#255: allocate the literals the way the platform frees them

[EdmondDantes]

Closes #255.

What was wrong

op_array_to_emalloc allocated opcodes and literals as one block and pointed op_array->literals at the tail of it. That is the layout pass_two() leaves on 64-bit, where destroy_op_array frees only op_array->opcodes.

Under ZEND_USE_ABS_CONST_ADDR — SIZEOF_SIZE_T == 4, i.e. 32-bit — the compiler keeps the two in separate allocations, and destroy_op_array frees both:

if (ZEND_USE_ABS_CONST_ADDR || !(op_array->fn_flags & ZEND_ACC_DONE_PASS_TWO)) {
    efree(op_array->literals);
}
efree(op_array->opcodes);

So on 32-bit every materialized closure with at least one literal handed the allocator a pointer into the middle of a block it never issued, and then freed the block a second time. The opline fix-up already branched on the macro; the allocation did not.

#251 made op_array_to_emalloc unconditional, which turned this from some closures into every closure that reaches a worker.

What changed

The copy follows the same rule as the compiler on both platforms: one block under !ZEND_USE_ABS_CONST_ADDR, where RT_CONSTANT addresses the literal by an int32_t offset from the opline and separate allocations would overflow it, two blocks otherwise.

Verification

No 32-bit runner exists here, so the configuration was reproduced on x86-64 by forcing ZEND_USE_ABS_JMP_ADDR and ZEND_USE_ABS_CONST_ADDR to 1 in Zend/zend_compile.h and rebuilding the engine. In that build:

• before the change: 35 of 102 thread_pool tests fail;
• after: 0.

In the ordinary 64-bit build the layout is unchanged: tests 1191 passed / 0 failed, fuzzy-tests/_generated 724 / 0. The one warning is the stale --XFAIL-- on 030, removed by #256.

Build: ZTS DEBUG, TrueAsync ABI v0.24.0.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!