODE_SNE:
+ case OPCODE_SSG:
channel_mask = inst->DstReg.WriteMask & dst_mask;
break;
case OPCODE_RCP:
static GLboolean
_mesa_remove_dead_code_global(struct gl_program *prog)
{
- GLboolean tempRead[MAX_PROGRAM_TEMPS][4];
+ GLboolean tempRead[REG_ALLOCATE_MAX_PROGRAM_TEMPS][4];
GLboolean *removeInst; /* per-instruction removal flag */
GLuint i, rem = 0, comp;
/*_mesa_print_program(prog);*/
}
- removeInst = (GLboolean *)
+ removeInst =
calloc(1, prog->NumInstructions * sizeof(GLboolean));
/* Determine which temps are read and written */
if (inst->SrcReg[j].File == PROGRAM_TEMPORARY) {
const GLuint index = inst->SrcReg[j].Index;
GLuint read_mask;
- ASSERT(index < MAX_PROGRAM_TEMPS);
+ ASSERT(index < REG_ALLOCATE_MAX_PROGRAM_TEMPS);
read_mask = get_src_arg_mask(inst, j, NO_MASK);
if (inst->SrcReg[j].RelAddr) {
/* check dst reg */
if (inst->DstReg.File == PROGRAM_TEMPORARY) {
const GLuint index = inst->DstReg.Index;
- ASSERT(index < MAX_PROGRAM_TEMPS);
+ ASSERT(index < REG_ALLOCATE_MAX_PROGRAM_TEMPS);
if (inst->DstReg.RelAddr) {
if (dbg)
switch (inst->Opcode) {
case OPCODE_BGNLOOP:
case OPCODE_BGNSUB:
- case OPCODE_BRA:
case OPCODE_CAL:
case OPCODE_CONT:
case OPCODE_IF:
switch (opcode) {
case OPCODE_BGNLOOP:
case OPCODE_BGNSUB:
- case OPCODE_BRA:
case OPCODE_CAL:
case OPCODE_CONT:
case OPCODE_IF:
mov->SrcReg[0].HasIndex2 == 0 &&
mov->SrcReg[0].RelAddr2 == 0 &&
mov->DstReg.RelAddr == 0 &&
- mov->DstReg.CondMask == COND_TR &&
- mov->SaturateMode == SATURATE_OFF;
+ mov->DstReg.CondMask == COND_TR;
}
{
return
can_downward_mov_be_modifed(mov) &&
- mov->DstReg.File == PROGRAM_TEMPORARY;
+ mov->DstReg.File == PROGRAM_TEMPORARY &&
+ mov->SaturateMode == SATURATE_OFF;
}
GLboolean *removeInst;
GLuint i, arg, rem = 0;
- removeInst = (GLboolean *)
+ removeInst =
calloc(1, prog->NumInstructions * sizeof(GLboolean));
for (i = 0; i < prog->NumInstructions; i++) {
if (mask != (inst->DstReg.WriteMask & mask))
return GL_FALSE;
+ inst->SaturateMode |= mov->SaturateMode;
+
/* Depending on the instruction, we may need to recompute the swizzles.
* Also, some other instructions (like TEX) are not linear. We will only
* consider completely active sources and destinations
_mesa_print_program(prog);
}
- removeInst = (GLboolean *)
+ removeInst =
calloc(1, prog->NumInstructions * sizeof(GLboolean));
/*
nesting--;
break;
case OPCODE_MOV:
- if (i > 0 && can_downward_mov_be_modifed(mov) && nesting == 0) {
+ if (i > 0 &&
+ can_downward_mov_be_modifed(mov) &&
+ mov->SrcReg[0].File == PROGRAM_TEMPORARY &&
+ nesting == 0)
+ {
/* see if this MOV can be removed */
const GLuint id = mov->SrcReg[0].Index;
if (prevInst->DstReg.File == PROGRAM_TEMPORARY &&
prevInst->DstReg.Index == id &&
prevInst->DstReg.RelAddr == 0 &&
- prevInst->DstReg.CondSrc == 0 &&
prevInst->DstReg.CondMask == COND_TR) {
const GLuint dst_mask = prevInst->DstReg.WriteMask;
struct interval_list
{
GLuint Num;
- struct interval Intervals[MAX_PROGRAM_TEMPS];
+ struct interval Intervals[REG_ALLOCATE_MAX_PROGRAM_TEMPS];
};
#endif
}
+struct loop_info
+{
+ GLuint Start, End; /**< Start, end instructions of loop */
+};
/**
* Update the intermediate interval info for register 'index' and
* instruction 'ic'.
*/
static void
-update_interval(GLint intBegin[], GLint intEnd[], GLuint index, GLuint ic)
+update_interval(GLint intBegin[], GLint intEnd[],
+ struct loop_info *loopStack, GLuint loopStackDepth,
+ GLuint index, GLuint ic)
{
- ASSERT(index < MAX_PROGRAM_TEMPS);
+ int i;
+ GLuint begin = ic;
+ GLuint end = ic;
+
+ /* If the register is used in a loop, extend its lifetime through the end
+ * of the outermost loop that doesn't contain its definition.
+ */
+ for (i = 0; i < loopStackDepth; i++) {
+ if (intBegin[index] < loopStack[i].Start) {
+ end = loopStack[i].End;
+ break;
+ }
+ }
+
+ /* Variables that are live at the end of a loop will also be live at the
+ * beginning, so an instruction inside of a loop should have its live
+ * interval begin at the start of the outermost loop.
+ */
+ if (loopStackDepth > 0 && ic > loopStack[0].Start && ic < loopStack[0].End) {
+ begin = loopStack[0].Start;
+ }
+
+ ASSERT(index < REG_ALLOCATE_MAX_PROGRAM_TEMPS);
if (intBegin[index] == -1) {
ASSERT(intEnd[index] == -1);
- intBegin[index] = intEnd[index] = ic;
+ intBegin[index] = begin;
+ intEnd[index] = end;
}
else {
- intEnd[index] = ic;
+ intEnd[index] = end;
}
}
GLboolean
_mesa_find_temp_intervals(const struct prog_instruction *instructions,
GLuint numInstructions,
- GLint intBegin[MAX_PROGRAM_TEMPS],
- GLint intEnd[MAX_PROGRAM_TEMPS])
+ GLint intBegin[REG_ALLOCATE_MAX_PROGRAM_TEMPS],
+ GLint intEnd[REG_ALLOCATE_MAX_PROGRAM_TEMPS])
{
- struct loop_info
- {
- GLuint Start, End; /**< Start, end instructions of loop */
- };
struct loop_info loopStack[MAX_LOOP_NESTING];
GLuint loopStackDepth = 0;
GLuint i;
- for (i = 0; i < MAX_PROGRAM_TEMPS; i++){
+ for (i = 0; i < REG_ALLOCATE_MAX_PROGRAM_TEMPS; i++){
intBegin[i] = intEnd[i] = -1;
}
const GLuint index = inst->SrcReg[j].Index;
if (inst->SrcReg[j].RelAddr)
return GL_FALSE;
- update_interval(intBegin, intEnd, index, i);
- if (loopStackDepth > 0) {
- /* extend temp register's interval to end of loop */
- GLuint loopEnd = loopStack[loopStackDepth - 1].End;
- update_interval(intBegin, intEnd, index, loopEnd);
- }
+ update_interval(intBegin, intEnd, loopStack, loopStackDepth,
+ index, i);
}
}
if (inst->DstReg.File == PROGRAM_TEMPORARY) {
const GLuint index = inst->DstReg.Index;
if (inst->DstReg.RelAddr)
return GL_FALSE;
- update_interval(intBegin, intEnd, index, i);
- if (loopStackDepth > 0) {
- /* extend temp register's interval to end of loop */
- GLuint loopEnd = loopStack[loopStackDepth - 1].End;
- update_interval(intBegin, intEnd, index, loopEnd);
- }
+ update_interval(intBegin, intEnd, loopStack, loopStackDepth,
+ index, i);
}
}
}
find_live_intervals(struct gl_program *prog,
struct interval_list *liveIntervals)
{
- GLint intBegin[MAX_PROGRAM_TEMPS], intEnd[MAX_PROGRAM_TEMPS];
+ GLint intBegin[REG_ALLOCATE_MAX_PROGRAM_TEMPS];
+ GLint intEnd[REG_ALLOCATE_MAX_PROGRAM_TEMPS];
GLuint i;
/*
/* Build live intervals list from intermediate arrays */
liveIntervals->Num = 0;
- for (i = 0; i < MAX_PROGRAM_TEMPS; i++) {
+ for (i = 0; i < REG_ALLOCATE_MAX_PROGRAM_TEMPS; i++) {
if (intBegin[i] >= 0) {
struct interval inv;
inv.Reg = i;
/** Scan the array of used register flags to find free entry */
static GLint
-alloc_register(GLboolean usedRegs[MAX_PROGRAM_TEMPS])
+alloc_register(GLboolean usedRegs[REG_ALLOCATE_MAX_PROGRAM_TEMPS])
{
GLuint k;
- for (k = 0; k < MAX_PROGRAM_TEMPS; k++) {
+ for (k = 0; k < REG_ALLOCATE_MAX_PROGRAM_TEMPS; k++) {
if (!usedRegs[k]) {
usedRegs[k] = GL_TRUE;
return k;
_mesa_reallocate_registers(struct gl_program *prog)
{
struct interval_list liveIntervals;
- GLint registerMap[MAX_PROGRAM_TEMPS];
- GLboolean usedRegs[MAX_PROGRAM_TEMPS];
+ GLint registerMap[REG_ALLOCATE_MAX_PROGRAM_TEMPS];
+ GLboolean usedRegs[REG_ALLOCATE_MAX_PROGRAM_TEMPS];
GLuint i;
GLint maxTemp = -1;
_mesa_print_program(prog);
}
- for (i = 0; i < MAX_PROGRAM_TEMPS; i++){
+ for (i = 0; i < REG_ALLOCATE_MAX_PROGRAM_TEMPS; i++){
registerMap[i] = -1;
usedRegs[i] = GL_FALSE;
}
#if 0
static void
-print_it(GLcontext *ctx, struct gl_program *program, const char *txt) {
+print_it(struct gl_context *ctx, struct gl_program *program, const char *txt) {
fprintf(stderr, "%s (%u inst):\n", txt, program->NumInstructions);
_mesa_print_program(program);
_mesa_print_program_parameters(ctx, program);
}
#endif
+/**
+ * This pass replaces CMP T0, T1 T2 T0 with MOV T0, T2 when the CMP
+ * instruction is the first instruction to write to register T0. The are
+ * several lowering passes done in GLSL IR (e.g. branches and
+ * relative addressing) that create a large number of conditional assignments
+ * that ir_to_mesa converts to CMP instructions like the one mentioned above.
+ *
+ * Here is why this conversion is safe:
+ * CMP T0, T1 T2 T0 can be expanded to:
+ * if (T1 < 0.0)
+ * MOV T0, T2;
+ * else
+ * MOV T0, T0;
+ *
+ * If (T1 < 0.0) evaluates to true then our replacement MOV T0, T2 is the same
+ * as the original program. If (T1 < 0.0) evaluates to false, executing
+ * MOV T0, T0 will store a garbage value in T0 since T0 is uninitialized.
+ * Therefore, it doesn't matter that we are replacing MOV T0, T0 with MOV T0, T2
+ * because any instruction that was going to read from T0 after this was going
+ * to read a garbage value anyway.
+ */
+static void
+_mesa_simplify_cmp(struct gl_program * program)
+{
+ GLuint tempWrites[REG_ALLOCATE_MAX_PROGRAM_TEMPS];
+ GLuint outputWrites[MAX_PROGRAM_OUTPUTS];
+ GLuint i;
+
+ if (dbg) {
+ printf("Optimize: Begin reads without writes\n");
+ _mesa_print_program(program);
+ }
+
+ for (i = 0; i < REG_ALLOCATE_MAX_PROGRAM_TEMPS; i++) {
+ tempWrites[i] = 0;
+ }
+
+ for (i = 0; i < MAX_PROGRAM_OUTPUTS; i++) {
+ outputWrites[i] = 0;
+ }
+
+ for (i = 0; i < program->NumInstructions; i++) {
+ struct prog_instruction *inst = program->Instructions + i;
+ GLuint prevWriteMask;
+
+ /* Give up if we encounter relative addressing or flow control. */
+ if (_mesa_is_flow_control_opcode(inst->Opcode) || inst->DstReg.RelAddr) {
+ return;
+ }
+
+ if (inst->DstReg.File == PROGRAM_OUTPUT) {
+ assert(inst->DstReg.Index < MAX_PROGRAM_OUTPUTS);
+ prevWriteMask = outputWrites[inst->DstReg.Index];
+ outputWrites[inst->DstReg.Index] |= inst->DstReg.WriteMask;
+ } else if (inst->DstReg.File == PROGRAM_TEMPORARY) {
+ assert(inst->DstReg.Index < REG_ALLOCATE_MAX_PROGRAM_TEMPS);
+ prevWriteMask = tempWrites[inst->DstReg.Index];
+ tempWrites[inst->DstReg.Index] |= inst->DstReg.WriteMask;
+ } else {
+ /* No other register type can be a destination register. */
+ continue;
+ }
+
+ /* For a CMP to be considered a conditional write, the destination
+ * register and source register two must be the same. */
+ if (inst->Opcode == OPCODE_CMP
+ && !(inst->DstReg.WriteMask & prevWriteMask)
+ && inst->SrcReg[2].File == inst->DstReg.File
+ && inst->SrcReg[2].Index == inst->DstReg.Index
+ && inst->DstReg.WriteMask == get_src_arg_mask(inst, 2, NO_MASK)) {
+
+ inst->Opcode = OPCODE_MOV;
+ inst->SrcReg[0] = inst->SrcReg[1];
+
+ /* Unused operands are expected to have the file set to
+ * PROGRAM_UNDEFINED. This is how _mesa_init_instructions initializes
+ * all of the sources.
+ */
+ inst->SrcReg[1].File = PROGRAM_UNDEFINED;
+ inst->SrcReg[1].Swizzle = SWIZZLE_NOOP;
+ inst->SrcReg[2].File = PROGRAM_UNDEFINED;
+ inst->SrcReg[2].Swizzle = SWIZZLE_NOOP;
+ }
+ }
+ if (dbg) {
+ printf("Optimize: End reads without writes\n");
+ _mesa_print_program(program);
+ }
+}
/**
* Apply optimizations to the given program to eliminate unnecessary
* instructions, temp regs, etc.
*/
void
-_mesa_optimize_program(GLcontext *ctx, struct gl_program *program)
+_mesa_optimize_program(struct gl_context *ctx, struct gl_program *program)
{
GLboolean any_change;
+ _mesa_simplify_cmp(program);
/* Stop when no modifications were output */
do {
any_change = GL_FALSE;
any_change = GL_TRUE;
if (_mesa_remove_dead_code_local(program))
any_change = GL_TRUE;
+
+ any_change = _mesa_constant_fold(program) || any_change;
_mesa_reallocate_registers(program);
} while (any_change);
}