danog/ir
1
0
Fork 0
mirror of https://github.com/danog/ir.git synced 2024-11-27 04:45:38 +01:00
Code Issues Packages Projects Releases Wiki Activity
d858872b90
ir/ir_sccp.c
Dmitry Stogov 142cdb4b8b SCCP cleanup and optimization
2022-11-09 13:24:37 +03:00

823 lines
23 KiB
C
Raw Blame History

/*
* IR - Lightweight JIT Compilation Framework
* (SCCP - Sparse Conditional Constant Propagation)
* Copyright (C) 2022 Zend by Perforce.
* Authors: Dmitry Stogov <dmitry@php.net>
*
* The SCCP algorithm is based on M. N. Wegman and F. K. Zadeck publication
* See: M. N. Wegman and F. K. Zadeck. "Constant propagation with conditional branches"
* ACM Transactions on Programming Languages and Systems, 13(2):181-210, April 1991
*/
#include "ir.h"
#include "ir_private.h"
#define IR_TOP IR_UNUSED
#define IR_BOTTOM IR_LAST_OP
#define IR_MAKE_TOP(ref) do {IR_ASSERT(ref > 0); _values[ref].optx = IR_TOP;} while (0)
#define IR_MAKE_BOTTOM(ref) do {IR_ASSERT(ref > 0); _values[ref].optx = IR_BOTTOM;} while (0)
#define IR_IS_TOP(ref) (ref >= 0 && _values[ref].optx == IR_TOP)
#define IR_IS_BOTTOM(ref) (ref >= 0 && _values[ref].optx == IR_BOTTOM)
#define IR_IS_FEASIBLE(ref) (ref >= 0 && _values[ref].optx != IR_TOP)
#define IR_COMBO_COPY_PROPAGATION 1
#if IR_COMBO_COPY_PROPAGATION
IR_ALWAYS_INLINE ir_ref ir_sccp_identity(ir_insn *_values, ir_ref a)
{
if (a > 0 && _values[a].op == IR_COPY) {
a = _values[a].op1;
IR_ASSERT(a <= 0 || _values[a].op != IR_COPY);
}
return a;
}
#endif
static ir_ref ir_sccp_fold(ir_ctx *ctx, ir_insn *_values, ir_ref res, uint32_t opt, ir_ref op1, ir_ref op2, ir_ref op3)
{
ir_insn *op1_insn, *op2_insn, *op3_insn, *insn;
#if IR_COMBO_COPY_PROPAGATION
op1 = ir_sccp_identity(_values, op1);
op2 = ir_sccp_identity(_values, op2);
op3 = ir_sccp_identity(_values, op3);
#endif
restart:
op1_insn = (op1 > 0 && IR_IS_CONST_OP(_values[op1].op)) ? _values + op1 : ctx->ir_base + op1;
op2_insn = (op2 > 0 && IR_IS_CONST_OP(_values[op2].op)) ? _values + op2 : ctx->ir_base + op2;
op3_insn = (op3 > 0 && IR_IS_CONST_OP(_values[op3].op)) ? _values + op3 : ctx->ir_base + op3;
switch (ir_folding(ctx, opt, op1, op2, op3, op1_insn, op2_insn, op3_insn)) {
case IR_FOLD_DO_RESTART:
opt = ctx->fold_insn.optx;
op1 = ctx->fold_insn.op1;
op2 = ctx->fold_insn.op2;
op3 = ctx->fold_insn.op3;
goto restart;
case IR_FOLD_DO_EMIT:
IR_MAKE_BOTTOM(res);
return 1;
case IR_FOLD_DO_COPY:
op1 = ctx->fold_insn.op1;
#if IR_COMBO_COPY_PROPAGATION
op1 = ir_sccp_identity(_values, op1);
#endif
insn = (op1 > 0 && IR_IS_CONST_OP(_values[op1].op)) ? _values + op1 : ctx->ir_base + op1;
if (IR_IS_CONST_OP(insn->op)) {
/* pass */
#if IR_COMBO_COPY_PROPAGATION
} else if (IR_IS_TOP(res)) {
_values[res].optx = IR_OPT(IR_COPY, insn->type);
_values[res].op1 = op1;
return 1;
} else if (_values[res].op == IR_COPY && _values[res].op1 == op1) {
return 0; /* not changed */
#endif
} else {
IR_MAKE_BOTTOM(res);
return 1;
}
break;
case IR_FOLD_DO_CONST:
insn = &ctx->fold_insn;
break;
default:
IR_ASSERT(0);
return 0;
}
if (IR_IS_TOP(res)) {
_values[res].optx = IR_OPT(insn->type, insn->type);
_values[res].val.u64 = insn->val.u64;
return 1;
} else if (_values[res].opt != IR_OPT(insn->type, insn->type) || _values[res].val.u64 != insn->val.u64) {
IR_MAKE_BOTTOM(res);
return 1;
}
return 0; /* not changed */
}
static bool ir_sccp_join_values(ir_ctx *ctx, ir_insn *_values, ir_ref a, ir_ref b)
{
ir_insn *v;
if (!IR_IS_BOTTOM(a) && !IR_IS_TOP(b)) {
b = ir_sccp_identity(_values, b);
v = IR_IS_CONST_REF(b) ? &ctx->ir_base[b] : &_values[b];
if (IR_IS_TOP(a)) {
#if IR_COMBO_COPY_PROPAGATION
if (v->op == IR_BOTTOM) {
_values[a].optx = IR_OPT(IR_COPY, ctx->ir_base[b].type);
_values[a].op1 = b;
return 1;
}
#endif
_values[a].optx = v->opt;
_values[a].val.u64 = v->val.u64;
return 1;
} else if (_values[a].opt == v->opt && _values[a].val.u64 == v->val.u64) {
/* pass */
#if IR_COMBO_COPY_PROPAGATION
} else if (_values[a].op == IR_COPY && _values[a].op1 == b) {
/* pass */
#endif
} else {
IR_MAKE_BOTTOM(a);
return 1;
}
}
return 0;
}
static bool ir_sccp_is_true(ir_ctx *ctx, ir_insn *_values, ir_ref a)
{
ir_insn *v = IR_IS_CONST_REF(a) ? &ctx->ir_base[a] : &_values[a];
if (v->type == IR_BOOL) {
return v->val.b;
} else if (IR_IS_TYPE_INT(v->type)) {
return v->val.i64 != 0;
} else if (v->type == IR_DOUBLE) {
return v->val.d != 0.0;
} else if (v->type == IR_FLOAT) {
return v->val.f != 0.0;
}
IR_ASSERT(0 && "NYI");
return 0;
}
static bool ir_sccp_is_equal(ir_ctx *ctx, ir_insn *_values, ir_ref a, ir_ref b)
{
ir_insn *v1 = IR_IS_CONST_REF(a) ? &ctx->ir_base[a] : &_values[a];
ir_insn *v2 = IR_IS_CONST_REF(b) ? &ctx->ir_base[b] : &_values[b];
return v1->val.u64 == v2->val.u64;
}
static void ir_sccp_remove_from_use_list(ir_ctx *ctx, ir_ref from, ir_ref ref)
{
ir_ref j, n, *p, *q, use;
ir_use_list *use_list = &ctx->use_lists[from];
ir_ref skip = 0;
n = use_list->count;
for (j = 0, p = q = &ctx->use_edges[use_list->refs]; j < n; j++, p++) {
use = *p;
if (use == ref) {
skip++;
} else {
if (p != q) {
*q = use;
}
q++;
}
}
use_list->count -= skip;
#if IR_COMBO_COPY_PROPAGATION
if (skip) {
do {
*q = IR_UNUSED;
q++;
} while (--skip);
}
#endif
}
#if IR_COMBO_COPY_PROPAGATION
static int ir_sccp_add_to_use_list(ir_ctx *ctx, ir_ref to, ir_ref ref)
{
ir_use_list *use_list = &ctx->use_lists[to];
ir_ref n = use_list->refs + use_list->count;
if (n < ctx->use_edges_count && ctx->use_edges[n] == IR_UNUSED) {
ctx->use_edges[n] = ref;
use_list->count++;
return 0;
} else {
/* Reallocate the whole edges buffer (this is inefficient) */
ctx->use_edges = ir_mem_realloc(ctx->use_edges, (ctx->use_edges_count + use_list->count + 1) * sizeof(ir_ref));
memcpy(ctx->use_edges + ctx->use_edges_count, ctx->use_edges + use_list->refs, use_list->count * sizeof(ir_ref));
use_list->refs = ctx->use_edges_count;
ctx->use_edges[use_list->refs + use_list->count] = ref;
use_list->count++;
ctx->use_edges_count += use_list->count;
return 1;
}
}
#endif
static void ir_sccp_make_nop(ir_ctx *ctx, ir_ref ref)
{
ir_ref j, n, *p;
ir_use_list *use_list = &ctx->use_lists[ref];
ir_insn *insn;
use_list->refs = 0;
use_list->count = 0;
insn = &ctx->ir_base[ref];
n = ir_input_edges_count(ctx, insn);
insn->optx = IR_NOP;
for (j = 1, p = insn->ops + j; j <= n; j++, p++) {
*p = IR_UNUSED;
}
}
static void ir_sccp_replace_insn(ir_ctx *ctx, ir_insn *_values, ir_ref ref, ir_ref new_ref, ir_bitqueue *worklist)
{
ir_ref j, n, *p, use, k, l;
ir_insn *insn;
ir_use_list *use_list;
IR_ASSERT(ref != new_ref);
insn = &ctx->ir_base[ref];
n = ir_input_edges_count(ctx, insn);
for (j = 1, p = insn->ops + 1; j <= n; j++, p++) {
ir_ref input = *p;
if (input > 0) {
ir_sccp_remove_from_use_list(ctx, input, ref);
/* schedule DCE */
if (worklist
&& ((IR_IS_FOLDABLE_OP(ctx->ir_base[input].op) && ctx->use_lists[input].count == 0)
|| ((ir_op_flags[ctx->ir_base[input].op] & (IR_OP_FLAG_MEM|IR_OP_FLAG_MEM_MASK)) == (IR_OP_FLAG_MEM|IR_OP_FLAG_MEM_LOAD)
&& ctx->use_lists[input].count == 1))) {
ir_bitqueue_add(worklist, input);
}
}
}
use_list = &ctx->use_lists[ref];
n = use_list->count;
for (j = 0, p = &ctx->use_edges[use_list->refs]; j < n; j++, p++) {
use = *p;
if (IR_IS_FEASIBLE(use)) {
insn = &ctx->ir_base[use];
l = ir_input_edges_count(ctx, insn);
for (k = 1; k <= l; k++) {
if (ir_insn_op(insn, k) == ref) {
ir_insn_set_op(insn, k, new_ref);
}
}
#if IR_COMBO_COPY_PROPAGATION
if (new_ref > 0 && IR_IS_BOTTOM(use)) {
if (ir_sccp_add_to_use_list(ctx, new_ref, use)) {
/* restore after reallocation */
use_list = &ctx->use_lists[ref];
n = use_list->count;
p = &ctx->use_edges[use_list->refs + j];
}
}
#endif
/* schedule folding */
if (worklist && _values[use].op == IR_BOTTOM) {
ir_bitqueue_add(worklist, use);
}
}
}
ir_sccp_make_nop(ctx, ref);
}
static void ir_sccp_fold2(ir_ctx *ctx, ir_insn *_values, ir_ref ref, ir_bitqueue *worklist)
{
uint32_t opt;
ir_ref op1, op2, op3;
ir_insn *op1_insn, *op2_insn, *op3_insn, *insn;
insn = &ctx->ir_base[ref];
opt = insn->opt;
op1 = insn->op1;
op2 = insn->op2;
op3 = insn->op3;
restart:
op1_insn = ctx->ir_base + op1;
op2_insn = ctx->ir_base + op2;
op3_insn = ctx->ir_base + op3;
switch (ir_folding(ctx, opt, op1, op2, op3, op1_insn, op2_insn, op3_insn)) {
case IR_FOLD_DO_RESTART:
opt = ctx->fold_insn.optx;
op1 = ctx->fold_insn.op1;
op2 = ctx->fold_insn.op2;
op3 = ctx->fold_insn.op3;
goto restart;
case IR_FOLD_DO_EMIT:
insn = &ctx->ir_base[ref];
if (insn->opt != ctx->fold_insn.opt
|| insn->op1 != ctx->fold_insn.op1
|| insn->op2 != ctx->fold_insn.op2
|| insn->op3 != ctx->fold_insn.op3) {
ir_use_list *use_list;
ir_ref n, j, *p, use;
insn->optx = ctx->fold_insn.opt;
insn->op1 = ctx->fold_insn.op1;
insn->op2 = ctx->fold_insn.op2;
insn->op3 = ctx->fold_insn.op3;
use_list = &ctx->use_lists[ref];
n = use_list->count;
for (j = 0, p = &ctx->use_edges[use_list->refs]; j < n; j++, p++) {
use = *p;
if (_values[use].op == IR_BOTTOM) {
ir_bitqueue_add(worklist, use);
}
}
}
break;
case IR_FOLD_DO_COPY:
op1 = ctx->fold_insn.op1;
ir_sccp_replace_insn(ctx, _values, ref, op1, worklist);
break;
case IR_FOLD_DO_CONST:
op1 = ir_const(ctx, ctx->fold_insn.val, ctx->fold_insn.type);
ir_sccp_replace_insn(ctx, _values, ref, op1, worklist);
break;
default:
IR_ASSERT(0);
break;
}
}
static void ir_sccp_replace_use(ir_ctx *ctx, ir_ref ref, ir_ref use, ir_ref new_use)
{
ir_use_list *use_list = &ctx->use_lists[ref];
ir_ref i, n, *p;
n = use_list->count;
for (i = 0, p = &ctx->use_edges[use_list->refs]; i < n; i++, p++) {
if (*p == use) {
*p = new_use;
break;
}
}
}
static void ir_sccp_remove_if(ir_ctx *ctx, ir_insn *_values, ir_ref ref, ir_ref dst)
{
ir_ref j, n, *p, use, next;
ir_insn *insn, *next_insn;
ir_use_list *use_list = &ctx->use_lists[ref];
insn = &ctx->ir_base[ref];
n = use_list->count;
for (j = 0, p = &ctx->use_edges[use_list->refs]; j < n; j++, p++) {
use = *p;
if (use == dst) {
next = ctx->use_edges[ctx->use_lists[use].refs];
next_insn = &ctx->ir_base[next];
/* remove IF and IF_TRUE/FALSE from double linked control list */
next_insn->op1 = insn->op1;
ir_sccp_replace_use(ctx, insn->op1, ref, next);
/* remove IF and IF_TRUE/FALSE instructions */
ir_sccp_make_nop(ctx, ref);
ir_sccp_make_nop(ctx, use);
break;
}
}
}
static void ir_sccp_remove_unfeasible_merge_inputs(ir_ctx *ctx, ir_insn *_values, ir_ref ref, ir_ref unfeasible_inputs)
{
ir_ref i, j, n, k, *p, use;
ir_insn *insn, *use_insn;
ir_use_list *use_list;
ir_bitset life_inputs;
insn = &ctx->ir_base[ref];
IR_ASSERT(insn->op == IR_MERGE || insn->op == IR_LOOP_BEGIN);
n = insn->inputs_count;
if (n == 0) {
n = 2;
}
if (n - unfeasible_inputs == 1) {
/* remove MERGE completely */
for (j = 1; j <= n; j++) {
ir_ref input = ir_insn_op(insn, j);
if (input && IR_IS_FEASIBLE(input)) {
ir_ref prev, next = IR_UNUSED;
ir_insn *next_insn = NULL, *input_insn = &ctx->ir_base[input];
IR_ASSERT(input_insn->op == IR_END || input_insn->op == IR_IJMP || input_insn->op == IR_UNREACHABLE);
prev = input_insn->op1;
use_list = &ctx->use_lists[ref];
for (k = 0, p = &ctx->use_edges[use_list->refs]; k < use_list->count; k++, p++) {
use = *p;
use_insn = &ctx->ir_base[use];
IR_ASSERT((use_insn->op != IR_PHI) && "PHI must be already removed");
if (ir_op_flags[use_insn->op] & IR_OP_FLAG_CONTROL) {
next = use;
next_insn = use_insn;
break;
}
}
IR_ASSERT(prev && next);
/* remove MERGE and input END from double linked control list */
next_insn->op1 = prev;
ir_sccp_replace_use(ctx, prev, input, next);
/* remove MERGE and input END instructions */
ir_sccp_make_nop(ctx, ref);
ir_sccp_make_nop(ctx, input);
break;
}
}
} else {
n = insn->inputs_count;
if (n == 0) {
n = 3;
}
i = 1;
life_inputs = ir_bitset_malloc(n + 1);
for (j = 1; j <= n; j++) {
ir_ref input = ir_insn_op(insn, j);
if (input) {
if (i != j) {
ir_insn_set_op(insn, i, input);
}
ir_bitset_incl(life_inputs, j);
i++;
}
}
i--;
if (i == 2) {
i = 0;
}
insn->inputs_count = i;
n++;
use_list = &ctx->use_lists[ref];
for (k = 0, p = &ctx->use_edges[use_list->refs]; k < use_list->count; k++, p++) {
use = *p;
use_insn = &ctx->ir_base[use];
if (use_insn->op == IR_PHI) {
i = 2;
for (j = 2; j <= n; j++) {
if (ir_bitset_in(life_inputs, j - 1)) {
IR_ASSERT(ir_insn_op(use_insn, j));
if (i != j) {
ir_insn_set_op(use_insn, i, ir_insn_op(use_insn, j));
}
i++;
} else {
IR_ASSERT(ir_insn_op(use_insn, j) <= 0);
}
}
}
}
ir_mem_free(life_inputs);
}
}
int ir_sccp(ir_ctx *ctx)
{
ir_ref i, j, n, *p, use;
ir_use_list *use_list;
ir_insn *insn, *use_insn, *value;
uint32_t flags;
ir_bitqueue worklist;
ir_insn *_values = ir_mem_calloc(ctx->insns_count, sizeof(ir_insn));
ctx->flags |= IR_OPT_IN_SCCP;
/* A bit modified SCCP algorith of M. N. Wegman and F. K. Zadeck */
ir_bitqueue_init(&worklist, ctx->insns_count);
worklist.pos = 0;
ir_bitset_incl(worklist.set, 1);
i = ctx->ir_base[1].op2;
while (i) {
ir_bitset_incl(worklist.set, i);
i = ctx->ir_base[i].op2;
}
while ((i = ir_bitqueue_pop(&worklist)) >= 0) {
insn = &ctx->ir_base[i];
flags = ir_op_flags[insn->op];
if (flags & IR_OP_FLAG_DATA) {
if (insn->op == IR_PHI) {
ir_insn *merge_insn = &ctx->ir_base[insn->op1];
bool changed = 0;
if (!IR_IS_FEASIBLE(insn->op1)) {
continue;
}
n = ir_variable_inputs_count(merge_insn) + 1;
if (n > 3 && _values[i].optx == IR_TOP) {
for (j = 0; j < (n>>2); j++) {
_values[i+j+1].optx = IR_BOTTOM; /* keep the tail of a long multislot instruction */
}
}
/* for all live merge inputs */
for (j = 1; j < n; j++) {
ir_ref merge_input = ir_insn_op(merge_insn, j);
IR_ASSERT(merge_input > 0);
if (_values[merge_input].optx != IR_TOP) {
ir_ref input = ir_insn_op(insn, j + 1);
if (input > 0 && IR_IS_TOP(input)) {
ir_bitqueue_add(&worklist, input);
} else if (ir_sccp_join_values(ctx, _values, i, input)) {
changed = 1;
}
}
}
if (!changed) {
continue;
}
} else if (EXPECTED(IR_IS_FOLDABLE_OP(insn->op))) {
bool may_benefit = 0;
bool has_top = 0;
n = IR_INPUT_EDGES_COUNT(flags);
IR_ASSERT(IR_IS_FIXED_INPUTS_COUNT(n));
for (p = insn->ops + 1; n > 0; p++, n--) {
ir_ref input = *p;
if (input > 0) {
if (_values[input].optx == IR_TOP) {
has_top = 1;
ir_bitqueue_add(&worklist, input);
} else if (_values[input].optx != IR_BOTTOM) {
/* Perform folding only if some of direct inputs
* is going to be replaced by a constant or copy.
* This approach may miss some folding optimizations
* dependent on indirect inputs. e.g reassociation.
*/
may_benefit = 1;
}
}
}
if (has_top) {
continue;
}
if (!may_benefit) {
IR_MAKE_BOTTOM(i);
} else if (!ir_sccp_fold(ctx, _values, i, insn->opt, insn->op1, insn->op2, insn->op3)) {
/* not changed */
continue;
}
} else {
IR_MAKE_BOTTOM(i);
}
} else if (flags & IR_OP_FLAG_BB_START) {
if (insn->op == IR_MERGE || insn->op == IR_LOOP_BEGIN) {
ir_ref unfeasible_inputs = 0;
n = ir_variable_inputs_count(insn);
if (n > 3 && _values[i].optx == IR_TOP) {
for (j = 0; j < (n>>2); j++) {
_values[i+j+1].optx = IR_BOTTOM; /* keep the tail of a long multislot instruction */
}
}
for (p = insn->ops + 1; n > 0; p++, n--) {
ir_ref input = *p;
IR_ASSERT(input > 0);
if (_values[input].optx == IR_TOP) {
unfeasible_inputs++;
}
}
if (unfeasible_inputs == 0) {
IR_MAKE_BOTTOM(i);
} else if (_values[i].op1 != unfeasible_inputs) {
_values[i].optx = IR_MERGE;
_values[i].op1 = unfeasible_inputs;
} else {
continue;
}
} else {
IR_ASSERT(insn->op == IR_START || insn->op == IR_ENTRY || IR_IS_FEASIBLE(insn->op1));
IR_MAKE_BOTTOM(i);
}
} else {
IR_ASSERT(insn->op1 > 0);
if (_values[insn->op1].optx == IR_TOP) {
/* control inpt is not feasible */
continue;
}
if (insn->op == IR_IF) {
if (IR_IS_TOP(insn->op2)) {
ir_bitqueue_add(&worklist, insn->op2);
continue;
}
if (!IR_IS_BOTTOM(insn->op2)) {
bool b = ir_sccp_is_true(ctx, _values, insn->op2);
use_list = &ctx->use_lists[i];
IR_ASSERT(use_list->count == 2);
p = &ctx->use_edges[use_list->refs];
use = *p;
use_insn = &ctx->ir_base[use];
IR_ASSERT(use_insn->op == IR_IF_TRUE || use_insn->op == IR_IF_FALSE);
if ((use_insn->op == IR_IF_TRUE) != b) {
use = *(p+1);
IR_ASSERT(ctx->ir_base[use].op == IR_IF_TRUE || ctx->ir_base[use].op == IR_IF_FALSE);
}
if (_values[i].optx == IR_TOP) {
_values[i].optx = IR_IF;
_values[i].op1 = use;
} else if (_values[i].optx != IR_IF || _values[i].op1 != use) {
IR_MAKE_BOTTOM(i);
}
if (!IR_IS_BOTTOM(use)) {
ir_bitqueue_add(&worklist, use);
}
continue;
}
IR_MAKE_BOTTOM(i);
} else if (insn->op == IR_SWITCH) {
if (IR_IS_TOP(insn->op2)) {
ir_bitqueue_add(&worklist, insn->op2);
continue;
}
if (!IR_IS_BOTTOM(insn->op2)) {
ir_ref use_case = IR_UNUSED;
use_list = &ctx->use_lists[i];
n = use_list->count;
for (j = 0, p = &ctx->use_edges[use_list->refs]; j < n; j++, p++) {
use = *p;
IR_ASSERT(use > 0);
use_insn = &ctx->ir_base[use];
if (use_insn->op == IR_CASE_VAL) {
if (ir_sccp_is_equal(ctx, _values, insn->op2, use_insn->op2)) {
use_case = use;
break;
}
} else if (use_insn->op == IR_CASE_DEFAULT) {
use_case = use;
}
}
if (use_case) {
use_insn = &ctx->ir_base[use_case];
if (_values[i].optx == IR_TOP) {
_values[i].optx = IR_IF;
_values[i].op1 = use_case;
} else if (_values[i].optx != IR_IF || _values[i].op1 != use_case) {
IR_MAKE_BOTTOM(i);
}
if (!IR_IS_BOTTOM(use_case)) {
ir_bitqueue_add(&worklist, use_case);
}
}
if (!IR_IS_BOTTOM(i)) {
continue;
}
}
IR_MAKE_BOTTOM(i);
} else if ((flags & (IR_OP_FLAG_MEM|IR_OP_FLAG_MEM_MASK)) == (IR_OP_FLAG_MEM|IR_OP_FLAG_MEM_LOAD)
&& ctx->use_lists[i].count == 1) {
/* dead load */
_values[i].optx = IR_LOAD;
} else {
IR_MAKE_BOTTOM(i);
/* control, call, load and strore instructions may have unprocessed inputs */
n = IR_INPUT_EDGES_COUNT(flags);
if (!IR_IS_FIXED_INPUTS_COUNT(n)
&& (n = ir_variable_inputs_count(insn)) > 3) {
for (j = 0; j < (n>>2); j++) {
_values[i+j+1].optx = IR_BOTTOM; /* keep the tail of a long multislot instruction */
}
for (j = 2, p = insn->ops + j; j <= n; j++, p++) {
IR_ASSERT(IR_OPND_KIND(flags, j) == IR_OPND_DATA || IR_OPND_KIND(flags, j) == IR_OPND_VAR);
use = *p;
if (use > 0 && UNEXPECTED(_values[use].optx == IR_TOP)) {
ir_bitqueue_add(&worklist, use);
}
}
} else if (n >= 2) {
IR_ASSERT(IR_OPND_KIND(flags, 2) == IR_OPND_DATA || IR_OPND_KIND(flags, 2) == IR_OPND_VAR);
use = insn->op2;
if (use > 0 && UNEXPECTED(_values[use].optx == IR_TOP)) {
ir_bitqueue_add(&worklist, use);
}
if (n > 2) {
IR_ASSERT(n == 3);
IR_ASSERT(IR_OPND_KIND(flags, 3) == IR_OPND_DATA || IR_OPND_KIND(flags, 3) == IR_OPND_VAR);
use = insn->op3;
if (use > 0 && UNEXPECTED(_values[use].optx == IR_TOP)) {
ir_bitqueue_add(&worklist, use);
}
}
}
}
}
use_list = &ctx->use_lists[i];
n = use_list->count;
for (p = &ctx->use_edges[use_list->refs]; n > 0; p++, n--) {
use = *p;
if (_values[use].optx != IR_BOTTOM) {
ir_bitqueue_add(&worklist, use);
}
}
}
#ifdef IR_DEBUG
if (ctx->flags & IR_DEBUG_SCCP) {
for (i = 1; i < ctx->insns_count; i++) {
if (IR_IS_CONST_OP(_values[i].op)) {
fprintf(stderr, "%d. CONST(", i);
ir_print_const(ctx, &_values[i], stderr);
fprintf(stderr, ")\n");
#if IR_COMBO_COPY_PROPAGATION
} else if (_values[i].op == IR_COPY) {
fprintf(stderr, "%d. COPY(%d)\n", i, _values[i].op1);
#endif
} else if (IR_IS_TOP(i)) {
fprintf(stderr, "%d. TOP\n", i);
} else if (_values[i].op == IR_IF) {
fprintf(stderr, "%d. IF(%d)\n", i, _values[i].op1);
} else if (_values[i].op == IR_MERGE) {
fprintf(stderr, "%d. MERGE(%d)\n", i, _values[i].op1);
} else if (!IR_IS_BOTTOM(i)) {
fprintf(stderr, "%d. %d\n", i, _values[i].op);
}
}
}
#endif
for (i = 1, value = _values + i; i < ctx->insns_count; value++, i++) {
if (value->op == IR_BOTTOM) {
continue;
} else if (IR_IS_CONST_OP(value->op)) {
/* replace instruction by constant */
j = ir_const(ctx, value->val, value->type);
ir_sccp_replace_insn(ctx, _values, i, j, &worklist);
#if IR_COMBO_COPY_PROPAGATION
} else if (value->op == IR_COPY) {
ir_sccp_replace_insn(ctx, _values, i, value->op1, &worklist);
#endif
} else if (value->op == IR_TOP) {
/* remove unreachable instruction */
insn = &ctx->ir_base[i];
if (ir_op_flags[insn->op] & (IR_OP_FLAG_DATA|IR_OP_FLAG_MEM)) {
if (insn->op != IR_PARAM && insn->op != IR_VAR) {
ir_sccp_make_nop(ctx, i);
}
} else {
if (ir_op_flags[insn->op] & IR_OP_FLAG_TERMINATOR) {
/* remove from terminators list */
ir_ref prev = ctx->ir_base[1].op1;
if (prev == i) {
ctx->ir_base[1].op1 = insn->op3;
} else {
while (prev) {
if (ctx->ir_base[prev].op3 == i) {
ctx->ir_base[prev].op3 = insn->op3;
break;
}
prev = ctx->ir_base[prev].op3;
}
}
}
ir_sccp_replace_insn(ctx, _values, i, IR_UNUSED, NULL);
}
} else if (value->op == IR_IF) {
/* remove one way IF/SWITCH */
ir_sccp_remove_if(ctx, _values, i, value->op1);
} else if (value->op == IR_MERGE) {
/* schedule merge to remove unfeasible MERGE inputs */
ir_bitqueue_add(&worklist, i);
} else if (value->op == IR_LOAD) {
/* schedule dead load elimination */
ir_bitqueue_add(&worklist, i);
}
}
while ((i = ir_bitqueue_pop(&worklist)) >= 0) {
if (_values[i].op == IR_MERGE) {
ir_sccp_remove_unfeasible_merge_inputs(ctx, _values, i, _values[i].op1);
} else {
insn = &ctx->ir_base[i];
if (IR_IS_FOLDABLE_OP(insn->op)) {
if (ctx->use_lists[i].count == 0) {
ir_sccp_replace_insn(ctx, _values, i, IR_UNUSED, &worklist);
} else {
ir_sccp_fold2(ctx, _values, i, &worklist);
}
} else if ((ir_op_flags[insn->op] & (IR_OP_FLAG_MEM|IR_OP_FLAG_MEM_MASK)) == (IR_OP_FLAG_MEM|IR_OP_FLAG_MEM_LOAD)
&& ctx->use_lists[i].count == 1) {
/* dead load */
ir_ref next = ctx->use_edges[ctx->use_lists[i].refs];
/* remove LOAD from double linked control list */
ctx->ir_base[next].op1 = insn->op1;
ir_sccp_replace_use(ctx, insn->op1, i, next);
insn->op1 = IR_UNUSED;
ir_sccp_replace_insn(ctx, _values, i, IR_UNUSED, &worklist);
}
}
}
ir_mem_free(_values);
ir_bitqueue_free(&worklist);
ctx->flags &= ~IR_OPT_IN_SCCP;
return 1;
}
Reference in New Issue View Git Blame Copy Permalink