danog/ir
1
0
Fork 0
mirror of https://github.com/danog/ir.git synced 2024-11-30 04:39:43 +01:00
Code Issues Packages Projects Releases Wiki Activity
1130c256ae
ir/ir_ra.c
Dmitry Stogov 1130c256ae Find optimal split position
2022-05-04 11:59:35 +03:00

1862 lines
50 KiB
C
Raw Blame History

#define _GNU_SOURCE
#include "ir.h"
#include "ir_private.h"
#include <stdlib.h>
#include "ir_x86.h"
#ifdef IR_DEBUG
uint32_t debug_regset = 0xffffffff; /* all 32 regisers */
#endif
/* RA - Register Allocation, Liveness, Coalescing and SSA Resolution */
int ir_assign_virtual_registers(ir_ctx *ctx)
{
uint32_t *vregs;
uint32_t vregs_count = 0;
int b, i, n;
ir_block *bb;
ir_insn *insn;
uint32_t flags;
/* Assign unique virtual register to each data node */
if (!ctx->prev_insn_len) {
ctx->prev_insn_len = ir_mem_malloc(ctx->insns_count * sizeof(uint32_t));
}
vregs = ir_mem_calloc(ctx->insns_count, sizeof(ir_ref));
for (b = 1, bb = ctx->cfg_blocks + b; b <= ctx->cfg_blocks_count; b++, bb++) {
n = b; /* The first insn of BB keeps BB number in prev_insn_len[] */
for (i = bb->start, insn = ctx->ir_base + i; i <= bb->end;) {
ctx->prev_insn_len[i] = n; /* The first insn of BB keeps BB number in prev_insn_len[] */
flags = ir_op_flags[insn->op];
if ((flags & IR_OP_FLAG_DATA) || ((flags & IR_OP_FLAG_MEM) && insn->type != IR_VOID)) {
if ((insn->op == IR_PARAM || insn->op == IR_VAR) && ctx->use_lists[i].count == 0) {
/* pass */
} else if (insn->op == IR_VAR && ctx->use_lists[i].count > 0) {
vregs[i] = ++vregs_count; /* for spill slot */
} else if (!ctx->rules || ir_needs_vreg(ctx, i)) {
vregs[i] = ++vregs_count;
}
}
n = ir_operands_count(ctx, insn);
n = 1 + (n >> 2); // support for multi-word instructions like MERGE and PHI
i += n;
insn += n;
}
}
ctx->vregs_count = vregs_count;
ctx->vregs = vregs;
return 1;
}
/* Lifetime intervals construction
*
* See "Linear Scan Register Allocation on SSA Form", Christian Wimmer and
* Michael Franz, CGO'10 (2010), Figure 4.
*/
static void ir_add_local_var(ir_ctx *ctx, int v, uint8_t type)
{
ir_live_interval *ival = ctx->live_intervals[v];
IR_ASSERT(!ival);
ival = ir_mem_malloc(sizeof(ir_live_interval));
IR_ASSERT(type != IR_VOID);
ival->type = type;
ival->reg = IR_REG_NONE;
ival->flags = 0;
ival->stack_spill_pos = 0; // not allocated
ival->range.start = 0;
ival->range.end = ctx->insns_count;
ival->range.next = NULL;
ival->use_pos = NULL;
ival->top = ival;
ival->next = NULL;
ctx->live_intervals[v] = ival;
}
static void ir_add_live_range(ir_ctx *ctx, int v, uint8_t type, ir_live_pos start, ir_live_pos end)
{
ir_live_interval *ival = ctx->live_intervals[v];
ir_live_range *p, *q, *next, *prev;
if (!ival) {
ival = ir_mem_malloc(sizeof(ir_live_interval));
IR_ASSERT(type != IR_VOID);
ival->type = type;
ival->reg = IR_REG_NONE;
ival->flags = 0;
ival->stack_spill_pos = 0; // not allocated
ival->range.start = start;
ival->range.end = end;
ival->range.next = NULL;
ival->use_pos = NULL;
ival->top = ival;
ival->next = NULL;
ctx->live_intervals[v] = ival;
return;
}
IR_ASSERT(type == IR_VOID || type == ival->type);
p = &ival->range;
prev = NULL;
while (p && end >= p->start) {
if (p->end >= start) {
if (start < p->start) {
p->start = start;
}
if (end > p->end) {
p->end = end;
/* merge with next */
next = p->next;
while (next && p->end >= next->start) {
if (next->end > p->end) {
p->end = next->end;
}
p->next = next->next;
/* list of deleted structures is keapt at ctx->unused_live_ranges for reuse */
next->next = ctx->unused_live_ranges;
ctx->unused_live_ranges = next;
next = p->next;
}
}
return;
}
prev = p;
p = prev->next;
}
if (ctx->unused_live_ranges) {
/* reuse */
q = ctx->unused_live_ranges;
ctx->unused_live_ranges = q->next;
} else {
q = ir_mem_malloc(sizeof(ir_live_range));
}
if (prev) {
prev->next = q;
} else {
q->start = ival->range.start;
q->end = ival->range.end;
q->next = ival->range.next;
p = q;
q = &ival->range;
}
q->start = start;
q->end = end;
q->next = p;
}
static void ir_add_fixed_live_range(ir_ctx *ctx, ir_reg reg, ir_live_pos start, ir_live_pos end)
{
int v = ctx->vregs_count + 1 + reg;
ir_live_interval *ival = ctx->live_intervals[v];
if (!ival) {
ival = ir_mem_malloc(sizeof(ir_live_interval));
ival->type = IR_VOID;
ival->reg = reg;
ival->flags = 0;
ival->stack_spill_pos = 0; // not allocated
ival->range.start = start;
ival->range.end = end;
ival->range.next = NULL;
ival->use_pos = NULL;
ival->top = ival;
ival->next = NULL;
ctx->live_intervals[v] = ival;
return;
}
ir_add_live_range(ctx, v, IR_VOID, start, end);
}
static void ir_fix_live_range(ir_ctx *ctx, int v, ir_live_pos old_start, ir_live_pos new_start)
{
ir_live_range *p = &ctx->live_intervals[v]->range;
while (p && p->start < old_start) {
p = p->next;
}
IR_ASSERT(p && p->start == old_start);
p->start = new_start;
}
static void ir_add_use_pos(ir_ctx *ctx, int v, ir_use_pos *use_pos)
{
ir_live_interval *ival = ctx->live_intervals[v];
ir_use_pos *prev = NULL;
ir_use_pos *p = ival->use_pos;
while (p && (p->pos < use_pos->pos ||
(p->pos == use_pos->pos &&
(use_pos->op_num == 0 || p->op_num < use_pos->op_num)))) {
prev = p;
p = p->next;
}
if (prev) {
use_pos->next = prev->next;
prev->next = use_pos;
} else {
use_pos->next = ival->use_pos;
ival->use_pos = use_pos;
}
}
static void ir_add_use(ir_ctx *ctx, int v, int op_num, ir_live_pos pos, ir_reg hint, ir_ref hint_ref)
{
ir_use_pos *use_pos;
use_pos = ir_mem_malloc(sizeof(ir_use_pos));
use_pos->op_num = op_num;
use_pos->hint = hint;
use_pos->hint_ref = hint_ref;
use_pos->pos = pos;
use_pos->flags = ctx->rules ? ir_get_use_flags(ctx, IR_LIVE_POS_TO_REF(pos), op_num) : 0;
ir_add_use_pos(ctx, v, use_pos);
}
int ir_compute_live_ranges(ir_ctx *ctx)
{
int i, j, k, n;
int b, succ;
uint32_t flags, len;
ir_insn *insn;
ir_block *bb, *succ_bb;
ir_bitset visited, live;
ir_bitset loops = NULL;
ir_bitset queue = NULL;
ir_reg reg;
if (!(ctx->flags & IR_LINEAR) || !ctx->vregs) {
return 0;
}
/* Compute Live Ranges */
visited = ir_bitset_malloc(ctx->cfg_blocks_count + 1);
len = ir_bitset_len(ctx->vregs_count + 1);
live = ir_bitset_malloc((ctx->cfg_blocks_count + 1) * len * 8 * sizeof(*live));
ctx->live_intervals = ir_mem_calloc(ctx->vregs_count + 1 + IR_REG_NUM, sizeof(ir_live_interval*));
for (b = ctx->cfg_blocks_count; b > 0; b--) {
bb = &ctx->cfg_blocks[b];
/* for each successor of b */
ir_bitset_incl(visited, b);
ir_bitset_clear(live, len);
for (i = 0; i < bb->successors_count; i++) {
succ = ctx->cfg_edges[bb->successors + i];
/* blocks must be ordered where all dominators of a block are before this block */
IR_ASSERT(ir_bitset_in(visited, succ) || bb->loop_header == succ);
/* live = union of successors.liveIn */
ir_bitset_union(live, live + (len * succ), len);
/* for each phi function phi of successor */
succ_bb = &ctx->cfg_blocks[succ];
if (succ_bb->predecessors_count > 1) {
ir_use_list *use_list = &ctx->use_lists[succ_bb->start];
k = 0;
for (j = 0; j < succ_bb->predecessors_count; j++) {
if (ctx->cfg_edges[succ_bb->predecessors + j] == b) {
k = j + 2;
break;
}
}
IR_ASSERT(k != 0);
for (j = 0; j < use_list->count; j++) {
insn = &ctx->ir_base[ctx->use_edges[use_list->refs + j]];
if (insn->op == IR_PHI) {
if (insn->ops[k] > 0) {
/* live.add(phi.inputOf(b)) */
IR_ASSERT(ctx->vregs[insn->ops[k]]);
ir_bitset_incl(live, ctx->vregs[insn->ops[k]]);
// TODO: ir_add_live_range() is used just to set ival->type
/* intervals[phi.inputOf(b)].addRange(b.from, b.to) */
ir_add_live_range(ctx, ctx->vregs[insn->ops[k]], insn->type,
IR_START_LIVE_POS_FROM_REF(bb->start),
IR_END_LIVE_POS_FROM_REF(bb->end));
}
}
}
}
}
/* for each opd in live */
IR_BITSET_FOREACH(live, len, i) {
/* intervals[opd].addRange(b.from, b.to) */
ir_add_live_range(ctx, i, IR_VOID,
IR_START_LIVE_POS_FROM_REF(bb->start),
IR_END_LIVE_POS_FROM_REF(bb->end));
} IR_BITSET_FOREACH_END();
/* for each operation op of b in reverse order */
for (i = bb->end; i > bb->start; i -= ctx->prev_insn_len[i]) {
insn = &ctx->ir_base[i];
flags = ir_op_flags[insn->op];
if ((flags & IR_OP_FLAG_DATA) || ((flags & IR_OP_FLAG_MEM) && insn->type != IR_VOID)) {
if (ctx->vregs[i]) {
if (ir_bitset_in(live, ctx->vregs[i])) {
if (insn->op != IR_PHI) {
ir_live_pos def_pos;
ir_ref hint_ref = 0;
reg = ctx->rules ? ir_uses_fixed_reg(ctx, i, 0) : IR_REG_NONE;
if (reg != IR_REG_NONE) {
def_pos = IR_SAVE_LIVE_POS_FROM_REF(i);
if (insn->op == IR_PARAM) {
/* parameter register must be kept before it's copied */
ir_add_fixed_live_range(ctx, reg,
IR_START_LIVE_POS_FROM_REF(bb->start), def_pos);
} else {
ir_add_fixed_live_range(ctx, reg,
IR_DEF_LIVE_POS_FROM_REF(i), def_pos);
}
} else if (ctx->rules && ir_result_reuses_op1_reg(ctx, i)) {
/* We add two uses to emulate move from op1 to res */
ir_add_use(ctx, ctx->vregs[i], 0, IR_DEF_LIVE_POS_FROM_REF(i), reg, hint_ref);
def_pos = IR_LOAD_LIVE_POS_FROM_REF(i);
hint_ref = insn->op1;
} else {
def_pos = IR_DEF_LIVE_POS_FROM_REF(i);
}
/* intervals[opd].setFrom(op.id) */
ir_fix_live_range(ctx, ctx->vregs[i],
IR_START_LIVE_POS_FROM_REF(bb->start), def_pos);
ir_add_use(ctx, ctx->vregs[i], 0, def_pos, reg, hint_ref);
} else {
ir_add_use(ctx, ctx->vregs[i], 0, IR_DEF_LIVE_POS_FROM_REF(i), IR_REG_NONE, 0);
}
/* live.remove(opd) */
ir_bitset_excl(live, ctx->vregs[i]);
} else if (insn->op == IR_VAR) {
if (ctx->use_lists[i].count > 0) {
ir_add_local_var(ctx, ctx->vregs[i], insn->type);
}
}
}
}
if (insn->op != IR_PHI) {
n = ir_input_edges_count(ctx, insn);
for (j = 1; j <= n; j++) {
if (IR_OPND_KIND(flags, j) == IR_OPND_DATA) {
ir_ref input = insn->ops[j];
if (input > 0 && ctx->vregs[input]) {
ir_live_pos use_pos;
if (ctx->rules && j == 1 && ir_result_reuses_op1_reg(ctx, i)) {
use_pos = IR_LOAD_LIVE_POS_FROM_REF(i);
reg = ctx->rules ? ir_uses_fixed_reg(ctx, i, j) : IR_REG_NONE;
if (reg != IR_REG_NONE) {
ir_add_fixed_live_range(ctx, reg,
use_pos, IR_USE_LIVE_POS_FROM_REF(i));
}
} else {
reg = ctx->rules ? ir_uses_fixed_reg(ctx, i, j) : IR_REG_NONE;
if (reg != IR_REG_NONE) {
use_pos = IR_LOAD_LIVE_POS_FROM_REF(i);
ir_add_fixed_live_range(ctx, reg,
use_pos, IR_USE_LIVE_POS_FROM_REF(i));
} else if (j > 1 && input == insn->op1 && ctx->rules && ir_result_reuses_op1_reg(ctx, i)) {
/* Input is the same as "op1" */
use_pos = IR_LOAD_LIVE_POS_FROM_REF(i);
} else {
use_pos = IR_USE_LIVE_POS_FROM_REF(i);
}
}
/* intervals[opd].addRange(b.from, op.id) */
ir_add_live_range(ctx, ctx->vregs[input], ctx->ir_base[input].type,
IR_START_LIVE_POS_FROM_REF(bb->start), use_pos);
ir_add_use(ctx, ctx->vregs[input], j, use_pos, reg, 0);
/* live.add(opd) */
ir_bitset_incl(live, ctx->vregs[input]);
}
}
}
/* CPU specific constraints */
if (ctx->rules) {
ir_regset regset = ir_get_scratch_regset(ctx, i);
if (regset != IR_REGSET_EMPTY) {
IR_REGSET_FOREACH(regset, reg) {
ir_add_fixed_live_range(ctx, reg,
IR_LOAD_LIVE_POS_FROM_REF(i), // TODO: LOAD instead of USE disables register usage for input
// this is necessary for DIV and MOD, but not for MUL
IR_DEF_LIVE_POS_FROM_REF(i));
} IR_REGSET_FOREACH_END();
}
}
}
}
/* if b is loop header */
if ((bb->flags & IR_BB_LOOP_HEADER)
&& !ir_bitset_empty(live, len)) {
/* variables live at loop header are alive at the whole loop body */
uint32_t bb_set_len = ir_bitset_len(ctx->cfg_blocks_count + 1);
int child;
ir_block *child_bb;
if (!loops) {
loops = ir_bitset_malloc(ctx->cfg_blocks_count + 1);
queue = ir_bitset_malloc(ctx->cfg_blocks_count + 1);
} else {
ir_bitset_clear(loops, bb_set_len);
ir_bitset_clear(queue, bb_set_len);
}
ir_bitset_incl(loops, b);
ir_bitset_incl(queue, b);
do {
child = ir_bitset_pop_first(queue, bb_set_len);
child_bb = &ctx->cfg_blocks[child];
IR_BITSET_FOREACH(live, len, i) {
ir_add_live_range(ctx, i, IR_VOID,
IR_START_LIVE_POS_FROM_REF(child_bb->start),
IR_END_LIVE_POS_FROM_REF(child_bb->end));
} IR_BITSET_FOREACH_END();
child = child_bb->dom_child;
while (child) {
child_bb = &ctx->cfg_blocks[child];
if (child_bb->loop_header && ir_bitset_in(loops, child_bb->loop_header)) {
ir_bitset_incl(queue, child);
if (child_bb->flags & IR_BB_LOOP_HEADER) {
ir_bitset_incl(loops, child);
}
}
child = child_bb->dom_next_child;
}
} while (!ir_bitset_empty(queue, bb_set_len));
}
/* b.liveIn = live */
ir_bitset_copy(live + (len * b), live, len);
}
if (loops) {
ir_mem_free(loops);
ir_mem_free(queue);
}
ir_mem_free(live);
ir_mem_free(visited);
return 1;
}
void ir_free_live_ranges(ir_live_range *live_range)
{
ir_live_range *p;
while (live_range) {
p = live_range;
live_range = live_range->next;
ir_mem_free(p);
}
}
void ir_free_live_intervals(ir_live_interval **live_intervals, int count)
{
uint32_t i;
ir_live_interval *ival;
for (i = 1; i <= count; i++) {
ival = live_intervals[i];
if (ival) {
if (ival->range.next) {
ir_free_live_ranges(ival->range.next);
}
ir_mem_free(ival);
}
}
ir_mem_free(live_intervals);
}
/* Live Ranges coalescing */
static ir_live_pos ir_vregs_overlap(ir_ctx *ctx, uint32_t r1, uint32_t r2)
{
ir_live_range *lrg1 = &ctx->live_intervals[r1]->range;
ir_live_range *lrg2 = &ctx->live_intervals[r2]->range;
while (1) {
if (lrg1->start < lrg2->end) {
if (lrg2->start < lrg1->end) {
return IR_MAX(lrg1->start, lrg2->start);
} else {
lrg1 = lrg1->next;
if (!lrg1) {
return 0;
}
}
} else {
lrg2 = lrg2->next;
if (!lrg2) {
return 0;
}
}
}
}
static void ir_vregs_join(ir_ctx *ctx, uint32_t r1, uint32_t r2)
{
ir_live_interval *ival = ctx->live_intervals[r2];
ir_live_range *live_range = &ival->range;
ir_live_range *next;
ir_use_pos *use_pos;
#if 0
fprintf(stderr, "COALESCE %d -> %d\n", r2, r1);
#endif
ir_add_live_range(ctx, r1, ival->type, live_range->start, live_range->end);
live_range = live_range->next;
while (live_range) {
ir_add_live_range(ctx, r1, ival->type, live_range->start, live_range->end);
next = live_range->next;
live_range->next = ctx->unused_live_ranges;
ctx->unused_live_ranges = live_range;
live_range = next;
} while (live_range);
use_pos = ival->use_pos;
while (use_pos) {
ir_use_pos *next_use_pos = use_pos->next;
if (ctx->vregs[use_pos->hint_ref] == r1) {
use_pos->hint_ref = 0;
}
ir_add_use_pos(ctx, r1, use_pos);
use_pos = next_use_pos;
}
ir_mem_free(ival);
ctx->live_intervals[r2] = NULL;
ctx->live_intervals[r1]->flags |= IR_LIVE_INTERVAL_COALESCED;
}
static bool ir_try_coalesce(ir_ctx *ctx, ir_ref from, ir_ref to)
{
ir_ref i;
int v1 = ctx->vregs[from];
int v2 = ctx->vregs[to];
if (v1 != v2 && !ir_vregs_overlap(ctx, v1, v2)) {
uint8_t f1 = ctx->live_intervals[v1]->flags;
uint8_t f2 = ctx->live_intervals[v2]->flags;
if ((f1 & IR_LIVE_INTERVAL_COALESCED) && !(f2 & IR_LIVE_INTERVAL_COALESCED)) {
ir_vregs_join(ctx, v1, v2);
ctx->vregs[to] = v1;
} else if ((f2 & IR_LIVE_INTERVAL_COALESCED) && !(f1 & IR_LIVE_INTERVAL_COALESCED)) {
ir_vregs_join(ctx, v2, v1);
ctx->vregs[from] = v2;
} else if (v1 < v2) {
ir_vregs_join(ctx, v1, v2);
if (f2 & IR_LIVE_INTERVAL_COALESCED) {
for (i = 0; i < ctx->insns_count; i++) {
if (ctx->vregs[i] == v2) {
ctx->vregs[i] = v1;
}
}
} else {
ctx->vregs[to] = v1;
}
} else {
ir_vregs_join(ctx, v2, v1);
if (f1 & IR_LIVE_INTERVAL_COALESCED) {
for (i = 0; i < ctx->insns_count; i++) {
if (ctx->vregs[i] == v1) {
ctx->vregs[i] = v2;
}
}
} else {
ctx->vregs[from] = v2;
}
}
return 1;
}
return 0;
}
static void ir_add_phi_move(ir_ctx *ctx, int b, ir_ref from, ir_ref to)
{
if (IR_IS_CONST_REF(from) || ctx->vregs[from] != ctx->vregs[to]) {
ctx->cfg_blocks[b].flags |= IR_BB_DESSA_MOVES;
#if 0
fprintf(stderr, "BB%d: MOV %d -> %d\n", b, from, to);
#endif
}
}
static int ir_block_cmp(const void *b1, const void *b2, void *data)
{
ir_ctx *ctx = data;
int d1 = ctx->cfg_blocks[*(ir_ref*)b1].loop_depth;
int d2 = ctx->cfg_blocks[*(ir_ref*)b2].loop_depth;
if (d1 > d2) {
return -1;
} else if (d1 == d2) {
return 0;
} else {
return 1;
}
}
static void ir_swap_operands(ir_ctx *ctx, ir_ref i, ir_insn *insn)
{
ir_live_pos pos = IR_USE_LIVE_POS_FROM_REF(i);
ir_live_pos load_pos = IR_LOAD_LIVE_POS_FROM_REF(i);
ir_live_interval *ival;
ir_live_range *r;
ir_use_pos *p, *p1 = NULL, *p2 = NULL;
ir_ref tmp;
tmp = insn->op1;
insn->op1 = insn->op2;
insn->op2 = tmp;
ival = ctx->live_intervals[ctx->vregs[insn->op1]];
p = ival->use_pos;
while (p) {
if (p->pos == pos) {
p->pos = load_pos;
p->op_num = 1;
p1 = p;
break;
}
p = p->next;
}
ival = ctx->live_intervals[ctx->vregs[i]];
p = ival->use_pos;
while (p) {
if (p->pos == load_pos) {
p->hint_ref = insn->op1;
break;
}
p = p->next;
}
if (insn->op2 > 0 && ctx->vregs[insn->op2]) {
ival = ctx->live_intervals[ctx->vregs[insn->op2]];
r = &ival->range;
while (r) {
if (r->end == load_pos) {
r->end = pos;
break;
}
r = r->next;
}
p = ival->use_pos;
while (p) {
if (p->pos == load_pos) {
p->pos = pos;
p->op_num = 2;
p2 = p;
break;
}
p = p->next;
}
}
if (p1 && p2) {
uint8_t tmp = p1->flags;
p1->flags = p2->flags;
p2->flags = tmp;
}
}
static int ir_hint_conflict(ir_ctx *ctx, ir_ref ref, int use, int def)
{
ir_use_pos *p;
ir_reg r1 = IR_REG_NONE;
ir_reg r2 = IR_REG_NONE;
p = ctx->live_intervals[use]->use_pos;
while (p) {
if (IR_LIVE_POS_TO_REF(p->pos) == ref) {
break;
}
if (p->hint != IR_REG_NONE) {
r1 = p->hint;
}
p = p->next;
}
p = ctx->live_intervals[def]->use_pos;
while (p) {
if (IR_LIVE_POS_TO_REF(p->pos) > ref) {
if (p->hint != IR_REG_NONE) {
r2 = p->hint;
break;
}
}
p = p->next;
}
return r1 != r2 && r1 != IR_REG_NONE && r2 != IR_REG_NONE;
}
static int ir_try_swap_operands(ir_ctx *ctx, ir_ref i, ir_insn *insn)
{
if (insn->op1 > 0
&& ctx->vregs[insn->op1] != ctx->vregs[i]
&& !ir_vregs_overlap(ctx, ctx->vregs[insn->op1], ctx->vregs[i])
&& !ir_hint_conflict(ctx, i, ctx->vregs[insn->op1], ctx->vregs[i])) {
/* pass */
} else if (insn->op2 > 0 && insn->op1 != insn->op2
&& (ir_op_flags[insn->op] & IR_OP_FLAG_COMMUTATIVE)) {
if (ctx->vregs[insn->op2] != ctx->vregs[i]) {
ir_live_pos pos = IR_USE_LIVE_POS_FROM_REF(i);
ir_live_pos load_pos = IR_LOAD_LIVE_POS_FROM_REF(i);
ir_live_interval *ival = ctx->live_intervals[ctx->vregs[insn->op2]];
ir_live_range *r = &ival->range;
while (r) {
if (r->end == pos) {
r->end = load_pos;
if (!ir_vregs_overlap(ctx, ctx->vregs[insn->op2], ctx->vregs[i])
&& !ir_hint_conflict(ctx, i, ctx->vregs[insn->op2], ctx->vregs[i])) {
ir_swap_operands(ctx, i, insn);
return 1;
} else {
r->end = pos;
}
break;
}
r = r->next;
}
}
}
return 0;
}
int ir_coalesce(ir_ctx *ctx)
{
int b, i, n, succ;
ir_ref *p, use, input, k, j;
ir_block *bb, *succ_bb;
ir_use_list *use_list;
ir_insn *insn;
uint32_t *offsets;
ir_worklist blocks;
bool compact = 0;
/* Collect a list of blocks which are predecossors to block with phi finctions */
ir_worklist_init(&blocks, ctx->cfg_blocks_count + 1);
for (b = 1, bb = &ctx->cfg_blocks[1]; b <= ctx->cfg_blocks_count; b++, bb++) {
if (bb->predecessors_count > 1) {
use_list = &ctx->use_lists[bb->start];
n = use_list->count;
for (i = 0, p = &ctx->use_edges[use_list->refs]; i < n; i++, p++) {
use = *p;
insn = &ctx->ir_base[use];
if (insn->op == IR_PHI) {
k = ir_input_edges_count(ctx, insn);
for (j = 2; j <= k; j++) {
ir_worklist_push(&blocks, ctx->cfg_edges[bb->predecessors + (j-2)]);
}
}
}
}
}
qsort_r(blocks.l.a.refs, ir_worklist_len(&blocks), sizeof(ir_ref), ir_block_cmp, ctx);
while (ir_worklist_len(&blocks)) {
b = ir_worklist_pop(&blocks);
bb = &ctx->cfg_blocks[b];
IR_ASSERT(bb->successors_count == 1);
succ = ctx->cfg_edges[bb->successors];
succ_bb = &ctx->cfg_blocks[succ];
IR_ASSERT(succ_bb->predecessors_count > 1);
k = 0;
for (j = 0; j < succ_bb->predecessors_count; j++) {
if (ctx->cfg_edges[succ_bb->predecessors + j] == b) {
k = j + 2;
break;
}
}
IR_ASSERT(k != 0);
use_list = &ctx->use_lists[succ_bb->start];
n = use_list->count;
for (i = 0, p = &ctx->use_edges[use_list->refs]; i < n; i++, p++) {
use = *p;
insn = &ctx->ir_base[use];
if (insn->op == IR_PHI) {
input = insn->ops[k];
if (input > 0) {
if (!ir_try_coalesce(ctx, input, use)) {
ir_add_phi_move(ctx, b, input, use);
} else {
compact = 1;
}
} else {
/* Move for constant input */
ir_add_phi_move(ctx, b, input, use);
}
}
}
}
ir_worklist_free(&blocks);
#if 1
if (ctx->rules) {
/* try to swap operands of commutative instructions for better register allocation */
for (b = 1, bb = &ctx->cfg_blocks[1]; b <= ctx->cfg_blocks_count; b++, bb++) {
for (i = bb->start, insn = ctx->ir_base + i; i <= bb->end;) {
if (ir_result_reuses_op1_reg(ctx, i)) {
if (insn->op2 > 0 && insn->op1 != insn->op2
&& (ir_op_flags[insn->op] & IR_OP_FLAG_COMMUTATIVE)) {
ir_try_swap_operands(ctx, i, insn);
}
// if (insn->op1 > 0) {
// ir_try_coalesce(ctx, insn->op1, i);
// }
// } else if (insn->op == IR_COPY) {
// if (insn->op1 > 0) {
// ir_try_coalesce(ctx, insn->op1, i);
// }
}
n = ir_operands_count(ctx, insn);
n = 1 + (n >> 2); // support for multi-word instructions like MERGE and PHI
i += n;
insn += n;
}
}
}
#endif
if (compact) {
#if 1
offsets = ir_mem_calloc(ctx->vregs_count + 1, sizeof(uint32_t));
for (i = 1, n = 1; i <= ctx->vregs_count; i++) {
if (ctx->live_intervals[i]) {
if (i != n) {
ctx->live_intervals[n] = ctx->live_intervals[i];
offsets[i] = i - n;
}
n++;
}
}
n--;
if (n != ctx->vregs_count) {
j = ctx->vregs_count - n;
for (i = n + 1; i <= ctx->vregs_count + IR_REG_NUM; i++) {
ctx->live_intervals[i] = ctx->live_intervals[i + j];
}
for (j = 1; j < ctx->insns_count; j++) {
if (ctx->vregs[j]) {
ctx->vregs[j] -= offsets[ctx->vregs[j]];
}
}
ctx->vregs_count = n;
}
ir_mem_free(offsets);
#endif
}
return 1;
}
/* SSA Deconstruction */
int ir_compute_dessa_moves(ir_ctx *ctx)
{
int b, i, n;
ir_ref j, k, *p, use;
ir_block *bb;
ir_use_list *use_list;
ir_insn *insn;
for (b = 1, bb = &ctx->cfg_blocks[1]; b <= ctx->cfg_blocks_count; b++, bb++) {
if (bb->predecessors_count > 1) {
use_list = &ctx->use_lists[bb->start];
n = use_list->count;
for (i = 0, p = &ctx->use_edges[use_list->refs]; i < n; i++, p++) {
use = *p;
insn = &ctx->ir_base[use];
if (insn->op == IR_PHI) {
k = ir_input_edges_count(ctx, insn);
for (j = 2; j <= k; j++) {
if (IR_IS_CONST_REF(insn->ops[j]) || ctx->vregs[insn->ops[j]] != ctx->vregs[use]) {
int pred = ctx->cfg_edges[bb->predecessors + (j-2)];
ctx->cfg_blocks[pred].flags |= IR_BB_DESSA_MOVES;
}
}
}
}
}
}
return 1;
}
int ir_gen_dessa_moves(ir_ctx *ctx, int b, emit_copy_t emit_copy)
{
int succ, j, k = 0, n = 0;
ir_block *bb, *succ_bb;
ir_use_list *use_list;
uint8_t *type;
uint32_t *loc, *pred;
uint32_t len;
ir_bitset todo, ready;
bb = &ctx->cfg_blocks[b];
if (!(bb->flags & IR_BB_DESSA_MOVES)) {
return 0;
}
IR_ASSERT(bb->successors_count == 1);
succ = ctx->cfg_edges[bb->successors];
succ_bb = &ctx->cfg_blocks[succ];
IR_ASSERT(succ_bb->predecessors_count > 1);
use_list = &ctx->use_lists[succ_bb->start];
for (j = 0; j < succ_bb->predecessors_count; j++) {
if (ctx->cfg_edges[succ_bb->predecessors + j] == b) {
k = j + 2;
break;
}
}
IR_ASSERT(k != 0);
type = ir_mem_calloc((ctx->vregs_count + 1), sizeof(uint8_t));
loc = ir_mem_calloc((ctx->vregs_count + 1) * 2, sizeof(uint32_t));
pred = loc + (ctx->vregs_count + 1);
len = ir_bitset_len(ctx->vregs_count + 1);
todo = ir_bitset_malloc(ctx->vregs_count + 1);
ready = ir_bitset_malloc(ctx->vregs_count + 1);
for (j = 0; j < use_list->count; j++) {
ir_ref ref = ctx->use_edges[use_list->refs + j];
ir_insn *insn = &ctx->ir_base[ref];
if (insn->op == IR_PHI) {
ir_ref input = insn->ops[k];
if (IR_IS_CONST_REF(input)) {
emit_copy(ctx, insn->type, input, ctx->vregs[ref]);
} else if (ctx->vregs[input] != ctx->vregs[ref]) {
loc[ctx->vregs[input]] = ctx->vregs[input];
pred[ctx->vregs[ref]] = ctx->vregs[input];
type[ctx->vregs[ref]] = insn->type;
ir_bitset_incl(todo, ctx->vregs[ref]);
n++;
}
}
}
IR_BITSET_FOREACH(todo, len, j) {
if (!loc[j]) {
ir_bitset_incl(ready, j);
}
} IR_BITSET_FOREACH_END();
while (!ir_bitset_empty(todo, len)) {
uint32_t a, b, c;
while (!ir_bitset_empty(ready, len)) {
b = ir_bitset_pop_first(ready, len);
a = pred[b];
c = loc[a];
emit_copy(ctx, type[b], c, b);
loc[a] = b;
if (a == c && pred[a]) {
ir_bitset_incl(ready, a);
}
}
b = ir_bitset_pop_first(todo, len);
if (b != loc[pred[b]]) {
emit_copy(ctx, type[b], b, 0);
loc[b] = 0;
ir_bitset_incl(ready, b);
}
}
ir_mem_free(ready);
ir_mem_free(todo);
ir_mem_free(loc);
ir_mem_free(type);
return 1;
}
/* Linear Scan Register Allocation
*
* See "Optimized Interval Splitting in a Linear Scan Register Allocator",
* Christian Wimmer VEE'10 (2005), Figure 2.
*/
typedef struct _ir_lsra_data {
uint32_t stack_frame_size;
} ir_lsra_data;
#ifdef IR_DEBUG
# define IR_LOG_LSRA(action, vreg, ival, comment) do { \
if (ctx->flags & IR_DEBUG_RA) { \
int _vreg = (vreg); \
ir_live_interval *_ival = (ival); \
ir_live_pos _start = _ival->range.start; \
ir_live_pos _end = ir_ival_end(_ival); \
fprintf(stderr, action " R%d [%d.%d...%d.%d)" comment "\n", \
_vreg, \
IR_LIVE_POS_TO_REF(_start), IR_LIVE_POS_TO_SUB_REF(_start), \
IR_LIVE_POS_TO_REF(_end), IR_LIVE_POS_TO_SUB_REF(_end)); \
} \
} while (0)
# define IR_LOG_LSRA_ASSIGN(action, vreg, ival, comment) do { \
if (ctx->flags & IR_DEBUG_RA) { \
int _vreg = (vreg); \
ir_live_interval *_ival = (ival); \
ir_live_pos _start = _ival->range.start; \
ir_live_pos _end = ir_ival_end(_ival); \
fprintf(stderr, action " R%d [%d.%d...%d.%d) to %s" comment "\n", \
_vreg, \
IR_LIVE_POS_TO_REF(_start), IR_LIVE_POS_TO_SUB_REF(_start), \
IR_LIVE_POS_TO_REF(_end), IR_LIVE_POS_TO_SUB_REF(_end), \
ir_reg_name(_ival->reg, _ival->type)); \
} \
} while (0)
# define IR_LOG_LSRA_SPLIT(vreg, ival, pos) do { \
if (ctx->flags & IR_DEBUG_RA) { \
int _vreg = (vreg); \
ir_live_interval *_ival = (ival); \
ir_live_pos _start = _ival->range.start; \
ir_live_pos _end = ir_ival_end(_ival); \
ir_live_pos _pos = (pos); \
fprintf(stderr, " ---- Split R%d [%d.%d...%d.%d) at %d.%d\n", \
_vreg, \
IR_LIVE_POS_TO_REF(_start), IR_LIVE_POS_TO_SUB_REF(_start), \
IR_LIVE_POS_TO_REF(_end), IR_LIVE_POS_TO_SUB_REF(_end), \
IR_LIVE_POS_TO_REF(_pos), IR_LIVE_POS_TO_SUB_REF(_pos)); \
} \
} while (0)
# define IR_LOG_LSRA_CONFLICT(action, vreg, ival, pos) do { \
if (ctx->flags & IR_DEBUG_RA) { \
int _vreg = (vreg); \
ir_live_interval *_ival = (ival); \
ir_live_pos _start = _ival->range.start; \
ir_live_pos _end = ir_ival_end(_ival); \
ir_live_pos _pos = (pos); \
fprintf(stderr, action " R%d [%d.%d...%d.%d) assigned to %s at %d.%d\n", \
_vreg, \
IR_LIVE_POS_TO_REF(_start), IR_LIVE_POS_TO_SUB_REF(_start), \
IR_LIVE_POS_TO_REF(_end), IR_LIVE_POS_TO_SUB_REF(_end), \
ir_reg_name(_ival->reg, _ival->type), \
IR_LIVE_POS_TO_REF(_pos), IR_LIVE_POS_TO_SUB_REF(_pos)); \
} \
} while (0)
#else
# define IR_LOG_LSRA(action, vreg, ival, comment)
# define IR_LOG_LSRA_ASSIGN(vreg, ival, comment)
# define IR_LOG_LSRA_SPLIT(vreg, ival, pos)
# define IR_LOG_LSRA_CONFLICT(action, vreg, ival, pos);
#endif
static ir_live_pos ir_ival_end(ir_live_interval *ival)
{
ir_live_range *live_range = &ival->range;
while (live_range->next) {
live_range = live_range->next;
}
return live_range->end;
}
static bool ir_ival_covers(ir_live_interval *ival, ir_live_pos position)
{
ir_live_range *live_range = &ival->range;
do {
if (position >= live_range->start && position < live_range->end) {
return 1;
}
live_range = live_range->next;
} while (live_range);
return 0;
}
static bool ir_ival_has_hole_between(ir_live_interval *ival, ir_live_pos from, ir_live_pos to)
{
ir_live_range *r = &ival->range;
while (r) {
if (from < r->start) {
return 1;
} else if (to <= r->end) {
return 0;
} else if (from >= r->end) {
return 1;
}
r = r->next;
}
return 0;
}
static ir_live_pos ir_last_use_pos_before(ir_live_interval *ival, ir_live_pos pos, uint8_t flags)
{
ir_live_pos ret = 0;
ir_use_pos *p = ival->use_pos;
while (p && p->pos <= pos) {
if (p->flags & flags) {
ret = p->pos;
}
p = p->next;
}
return ret;
}
static ir_live_pos ir_first_use_pos_after(ir_live_interval *ival, ir_live_pos pos, uint8_t flags)
{
ir_use_pos *p = ival->use_pos;
while (p && p->pos <= pos) {
p = p->next;
}
while (p && !(p->flags & flags)) {
p = p->next;
}
return p ? p->pos : 0x7fffffff;
}
static ir_block *ir_block_from_live_pos(ir_ctx *ctx, ir_live_pos pos)
{
int b;
ir_block *bb;
ir_ref ref = IR_LIVE_POS_TO_REF(pos);
// TODO: use binary search or map
for (b = 1, bb = ctx->cfg_blocks + 1; b <= ctx->cfg_blocks_count; b++, bb++) {
if (ref >= bb->start && ref <= bb->end) {
return bb;
}
}
IR_ASSERT(0);
}
static ir_live_pos ir_find_optimal_split_position(ir_ctx *ctx, int v, ir_live_interval *ival, ir_live_pos min_pos, ir_live_pos max_pos)
{
ir_block *min_bb, *max_bb;
if (min_pos == max_pos) {
return max_pos;
}
IR_ASSERT(min_pos < max_pos);
IR_ASSERT(min_pos >= ival->range.start);
IR_ASSERT(max_pos < ir_ival_end(ival));
min_bb = ir_block_from_live_pos(ctx, min_pos);
max_bb = ir_block_from_live_pos(ctx, max_pos);
if (min_bb == max_bb
|| ir_ival_has_hole_between(ival, min_pos, max_pos)) { // TODO: ???
return max_pos;
}
if (min_bb->loop_depth < max_bb->loop_depth) {
/* Split at the end of the loop entry */
do {
if (max_bb->loop_header) {
max_bb = &ctx->cfg_blocks[max_bb->loop_header];
}
max_bb = &ctx->cfg_blocks[ctx->cfg_edges[max_bb->predecessors]];
IR_ASSERT(ir_ival_covers(ival, IR_DEF_LIVE_POS_FROM_REF(max_bb->end)));
} while (min_bb->loop_depth < max_bb->loop_depth);
return IR_DEF_LIVE_POS_FROM_REF(max_bb->end);
}
IR_ASSERT(min_bb->loop_depth == max_bb->loop_depth); // TODO: Can "min_bb" be in a deeper loop than "max_bb" ???
return IR_LOAD_LIVE_POS_FROM_REF(max_bb->start);
}
static ir_live_interval *ir_split_interval_at(ir_ctx *ctx, int v, ir_live_interval *ival, ir_live_pos pos)
{
ir_live_interval *child;
ir_live_range *p, *prev;
ir_use_pos *use_pos, *prev_use_pos;
IR_LOG_LSRA_SPLIT(v, ival, pos);
IR_ASSERT(pos > ival->range.start);
p = &ival->range;
prev = NULL;
while (p && pos >= p->end) {
prev = p;
p = prev->next;
}
IR_ASSERT(p);
if (pos < p->start) {
/* split between ranges */
pos = p->start;
}
use_pos = ival->use_pos;
prev_use_pos = NULL;
if (p->start == pos) {
while (use_pos && pos > use_pos->pos) {
prev_use_pos = use_pos;
use_pos = use_pos->next;
}
} else {
while (use_pos && pos >= use_pos->pos) {
prev_use_pos = use_pos;
use_pos = use_pos->next;
}
}
child = ir_mem_malloc(sizeof(ir_live_interval));
child->type = ival->type;
child->reg = IR_REG_NONE;
child->flags = 0;
child->stack_spill_pos = 0; // not allocated
child->range.start = pos;
child->range.end = p->end;
child->range.next = p->next;
child->use_pos = prev_use_pos ? prev_use_pos->next : use_pos;
child->top = ival->top;
child->next = ival->next;
ival->next = child;
if (pos == p->start) {
prev->next = NULL;
} else {
p->end = pos;
p->next = NULL;
}
if (prev_use_pos) {
prev_use_pos->next = NULL;
} else {
ival->use_pos = NULL;
}
return child;
}
static void ir_allocate_spill_slot(ir_ctx *ctx, int current, ir_lsra_data *data)
{
ir_live_interval *ival = ctx->live_intervals[current]->top;
if (ival->stack_spill_pos == 0) {
data->stack_frame_size += 8; // ir_type_size[insn->type]; // TODO: alignment
ival->stack_spill_pos = data->stack_frame_size;
}
}
static ir_reg ir_try_allocate_preferred_reg(ir_ctx *ctx, ir_live_interval *ival, ir_regset available, ir_live_pos *freeUntilPos)
{
ir_use_pos *use_pos;
use_pos = ival->use_pos;
while (use_pos) {
if (use_pos->hint >= 0 && IR_REGSET_IN(available, use_pos->hint)) {
if (ir_ival_end(ival) <= freeUntilPos[use_pos->hint]) {
/* register available for the whole interval */
return use_pos->hint;
}
}
use_pos = use_pos->next;
}
use_pos = ival->use_pos;
while (use_pos) {
if (use_pos->hint_ref) {
ir_reg reg = ctx->live_intervals[ctx->vregs[use_pos->hint_ref]]->reg;
if (reg >= 0) {
if (ir_ival_end(ival) <= freeUntilPos[reg]) {
/* register available for the whole interval */
return reg;
}
}
}
use_pos = use_pos->next;
}
return IR_REG_NONE;
}
static void ir_add_to_unhandled(ir_ctx *ctx, ir_list *unhandled, int current)
{
ir_live_pos pos = ctx->live_intervals[current]->range.start;
if (ir_list_len(unhandled) == 0 || pos < ctx->live_intervals[ir_list_peek(unhandled)]->range.start) {
ir_list_push(unhandled, current);
} else {
uint32_t i = ir_list_len(unhandled);
while (i > 0) {
i--;
if (pos < ctx->live_intervals[ir_list_at(unhandled, i)]->range.start) {
i++;
break;
}
}
ir_list_insert(unhandled, i, current);
}
}
static ir_reg ir_try_allocate_free_reg(ir_ctx *ctx, int current, uint32_t len, ir_bitset active, ir_bitset inactive)
{
ir_live_pos freeUntilPos[IR_REG_NUM];
int i, reg;
ir_live_pos pos, next;
ir_live_interval *ival = ctx->live_intervals[current];
ir_regset available;
if (IR_IS_TYPE_FP(ival->type)) {
available = IR_REGSET_FP;
/* set freeUntilPos of all physical registers to maxInt */
for (i = IR_REG_FP_FIRST; i <= IR_REG_FP_LAST; i++) {
freeUntilPos[i] = 0x7fffffff;
}
} else {
available = IR_REGSET_GP;
if (ctx->flags & IR_USE_FRAME_POINTER) {
IR_REGSET_EXCL(available, IR_REG_FRAME_POINTER);
}
/* set freeUntilPos of all physical registers to maxInt */
for (i = IR_REG_GP_FIRST; i <= IR_REG_GP_LAST; i++) {
freeUntilPos[i] = 0x7fffffff;
}
}
#ifdef IR_DEBUG
available &= debug_regset;
#endif
/* for each interval it in active */
IR_BITSET_FOREACH(active, len, i) {
/* freeUntilPos[it.reg] = 0 */
reg = ctx->live_intervals[i]->reg;
IR_ASSERT(reg >= 0);
if (IR_REGSET_IN(available, reg)) {
freeUntilPos[reg] = 0;
}
} IR_BITSET_FOREACH_END();
/* for each interval it in inactive intersecting with current
*
* This loop is not necessary for program in SSA form (see LSRA on SSA fig. 6),
* but it is still necessary after coalescing and splitting
*/
IR_BITSET_FOREACH(inactive, len, i) {
/* freeUntilPos[it.reg] = next intersection of it with current */
reg = ctx->live_intervals[i]->reg;
IR_ASSERT(reg >= 0);
if (IR_REGSET_IN(available, reg)) {
next = ir_vregs_overlap(ctx, current, i);
if (next && next < freeUntilPos[reg]) {
freeUntilPos[reg] = next;
}
}
} IR_BITSET_FOREACH_END();
/* Try to use hint */
reg = ir_try_allocate_preferred_reg(ctx, ival, available, freeUntilPos);
if (reg != IR_REG_NONE) {
ival->reg = reg;
IR_LOG_LSRA_ASSIGN(" ---- Assign", current, ival, " (hint available without spilling)");
return reg;
}
/* reg = register with highest freeUntilPos */
reg = IR_REGSET_FIRST(available);
IR_REGSET_EXCL(available, reg);
pos = freeUntilPos[reg];
IR_REGSET_FOREACH(available, i) {
if (freeUntilPos[i] > pos) {
pos = freeUntilPos[i];
reg = i;
} else if (freeUntilPos[i] == pos
&& !IR_REGSET_IN(IR_REGSET_SCRATCH, reg)
&& IR_REGSET_IN(IR_REGSET_SCRATCH, i)) {
/* prefer caller-saved registers to avoid save/restore in prologue/epilogue */
pos = freeUntilPos[i];
reg = i;
}
} IR_REGSET_FOREACH_END();
if (!pos) {
/* no register available without spilling */
return IR_REG_NONE;
} else if (ir_ival_end(ival) <= pos) {
/* register available for the whole interval */
ival->reg = reg;
IR_LOG_LSRA_ASSIGN(" ---- Assign", current, ival, " (available without spilling)");
return reg;
} else if (pos > ival->range.start) {
/* register available for the first part of the interval */
/* split current before freeUntilPos[reg] */
ir_live_pos split_pos = ir_last_use_pos_before(ival, pos,
IR_USE_MUST_BE_IN_REG | IR_USE_SHOULD_BE_IN_REG);
if (split_pos > ival->range.start) {
split_pos = ir_find_optimal_split_position(ctx, current, ival, split_pos, pos);
ir_split_interval_at(ctx, current, ival, split_pos);
ival->reg = reg;
IR_LOG_LSRA_ASSIGN(" ---- Assign", current, ival, " (available without spilling for the first part)");
return reg;
}
}
return IR_REG_NONE;
}
static ir_reg ir_allocate_blocked_reg(ir_ctx *ctx, int current, uint32_t len, ir_bitset active, ir_bitset inactive, ir_list *unhandled)
{
ir_live_pos nextUsePos[IR_REG_NUM];
ir_live_pos blockPos[IR_REG_NUM];
int i, reg;
ir_live_pos pos, next_use_pos;
ir_live_interval *ival = ctx->live_intervals[current];
ir_use_pos *use_pos;
ir_regset available;
use_pos = ival->use_pos;
while (use_pos && !(use_pos->flags & IR_USE_MUST_BE_IN_REG)) {
use_pos = use_pos->next;
}
if (!use_pos) {
/* spill */
IR_LOG_LSRA(" ---- Spill", current, ival, " (no use pos that must be in reg)");
return IR_REG_NONE;
}
next_use_pos = use_pos->pos;
if (IR_IS_TYPE_FP(ival->type)) {
available = IR_REGSET_FP;
/* set nextUsePos of all physical registers to maxInt */
for (i = IR_REG_FP_FIRST; i <= IR_REG_FP_LAST; i++) {
nextUsePos[i] = 0x7fffffff;
blockPos[i] = 0x7fffffff;
}
} else {
available = IR_REGSET_GP;
if (ctx->flags & IR_USE_FRAME_POINTER) {
IR_REGSET_EXCL(available, IR_REG_FRAME_POINTER);
}
/* set nextUsePos of all physical registers to maxInt */
for (i = IR_REG_GP_FIRST; i <= IR_REG_GP_LAST; i++) {
nextUsePos[i] = 0x7fffffff;
blockPos[i] = 0x7fffffff;
}
}
#ifdef IR_DEBUG
available &= debug_regset;
#endif
/* for each interval it in active */
IR_BITSET_FOREACH(active, len, i) {
/* nextUsePos[it.reg] = next use of it after start of current */
reg = ctx->live_intervals[i]->reg;
IR_ASSERT(reg >= 0);
if (IR_REGSET_IN(available, reg)) {
// TODO: intervals that can't be spilled should be handled as fixed
if (ctx->live_intervals[i]->type == IR_VOID) {
/* fixed intervals */
blockPos[reg] = nextUsePos[reg] = 0;
} else {
pos = ir_first_use_pos_after(ctx->live_intervals[i], ival->range.start,
IR_USE_MUST_BE_IN_REG /* | IR_USE_SHOULD_BE_IN_REG */);
if (pos < nextUsePos[reg]) {
nextUsePos[reg] = pos;
}
}
}
} IR_BITSET_FOREACH_END();
/* for each interval it in inactive intersecting with current */
IR_BITSET_FOREACH(inactive, len, i) {
/* freeUntilPos[it.reg] = next intersection of it with current */
reg = ctx->live_intervals[i]->reg;
IR_ASSERT(reg >= 0);
if (IR_REGSET_IN(available, reg)) {
ir_live_pos overlap = ir_vregs_overlap(ctx, current, i);
if (overlap) {
if (ctx->live_intervals[i]->type == IR_VOID) {
/* fixed intervals */
if (overlap < nextUsePos[reg]) {
nextUsePos[reg] = overlap;
}
if (overlap < blockPos[reg]) {
blockPos[reg] = overlap;
}
} else {
pos = ir_first_use_pos_after(ctx->live_intervals[i], ival->range.start,
IR_USE_MUST_BE_IN_REG /* | IR_USE_SHOULD_BE_IN_REG */);
if (pos < nextUsePos[reg]) {
nextUsePos[reg] = pos;
}
}
}
}
} IR_BITSET_FOREACH_END();
// TODO: support for register hinting
/* reg = register with highest nextUsePos */
reg = IR_REGSET_FIRST(available);
IR_REGSET_EXCL(available, reg);
pos = nextUsePos[reg];
IR_REGSET_FOREACH(available, i) {
if (nextUsePos[i] > pos) {
pos = nextUsePos[i];
reg = i;
}
} IR_REGSET_FOREACH_END();
/* if first usage of current is after nextUsePos[reg] then */
if (next_use_pos > pos) {
/* all other intervals are used before current, so it is best to spill current itself */
/* assign spill slot to current */
/* split current before its first use position that requires a register */
ir_live_pos split_pos;
if (next_use_pos == ival->range.start) {
IR_ASSERT(use_pos && use_pos->op_num == 0);
/* split right after definition */
split_pos = next_use_pos + 1;
} else {
split_pos = ir_find_optimal_split_position(ctx, current, ival, ival->range.start, next_use_pos - 1);
}
if (split_pos > ival->range.start) {
IR_LOG_LSRA(" ---- Conflict with others", current, ival, " (all others are used before)");
ir_live_interval *child = ir_split_interval_at(ctx, current, ival, split_pos);
IR_LOG_LSRA(" ---- Spill", current, ival, "");
ctx->live_intervals[current] = child;
ir_add_to_unhandled(ctx, unhandled, current);
IR_LOG_LSRA(" ---- Queue", current, child, "");
return IR_REG_NONE;
}
}
if (ir_ival_end(ival) > blockPos[reg]) {
/* spilling make a register free only for the first part of current */
IR_LOG_LSRA(" ---- Conflict with others", current, ival, " (spilling make a register free only for the first part)");
/* split current at optimal position before block_pos[reg] */
ir_live_pos split_pos = ir_last_use_pos_before(ival, blockPos[reg] + 1,
IR_USE_MUST_BE_IN_REG | IR_USE_SHOULD_BE_IN_REG);
split_pos = ir_find_optimal_split_position(ctx, current, ival, split_pos, blockPos[reg]);
ir_split_interval_at(ctx, current, ival, split_pos);
}
/* spill intervals that currently block reg */
IR_BITSET_FOREACH(active, len, i) {
ir_live_interval *other = ctx->live_intervals[i];
ir_live_pos split_pos;
if (reg == other->reg) {
/* split active interval for reg at position */
ir_live_pos overlap = ir_vregs_overlap(ctx, current, i);
if (overlap) {
IR_ASSERT(other->type != IR_VOID);
IR_LOG_LSRA_CONFLICT(" ---- Conflict with active", i, other, overlap);
split_pos = ir_last_use_pos_before(other, ival->range.start, IR_USE_MUST_BE_IN_REG);
if (split_pos < ival->range.start) {
split_pos++; // TODO: ???
}
if (split_pos > other->range.start) {
split_pos = ir_find_optimal_split_position(ctx, i, other, split_pos, ival->range.start);
ir_live_interval *child = ir_split_interval_at(ctx, i, other, split_pos);
ir_bitset_excl(active, i);
IR_LOG_LSRA(" ---- Finish", i, other, "");
other = child;
} else {
other->reg = IR_REG_NONE;
ir_bitset_excl(active, i);
IR_LOG_LSRA(" ---- Spill and Finish", i, other, " (it must not be in reg)");
}
split_pos = ir_first_use_pos_after(other, ival->range.start, IR_USE_MUST_BE_IN_REG | IR_USE_SHOULD_BE_IN_REG) - 1;
if (split_pos >= ir_ival_end(other)) {
other = other->next;
} else if (split_pos > other->range.start) {
split_pos = ir_find_optimal_split_position(ctx, i, other, ival->range.start, split_pos);
ir_live_interval *child = ir_split_interval_at(ctx, i, other, split_pos);
IR_LOG_LSRA(" ---- Spill", i, other, "");
other = child;
} else {
// TODO: this may cause enless loop
}
if (other) {
ctx->live_intervals[i] = other;
ir_add_to_unhandled(ctx, unhandled, i);
IR_LOG_LSRA(" ---- Queue", i, other, "");
}
}
break;
}
} IR_BITSET_FOREACH_END();
/* split any inactive interval for reg at the end of its lifetime hole */
IR_BITSET_FOREACH(inactive, len, i) {
/* freeUntilPos[it.reg] = next intersection of it with current */
ir_live_interval *other = ctx->live_intervals[i];
if (reg == other->reg) {
ir_live_pos overlap = ir_vregs_overlap(ctx, current, i);
if (overlap) {
IR_ASSERT(other->type != IR_VOID);
IR_LOG_LSRA_CONFLICT(" ---- Conflict with inactive", i, other, overlap);
// TODO: optimal split position (this case is not tested)
ir_split_interval_at(ctx, i, other, overlap);
}
}
} IR_BITSET_FOREACH_END();
/* current.reg = reg */
ival->reg = reg;
IR_LOG_LSRA_ASSIGN(" ---- Assign", current, ival, " (after splitting others)");
return reg;
}
static int ir_live_range_cmp(const void *r1, const void *r2, void *data)
{
ir_ctx *ctx = data;
ir_live_range *lrg1 = &ctx->live_intervals[*(ir_ref*)r1]->range;
ir_live_range *lrg2 = &ctx->live_intervals[*(ir_ref*)r2]->range;
return lrg2->start - lrg1->start;
}
static int ir_fix_dessa_tmps(ir_ctx *ctx, uint8_t type, int from, int to)
{
if (to == 0) {
ir_block *bb = ctx->data;
ir_reg reg;
if (IR_IS_TYPE_INT(type)) {
reg = IR_REG_R0; // TODO: Temporary register
} else if (IR_IS_TYPE_FP(type)) {
reg = IR_REG_XMM0; // TODO: Temporary register
} else {
IR_ASSERT(0);
return 0;
}
ir_add_fixed_live_range(ctx, reg,
IR_START_LIVE_POS_FROM_REF(bb->end),
IR_END_LIVE_POS_FROM_REF(bb->end));
}
return 1;
}
static int ir_linear_scan(ir_ctx *ctx)
{
int b;
ir_block *bb;
ir_list unhandled;
ir_bitset active, inactive;
ir_live_interval *ival;
int current, i;
uint32_t len;
ir_live_pos position;
ir_reg reg;
ir_lsra_data data;
if (!ctx->live_intervals) {
return 0;
}
/* Add fixed intervals for temporary registers used for DESSA moves */
for (b = 1, bb = &ctx->cfg_blocks[1]; b <= ctx->cfg_blocks_count; b++, bb++) {
if (bb->flags & IR_BB_DESSA_MOVES) {
ctx->data = bb;
ir_gen_dessa_moves(ctx, b, ir_fix_dessa_tmps);
}
}
ctx->data = &data;
data.stack_frame_size = 0;
ir_list_init(&unhandled, ctx->vregs_count + 1);
len = ir_bitset_len(ctx->vregs_count + 1 + IR_REG_NUM);
active = ir_bitset_malloc(ctx->vregs_count + 1 + IR_REG_NUM);
inactive = ir_bitset_malloc(ctx->vregs_count + 1 + IR_REG_NUM);
for (i = 1; i <= ctx->vregs_count; i++) {
if (ctx->live_intervals[i] && ctx->live_intervals[i]->range.start > 0) {
ir_list_push(&unhandled, i);
}
}
for (i = ctx->vregs_count + 1; i <= ctx->vregs_count + IR_REG_NUM; i++) {
if (ctx->live_intervals[i]) {
ir_bitset_incl(inactive, i);
}
}
qsort_r(unhandled.a.refs, ir_list_len(&unhandled), sizeof(ir_ref), ir_live_range_cmp, ctx);
#ifdef IR_DEBUG
if (ctx->flags & IR_DEBUG_RA) {
fprintf(stderr, "----\n");
ir_dump_live_ranges(ctx, stderr);
fprintf(stderr, "---- Start LSRA\n");
}
#endif
while (1) {
if (ir_list_len(&unhandled) == 0) {
position = 0x7fffffff;
IR_BITSET_FOREACH(active, len, i) {
ival = ctx->live_intervals[i];
if (ival->next) {
if (ival->next->range.start < position) {
position = ival->next->range.start;
current = i;
}
}
} IR_BITSET_FOREACH_END();
if (position < 0x7fffffff) {
ir_bitset_excl(active, current);
ival = ctx->live_intervals[current] = ctx->live_intervals[current]->next;
} else {
break;
}
} else {
current = ir_list_pop(&unhandled);
ival = ctx->live_intervals[current];
position = ival->range.start;
}
IR_LOG_LSRA(" ---- Processing", current, ival, "...");
/* for each interval i in active */
IR_BITSET_FOREACH(active, len, i) {
ival = ctx->live_intervals[i];
if (ir_ival_end(ival) <= position) {
/* move i from active to handled */
ir_bitset_excl(active, i);
if (ival->next) {
ctx->live_intervals[i] = ival->next;
ir_add_to_unhandled(ctx, &unhandled, i);
}
} else if (!ir_ival_covers(ival, position)) {
/* move i from active to inactive */
ir_bitset_excl(active, i);
ir_bitset_incl(inactive, i);
}
} IR_BITSET_FOREACH_END();
/* for each interval i in inactive */
IR_BITSET_FOREACH(inactive, len, i) {
ival = ctx->live_intervals[i];
if (ir_ival_end(ival) <= position) {
/* move i from inactive to handled */
ir_bitset_excl(inactive, i);
if (ival->next) {
ctx->live_intervals[i] = ival->next;
ir_add_to_unhandled(ctx, &unhandled, i);
}
} else if (ir_ival_covers(ival, position)) {
/* move i from active to inactive */
ir_bitset_excl(inactive, i);
ir_bitset_incl(active, i);
}
} IR_BITSET_FOREACH_END();
#if 1 && IR_DEBUG
ival = ctx->live_intervals[current];
ir_insn *insn = &ctx->ir_base[IR_LIVE_POS_TO_REF(ival->range.start)];
if (insn->op == IR_VLOAD) {
ir_insn *var = &ctx->ir_base[insn->op2];
IR_ASSERT(var->op == IR_VAR);
if (strcmp(ir_get_str(ctx, var->op2), "_spill_") == 0) {
if (ctx->live_intervals[ctx->vregs[insn->op2]]->stack_spill_pos) {
ctx->live_intervals[current]->stack_spill_pos =
ctx->live_intervals[ctx->vregs[insn->op2]]->stack_spill_pos;
} else {
ir_allocate_spill_slot(ctx, current, &data);
ctx->live_intervals[ctx->vregs[insn->op2]]->stack_spill_pos =
ctx->live_intervals[current]->stack_spill_pos;
}
continue;
}
}
insn = &ctx->ir_base[IR_LIVE_POS_TO_REF(ival->range.end)];
if (insn->op == IR_VSTORE) {
ir_insn *var = &ctx->ir_base[insn->op2];
IR_ASSERT(var->op == IR_VAR);
if (strcmp(ir_get_str(ctx, var->op2), "_spill_") == 0) {
if (ctx->live_intervals[ctx->vregs[insn->op2]]->stack_spill_pos) {
ctx->live_intervals[current]->stack_spill_pos =
ctx->live_intervals[ctx->vregs[insn->op2]]->stack_spill_pos;
} else {
ir_allocate_spill_slot(ctx, current, &data);
ctx->live_intervals[ctx->vregs[insn->op2]]->stack_spill_pos =
ctx->live_intervals[current]->stack_spill_pos;
}
continue;
}
}
#endif
reg = ir_try_allocate_free_reg(ctx, current, len, active, inactive);
if (reg == IR_REG_NONE) {
reg = ir_allocate_blocked_reg(ctx, current, len, active, inactive, &unhandled);
}
if (reg != IR_REG_NONE) {
if (ctx->live_intervals[current]->reg != IR_REG_NONE) {
ir_bitset_incl(active, current);
}
}
}
#ifdef IR_DEBUG
/* all intervals must be processed */
IR_BITSET_FOREACH(active, len, i) {
ival = ctx->live_intervals[i];
IR_ASSERT(!ival->next);
} IR_BITSET_FOREACH_END();
IR_BITSET_FOREACH(inactive, len, i) {
ival = ctx->live_intervals[i];
IR_ASSERT(!ival->next);
} IR_BITSET_FOREACH_END();
#endif
ir_mem_free(inactive);
ir_mem_free(active);
ir_list_free(&unhandled);
for (i = 1; i <= ctx->vregs_count; i++) {
ival = ctx->live_intervals[i];
if (ival) {
ival = ival->top;
ctx->live_intervals[i] = ival;
if (ival->next || ival->reg == IR_REG_NONE) {
ir_allocate_spill_slot(ctx, i, &data);
}
}
}
#ifdef IR_DEBUG
if (ctx->flags & IR_DEBUG_RA) {
fprintf(stderr, "---- Finish LSRA\n");
ir_dump_live_ranges(ctx, stderr);
fprintf(stderr, "----\n");
}
#endif
return 1;
}
int ir_reg_alloc(ir_ctx *ctx)
{
return ir_linear_scan(ctx);
}
Reference in New Issue View Git Blame Copy Permalink