stage2 大量优化代码使其自举编译时间大幅缩短
This commit is contained in:
zzy committed
1 parent
e02b0c633a
commit
846f441e3a
38 files changed
+2418
-835
No files matched your search
+165
@@ -0,0 +1,165 @@
|
||||
/* vm_opt.c - 字节码（VM 层）优化 pass
|
||||
*
|
||||
* 从 stage0/spl_mcode.c 迁出：对已生成的 spl_prog_t 做 prog->prog 优化。
|
||||
* 默认不启用（由 splc0 --opt 驱动）。 */
|
||||
|
||||
#include "vm_opt.h"
|
||||
|
||||
#include "../stage0/include/core_map.h"
|
||||
|
||||
#include <stdlib.h>
|
||||
|
||||
/* 消除“store 后立刻 load 同一槽位”的往返。
|
||||
*
|
||||
* 模式（5 条）：laddr X; swap; store k; laddr X; load k
|
||||
* 变换（3 条）：laddr X; pick 1; store k
|
||||
* [.., v] -> laddr X => [.., v, X]
|
||||
* -> pick 1 => [.., v, X, v]
|
||||
* -> store k => [.., v] （存 v 到 X，同时栈上保留 v）
|
||||
* 与原来 store 后再 load 的效果等价，但去掉 2 条指令与一次内存读。
|
||||
*
|
||||
* 会真的删除指令，因此必须重定位：相对跳转、函数表 address、
|
||||
* 以及作为函数指针的 `push ptr <addr>`。仅当被删/被改写的指令不是
|
||||
* 跳转目标或函数入口时才变换（已收集 mark）。 */
|
||||
static usize peep_hash_int(usize k) { return MAP_HASH_INT(k); }
|
||||
static int peep_cmp_int(usize a, usize b) { return MAP_CMP_INT(a, b); }
|
||||
|
||||
void spl_prog_peephole(spl_prog_t *prog) {
|
||||
if (!prog)
|
||||
return;
|
||||
usize n = vec_size(prog->insns);
|
||||
if (n < 5)
|
||||
return;
|
||||
spl_vm_ins_t *ins = prog->insns.data;
|
||||
|
||||
unsigned char *mark = (unsigned char *)calloc(n, 1); /* 跳转目标/函数入口，受保护 */
|
||||
unsigned char *rem = (unsigned char *)calloc(n, 1);
|
||||
if (!mark || !rem) {
|
||||
free(mark);
|
||||
free(rem);
|
||||
return;
|
||||
}
|
||||
|
||||
for (usize i = 0; i < n; i++) {
|
||||
spl_vm_ins_t *x = &ins[i];
|
||||
if (x->opcode == SPL_JMP || x->opcode == SPL_BZ || x->opcode == SPL_BNZ) {
|
||||
isize t = (isize)i + 1 + (isize)x->imm;
|
||||
if (t >= 0 && (usize)t < n)
|
||||
mark[(usize)t] = 1;
|
||||
}
|
||||
}
|
||||
vec_for(prog->funcs, fi) {
|
||||
spl_vm_func_t *f = &vec_at(prog->funcs, fi);
|
||||
if (f->address < n)
|
||||
mark[f->address] = 1;
|
||||
}
|
||||
|
||||
for (usize i = 0; i + 4 < n; i++) {
|
||||
spl_vm_ins_t *a = &ins[i];
|
||||
spl_vm_ins_t *b = &ins[i + 1];
|
||||
spl_vm_ins_t *c = &ins[i + 2];
|
||||
spl_vm_ins_t *d = &ins[i + 3];
|
||||
spl_vm_ins_t *e = &ins[i + 4];
|
||||
if (a->opcode == SPL_LADDR && b->opcode == SPL_SWAP && c->opcode == SPL_STORE &&
|
||||
d->opcode == SPL_LADDR && e->opcode == SPL_LOAD && a->imm == d->imm &&
|
||||
c->type == e->type) {
|
||||
if (mark[i + 1] || mark[i + 3] || mark[i + 4])
|
||||
continue;
|
||||
b->opcode = SPL_PICK;
|
||||
b->type = SPL_VOID;
|
||||
b->imm = 1;
|
||||
rem[i + 3] = 1;
|
||||
rem[i + 4] = 1;
|
||||
i += 4;
|
||||
}
|
||||
}
|
||||
|
||||
usize kept = 0;
|
||||
for (usize i = 0; i < n; i++)
|
||||
if (!rem[i])
|
||||
kept++;
|
||||
if (kept == n) {
|
||||
free(mark);
|
||||
free(rem);
|
||||
return;
|
||||
}
|
||||
|
||||
spl_vm_ins_t *fresh = (spl_vm_ins_t *)malloc(sizeof(spl_vm_ins_t) * (kept ? kept : 1));
|
||||
usize *map = (usize *)malloc(sizeof(usize) * (n + 1)); /* old index -> new index */
|
||||
usize *old_of_new = (usize *)malloc(sizeof(usize) * (kept ? kept : 1));
|
||||
if (!fresh || !map || !old_of_new) {
|
||||
free(mark);
|
||||
free(rem);
|
||||
free(fresh);
|
||||
free(map);
|
||||
free(old_of_new);
|
||||
return;
|
||||
}
|
||||
|
||||
usize out = 0;
|
||||
for (usize i = 0; i < n; i++) {
|
||||
if (!rem[i]) {
|
||||
map[i] = out;
|
||||
old_of_new[out] = i;
|
||||
fresh[out] = ins[i];
|
||||
out++;
|
||||
}
|
||||
}
|
||||
map[n] = kept;
|
||||
for (usize i = n; i-- > 0;)
|
||||
if (rem[i])
|
||||
map[i] = map[i + 1];
|
||||
|
||||
/* 函数地址 old->new */
|
||||
MAP(usize, usize) addr_map;
|
||||
map_init(addr_map, peep_hash_int, peep_cmp_int);
|
||||
vec_for(prog->funcs, fi) {
|
||||
spl_vm_func_t *f = &vec_at(prog->funcs, fi);
|
||||
usize old_a = f->address;
|
||||
if (old_a < n) {
|
||||
usize new_a = map[old_a];
|
||||
map_put(addr_map, old_a, new_a);
|
||||
f->address = new_a;
|
||||
}
|
||||
}
|
||||
|
||||
/* 修正相对跳转与函数指针常量 */
|
||||
for (usize j = 0; j < kept; j++) {
|
||||
spl_vm_ins_t *x = &fresh[j];
|
||||
if (x->opcode == SPL_JMP || x->opcode == SPL_BZ || x->opcode == SPL_BNZ) {
|
||||
isize t = (isize)old_of_new[j] + 1 + (isize)x->imm;
|
||||
if (t >= 0 && (usize)t <= n) {
|
||||
usize nt = map[(usize)t];
|
||||
x->imm = (spl_vm_val_t)(isize)((isize)nt - (isize)(j + 1));
|
||||
}
|
||||
} else if (x->opcode == SPL_PUSH && x->type == SPL_PTR && x->imm != 0) {
|
||||
usize nv = 0;
|
||||
if (map_get(addr_map, (usize)x->imm, &nv))
|
||||
x->imm = nv;
|
||||
}
|
||||
}
|
||||
map_free(addr_map);
|
||||
|
||||
/* 指令删除后函数长度变化，重算 ninsns（函数按地址升序排列）。 */
|
||||
{
|
||||
usize nf = vec_size(prog->funcs);
|
||||
for (usize i = 0; i < nf; i++) {
|
||||
spl_vm_func_t *f = &vec_at(prog->funcs, i);
|
||||
usize next = (i + 1 < nf) ? vec_at(prog->funcs, i + 1).address : kept;
|
||||
f->ninsns = next > f->address ? (spl_vm_val_t)(next - f->address) : 0;
|
||||
}
|
||||
}
|
||||
|
||||
free(prog->insns.data);
|
||||
prog->insns.data = fresh;
|
||||
prog->insns.size = kept;
|
||||
prog->insns.cap = kept;
|
||||
|
||||
free(mark);
|
||||
free(rem);
|
||||
free(map);
|
||||
free(old_of_new);
|
||||
}
|
||||
|
||||
/* 优化 pass 流水线（纯 prog->prog）：当前 P1，后续可追加。 */
|
||||
void spl_prog_opt(spl_prog_t *prog) { spl_prog_peephole(prog); }
|
||||
Reference in new issue
Block a user