static void projective_divide(struct r300_fragment_program_compiler *compiler,
			      struct rc_instruction *inst)
{
	struct rc_instruction *inst_mul, *inst_rcp;

	unsigned temp = rc_find_free_temporary(&compiler->Base);

	inst_rcp = rc_insert_new_instruction(&compiler->Base, inst->Prev);
	inst_rcp->U.I.Opcode = RC_OPCODE_RCP;
	inst_rcp->U.I.DstReg.File = RC_FILE_TEMPORARY;
	inst_rcp->U.I.DstReg.Index = temp;
	inst_rcp->U.I.DstReg.WriteMask = RC_MASK_W;
	inst_rcp->U.I.SrcReg[0] = inst->U.I.SrcReg[0];
	/* Because the input can be arbitrarily swizzled,
	 * read the component mapped to W. */
	inst_rcp->U.I.SrcReg[0].Swizzle =
		RC_MAKE_SWIZZLE_SMEAR(GET_SWZ(inst->U.I.SrcReg[0].Swizzle, 3));

	inst_mul = rc_insert_new_instruction(&compiler->Base, inst->Prev);
	inst_mul->U.I.Opcode = RC_OPCODE_MUL;
	inst_mul->U.I.DstReg.File = RC_FILE_TEMPORARY;
	inst_mul->U.I.DstReg.Index = temp;
	inst_mul->U.I.SrcReg[0] = inst->U.I.SrcReg[0];
	inst_mul->U.I.SrcReg[1].File = RC_FILE_TEMPORARY;
	inst_mul->U.I.SrcReg[1].Index = temp;
	inst_mul->U.I.SrcReg[1].Swizzle = RC_SWIZZLE_WWWW;

	reset_srcreg(&inst->U.I.SrcReg[0]);
	inst->U.I.Opcode = RC_OPCODE_TEX;
	inst->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
	inst->U.I.SrcReg[0].Index = temp;
}
/**
 * The FACE input in hardware contains 1 if it's a back face, 0 otherwise.
 * Gallium and OpenGL define it the other way around.
 *
 * So let's just negate FACE at the beginning of the shader and rewrite the rest
 * of the shader to read from the newly allocated temporary.
 */
void rc_transform_fragment_face(struct radeon_compiler *c, unsigned face)
{
	unsigned tempregi = rc_find_free_temporary(c);
	struct rc_instruction *inst_add;
	struct rc_instruction *inst;

	/* perspective divide */
	inst_add = rc_insert_new_instruction(c, &c->Program.Instructions);
	inst_add->U.I.Opcode = RC_OPCODE_ADD;

	inst_add->U.I.DstReg.File = RC_FILE_TEMPORARY;
	inst_add->U.I.DstReg.Index = tempregi;
	inst_add->U.I.DstReg.WriteMask = RC_MASK_X;

	inst_add->U.I.SrcReg[0].File = RC_FILE_NONE;
	inst_add->U.I.SrcReg[0].Swizzle = RC_SWIZZLE_1111;

	inst_add->U.I.SrcReg[1].File = RC_FILE_INPUT;
	inst_add->U.I.SrcReg[1].Index = face;
	inst_add->U.I.SrcReg[1].Swizzle = RC_SWIZZLE_XXXX;
	inst_add->U.I.SrcReg[1].Negate = RC_MASK_XYZW;

	for (inst = inst_add->Next; inst != &c->Program.Instructions; inst = inst->Next) {
		const struct rc_opcode_info * opcode = rc_get_opcode_info(inst->U.I.Opcode);
		unsigned i;

		for(i = 0; i < opcode->NumSrcRegs; i++) {
			if (inst->U.I.SrcReg[i].File == RC_FILE_INPUT &&
			    inst->U.I.SrcReg[i].Index == face) {
				inst->U.I.SrcReg[i].File = RC_FILE_TEMPORARY;
				inst->U.I.SrcReg[i].Index = tempregi;
			}
		}
	}
}
Beispiel #3
0
/**
 * R3xx-R4xx vertex engine does not support the Absolute source operand modifier
 * and the Saturate opcode modifier. Only Absolute is currently transformed.
 */
static int transform_nonnative_modifiers(
	struct radeon_compiler *c,
	struct rc_instruction *inst,
	void* unused)
{
	const struct rc_opcode_info *opcode = rc_get_opcode_info(inst->U.I.Opcode);
	unsigned i;

	/* Transform ABS(a) to MAX(a, -a). */
	for (i = 0; i < opcode->NumSrcRegs; i++) {
		if (inst->U.I.SrcReg[i].Abs) {
			struct rc_instruction *new_inst;
			unsigned temp;

			inst->U.I.SrcReg[i].Abs = 0;

			temp = rc_find_free_temporary(c);

			new_inst = rc_insert_new_instruction(c, inst->Prev);
			new_inst->U.I.Opcode = RC_OPCODE_MAX;
			new_inst->U.I.DstReg.File = RC_FILE_TEMPORARY;
			new_inst->U.I.DstReg.Index = temp;
			new_inst->U.I.SrcReg[0] = inst->U.I.SrcReg[i];
			new_inst->U.I.SrcReg[1] = inst->U.I.SrcReg[i];
			new_inst->U.I.SrcReg[1].Negate ^= RC_MASK_XYZW;

			memset(&inst->U.I.SrcReg[i], 0, sizeof(inst->U.I.SrcReg[i]));
			inst->U.I.SrcReg[i].File = RC_FILE_TEMPORARY;
			inst->U.I.SrcReg[i].Index = temp;
			inst->U.I.SrcReg[i].Swizzle = RC_SWIZZLE_XYZW;
		}
	}
	return 1;
}
Beispiel #4
0
static void transform_negative_addressing(struct r300_vertex_program_compiler *c,
					  struct rc_instruction *arl,
					  struct rc_instruction *end,
					  int min_offset)
{
	struct rc_instruction *inst, *add;
	unsigned const_swizzle;

	/* Transform ARL */
	add = rc_insert_new_instruction(&c->Base, arl->Prev);
	add->U.I.Opcode = RC_OPCODE_ADD;
	add->U.I.DstReg.File = RC_FILE_TEMPORARY;
	add->U.I.DstReg.Index = rc_find_free_temporary(&c->Base);
	add->U.I.DstReg.WriteMask = RC_MASK_X;
	add->U.I.SrcReg[0] = arl->U.I.SrcReg[0];
	add->U.I.SrcReg[1].File = RC_FILE_CONSTANT;
	add->U.I.SrcReg[1].Index = rc_constants_add_immediate_scalar(&c->Base.Program.Constants,
								     min_offset, &const_swizzle);
	add->U.I.SrcReg[1].Swizzle = const_swizzle;

	arl->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
	arl->U.I.SrcReg[0].Index = add->U.I.DstReg.Index;
	arl->U.I.SrcReg[0].Swizzle = RC_SWIZZLE_XXXX;

	/* Rewrite offsets up to and excluding inst. */
	for (inst = arl->Next; inst != end; inst = inst->Next) {
		const struct rc_opcode_info * opcode = rc_get_opcode_info(inst->U.I.Opcode);

		for (unsigned i = 0; i < opcode->NumSrcRegs; i++)
			if (inst->U.I.SrcReg[i].RelAddr)
				inst->U.I.SrcReg[i].Index -= min_offset;
	}
}
Beispiel #5
0
static void transform_instruction(struct tgsi_to_rc * ttr, struct tgsi_full_instruction * src)
{
    struct rc_instruction * dst;
    int i;

    if (src->Instruction.Opcode == TGSI_OPCODE_END)
        return;

    dst = rc_insert_new_instruction(ttr->compiler, ttr->compiler->Program.Instructions.Prev);
    dst->U.I.Opcode = translate_opcode(src->Instruction.Opcode);
    dst->U.I.SaturateMode = translate_saturate(src->Instruction.Saturate);

    if (src->Instruction.NumDstRegs)
        transform_dstreg(ttr, &dst->U.I.DstReg, &src->FullDstRegisters[0]);

    for(i = 0; i < src->Instruction.NumSrcRegs; ++i) {
        if (src->FullSrcRegisters[i].SrcRegister.File == TGSI_FILE_SAMPLER)
            dst->U.I.TexSrcUnit = src->FullSrcRegisters[i].SrcRegister.Index;
        else
            transform_srcreg(ttr, &dst->U.I.SrcReg[i], &src->FullSrcRegisters[i]);
    }

    /* Texturing. */
    transform_texture(dst, src->InstructionExtTexture);
}
static void lower_texture_rect(struct r300_fragment_program_compiler *compiler,
							   struct rc_instruction *inst)
{
	struct rc_instruction *inst_rect;
	unsigned temp = rc_find_free_temporary(&compiler->Base);

	if (inst->U.I.TexSrcTarget == RC_TEXTURE_RECT ||
		compiler->state.unit[inst->U.I.TexSrcUnit].non_normalized_coords) {
		inst_rect = rc_insert_new_instruction(&compiler->Base, inst->Prev);

		inst_rect->U.I.Opcode = RC_OPCODE_MUL;
		inst_rect->U.I.DstReg.File = RC_FILE_TEMPORARY;
		inst_rect->U.I.DstReg.Index = temp;
		inst_rect->U.I.SrcReg[0] = inst->U.I.SrcReg[0];
		inst_rect->U.I.SrcReg[1].File = RC_FILE_CONSTANT;
		inst_rect->U.I.SrcReg[1].Index =
				rc_constants_add_state(&compiler->Base.Program.Constants,
									   RC_STATE_R300_TEXRECT_FACTOR, inst->U.I.TexSrcUnit);

		reset_srcreg(&inst->U.I.SrcReg[0]);
		inst->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
		inst->U.I.SrcReg[0].Index = temp;

		inst->U.I.TexSrcTarget = RC_TEXTURE_2D;
	}
}
Beispiel #7
0
void add_instruction(struct radeon_compiler *c, const char * inst_string)
{
	struct rc_instruction * new_inst =
		rc_insert_new_instruction(c, c->Program.Instructions.Prev);

	parse_rc_normal_instruction(new_inst, inst_string);

}
static void rewrite_source(struct radeon_compiler * c,
		struct rc_instruction * inst, unsigned src)
{
	struct rc_swizzle_split split;
	unsigned int tempreg = rc_find_free_temporary(c);
	unsigned int usemask;

	usemask = 0;
	for(unsigned int chan = 0; chan < 4; ++chan) {
		if (GET_SWZ(inst->U.I.SrcReg[src].Swizzle, chan) != RC_SWIZZLE_UNUSED)
			usemask |= 1 << chan;
	}

	c->SwizzleCaps->Split(inst->U.I.SrcReg[src], usemask, &split);

	for(unsigned int phase = 0; phase < split.NumPhases; ++phase) {
		struct rc_instruction * mov = rc_insert_new_instruction(c, inst->Prev);
		unsigned int phase_refmask;
		unsigned int masked_negate;

		mov->U.I.Opcode = RC_OPCODE_MOV;
		mov->U.I.DstReg.File = RC_FILE_TEMPORARY;
		mov->U.I.DstReg.Index = tempreg;
		mov->U.I.DstReg.WriteMask = split.Phase[phase];
		mov->U.I.SrcReg[0] = inst->U.I.SrcReg[src];
		mov->U.I.PreSub = inst->U.I.PreSub;

		phase_refmask = 0;
		for(unsigned int chan = 0; chan < 4; ++chan) {
			if (!GET_BIT(split.Phase[phase], chan))
				SET_SWZ(mov->U.I.SrcReg[0].Swizzle, chan, RC_SWIZZLE_UNUSED);
			else
				phase_refmask |= 1 << GET_SWZ(mov->U.I.SrcReg[0].Swizzle, chan);
		}

		phase_refmask &= RC_MASK_XYZW;

		masked_negate = split.Phase[phase] & mov->U.I.SrcReg[0].Negate;
		if (masked_negate == 0)
			mov->U.I.SrcReg[0].Negate = 0;
		else if (masked_negate == split.Phase[phase])
			mov->U.I.SrcReg[0].Negate = RC_MASK_XYZW;

	}

	inst->U.I.SrcReg[src].File = RC_FILE_TEMPORARY;
	inst->U.I.SrcReg[src].Index = tempreg;
	inst->U.I.SrcReg[src].Swizzle = 0;
	inst->U.I.SrcReg[src].Negate = RC_MASK_NONE;
	inst->U.I.SrcReg[src].Abs = 0;
	for(unsigned int chan = 0; chan < 4; ++chan) {
		SET_SWZ(inst->U.I.SrcReg[src].Swizzle, chan,
				GET_BIT(usemask, chan) ? chan : RC_SWIZZLE_UNUSED);
	}
}
Beispiel #9
0
/**
 * Vertex engine cannot read two inputs or two constants at the same time.
 * Introduce intermediate MOVs to temporary registers to account for this.
 */
static int transform_source_conflicts(
	struct radeon_compiler *c,
	struct rc_instruction* inst,
	void* unused)
{
	const struct rc_opcode_info * opcode = rc_get_opcode_info(inst->U.I.Opcode);

	if (opcode->NumSrcRegs == 3) {
		if (t_src_conflict(inst->U.I.SrcReg[1], inst->U.I.SrcReg[2])
		    || t_src_conflict(inst->U.I.SrcReg[0], inst->U.I.SrcReg[2])) {
			int tmpreg = rc_find_free_temporary(c);
			struct rc_instruction * inst_mov = rc_insert_new_instruction(c, inst->Prev);
			inst_mov->U.I.Opcode = RC_OPCODE_MOV;
			inst_mov->U.I.DstReg.File = RC_FILE_TEMPORARY;
			inst_mov->U.I.DstReg.Index = tmpreg;
			inst_mov->U.I.SrcReg[0] = inst->U.I.SrcReg[2];

			reset_srcreg(&inst->U.I.SrcReg[2]);
			inst->U.I.SrcReg[2].File = RC_FILE_TEMPORARY;
			inst->U.I.SrcReg[2].Index = tmpreg;
		}
	}

	if (opcode->NumSrcRegs >= 2) {
		if (t_src_conflict(inst->U.I.SrcReg[1], inst->U.I.SrcReg[0])) {
			int tmpreg = rc_find_free_temporary(c);
			struct rc_instruction * inst_mov = rc_insert_new_instruction(c, inst->Prev);
			inst_mov->U.I.Opcode = RC_OPCODE_MOV;
			inst_mov->U.I.DstReg.File = RC_FILE_TEMPORARY;
			inst_mov->U.I.DstReg.Index = tmpreg;
			inst_mov->U.I.SrcReg[0] = inst->U.I.SrcReg[1];

			reset_srcreg(&inst->U.I.SrcReg[1]);
			inst->U.I.SrcReg[1].File = RC_FILE_TEMPORARY;
			inst->U.I.SrcReg[1].Index = tmpreg;
		}
	}

	return 1;
}
Beispiel #10
0
static struct rc_instruction *emit1(
	struct radeon_compiler * c, struct rc_instruction * after,
	rc_opcode Opcode, rc_saturate_mode Saturate, struct rc_dst_register DstReg,
	struct rc_src_register SrcReg)
{
	struct rc_instruction *fpi = rc_insert_new_instruction(c, after);

	fpi->U.I.Opcode = Opcode;
	fpi->U.I.SaturateMode = Saturate;
	fpi->U.I.DstReg = DstReg;
	fpi->U.I.SrcReg[0] = SrcReg;
	return fpi;
}
Beispiel #11
0
/**
 * The NV_vertex_program spec mandates that all registers be
 * initialized to zero. We do this here unconditionally.
 *
 * \note We rely on dead-code elimination in the compiler.
 */
static void initialize_NV_registers(struct radeon_compiler * compiler)
{
	unsigned int reg;
	struct rc_instruction * inst;

	for(reg = 0; reg < 12; ++reg) {
		inst = rc_insert_new_instruction(compiler, &compiler->Program.Instructions);
		inst->U.I.Opcode = RC_OPCODE_MOV;
		inst->U.I.DstReg.File = RC_FILE_TEMPORARY;
		inst->U.I.DstReg.Index = reg;
		inst->U.I.SrcReg[0].File = RC_FILE_NONE;
		inst->U.I.SrcReg[0].Swizzle = RC_SWIZZLE_0000;
	}

	inst = rc_insert_new_instruction(compiler, &compiler->Program.Instructions);
	inst->U.I.Opcode = RC_OPCODE_ARL;
	inst->U.I.DstReg.File = RC_FILE_ADDRESS;
	inst->U.I.DstReg.Index = 0;
	inst->U.I.DstReg.WriteMask = WRITEMASK_X;
	inst->U.I.SrcReg[0].File = RC_FILE_NONE;
	inst->U.I.SrcReg[0].Swizzle = RC_SWIZZLE_0000;
}
/**
 * Rewrite the program such that a given output is duplicated.
 */
void rc_copy_output(struct radeon_compiler * c, unsigned output, unsigned dup_output)
{
	unsigned tempreg = rc_find_free_temporary(c);
	struct rc_instruction * inst;

	for(inst = c->Program.Instructions.Next; inst != &c->Program.Instructions; inst = inst->Next) {
		const struct rc_opcode_info * opcode = rc_get_opcode_info(inst->U.I.Opcode);

		if (opcode->HasDstReg) {
			if (inst->U.I.DstReg.File == RC_FILE_OUTPUT && inst->U.I.DstReg.Index == output) {
				inst->U.I.DstReg.File = RC_FILE_TEMPORARY;
				inst->U.I.DstReg.Index = tempreg;
			}
		}
	}

	inst = rc_insert_new_instruction(c, c->Program.Instructions.Prev);
	inst->U.I.Opcode = RC_OPCODE_MOV;
	inst->U.I.DstReg.File = RC_FILE_OUTPUT;
	inst->U.I.DstReg.Index = output;

	inst->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
	inst->U.I.SrcReg[0].Index = tempreg;
	inst->U.I.SrcReg[0].Swizzle = RC_SWIZZLE_XYZW;

	inst = rc_insert_new_instruction(c, c->Program.Instructions.Prev);
	inst->U.I.Opcode = RC_OPCODE_MOV;
	inst->U.I.DstReg.File = RC_FILE_OUTPUT;
	inst->U.I.DstReg.Index = dup_output;

	inst->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
	inst->U.I.SrcReg[0].Index = tempreg;
	inst->U.I.SrcReg[0].Swizzle = RC_SWIZZLE_XYZW;

	c->Program.OutputsWritten |= 1 << dup_output;
}
static struct rc_instruction *emit1(
	struct radeon_compiler * c, struct rc_instruction * after,
	rc_opcode Opcode, struct rc_sub_instruction * base,
	struct rc_dst_register DstReg, struct rc_src_register SrcReg)
{
	struct rc_instruction *fpi = rc_insert_new_instruction(c, after);

	if (base) {
		memcpy(&fpi->U.I, base, sizeof(struct rc_sub_instruction));
	}

	fpi->U.I.Opcode = Opcode;
	fpi->U.I.DstReg = DstReg;
	fpi->U.I.SrcReg[0] = SrcReg;
	return fpi;
}
/**
 * Emit all ready texture instructions in a single block.
 *
 * Emit as a single block to (hopefully) sample many textures in parallel,
 * and to avoid hardware indirections on R300.
 */
static void emit_all_tex(struct schedule_state * s, struct rc_instruction * before)
{
	struct schedule_instruction *readytex;
	struct rc_instruction * inst_begin;

	assert(s->ReadyTEX);
	notify_sem_wait(s);

	/* Node marker for R300 */
	inst_begin = rc_insert_new_instruction(s->C, before->Prev);
	inst_begin->U.I.Opcode = RC_OPCODE_BEGIN_TEX;

	/* Link texture instructions back in */
	readytex = s->ReadyTEX;
	while(readytex) {
		rc_insert_instruction(before->Prev, readytex->Instruction);
		DBG("%i: commit TEX reads\n", readytex->Instruction->IP);

		/* All of the TEX instructions in the same TEX block have
		 * their source registers read from before any of the
		 * instructions in that block write to their destination
		 * registers.  This means that when we commit a TEX
		 * instruction, any other TEX instruction that wants to write
		 * to one of the committed instruction's source register can be
		 * marked as ready and should be emitted in the same TEX
		 * block. This prevents the following sequence from being
		 * emitted in two different TEX blocks:
		 * 0: TEX temp[0].xyz, temp[1].xy__, 2D[0];
		 * 1: TEX temp[1].xyz, temp[2].xy__, 2D[0];
		 */
		commit_update_reads(s, readytex);
		readytex = readytex->NextReady;
	}
	readytex = s->ReadyTEX;
	s->ReadyTEX = 0;
	while(readytex){
		DBG("%i: commit TEX writes\n", readytex->Instruction->IP);
		commit_update_writes(s, readytex);
		/* Set semaphore bits for last TEX instruction in the block */
		if (!readytex->NextReady) {
			readytex->Instruction->U.I.TexSemAcquire = 1;
			readytex->Instruction->U.I.TexSemWait = 1;
		}
		rc_list_add(&s->PendingTEX, rc_list(&s->C->Pool, readytex));
		readytex = readytex->NextReady;
	}
}
Beispiel #15
0
static void addArtificialOutputs(struct r300_vertex_program_compiler * compiler)
{
	int i;

	for(i = 0; i < 32; ++i) {
		if ((compiler->RequiredOutputs & (1 << i)) &&
		    !(compiler->Base.Program.OutputsWritten & (1 << i))) {
			struct rc_instruction * inst = rc_insert_new_instruction(&compiler->Base, compiler->Base.Program.Instructions.Prev);
			inst->U.I.Opcode = RC_OPCODE_MOV;

			inst->U.I.DstReg.File = RC_FILE_OUTPUT;
			inst->U.I.DstReg.Index = i;
			inst->U.I.DstReg.WriteMask = RC_MASK_XYZW;

			inst->U.I.SrcReg[0].File = RC_FILE_CONSTANT;
			inst->U.I.SrcReg[0].Index = 0;
			inst->U.I.SrcReg[0].Swizzle = RC_SWIZZLE_XYZW;

			compiler->Base.Program.OutputsWritten |= 1 << i;
		}
	}
}
static void scale_texcoords(struct r300_fragment_program_compiler *compiler,
			    struct rc_instruction *inst,
			    unsigned state_constant)
{
	struct rc_instruction *inst_mov;

	unsigned temp = rc_find_free_temporary(&compiler->Base);

	inst_mov = rc_insert_new_instruction(&compiler->Base, inst->Prev);

	inst_mov->U.I.Opcode = RC_OPCODE_MUL;
	inst_mov->U.I.DstReg.File = RC_FILE_TEMPORARY;
	inst_mov->U.I.DstReg.Index = temp;
	inst_mov->U.I.SrcReg[0] = inst->U.I.SrcReg[0];
	inst_mov->U.I.SrcReg[1].File = RC_FILE_CONSTANT;
	inst_mov->U.I.SrcReg[1].Index =
			rc_constants_add_state(&compiler->Base.Program.Constants,
					       state_constant, inst->U.I.TexSrcUnit);

	reset_srcreg(&inst->U.I.SrcReg[0]);
	inst->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
	inst->U.I.SrcReg[0].Index = temp;
}
/**
 * Introduce standard code fragment to deal with fragment.position.
 */
void rc_transform_fragment_wpos(struct radeon_compiler * c, unsigned wpos, unsigned new_input,
                                int full_vtransform)
{
	unsigned tempregi = rc_find_free_temporary(c);
	struct rc_instruction * inst_rcp;
	struct rc_instruction * inst_mul;
	struct rc_instruction * inst_mad;
	struct rc_instruction * inst;

	c->Program.InputsRead &= ~(1 << wpos);
	c->Program.InputsRead |= 1 << new_input;

	/* perspective divide */
	inst_rcp = rc_insert_new_instruction(c, &c->Program.Instructions);
	inst_rcp->U.I.Opcode = RC_OPCODE_RCP;

	inst_rcp->U.I.DstReg.File = RC_FILE_TEMPORARY;
	inst_rcp->U.I.DstReg.Index = tempregi;
	inst_rcp->U.I.DstReg.WriteMask = RC_MASK_W;

	inst_rcp->U.I.SrcReg[0].File = RC_FILE_INPUT;
	inst_rcp->U.I.SrcReg[0].Index = new_input;
	inst_rcp->U.I.SrcReg[0].Swizzle = RC_SWIZZLE_WWWW;

	inst_mul = rc_insert_new_instruction(c, inst_rcp);
	inst_mul->U.I.Opcode = RC_OPCODE_MUL;

	inst_mul->U.I.DstReg.File = RC_FILE_TEMPORARY;
	inst_mul->U.I.DstReg.Index = tempregi;
	inst_mul->U.I.DstReg.WriteMask = RC_MASK_XYZ;

	inst_mul->U.I.SrcReg[0].File = RC_FILE_INPUT;
	inst_mul->U.I.SrcReg[0].Index = new_input;

	inst_mul->U.I.SrcReg[1].File = RC_FILE_TEMPORARY;
	inst_mul->U.I.SrcReg[1].Index = tempregi;
	inst_mul->U.I.SrcReg[1].Swizzle = RC_SWIZZLE_WWWW;

	/* viewport transformation */
	inst_mad = rc_insert_new_instruction(c, inst_mul);
	inst_mad->U.I.Opcode = RC_OPCODE_MAD;

	inst_mad->U.I.DstReg.File = RC_FILE_TEMPORARY;
	inst_mad->U.I.DstReg.Index = tempregi;
	inst_mad->U.I.DstReg.WriteMask = RC_MASK_XYZ;

	inst_mad->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
	inst_mad->U.I.SrcReg[0].Index = tempregi;
	inst_mad->U.I.SrcReg[0].Swizzle = RC_SWIZZLE_XYZ0;

	inst_mad->U.I.SrcReg[1].File = RC_FILE_CONSTANT;
	inst_mad->U.I.SrcReg[1].Swizzle = RC_SWIZZLE_XYZ0;

	inst_mad->U.I.SrcReg[2].File = RC_FILE_CONSTANT;
	inst_mad->U.I.SrcReg[2].Swizzle = RC_SWIZZLE_XYZ0;

	if (full_vtransform) {
		inst_mad->U.I.SrcReg[1].Index = rc_constants_add_state(&c->Program.Constants, RC_STATE_R300_VIEWPORT_SCALE, 0);
		inst_mad->U.I.SrcReg[2].Index = rc_constants_add_state(&c->Program.Constants, RC_STATE_R300_VIEWPORT_OFFSET, 0);
	} else {
		inst_mad->U.I.SrcReg[1].Index =
		inst_mad->U.I.SrcReg[2].Index = rc_constants_add_state(&c->Program.Constants, RC_STATE_R300_WINDOW_DIMENSION, 0);
	}

	for (inst = inst_mad->Next; inst != &c->Program.Instructions; inst = inst->Next) {
		const struct rc_opcode_info * opcode = rc_get_opcode_info(inst->U.I.Opcode);
		unsigned i;

		for(i = 0; i < opcode->NumSrcRegs; i++) {
			if (inst->U.I.SrcReg[i].File == RC_FILE_INPUT &&
			    inst->U.I.SrcReg[i].Index == wpos) {
				inst->U.I.SrcReg[i].File = RC_FILE_TEMPORARY;
				inst->U.I.SrcReg[i].Index = tempregi;
			}
		}
	}
}
/**
 * Transform TEX, TXP, TXB, and KIL instructions in the following ways:
 *  - implement texture compare (shadow extensions)
 *  - extract non-native source / destination operands
 *  - premultiply texture coordinates for RECT
 *  - extract operand swizzles
 *  - introduce a temporary register when write masks are needed
 */
int radeonTransformTEX(
	struct radeon_compiler * c,
	struct rc_instruction * inst,
	void* data)
{
	struct r300_fragment_program_compiler *compiler =
		(struct r300_fragment_program_compiler*)data;

	if (inst->U.I.Opcode != RC_OPCODE_TEX &&
		inst->U.I.Opcode != RC_OPCODE_TXB &&
		inst->U.I.Opcode != RC_OPCODE_TXP &&
		inst->U.I.Opcode != RC_OPCODE_KIL)
		return 0;

	/* ARB_shadow & EXT_shadow_funcs */
	if (inst->U.I.Opcode != RC_OPCODE_KIL &&
		((c->Program.ShadowSamplers & (1 << inst->U.I.TexSrcUnit)) ||
		 (compiler->state.unit[inst->U.I.TexSrcUnit].compare_mode_enabled))) {
		rc_compare_func comparefunc = compiler->state.unit[inst->U.I.TexSrcUnit].texture_compare_func;

		if (comparefunc == RC_COMPARE_FUNC_NEVER || comparefunc == RC_COMPARE_FUNC_ALWAYS) {
			inst->U.I.Opcode = RC_OPCODE_MOV;

			if (comparefunc == RC_COMPARE_FUNC_ALWAYS) {
				inst->U.I.SrcReg[0].File = RC_FILE_NONE;
				inst->U.I.SrcReg[0].Swizzle = RC_SWIZZLE_1111;
			} else {
				inst->U.I.SrcReg[0] = shadow_ambient(compiler, inst->U.I.TexSrcUnit);
			}

			return 1;
		} else {
			rc_compare_func comparefunc = compiler->state.unit[inst->U.I.TexSrcUnit].texture_compare_func;
			struct rc_instruction * inst_rcp = NULL;
			struct rc_instruction * inst_mad;
			struct rc_instruction * inst_cmp;
			unsigned tmp_texsample;
			unsigned tmp_sum;
			unsigned tmp_recip_w = 0;
			int pass, fail, tex;

			/* Save the output register. */
			struct rc_dst_register output_reg = inst->U.I.DstReg;

			/* Redirect TEX to a new temp. */
			tmp_texsample = rc_find_free_temporary(c);
			inst->U.I.DstReg.File = RC_FILE_TEMPORARY;
			inst->U.I.DstReg.Index = tmp_texsample;
			inst->U.I.DstReg.WriteMask = RC_MASK_XYZW;

			if (inst->U.I.Opcode == RC_OPCODE_TXP) {
				tmp_recip_w = rc_find_free_temporary(c);

				/* Compute 1/W. */
				inst_rcp = rc_insert_new_instruction(c, inst);
				inst_rcp->U.I.Opcode = RC_OPCODE_RCP;
				inst_rcp->U.I.DstReg.File = RC_FILE_TEMPORARY;
				inst_rcp->U.I.DstReg.Index = tmp_recip_w;
				inst_rcp->U.I.DstReg.WriteMask = RC_MASK_W;
				inst_rcp->U.I.SrcReg[0] = inst->U.I.SrcReg[0];
				inst_rcp->U.I.SrcReg[0].Swizzle =
					RC_MAKE_SWIZZLE_SMEAR(GET_SWZ(inst->U.I.SrcReg[0].Swizzle, 3));
			}

			/* Perspective-divide Z by W (if it's TXP) and add the texture sample (see below). */
			tmp_sum = rc_find_free_temporary(c);
			inst_mad = rc_insert_new_instruction(c, inst_rcp ? inst_rcp : inst);
			inst_mad->U.I.DstReg.File = RC_FILE_TEMPORARY;
			inst_mad->U.I.DstReg.Index = tmp_sum;
			inst_mad->U.I.SrcReg[0] = inst->U.I.SrcReg[0];
			inst_mad->U.I.SrcReg[0].Swizzle =
				RC_MAKE_SWIZZLE_SMEAR(GET_SWZ(inst->U.I.SrcReg[0].Swizzle, 2));
			if (inst->U.I.Opcode == RC_OPCODE_TXP) {
				inst_mad->U.I.Opcode = RC_OPCODE_MAD;
				inst_mad->U.I.SrcReg[1].File = RC_FILE_TEMPORARY;
				inst_mad->U.I.SrcReg[1].Index = tmp_recip_w;
				inst_mad->U.I.SrcReg[1].Swizzle = RC_SWIZZLE_WWWW;
				tex = 2;
			} else {
				inst_mad->U.I.Opcode = RC_OPCODE_ADD;
				tex = 1;
			}
			inst_mad->U.I.SrcReg[tex].File = RC_FILE_TEMPORARY;
			inst_mad->U.I.SrcReg[tex].Index = tmp_texsample;
			inst_mad->U.I.SrcReg[tex].Swizzle = compiler->state.unit[inst->U.I.TexSrcUnit].depth_texture_swizzle;

			/* Fake EQUAL/NOTEQUAL, it seems to pass some tests suprisingly. */
			if (comparefunc == RC_COMPARE_FUNC_EQUAL) {
				comparefunc = RC_COMPARE_FUNC_GEQUAL;
			} else if (comparefunc == RC_COMPARE_FUNC_NOTEQUAL) {
				comparefunc = RC_COMPARE_FUNC_LESS;
			}

			/* Recall that SrcReg[0] is r, SrcReg[tex] is tex and:
			 *   LESS:    r  < tex  <=>      -tex+r < 0
			 *   GEQUAL:  r >= tex  <=> not (-tex+r < 0)
			 *   GREATER: r  > tex  <=>       tex-r < 0
			 *   LEQUAL:  r <= tex  <=> not ( tex-r < 0)
			 *
			 * This negates either r or tex: */
			if (comparefunc == RC_COMPARE_FUNC_LESS || comparefunc == RC_COMPARE_FUNC_GEQUAL)
				inst_mad->U.I.SrcReg[tex].Negate = inst_mad->U.I.SrcReg[tex].Negate ^ RC_MASK_XYZW;
			else
				inst_mad->U.I.SrcReg[0].Negate = inst_mad->U.I.SrcReg[0].Negate ^ RC_MASK_XYZW;

			/* This negates the whole expresion: */
			if (comparefunc == RC_COMPARE_FUNC_LESS || comparefunc == RC_COMPARE_FUNC_GREATER) {
				pass = 1;
				fail = 2;
			} else {
				pass = 2;
				fail = 1;
			}

			inst_cmp = rc_insert_new_instruction(c, inst_mad);
			inst_cmp->U.I.Opcode = RC_OPCODE_CMP;
			inst_cmp->U.I.DstReg = output_reg;
			inst_cmp->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
			inst_cmp->U.I.SrcReg[0].Index = tmp_sum;
			inst_cmp->U.I.SrcReg[pass].File = RC_FILE_NONE;
			inst_cmp->U.I.SrcReg[pass].Swizzle = RC_SWIZZLE_1111;
			inst_cmp->U.I.SrcReg[fail] = shadow_ambient(compiler, inst->U.I.TexSrcUnit);

			assert(tmp_texsample != tmp_sum && tmp_sum != tmp_recip_w);
		}
	}

	/* Divide by W if needed. */
	if (inst->U.I.Opcode == RC_OPCODE_TXP &&
	    (compiler->state.unit[inst->U.I.TexSrcUnit].wrap_mode == RC_WRAP_REPEAT ||
             compiler->state.unit[inst->U.I.TexSrcUnit].wrap_mode == RC_WRAP_MIRRORED_REPEAT)) {
		projective_divide(compiler, inst);
	}

	/* Texture wrap modes don't work on NPOT textures or texrects.
	 *
	 * The game plan is simple. We have two flags, fake_npot and
	 * non_normalized_coords, as well as a tex target. The RECT tex target
	 * will make the emitted code use non-scaled texcoords.
	 *
	 * Non-wrapped/clamped texcoords with NPOT are free in HW. Repeat and
	 * mirroring are not. If we need to repeat, we do:
	 *
	 * MUL temp, texcoord, <scaling factor constant>
	 * FRC temp, temp ; Discard integer portion of coords
	 *
	 * This gives us coords in [0, 1].
	 *
	 * Mirroring is trickier. We're going to start out like repeat:
	 *
	 * MUL temp, texcoord, <scaling factor constant> ; De-mirror across axes
	 * MUL temp, temp, 0.5 ; Pattern repeats in [0, 2]
	 *                            ; so scale to [0, 1]
	 * FRC temp, temp ; Make the pattern repeat
	 * MAD temp, temp, 2, -1 ; Move the pattern to [-1, 1]
	 * ADD temp, 1, -abs(temp) ; Now comes a neat trick: use abs to mirror the pattern.
	 *				; The pattern is backwards, so reverse it (1-x).
	 *
	 * This gives us coords in [0, 1].
	 *
	 * ~ C & M. ;)
	 */
	if (inst->U.I.Opcode != RC_OPCODE_KIL &&
		(inst->U.I.TexSrcTarget == RC_TEXTURE_RECT ||
			compiler->state.unit[inst->U.I.TexSrcUnit].fake_npot ||
			compiler->state.unit[inst->U.I.TexSrcUnit].non_normalized_coords)) {
		rc_wrap_mode wrapmode = compiler->state.unit[inst->U.I.TexSrcUnit].wrap_mode;

		/* R300 cannot sample from rectangles. */
		if (!c->is_r500) {
			lower_texture_rect(compiler, inst);
		}

		if (compiler->state.unit[inst->U.I.TexSrcUnit].fake_npot &&
			wrapmode != RC_WRAP_NONE) {
			struct rc_instruction *inst_mov;
			unsigned temp = rc_find_free_temporary(c);

			/* For NPOT fallback, we need normalized coordinates anyway. */
			if (c->is_r500) {
				lower_texture_rect(compiler, inst);
			}

			if (wrapmode == RC_WRAP_REPEAT) {
				/* Both instructions will be paired up. */
				struct rc_instruction *inst_frc = rc_insert_new_instruction(c, inst->Prev);

				inst_frc->U.I.Opcode = RC_OPCODE_FRC;
				inst_frc->U.I.DstReg.File = RC_FILE_TEMPORARY;
				inst_frc->U.I.DstReg.Index = temp;
				inst_frc->U.I.DstReg.WriteMask = RC_MASK_XYZ;
				inst_frc->U.I.SrcReg[0] = inst->U.I.SrcReg[0];
			} else if (wrapmode == RC_WRAP_MIRRORED_REPEAT) {
				/*
				 * Function:
				 *   f(v) = 1 - abs(frac(v * 0.5) * 2 - 1)
				 *
				 * Code:
				 *   MUL temp, src0, 0.5
				 *   FRC temp, temp
				 *   MAD temp, temp, 2, -1
				 *   ADD temp, 1, -abs(temp)
				 */

				struct rc_instruction *inst_mul, *inst_frc, *inst_mad, *inst_add;
				unsigned two, two_swizzle;

				inst_mul = rc_insert_new_instruction(c, inst->Prev);

				inst_mul->U.I.Opcode = RC_OPCODE_MUL;
				inst_mul->U.I.DstReg.File = RC_FILE_TEMPORARY;
				inst_mul->U.I.DstReg.Index = temp;
				inst_mul->U.I.DstReg.WriteMask = RC_MASK_XYZ;
				inst_mul->U.I.SrcReg[0] = inst->U.I.SrcReg[0];
				inst_mul->U.I.SrcReg[1].Swizzle = RC_SWIZZLE_HHHH;

				inst_frc = rc_insert_new_instruction(c, inst->Prev);

				inst_frc->U.I.Opcode = RC_OPCODE_FRC;
				inst_frc->U.I.DstReg.File = RC_FILE_TEMPORARY;
				inst_frc->U.I.DstReg.Index = temp;
				inst_frc->U.I.DstReg.WriteMask = RC_MASK_XYZ;
				inst_frc->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
				inst_frc->U.I.SrcReg[0].Index = temp;
				inst_frc->U.I.SrcReg[0].Swizzle = RC_SWIZZLE_XYZ0;

				two = rc_constants_add_immediate_scalar(&c->Program.Constants, 2, &two_swizzle);
				inst_mad = rc_insert_new_instruction(c, inst->Prev);

				inst_mad->U.I.Opcode = RC_OPCODE_MAD;
				inst_mad->U.I.DstReg.File = RC_FILE_TEMPORARY;
				inst_mad->U.I.DstReg.Index = temp;
				inst_mad->U.I.DstReg.WriteMask = RC_MASK_XYZ;
				inst_mad->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
				inst_mad->U.I.SrcReg[0].Index = temp;
				inst_mad->U.I.SrcReg[0].Swizzle = RC_SWIZZLE_XYZ0;
				inst_mad->U.I.SrcReg[1].File = RC_FILE_CONSTANT;
				inst_mad->U.I.SrcReg[1].Index = two;
				inst_mad->U.I.SrcReg[1].Swizzle = two_swizzle;
				inst_mad->U.I.SrcReg[2].Swizzle = RC_SWIZZLE_1111;
				inst_mad->U.I.SrcReg[2].Negate = RC_MASK_XYZ;

				inst_add = rc_insert_new_instruction(c, inst->Prev);

				inst_add->U.I.Opcode = RC_OPCODE_ADD;
				inst_add->U.I.DstReg.File = RC_FILE_TEMPORARY;
				inst_add->U.I.DstReg.Index = temp;
				inst_add->U.I.DstReg.WriteMask = RC_MASK_XYZ;
				inst_add->U.I.SrcReg[0].Swizzle = RC_SWIZZLE_1111;
				inst_add->U.I.SrcReg[1].File = RC_FILE_TEMPORARY;
				inst_add->U.I.SrcReg[1].Index = temp;
				inst_add->U.I.SrcReg[1].Swizzle = RC_SWIZZLE_XYZ0;
				inst_add->U.I.SrcReg[1].Abs = 1;
				inst_add->U.I.SrcReg[1].Negate = RC_MASK_XYZ;
			} else if (wrapmode == RC_WRAP_MIRRORED_CLAMP) {
				/*
				 * Mirrored clamp modes are bloody simple, we just use abs
				 * to mirror [0, 1] into [-1, 0]. This works for
				 * all modes i.e. CLAMP, CLAMP_TO_EDGE, and CLAMP_TO_BORDER.
				 */
				struct rc_instruction *inst_mov;

				inst_mov = rc_insert_new_instruction(c, inst->Prev);

				inst_mov->U.I.Opcode = RC_OPCODE_MOV;
				inst_mov->U.I.DstReg.File = RC_FILE_TEMPORARY;
				inst_mov->U.I.DstReg.Index = temp;
				inst_mov->U.I.DstReg.WriteMask = RC_MASK_XYZ;
				inst_mov->U.I.SrcReg[0] = inst->U.I.SrcReg[0];
				inst_mov->U.I.SrcReg[0].Abs = 1;
			}

			/* Preserve W for TXP/TXB. */
			inst_mov = rc_insert_new_instruction(c, inst->Prev);

			inst_mov->U.I.Opcode = RC_OPCODE_MOV;
			inst_mov->U.I.DstReg.File = RC_FILE_TEMPORARY;
			inst_mov->U.I.DstReg.Index = temp;
			inst_mov->U.I.DstReg.WriteMask = RC_MASK_W;
			inst_mov->U.I.SrcReg[0] = inst->U.I.SrcReg[0];

			reset_srcreg(&inst->U.I.SrcReg[0]);
			inst->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
			inst->U.I.SrcReg[0].Index = temp;
		}
	}

	/* Cannot write texture to output registers (all chips) or with masks (non-r500) */
	if (inst->U.I.Opcode != RC_OPCODE_KIL &&
		(inst->U.I.DstReg.File != RC_FILE_TEMPORARY ||
		 (!c->is_r500 && inst->U.I.DstReg.WriteMask != RC_MASK_XYZW))) {
		struct rc_instruction * inst_mov = rc_insert_new_instruction(c, inst);

		inst_mov->U.I.Opcode = RC_OPCODE_MOV;
		inst_mov->U.I.DstReg = inst->U.I.DstReg;
		inst_mov->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
		inst_mov->U.I.SrcReg[0].Index = rc_find_free_temporary(c);

		inst->U.I.DstReg.File = RC_FILE_TEMPORARY;
		inst->U.I.DstReg.Index = inst_mov->U.I.SrcReg[0].Index;
		inst->U.I.DstReg.WriteMask = RC_MASK_XYZW;
	}

	/* Cannot read texture coordinate from constants file */
	if (inst->U.I.SrcReg[0].File != RC_FILE_TEMPORARY && inst->U.I.SrcReg[0].File != RC_FILE_INPUT) {
		struct rc_instruction * inst_mov = rc_insert_new_instruction(c, inst->Prev);

		inst_mov->U.I.Opcode = RC_OPCODE_MOV;
		inst_mov->U.I.DstReg.File = RC_FILE_TEMPORARY;
		inst_mov->U.I.DstReg.Index = rc_find_free_temporary(c);
		inst_mov->U.I.SrcReg[0] = inst->U.I.SrcReg[0];

		reset_srcreg(&inst->U.I.SrcReg[0]);
		inst->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
		inst->U.I.SrcReg[0].Index = inst_mov->U.I.DstReg.Index;
	}

	return 1;
}
/**
 * Transform TEX, TXP, TXB, and KIL instructions in the following ways:
 *  - implement texture compare (shadow extensions)
 *  - extract non-native source / destination operands
 *  - premultiply texture coordinates for RECT
 *  - extract operand swizzles
 *  - introduce a temporary register when write masks are needed
 */
int radeonTransformTEX(
	struct radeon_compiler * c,
	struct rc_instruction * inst,
	void* data)
{
	struct r300_fragment_program_compiler *compiler =
		(struct r300_fragment_program_compiler*)data;
	rc_wrap_mode wrapmode = compiler->state.unit[inst->U.I.TexSrcUnit].wrap_mode;
	int is_rect = inst->U.I.TexSrcTarget == RC_TEXTURE_RECT ||
		      compiler->state.unit[inst->U.I.TexSrcUnit].non_normalized_coords;

	if (inst->U.I.Opcode != RC_OPCODE_TEX &&
		inst->U.I.Opcode != RC_OPCODE_TXB &&
		inst->U.I.Opcode != RC_OPCODE_TXP &&
		inst->U.I.Opcode != RC_OPCODE_TXD &&
		inst->U.I.Opcode != RC_OPCODE_TXL &&
		inst->U.I.Opcode != RC_OPCODE_KIL)
		return 0;

	/* ARB_shadow & EXT_shadow_funcs */
	if (inst->U.I.Opcode != RC_OPCODE_KIL &&
		((c->Program.ShadowSamplers & (1 << inst->U.I.TexSrcUnit)) ||
		 (compiler->state.unit[inst->U.I.TexSrcUnit].compare_mode_enabled))) {
		rc_compare_func comparefunc = compiler->state.unit[inst->U.I.TexSrcUnit].texture_compare_func;

		if (comparefunc == RC_COMPARE_FUNC_NEVER || comparefunc == RC_COMPARE_FUNC_ALWAYS) {
			inst->U.I.Opcode = RC_OPCODE_MOV;

			if (comparefunc == RC_COMPARE_FUNC_ALWAYS) {
				inst->U.I.SrcReg[0] = shadow_pass_value(compiler, inst->U.I.TexSrcUnit);
			} else {
				inst->U.I.SrcReg[0] = shadow_fail_value(compiler, inst->U.I.TexSrcUnit);
			}

			return 1;
		} else {
			struct rc_instruction * inst_rcp = NULL;
			struct rc_instruction *inst_mul, *inst_add, *inst_cmp;
			unsigned tmp_texsample;
			unsigned tmp_sum;
			int pass, fail;

			/* Save the output register. */
			struct rc_dst_register output_reg = inst->U.I.DstReg;
			unsigned saturate_mode = inst->U.I.SaturateMode;

			/* Redirect TEX to a new temp. */
			tmp_texsample = rc_find_free_temporary(c);
			inst->U.I.SaturateMode = 0;
			inst->U.I.DstReg.File = RC_FILE_TEMPORARY;
			inst->U.I.DstReg.Index = tmp_texsample;
			inst->U.I.DstReg.WriteMask = RC_MASK_XYZW;

			tmp_sum = rc_find_free_temporary(c);

			if (inst->U.I.Opcode == RC_OPCODE_TXP) {
				/* Compute 1/W. */
				inst_rcp = rc_insert_new_instruction(c, inst);
				inst_rcp->U.I.Opcode = RC_OPCODE_RCP;
				inst_rcp->U.I.DstReg.File = RC_FILE_TEMPORARY;
				inst_rcp->U.I.DstReg.Index = tmp_sum;
				inst_rcp->U.I.DstReg.WriteMask = RC_MASK_W;
				inst_rcp->U.I.SrcReg[0] = inst->U.I.SrcReg[0];
				inst_rcp->U.I.SrcReg[0].Swizzle =
					RC_MAKE_SWIZZLE_SMEAR(GET_SWZ(inst->U.I.SrcReg[0].Swizzle, 3));
			}

			/* Divide Z by W (if it's TXP) and saturate. */
			inst_mul = rc_insert_new_instruction(c, inst_rcp ? inst_rcp : inst);
			inst_mul->U.I.Opcode = inst->U.I.Opcode == RC_OPCODE_TXP ? RC_OPCODE_MUL : RC_OPCODE_MOV;
			inst_mul->U.I.DstReg.File = RC_FILE_TEMPORARY;
			inst_mul->U.I.DstReg.Index = tmp_sum;
			inst_mul->U.I.DstReg.WriteMask = RC_MASK_W;
			inst_mul->U.I.SaturateMode = RC_SATURATE_ZERO_ONE;
			inst_mul->U.I.SrcReg[0] = inst->U.I.SrcReg[0];
			inst_mul->U.I.SrcReg[0].Swizzle =
				RC_MAKE_SWIZZLE_SMEAR(GET_SWZ(inst->U.I.SrcReg[0].Swizzle, 2));
			if (inst->U.I.Opcode == RC_OPCODE_TXP) {
				inst_mul->U.I.SrcReg[1].File = RC_FILE_TEMPORARY;
				inst_mul->U.I.SrcReg[1].Index = tmp_sum;
				inst_mul->U.I.SrcReg[1].Swizzle = RC_SWIZZLE_WWWW;
			}

			/* Add the depth texture value. */
			inst_add = rc_insert_new_instruction(c, inst_mul);
			inst_add->U.I.Opcode = RC_OPCODE_ADD;
			inst_add->U.I.DstReg.File = RC_FILE_TEMPORARY;
			inst_add->U.I.DstReg.Index = tmp_sum;
			inst_add->U.I.DstReg.WriteMask = RC_MASK_W;
			inst_add->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
			inst_add->U.I.SrcReg[0].Index = tmp_sum;
			inst_add->U.I.SrcReg[0].Swizzle = RC_SWIZZLE_WWWW;
			inst_add->U.I.SrcReg[1].File = RC_FILE_TEMPORARY;
			inst_add->U.I.SrcReg[1].Index = tmp_texsample;
			inst_add->U.I.SrcReg[1].Swizzle = RC_SWIZZLE_XXXX;

			/* Note that SrcReg[0] is r, SrcReg[1] is tex and:
			 *   LESS:    r  < tex  <=>      -tex+r < 0
			 *   GEQUAL:  r >= tex  <=> not (-tex+r < 0)
			 *   GREATER: r  > tex  <=>       tex-r < 0
			 *   LEQUAL:  r <= tex  <=> not ( tex-r < 0)
			 *   EQUAL:   GEQUAL
			 *   NOTEQUAL:LESS
			 */

			/* This negates either r or tex: */
			if (comparefunc == RC_COMPARE_FUNC_LESS || comparefunc == RC_COMPARE_FUNC_GEQUAL ||
			    comparefunc == RC_COMPARE_FUNC_EQUAL || comparefunc == RC_COMPARE_FUNC_NOTEQUAL)
				inst_add->U.I.SrcReg[1].Negate = inst_add->U.I.SrcReg[1].Negate ^ RC_MASK_XYZW;
			else
				inst_add->U.I.SrcReg[0].Negate = inst_add->U.I.SrcReg[0].Negate ^ RC_MASK_XYZW;

			/* This negates the whole expresion: */
			if (comparefunc == RC_COMPARE_FUNC_LESS || comparefunc == RC_COMPARE_FUNC_GREATER ||
			    comparefunc == RC_COMPARE_FUNC_NOTEQUAL) {
				pass = 1;
				fail = 2;
			} else {
				pass = 2;
				fail = 1;
			}

			inst_cmp = rc_insert_new_instruction(c, inst_add);
			inst_cmp->U.I.Opcode = RC_OPCODE_CMP;
			inst_cmp->U.I.SaturateMode = saturate_mode;
			inst_cmp->U.I.DstReg = output_reg;
			inst_cmp->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
			inst_cmp->U.I.SrcReg[0].Index = tmp_sum;
			inst_cmp->U.I.SrcReg[0].Swizzle =
					combine_swizzles(RC_SWIZZLE_WWWW,
							 compiler->state.unit[inst->U.I.TexSrcUnit].texture_swizzle);
			inst_cmp->U.I.SrcReg[pass] = shadow_pass_value(compiler, inst->U.I.TexSrcUnit);
			inst_cmp->U.I.SrcReg[fail] = shadow_fail_value(compiler, inst->U.I.TexSrcUnit);

			assert(tmp_texsample != tmp_sum);
		}
	}

	/* R300 cannot sample from rectangles and the wrap mode fallback needs
	 * normalized coordinates anyway. */
	if (inst->U.I.Opcode != RC_OPCODE_KIL &&
	    is_rect && (!c->is_r500 || wrapmode != RC_WRAP_NONE)) {
		scale_texcoords(compiler, inst, RC_STATE_R300_TEXRECT_FACTOR);
		inst->U.I.TexSrcTarget = RC_TEXTURE_2D;
	}

	/* Divide by W if needed. */
	if (inst->U.I.Opcode == RC_OPCODE_TXP &&
	    (wrapmode == RC_WRAP_REPEAT || wrapmode == RC_WRAP_MIRRORED_REPEAT ||
	     compiler->state.unit[inst->U.I.TexSrcUnit].clamp_and_scale_before_fetch)) {
		projective_divide(compiler, inst);
	}

	/* Texture wrap modes don't work on NPOT textures.
	 *
	 * Non-wrapped/clamped texcoords with NPOT are free in HW. Repeat and
	 * mirroring are not. If we need to repeat, we do:
	 *
	 * MUL temp, texcoord, <scaling factor constant>
	 * FRC temp, temp ; Discard integer portion of coords
	 *
	 * This gives us coords in [0, 1].
	 *
	 * Mirroring is trickier. We're going to start out like repeat:
	 *
	 * MUL temp, texcoord, <scaling factor constant> ; De-mirror across axes
	 * MUL temp, temp, 0.5 ; Pattern repeats in [0, 2]
	 *                            ; so scale to [0, 1]
	 * FRC temp, temp ; Make the pattern repeat
	 * MAD temp, temp, 2, -1 ; Move the pattern to [-1, 1]
	 * ADD temp, 1, -abs(temp) ; Now comes a neat trick: use abs to mirror the pattern.
	 *				; The pattern is backwards, so reverse it (1-x).
	 *
	 * This gives us coords in [0, 1].
	 *
	 * ~ C & M. ;)
	 */
	if (inst->U.I.Opcode != RC_OPCODE_KIL &&
	    wrapmode != RC_WRAP_NONE) {
		struct rc_instruction *inst_mov;
		unsigned temp = rc_find_free_temporary(c);

		if (wrapmode == RC_WRAP_REPEAT) {
			/* Both instructions will be paired up. */
			struct rc_instruction *inst_frc = rc_insert_new_instruction(c, inst->Prev);

			inst_frc->U.I.Opcode = RC_OPCODE_FRC;
			inst_frc->U.I.DstReg.File = RC_FILE_TEMPORARY;
			inst_frc->U.I.DstReg.Index = temp;
			inst_frc->U.I.DstReg.WriteMask = RC_MASK_XYZ;
			inst_frc->U.I.SrcReg[0] = inst->U.I.SrcReg[0];
		} else if (wrapmode == RC_WRAP_MIRRORED_REPEAT) {
			/*
			 * Function:
			 *   f(v) = 1 - abs(frac(v * 0.5) * 2 - 1)
			 *
			 * Code:
			 *   MUL temp, src0, 0.5
			 *   FRC temp, temp
			 *   MAD temp, temp, 2, -1
			 *   ADD temp, 1, -abs(temp)
			 */

			struct rc_instruction *inst_mul, *inst_frc, *inst_mad, *inst_add;
			unsigned two, two_swizzle;

			inst_mul = rc_insert_new_instruction(c, inst->Prev);

			inst_mul->U.I.Opcode = RC_OPCODE_MUL;
			inst_mul->U.I.DstReg.File = RC_FILE_TEMPORARY;
			inst_mul->U.I.DstReg.Index = temp;
			inst_mul->U.I.DstReg.WriteMask = RC_MASK_XYZ;
			inst_mul->U.I.SrcReg[0] = inst->U.I.SrcReg[0];
			inst_mul->U.I.SrcReg[1].Swizzle = RC_SWIZZLE_HHHH;

			inst_frc = rc_insert_new_instruction(c, inst->Prev);

			inst_frc->U.I.Opcode = RC_OPCODE_FRC;
			inst_frc->U.I.DstReg.File = RC_FILE_TEMPORARY;
			inst_frc->U.I.DstReg.Index = temp;
			inst_frc->U.I.DstReg.WriteMask = RC_MASK_XYZ;
			inst_frc->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
			inst_frc->U.I.SrcReg[0].Index = temp;
			inst_frc->U.I.SrcReg[0].Swizzle = RC_SWIZZLE_XYZ0;

			two = rc_constants_add_immediate_scalar(&c->Program.Constants, 2, &two_swizzle);
			inst_mad = rc_insert_new_instruction(c, inst->Prev);

			inst_mad->U.I.Opcode = RC_OPCODE_MAD;
			inst_mad->U.I.DstReg.File = RC_FILE_TEMPORARY;
			inst_mad->U.I.DstReg.Index = temp;
			inst_mad->U.I.DstReg.WriteMask = RC_MASK_XYZ;
			inst_mad->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
			inst_mad->U.I.SrcReg[0].Index = temp;
			inst_mad->U.I.SrcReg[0].Swizzle = RC_SWIZZLE_XYZ0;
			inst_mad->U.I.SrcReg[1].File = RC_FILE_CONSTANT;
			inst_mad->U.I.SrcReg[1].Index = two;
			inst_mad->U.I.SrcReg[1].Swizzle = two_swizzle;
			inst_mad->U.I.SrcReg[2].Swizzle = RC_SWIZZLE_1111;
			inst_mad->U.I.SrcReg[2].Negate = RC_MASK_XYZ;

			inst_add = rc_insert_new_instruction(c, inst->Prev);

			inst_add->U.I.Opcode = RC_OPCODE_ADD;
			inst_add->U.I.DstReg.File = RC_FILE_TEMPORARY;
			inst_add->U.I.DstReg.Index = temp;
			inst_add->U.I.DstReg.WriteMask = RC_MASK_XYZ;
			inst_add->U.I.SrcReg[0].Swizzle = RC_SWIZZLE_1111;
			inst_add->U.I.SrcReg[1].File = RC_FILE_TEMPORARY;
			inst_add->U.I.SrcReg[1].Index = temp;
			inst_add->U.I.SrcReg[1].Swizzle = RC_SWIZZLE_XYZ0;
			inst_add->U.I.SrcReg[1].Abs = 1;
			inst_add->U.I.SrcReg[1].Negate = RC_MASK_XYZ;
		} else if (wrapmode == RC_WRAP_MIRRORED_CLAMP) {
			/*
			 * Mirrored clamp modes are bloody simple, we just use abs
			 * to mirror [0, 1] into [-1, 0]. This works for
			 * all modes i.e. CLAMP, CLAMP_TO_EDGE, and CLAMP_TO_BORDER.
			 */
			struct rc_instruction *inst_mov;

			inst_mov = rc_insert_new_instruction(c, inst->Prev);

			inst_mov->U.I.Opcode = RC_OPCODE_MOV;
			inst_mov->U.I.DstReg.File = RC_FILE_TEMPORARY;
			inst_mov->U.I.DstReg.Index = temp;
			inst_mov->U.I.DstReg.WriteMask = RC_MASK_XYZ;
			inst_mov->U.I.SrcReg[0] = inst->U.I.SrcReg[0];
			inst_mov->U.I.SrcReg[0].Abs = 1;
		}

		/* Preserve W for TXP/TXB. */
		inst_mov = rc_insert_new_instruction(c, inst->Prev);

		inst_mov->U.I.Opcode = RC_OPCODE_MOV;
		inst_mov->U.I.DstReg.File = RC_FILE_TEMPORARY;
		inst_mov->U.I.DstReg.Index = temp;
		inst_mov->U.I.DstReg.WriteMask = RC_MASK_W;
		inst_mov->U.I.SrcReg[0] = inst->U.I.SrcReg[0];

		reset_srcreg(&inst->U.I.SrcReg[0]);
		inst->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
		inst->U.I.SrcReg[0].Index = temp;
	}

	/* NPOT -> POT conversion for 3D textures. */
	if (inst->U.I.Opcode != RC_OPCODE_KIL &&
	    compiler->state.unit[inst->U.I.TexSrcUnit].clamp_and_scale_before_fetch) {
		struct rc_instruction *inst_mov;
		unsigned temp = rc_find_free_temporary(c);

		/* Saturate XYZ. */
		inst_mov = rc_insert_new_instruction(c, inst->Prev);
		inst_mov->U.I.Opcode = RC_OPCODE_MOV;
		inst_mov->U.I.SaturateMode = RC_SATURATE_ZERO_ONE;
		inst_mov->U.I.DstReg.File = RC_FILE_TEMPORARY;
		inst_mov->U.I.DstReg.Index = temp;
		inst_mov->U.I.DstReg.WriteMask = RC_MASK_XYZ;
		inst_mov->U.I.SrcReg[0] = inst->U.I.SrcReg[0];

		/* Copy W. */
		inst_mov = rc_insert_new_instruction(c, inst->Prev);
		inst_mov->U.I.Opcode = RC_OPCODE_MOV;
		inst_mov->U.I.DstReg.File = RC_FILE_TEMPORARY;
		inst_mov->U.I.DstReg.Index = temp;
		inst_mov->U.I.DstReg.WriteMask = RC_MASK_W;
		inst_mov->U.I.SrcReg[0] = inst->U.I.SrcReg[0];

		reset_srcreg(&inst->U.I.SrcReg[0]);
		inst->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
		inst->U.I.SrcReg[0].Index = temp;

		scale_texcoords(compiler, inst, RC_STATE_R300_TEXSCALE_FACTOR);
	}

	/* Convert SNORM-encoded ATI1N sampled as UNORM to SNORM.
	 * Formula: dst = tex > 0.5 ? tex*2-2 : tex*2
	 */
	if (inst->U.I.Opcode != RC_OPCODE_KIL &&
	    compiler->state.unit[inst->U.I.TexSrcUnit].convert_unorm_to_snorm) {
		unsigned two, two_swizzle;
		struct rc_instruction *inst_mul, *inst_mad, *inst_cnd;

		two = rc_constants_add_immediate_scalar(&c->Program.Constants, 2.35, &two_swizzle);

		inst_mul = rc_insert_new_instruction(c, inst);
		inst_mul->U.I.Opcode = RC_OPCODE_MUL;
		inst_mul->U.I.DstReg.File = RC_FILE_TEMPORARY;
		inst_mul->U.I.DstReg.Index = rc_find_free_temporary(c);
		inst_mul->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
		inst_mul->U.I.SrcReg[0].Index = rc_find_free_temporary(c); /* redirected TEX output */
		inst_mul->U.I.SrcReg[1].File = RC_FILE_CONSTANT; /* 2 */
		inst_mul->U.I.SrcReg[1].Index = two;
		inst_mul->U.I.SrcReg[1].Swizzle = two_swizzle;

		inst_mad = rc_insert_new_instruction(c, inst_mul);
		inst_mad->U.I.Opcode = RC_OPCODE_MAD;
		inst_mad->U.I.DstReg.File = RC_FILE_TEMPORARY;
		inst_mad->U.I.DstReg.Index = rc_find_free_temporary(c);
		inst_mad->U.I.SrcReg[0] = inst_mul->U.I.SrcReg[0]; /* redirected TEX output */
		inst_mad->U.I.SrcReg[1] = inst_mul->U.I.SrcReg[1]; /* 2 */
		inst_mad->U.I.SrcReg[2] = inst_mul->U.I.SrcReg[1]; /* 2 */
		inst_mad->U.I.SrcReg[2].Negate = RC_MASK_XYZW;

		inst_cnd = rc_insert_new_instruction(c, inst_mad);
		inst_cnd->U.I.Opcode = RC_OPCODE_CND;
		inst_cnd->U.I.SaturateMode = inst->U.I.SaturateMode;
		inst_cnd->U.I.DstReg = inst->U.I.DstReg;
		inst_cnd->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
		inst_cnd->U.I.SrcReg[0].Index = inst_mad->U.I.DstReg.Index;
		inst_cnd->U.I.SrcReg[0].Swizzle = compiler->state.unit[inst->U.I.TexSrcUnit].texture_swizzle;
		inst_cnd->U.I.SrcReg[1].File = RC_FILE_TEMPORARY;
		inst_cnd->U.I.SrcReg[1].Index = inst_mul->U.I.DstReg.Index;
		inst_cnd->U.I.SrcReg[1].Swizzle = compiler->state.unit[inst->U.I.TexSrcUnit].texture_swizzle;
		inst_cnd->U.I.SrcReg[2] = inst_mul->U.I.SrcReg[0]; /* redirected TEX output */

		inst->U.I.SaturateMode = 0;
		inst->U.I.DstReg.File = RC_FILE_TEMPORARY;
		inst->U.I.DstReg.Index = inst_mul->U.I.SrcReg[0].Index;
		inst->U.I.DstReg.WriteMask = RC_MASK_XYZW;
	}

	/* Cannot write texture to output registers or with saturate (all chips),
	 * or with masks (non-r500). */
	if (inst->U.I.Opcode != RC_OPCODE_KIL &&
		(inst->U.I.DstReg.File != RC_FILE_TEMPORARY ||
		 inst->U.I.SaturateMode ||
		 (!c->is_r500 && inst->U.I.DstReg.WriteMask != RC_MASK_XYZW))) {
		struct rc_instruction * inst_mov = rc_insert_new_instruction(c, inst);

		inst_mov->U.I.Opcode = RC_OPCODE_MOV;
		inst_mov->U.I.SaturateMode = inst->U.I.SaturateMode;
		inst_mov->U.I.DstReg = inst->U.I.DstReg;
		inst_mov->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
		inst_mov->U.I.SrcReg[0].Index = rc_find_free_temporary(c);

		inst->U.I.SaturateMode = 0;
		inst->U.I.DstReg.File = RC_FILE_TEMPORARY;
		inst->U.I.DstReg.Index = inst_mov->U.I.SrcReg[0].Index;
		inst->U.I.DstReg.WriteMask = RC_MASK_XYZW;
	}

	/* Cannot read texture coordinate from constants file */
	if (inst->U.I.SrcReg[0].File != RC_FILE_TEMPORARY && inst->U.I.SrcReg[0].File != RC_FILE_INPUT) {
		struct rc_instruction * inst_mov = rc_insert_new_instruction(c, inst->Prev);

		inst_mov->U.I.Opcode = RC_OPCODE_MOV;
		inst_mov->U.I.DstReg.File = RC_FILE_TEMPORARY;
		inst_mov->U.I.DstReg.Index = rc_find_free_temporary(c);
		inst_mov->U.I.SrcReg[0] = inst->U.I.SrcReg[0];

		reset_srcreg(&inst->U.I.SrcReg[0]);
		inst->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
		inst->U.I.SrcReg[0].Index = inst_mov->U.I.DstReg.Index;
	}

	return 1;
}