static void rgb_to_alpha_remap (
struct rc_instruction * inst,
struct rc_pair_instruction_arg * arg,
rc_register_file old_file,
rc_swizzle old_swz,
unsigned int new_index)
{
int new_src_index;
unsigned int i;
for (i = 0; i < 3; i++) {
if (get_swz(arg->Swizzle, i) == old_swz) {
SET_SWZ(arg->Swizzle, i, RC_SWIZZLE_W);
}
}
new_src_index = rc_pair_alloc_source(&inst->U.P, 0, 1,
old_file, new_index);
/* This conversion is not possible, we must have made a mistake in
* is_rgb_to_alpha_possible. */
if (new_src_index < 0) {
assert(0);
return;
}
arg->Source = new_src_index;
}
static int destructive_merge_instructions(
struct rc_pair_instruction * rgb,
struct rc_pair_instruction * alpha)
{
const struct rc_opcode_info * opcode;
assert(rgb->Alpha.Opcode == RC_OPCODE_NOP);
assert(alpha->RGB.Opcode == RC_OPCODE_NOP);
/* Presubtract registers need to be merged first so that registers
* needed by the presubtract operation can be placed in src0 and/or
* src1. */
/* Merge the rgb presubtract registers. */
const struct rc_opcode_info * rgb_info =
rc_get_opcode_info(rgb->RGB.Opcode);
if (alpha->RGB.Src[RC_PAIR_PRESUB_SRC].Used) {
unsigned int srcp_src;
unsigned int srcp_regs;
if (rgb->RGB.Src[RC_PAIR_PRESUB_SRC].Used)
return 0;
srcp_regs = rc_presubtract_src_reg_count(
alpha->RGB.Src[RC_PAIR_PRESUB_SRC].Index);
for(srcp_src = 0; srcp_src < srcp_regs; srcp_src++) {
unsigned int arg;
int free_source;
unsigned int one_way = 0;
struct radeon_pair_instruction_source srcp =
alpha->RGB.Src[srcp_src];
struct radeon_pair_instruction_source temp;
/* 2nd arg of 1 means this is an rgb source.
* 3rd arg of 0 means this is not an alpha source. */
free_source = rc_pair_alloc_source(rgb, 1, 0,
srcp.File, srcp.Index);
/* If free_source < 0 then there are no free source
* slots. */
if (free_source < 0)
return 0;
temp = rgb->RGB.Src[srcp_src];
rgb->RGB.Src[srcp_src] = rgb->RGB.Src[free_source];
/* srcp needs src0 and src1 to be the same */
if (free_source < srcp_src) {
if (!temp.Used)
continue;
free_source = rc_pair_alloc_source(rgb, 1, 0,
srcp.File, srcp.Index);
one_way = 1;
} else {
rgb->RGB.Src[free_source] = temp;
}
/* If free_source == srcp_src, then the presubtract
* source is already in the correct place. */
if (free_source == srcp_src)
continue;
/* Shuffle the sources, so we can put the
* presubtract source in the correct place. */
for (arg = 0; arg < rgb_info->NumSrcRegs; arg++) {
/*If this arg does not read from an rgb source,
* do nothing. */
if (rc_source_type_that_arg_reads(
rgb->RGB.Arg[arg].Source,
rgb->RGB.Arg[arg].Swizzle, 3)
!= RC_PAIR_SOURCE_RGB) {
continue;
}
if (rgb->RGB.Arg[arg].Source == srcp_src)
rgb->RGB.Arg[arg].Source = free_source;
/* We need to do this just in case register
* is one of the sources already, but in the
* wrong spot. */
else if(rgb->RGB.Arg[arg].Source == free_source
&& !one_way) {
rgb->RGB.Arg[arg].Source = srcp_src;
}
}
}
}
/* Merge the alpha presubtract registers */
if (alpha->Alpha.Src[RC_PAIR_PRESUB_SRC].Used) {
unsigned int srcp_src;
unsigned int srcp_regs;
if(rgb->Alpha.Src[RC_PAIR_PRESUB_SRC].Used)
return 0;
srcp_regs = rc_presubtract_src_reg_count(
alpha->Alpha.Src[RC_PAIR_PRESUB_SRC].Index);
for(srcp_src = 0; srcp_src < srcp_regs; srcp_src++) {
unsigned int arg;
int free_source;
unsigned int one_way = 0;
struct radeon_pair_instruction_source srcp =
alpha->Alpha.Src[srcp_src];
struct radeon_pair_instruction_source temp;
/* 2nd arg of 0 means this is not an rgb source.
* 3rd arg of 1 means this is an alpha source. */
free_source = rc_pair_alloc_source(rgb, 0, 1,
srcp.File, srcp.Index);
/* If free_source < 0 then there are no free source
* slots. */
if (free_source < 0)
return 0;
temp = rgb->Alpha.Src[srcp_src];
rgb->Alpha.Src[srcp_src] = rgb->Alpha.Src[free_source];
/* srcp needs src0 and src1 to be the same. */
if (free_source < srcp_src) {
if (!temp.Used)
continue;
free_source = rc_pair_alloc_source(rgb, 0, 1,
temp.File, temp.Index);
one_way = 1;
} else {
rgb->Alpha.Src[free_source] = temp;
}
/* If free_source == srcp_src, then the presubtract
* source is already in the correct place. */
if (free_source == srcp_src)
continue;
/* Shuffle the sources, so we can put the
* presubtract source in the correct place. */
for(arg = 0; arg < rgb_info->NumSrcRegs; arg++) {
/*If this arg does not read from an alpha
* source, do nothing. */
if (rc_source_type_that_arg_reads(
rgb->RGB.Arg[arg].Source,
rgb->RGB.Arg[arg].Swizzle, 3)
!= RC_PAIR_SOURCE_ALPHA) {
continue;
}
if (rgb->RGB.Arg[arg].Source == srcp_src)
rgb->RGB.Arg[arg].Source = free_source;
else if (rgb->RGB.Arg[arg].Source == free_source
&& !one_way) {
rgb->RGB.Arg[arg].Source = srcp_src;
}
}
}
}
/* Copy alpha args into rgb */
opcode = rc_get_opcode_info(alpha->Alpha.Opcode);
for(unsigned int arg = 0; arg < opcode->NumSrcRegs; ++arg) {
unsigned int srcrgb = 0;
unsigned int srcalpha = 0;
unsigned int oldsrc = alpha->Alpha.Arg[arg].Source;
rc_register_file file = 0;
unsigned int index = 0;
if (alpha->Alpha.Arg[arg].Swizzle < 3) {
srcrgb = 1;
file = alpha->RGB.Src[oldsrc].File;
index = alpha->RGB.Src[oldsrc].Index;
} else if (alpha->Alpha.Arg[arg].Swizzle < 4) {
srcalpha = 1;
file = alpha->Alpha.Src[oldsrc].File;
index = alpha->Alpha.Src[oldsrc].Index;
}
int source = rc_pair_alloc_source(rgb, srcrgb, srcalpha, file, index);
if (source < 0)
return 0;
rgb->Alpha.Arg[arg].Source = source;
rgb->Alpha.Arg[arg].Swizzle = alpha->Alpha.Arg[arg].Swizzle;
rgb->Alpha.Arg[arg].Abs = alpha->Alpha.Arg[arg].Abs;
rgb->Alpha.Arg[arg].Negate = alpha->Alpha.Arg[arg].Negate;
}
/* Copy alpha opcode into rgb */
rgb->Alpha.Opcode = alpha->Alpha.Opcode;
rgb->Alpha.DestIndex = alpha->Alpha.DestIndex;
rgb->Alpha.WriteMask = alpha->Alpha.WriteMask;
rgb->Alpha.OutputWriteMask = alpha->Alpha.OutputWriteMask;
rgb->Alpha.DepthWriteMask = alpha->Alpha.DepthWriteMask;
rgb->Alpha.Saturate = alpha->Alpha.Saturate;
/* Merge ALU result writing */
if (alpha->WriteALUResult) {
if (rgb->WriteALUResult)
return 0;
rgb->WriteALUResult = alpha->WriteALUResult;
rgb->ALUResultCompare = alpha->ALUResultCompare;
}
return 1;
}
/* This is a helper function for destructive_merge_instructions(). It helps
* merge presubtract sources from two instructions and makes sure the
* presubtract sources end up in the correct spot. This function assumes that
* dst_full is an rgb instruction, meaning that it has a vector instruction(rgb)
* but no scalar instruction (alpha).
* @return 0 if merging the presubtract sources fails.
* @retrun 1 if merging the presubtract sources succeeds.
*/
static int merge_presub_sources(
struct rc_pair_instruction * dst_full,
struct rc_pair_sub_instruction src,
unsigned int type)
{
unsigned int srcp_src, srcp_regs, is_rgb, is_alpha;
struct rc_pair_sub_instruction * dst_sub;
const struct rc_opcode_info * info;
assert(dst_full->Alpha.Opcode == RC_OPCODE_NOP);
switch(type) {
case RC_SOURCE_RGB:
is_rgb = 1;
is_alpha = 0;
dst_sub = &dst_full->RGB;
break;
case RC_SOURCE_ALPHA:
is_rgb = 0;
is_alpha = 1;
dst_sub = &dst_full->Alpha;
break;
default:
assert(0);
return 0;
}
info = rc_get_opcode_info(dst_full->RGB.Opcode);
if (dst_sub->Src[RC_PAIR_PRESUB_SRC].Used)
return 0;
srcp_regs = rc_presubtract_src_reg_count(
src.Src[RC_PAIR_PRESUB_SRC].Index);
for(srcp_src = 0; srcp_src < srcp_regs; srcp_src++) {
unsigned int arg;
int free_source;
unsigned int one_way = 0;
struct rc_pair_instruction_source srcp = src.Src[srcp_src];
struct rc_pair_instruction_source temp;
free_source = rc_pair_alloc_source(dst_full, is_rgb, is_alpha,
srcp.File, srcp.Index);
/* If free_source < 0 then there are no free source
* slots. */
if (free_source < 0)
return 0;
temp = dst_sub->Src[srcp_src];
dst_sub->Src[srcp_src] = dst_sub->Src[free_source];
/* srcp needs src0 and src1 to be the same */
if (free_source < srcp_src) {
if (!temp.Used)
continue;
free_source = rc_pair_alloc_source(dst_full, is_rgb,
is_alpha, temp.File, temp.Index);
if (free_source < 0)
return 0;
one_way = 1;
} else {
dst_sub->Src[free_source] = temp;
}
/* If free_source == srcp_src, then the presubtract
* source is already in the correct place. */
if (free_source == srcp_src)
continue;
/* Shuffle the sources, so we can put the
* presubtract source in the correct place. */
for(arg = 0; arg < info->NumSrcRegs; arg++) {
/*If this arg does not read from an rgb source,
* do nothing. */
if (!(rc_source_type_swz(dst_full->RGB.Arg[arg].Swizzle,
3) & type)) {
continue;
}
if (dst_full->RGB.Arg[arg].Source == srcp_src)
dst_full->RGB.Arg[arg].Source = free_source;
/* We need to do this just in case register
* is one of the sources already, but in the
* wrong spot. */
else if(dst_full->RGB.Arg[arg].Source == free_source
&& !one_way) {
dst_full->RGB.Arg[arg].Source = srcp_src;
}
}
}
return 1;
}
/* This function assumes that rgb.Alpha and alpha.RGB are unused */
static int destructive_merge_instructions(
struct rc_pair_instruction * rgb,
struct rc_pair_instruction * alpha)
{
const struct rc_opcode_info * opcode;
assert(rgb->Alpha.Opcode == RC_OPCODE_NOP);
assert(alpha->RGB.Opcode == RC_OPCODE_NOP);
/* Presubtract registers need to be merged first so that registers
* needed by the presubtract operation can be placed in src0 and/or
* src1. */
/* Merge the rgb presubtract registers. */
if (alpha->RGB.Src[RC_PAIR_PRESUB_SRC].Used) {
if (!merge_presub_sources(rgb, alpha->RGB, RC_SOURCE_RGB)) {
return 0;
}
}
/* Merge the alpha presubtract registers */
if (alpha->Alpha.Src[RC_PAIR_PRESUB_SRC].Used) {
if(!merge_presub_sources(rgb, alpha->Alpha, RC_SOURCE_ALPHA)){
return 0;
}
}
/* Copy alpha args into rgb */
opcode = rc_get_opcode_info(alpha->Alpha.Opcode);
for(unsigned int arg = 0; arg < opcode->NumSrcRegs; ++arg) {
unsigned int srcrgb = 0;
unsigned int srcalpha = 0;
unsigned int oldsrc = alpha->Alpha.Arg[arg].Source;
rc_register_file file = 0;
unsigned int index = 0;
int source;
if (alpha->Alpha.Arg[arg].Swizzle < 3) {
srcrgb = 1;
file = alpha->RGB.Src[oldsrc].File;
index = alpha->RGB.Src[oldsrc].Index;
} else if (alpha->Alpha.Arg[arg].Swizzle < 4) {
srcalpha = 1;
file = alpha->Alpha.Src[oldsrc].File;
index = alpha->Alpha.Src[oldsrc].Index;
}
source = rc_pair_alloc_source(rgb, srcrgb, srcalpha, file, index);
if (source < 0)
return 0;
rgb->Alpha.Arg[arg].Source = source;
rgb->Alpha.Arg[arg].Swizzle = alpha->Alpha.Arg[arg].Swizzle;
rgb->Alpha.Arg[arg].Abs = alpha->Alpha.Arg[arg].Abs;
rgb->Alpha.Arg[arg].Negate = alpha->Alpha.Arg[arg].Negate;
}
/* Copy alpha opcode into rgb */
rgb->Alpha.Opcode = alpha->Alpha.Opcode;
rgb->Alpha.DestIndex = alpha->Alpha.DestIndex;
rgb->Alpha.WriteMask = alpha->Alpha.WriteMask;
rgb->Alpha.OutputWriteMask = alpha->Alpha.OutputWriteMask;
rgb->Alpha.DepthWriteMask = alpha->Alpha.DepthWriteMask;
rgb->Alpha.Saturate = alpha->Alpha.Saturate;
/* Merge ALU result writing */
if (alpha->WriteALUResult) {
if (rgb->WriteALUResult)
return 0;
rgb->WriteALUResult = alpha->WriteALUResult;
rgb->ALUResultCompare = alpha->ALUResultCompare;
}
return 1;
}
/**
* Fill the given ALU instruction's opcodes and source operands into the given pair,
* if possible.
*/
static void set_pair_instruction(struct r300_fragment_program_compiler *c,
struct rc_pair_instruction * pair,
struct rc_sub_instruction * inst)
{
memset(pair, 0, sizeof(struct rc_pair_instruction));
int needrgb, needalpha, istranscendent;
classify_instruction(inst, &needrgb, &needalpha, &istranscendent);
if (needrgb) {
if (istranscendent)
pair->RGB.Opcode = RC_OPCODE_REPL_ALPHA;
else
pair->RGB.Opcode = inst->Opcode;
if (inst->SaturateMode == RC_SATURATE_ZERO_ONE)
pair->RGB.Saturate = 1;
}
if (needalpha) {
pair->Alpha.Opcode = inst->Opcode;
if (inst->SaturateMode == RC_SATURATE_ZERO_ONE)
pair->Alpha.Saturate = 1;
}
const struct rc_opcode_info * opcode = rc_get_opcode_info(inst->Opcode);
int nargs = opcode->NumSrcRegs;
int i;
/* Special case for DDX/DDY (MDH/MDV). */
if (inst->Opcode == RC_OPCODE_DDX || inst->Opcode == RC_OPCODE_DDY) {
nargs++;
}
for(i = 0; i < opcode->NumSrcRegs; ++i) {
int source;
if (needrgb && !istranscendent) {
unsigned int srcrgb = 0;
unsigned int srcalpha = 0;
int j;
for(j = 0; j < 3; ++j) {
unsigned int swz = GET_SWZ(inst->SrcReg[i].Swizzle, j);
if (swz < 3)
srcrgb = 1;
else if (swz < 4)
srcalpha = 1;
}
source = rc_pair_alloc_source(pair, srcrgb, srcalpha,
inst->SrcReg[i].File, inst->SrcReg[i].Index);
pair->RGB.Arg[i].Source = source;
pair->RGB.Arg[i].Swizzle = inst->SrcReg[i].Swizzle & 0x1ff;
pair->RGB.Arg[i].Abs = inst->SrcReg[i].Abs;
pair->RGB.Arg[i].Negate = !!(inst->SrcReg[i].Negate & (RC_MASK_X | RC_MASK_Y | RC_MASK_Z));
}
if (needalpha) {
unsigned int srcrgb = 0;
unsigned int srcalpha = 0;
unsigned int swz = GET_SWZ(inst->SrcReg[i].Swizzle, istranscendent ? 0 : 3);
if (swz < 3)
srcrgb = 1;
else if (swz < 4)
srcalpha = 1;
source = rc_pair_alloc_source(pair, srcrgb, srcalpha,
inst->SrcReg[i].File, inst->SrcReg[i].Index);
pair->Alpha.Arg[i].Source = source;
pair->Alpha.Arg[i].Swizzle = swz;
pair->Alpha.Arg[i].Abs = inst->SrcReg[i].Abs;
pair->Alpha.Arg[i].Negate = !!(inst->SrcReg[i].Negate & RC_MASK_W);
}
}
/* Destination handling */
if (inst->DstReg.File == RC_FILE_OUTPUT) {
if (inst->DstReg.Index == c->OutputDepth) {
pair->Alpha.DepthWriteMask |= GET_BIT(inst->DstReg.WriteMask, 3);
} else {
for (i = 0; i < 4; i++) {
if (inst->DstReg.Index == c->OutputColor[i]) {
pair->RGB.Target = i;
pair->Alpha.Target = i;
pair->RGB.OutputWriteMask |=
inst->DstReg.WriteMask & RC_MASK_XYZ;
pair->Alpha.OutputWriteMask |=
GET_BIT(inst->DstReg.WriteMask, 3);
break;
}
}
}
} else {
if (needrgb) {
pair->RGB.DestIndex = inst->DstReg.Index;
pair->RGB.WriteMask |= inst->DstReg.WriteMask & RC_MASK_XYZ;
}
if (needalpha) {
pair->Alpha.DestIndex = inst->DstReg.Index;
pair->Alpha.WriteMask |= GET_BIT(inst->DstReg.WriteMask, 3);
}
}
if (inst->WriteALUResult) {
pair->WriteALUResult = inst->WriteALUResult;
pair->ALUResultCompare = inst->ALUResultCompare;
}
}