static unsigned int loop_max_possible_iterations(struct radeon_compiler *c,
struct loop_info * loop)
{
unsigned int total_i = rc_recompute_ips(c);
unsigned int loop_i = (loop->EndLoop->IP - loop->BeginLoop->IP) - 1;
/* +1 because the program already has one iteration of the loop. */
return 1 + ((c->max_alu_insts - total_i) / loop_i);
}
static void compute_live_intervals(struct regalloc_state * s)
{
rc_recompute_ips(s->C);
for(struct rc_instruction * inst = s->C->Program.Instructions.Next;
inst != &s->C->Program.Instructions;
inst = inst->Next) {
rc_for_all_reads(inst, scan_callback, s);
rc_for_all_writes(inst, scan_callback, s);
}
}
/**
* This function renames registers in an attempt to get the code close to
* SSA form. After this function has completed, most of the register are only
* written to one time, with a few exceptions.
*
* This function assumes all the instructions are still of type
* RC_INSTRUCTION_NORMAL.
*/
void rc_rename_regs(struct radeon_compiler *c, void *user)
{
unsigned int i, used_length;
int new_index;
struct rc_instruction * inst;
struct rc_reader_data reader_data;
unsigned char * used;
/* XXX Remove this once the register allocation works with flow control. */
for(inst = c->Program.Instructions.Next;
inst != &c->Program.Instructions;
inst = inst->Next) {
if (inst->U.I.Opcode == RC_OPCODE_BGNLOOP)
return;
}
used_length = 2 * rc_recompute_ips(c);
used = memory_pool_malloc(&c->Pool, sizeof(unsigned char) * used_length);
memset(used, 0, sizeof(unsigned char) * used_length);
rc_get_used_temporaries(c, used, used_length);
for(inst = c->Program.Instructions.Next;
inst != &c->Program.Instructions;
inst = inst->Next) {
if (inst->U.I.DstReg.File != RC_FILE_TEMPORARY)
continue;
reader_data.ExitOnAbort = 1;
rc_get_readers(c, inst, &reader_data, NULL, NULL, NULL);
if (reader_data.Abort || reader_data.ReaderCount == 0)
continue;
new_index = rc_find_free_temporary_list(c, used, used_length,
RC_MASK_XYZW);
if (new_index < 0) {
rc_error(c, "Ran out of temporary registers\n");
return;
}
reader_data.Writer->U.I.DstReg.Index = new_index;
for(i = 0; i < reader_data.ReaderCount; i++) {
reader_data.Readers[i].U.I.Src->Index = new_index;
}
}
}
static void compute_live_intervals(struct radeon_compiler *c,
struct regalloc_state *s)
{
memset(s, 0, sizeof(*s));
s->C = c;
s->NumHwTemporaries = c->max_temp_regs;
s->HwTemporary =
memory_pool_malloc(&c->Pool,
s->NumHwTemporaries * sizeof(struct hardware_register));
memset(s->HwTemporary, 0, s->NumHwTemporaries * sizeof(struct hardware_register));
rc_recompute_ips(s->C);
for(struct rc_instruction * inst = s->C->Program.Instructions.Next;
inst != &s->C->Program.Instructions;
inst = inst->Next) {
/* For all instructions inside of a loop, the ENDLOOP
* instruction is used as the end of the live interval. */
if (inst->U.I.Opcode == RC_OPCODE_BGNLOOP && !s->end_loop) {
int loops = 1;
struct rc_instruction * tmp;
for(tmp = inst->Next;
tmp != &s->C->Program.Instructions;
tmp = tmp->Next) {
if (tmp->U.I.Opcode == RC_OPCODE_BGNLOOP) {
loops++;
} else if (tmp->U.I.Opcode
== RC_OPCODE_ENDLOOP) {
if(!--loops) {
s->end_loop = tmp->IP;
break;
}
}
}
}
if (inst->IP == s->end_loop)
s->end_loop = 0;
rc_for_all_reads_mask(inst, scan_callback, s);
rc_for_all_writes_mask(inst, scan_callback, s);
}
}
static void allocate_temporary_registers(struct radeon_compiler *c, void *user)
{
struct r300_vertex_program_compiler *compiler = (struct r300_vertex_program_compiler*)c;
struct rc_instruction *inst;
struct rc_instruction *end_loop = NULL;
unsigned int num_orig_temps = 0;
char hwtemps[RC_REGISTER_MAX_INDEX];
struct temporary_allocation * ta;
unsigned int i, j;
memset(hwtemps, 0, sizeof(hwtemps));
rc_recompute_ips(c);
/* Pass 1: Count original temporaries. */
for(inst = compiler->Base.Program.Instructions.Next; inst != &compiler->Base.Program.Instructions; inst = inst->Next) {
const struct rc_opcode_info * opcode = rc_get_opcode_info(inst->U.I.Opcode);
for (i = 0; i < opcode->NumSrcRegs; ++i) {
if (inst->U.I.SrcReg[i].File == RC_FILE_TEMPORARY) {
if (inst->U.I.SrcReg[i].Index >= num_orig_temps)
num_orig_temps = inst->U.I.SrcReg[i].Index + 1;
}
}
if (opcode->HasDstReg) {
if (inst->U.I.DstReg.File == RC_FILE_TEMPORARY) {
if (inst->U.I.DstReg.Index >= num_orig_temps)
num_orig_temps = inst->U.I.DstReg.Index + 1;
}
}
}
ta = (struct temporary_allocation*)memory_pool_malloc(&compiler->Base.Pool,
sizeof(struct temporary_allocation) * num_orig_temps);
memset(ta, 0, sizeof(struct temporary_allocation) * num_orig_temps);
/* Pass 2: Determine original temporary lifetimes */
for(inst = compiler->Base.Program.Instructions.Next; inst != &compiler->Base.Program.Instructions; inst = inst->Next) {
const struct rc_opcode_info * opcode = rc_get_opcode_info(inst->U.I.Opcode);
/* Instructions inside of loops need to use the ENDLOOP
* instruction as their LastRead. */
if (!end_loop && inst->U.I.Opcode == RC_OPCODE_BGNLOOP) {
int endloops = 1;
struct rc_instruction * ptr;
for(ptr = inst->Next;
ptr != &compiler->Base.Program.Instructions;
ptr = ptr->Next){
if (ptr->U.I.Opcode == RC_OPCODE_BGNLOOP) {
endloops++;
} else if (ptr->U.I.Opcode == RC_OPCODE_ENDLOOP) {
endloops--;
if (endloops <= 0) {
end_loop = ptr;
break;
}
}
}
}
if (inst == end_loop) {
end_loop = NULL;
continue;
}
for (i = 0; i < opcode->NumSrcRegs; ++i) {
if (inst->U.I.SrcReg[i].File == RC_FILE_TEMPORARY) {
ta[inst->U.I.SrcReg[i].Index].LastRead = end_loop ? end_loop : inst;
}
}
}
/* Pass 3: Register allocation */
for(inst = compiler->Base.Program.Instructions.Next; inst != &compiler->Base.Program.Instructions; inst = inst->Next) {
const struct rc_opcode_info * opcode = rc_get_opcode_info(inst->U.I.Opcode);
for (i = 0; i < opcode->NumSrcRegs; ++i) {
if (inst->U.I.SrcReg[i].File == RC_FILE_TEMPORARY) {
unsigned int orig = inst->U.I.SrcReg[i].Index;
inst->U.I.SrcReg[i].Index = ta[orig].HwTemp;
if (ta[orig].Allocated && inst == ta[orig].LastRead)
hwtemps[ta[orig].HwTemp] = 0;
}
}
if (opcode->HasDstReg) {
if (inst->U.I.DstReg.File == RC_FILE_TEMPORARY) {
unsigned int orig = inst->U.I.DstReg.Index;
if (!ta[orig].Allocated) {
for(j = 0; j < c->max_temp_regs; ++j) {
if (!hwtemps[j])
break;
}
ta[orig].Allocated = 1;
ta[orig].HwTemp = j;
hwtemps[ta[orig].HwTemp] = 1;
}
inst->U.I.DstReg.Index = ta[orig].HwTemp;
}
}
}
}