Skip to content

Fix GH-19065: Long match statement can segfault compiler during recursive SSA renaming #19083

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
wants to merge 3 commits into from
Closed
Changes from 1 commit
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
101 changes: 81 additions & 20 deletions Zend/Optimizer/zend_ssa.c
Original file line number Diff line number Diff line change
Expand Up @@ -22,6 +22,7 @@
#include "zend_ssa.h"
#include "zend_dump.h"
#include "zend_inference.h"
#include "zend_worklist.h"
#include "Optimizer/zend_optimizer_internal.h"

static bool dominates(const zend_basic_block *blocks, int a, int b) {
Expand Down Expand Up @@ -787,23 +788,14 @@ ZEND_API int zend_ssa_rename_op(const zend_op_array *op_array, const zend_op *op
}
/* }}} */

static zend_result zend_ssa_rename(const zend_op_array *op_array, uint32_t build_flags, zend_ssa *ssa, int *var, int n) /* {{{ */
static void zend_ssa_rename_in_block(const zend_op_array *op_array, uint32_t build_flags, zend_ssa *ssa, int *var, int n) /* {{{ */
{
zend_basic_block *blocks = ssa->cfg.blocks;
zend_ssa_block *ssa_blocks = ssa->blocks;
zend_ssa_op *ssa_ops = ssa->ops;
int ssa_vars_count = ssa->vars_count;
int i, j;
zend_op *opline, *end;
int *tmp = NULL;
ALLOCA_FLAG(use_heap = 0);

// FIXME: Can we optimize this copying out in some cases?
if (blocks[n].next_child >= 0) {
tmp = do_alloca(sizeof(int) * (op_array->last_var + op_array->T), use_heap);
memcpy(tmp, var, sizeof(int) * (op_array->last_var + op_array->T));
var = tmp;
}

if (ssa_blocks[n].phis) {
zend_ssa_phi *phi = ssa_blocks[n].phis;
Expand Down Expand Up @@ -887,22 +879,91 @@ static zend_result zend_ssa_rename(const zend_op_array *op_array, uint32_t build
}

ssa->vars_count = ssa_vars_count;
}
/* }}} */

j = blocks[n].children;
while (j >= 0) {
// FIXME: Tail call optimization?
if (zend_ssa_rename(op_array, build_flags, ssa, var, j) == FAILURE)
return FAILURE;
j = blocks[j].next_child;
}
static zend_result zend_ssa_rename(const zend_op_array *op_array, uint32_t build_flags, zend_ssa *ssa, int *var, int n)
{
/* The worklist contains block numbers, encoded as positive or negative value.
* Positive values indicate that the variable rename still needs to happen for the block.
* Negative values indicate the variable rename was done and all children were handled too.
* In that case, we will clean up.
* Because block 0 is valid, we bias the block numbers by adding 1 such that we can distinguish
* positive and negative values in all cases. */
zend_worklist_stack work;
ALLOCA_FLAG(work_use_heap);
ZEND_WORKLIST_STACK_ALLOCA(&work, ssa->cfg.blocks_count, work_use_heap);
zend_worklist_stack_push(&work, n + 1);

/* This is used together with `save_vars` to backtrack the right version of the renamed variables to use. */
ALLOCA_FLAG(save_positions_use_heap);
ALLOCA_FLAG(save_vars_use_heap);
unsigned int save_vars_top = 0;
unsigned int *save_positions = do_alloca(sizeof(unsigned int) * ssa->cfg.blocks_count, save_positions_use_heap);
int **save_vars = do_alloca(sizeof(int *) * (ssa->cfg.blocks_count + 1), save_vars_use_heap);
save_vars[0] = var;

while (work.len) {
n = zend_worklist_stack_pop(&work);

/* Enter state: perform SSA variable rename */
if (n > 0) {
n--;

/* Push backtrack state */
zend_worklist_stack_push(&work, -(n + 1));

// FIXME: Can we optimize this copying out in some cases?
int *new_var;
if (ssa->cfg.blocks[n].next_child >= 0) {
new_var = emalloc(sizeof(int) * (op_array->last_var + op_array->T));
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

We could still use alloca here (with a shared limit), as we free new_vars in reverse allocation order. Or an arena.

I think we need only one allocation per depth level, too, as the lifetime of each emalloc() here do not overlap for two blocks of the same depth. This is probably out of the scope of this bug fix however, but maybe we could do something like this:

if (ssa->cfg.blocks[n].next_child >= 0) {
    int level = ssa->cfg.blocks[n].level;
    if (save_vars[level] == save_vars[level-1]) {
        save_vars[level] = alloc(...);
    }
    new_vars = save_vars[level];
    memcpy(new_vars, save_vars[level-1]);
} else {
    new_vars = save_vars[level];
}

// When pushing children:
save_vars[level+1] = save_vars[level];

Then we don't need backtracking, and we can remove save_positions. Before leaving the function we only need to free any save_vars != vars.

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I'll leave that for master, I don't want to risk making too much complicated changes on stable versions

memcpy(new_var, save_vars[save_vars_top], sizeof(int) * (op_array->last_var + op_array->T));
save_positions[n] = save_vars_top++;
save_vars[save_vars_top] = new_var;
} else {
new_var = save_vars[save_vars_top];
save_positions[n] = save_vars_top;
}

zend_ssa_rename_in_block(op_array, build_flags, ssa, new_var, n);

if (tmp) {
free_alloca(tmp, use_heap);
/* Push children in enter state */
unsigned int child_count = 0;
int len_prior = work.len;
int j = ssa->cfg.blocks[n].children;
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

For leaves, can we jump directly to backtrack, and avoid pushing/poping -(n + 1)?

while (j >= 0) {
zend_worklist_stack_push(&work, j + 1);
j = ssa->cfg.blocks[j].next_child;
child_count++;
}

/* Reverse block order to maintain SSA variable number order given in previous PHP versions,
* but the data structure doesn't allow reverse dominator tree traversal. */
for (unsigned int i = 0; i < child_count / 2; i++) {
int tmp = work.buf[len_prior + i];
work.buf[len_prior + i] = work.buf[work.len - 1 - i];
work.buf[work.len - 1 - i] = tmp;
}
}
/* Leave state: backtrack */
else {
n = -n;
n--;

for (unsigned int i = save_vars_top, p = save_positions[n]; i > p; i--) {
efree(save_vars[i]);
}

save_vars_top = save_positions[n];
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I think this always frees exactly one save_vars or nothing: save_vars_top will always be either save_positions[n]+1 (when we have allocated vars for this block) or save_positions[n] (when we did not) because any increments when visiting children will have been reverted at this point.

This could be simplified to

if (save_vars_top != save_positions[n]) {
    ZEND_ASSERT(save_vars_top == save_positions[n]+1);
    efree(save_vars[save_vars_top]);
    save_vars_top--;
}

save_positions could be replaced by a bitset indicating whether we had allocated save_vars for the block:

if (zend_bitset_in(allocated, n)) {
    efree(save_vars[save_vars_top]);
    save_vars_top--;
}

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

We can even avoid a bitset by repeating the creation condition

}
}

free_alloca(save_vars, save_vars_use_heap);
free_alloca(save_positions, save_positions_use_heap);
ZEND_WORKLIST_STACK_FREE_ALLOCA(&work, work_use_heap);

return SUCCESS;
}
/* }}} */

ZEND_API zend_result zend_build_ssa(zend_arena **arena, const zend_script *script, const zend_op_array *op_array, uint32_t build_flags, zend_ssa *ssa) /* {{{ */
{
Expand Down
Loading