7. Scanning

Scanning is the process of identifying the references in a block of memory and “fixing” them. It’s the process at the heart of the Memory Pool System, and the most critical of the memory management functions that have to be implemented by the client program.

Scanning is performed for two tasks: during tracing, blocks are scanned in order to follow references, and so determine which blocks are reachable and which are not. After objects have been moved in memory, blocks are scanned in order to identify references that need to be updated to point to the new locations of these objects. Both tasks use the same scanning protocol, described here.

7.1. Scanning protocol

There are several types of scanning functions (the scan method in an object format, of type mps_fmt_scan_t, and root scanning functions of various types) but all take a scan state argument of type mps_ss_t, and a description of a region to be scanned. They must carry out the following steps:

1. Call the macro MPS_SCAN_BEGIN() on the scan state.

2. For each reference in the region:

   1. Call MPS_FIX1(), passing the scan state and the reference.

   2. If MPS_FIX1() returns false, the reference is not of interest to the MPS. Proceed to the next reference in the region.

   3. If MPS_FIX1() returns true, the reference is of interest to the MPS. Call MPS_FIX2(), passing the scan state and a pointer to a location containing the reference.

   4. If MPS_FIX2() returns a result code other than MPS_RES_OK(), return this result code from the scanning function as soon as practicable.

   5. If MPS_FIX2() returns MPS_RES_OK, it may have updated the reference. If necessary, make sure that the updated reference is stored back to the region being scanned.

3. Call the macro MPS_SCAN_END() on the scan state.

4. Return MPS_RES_OK.

This description of the protocol simplifies a number of important details, which are covered in the following sections.

7.2. Tagged references

If your references are tagged (or otherwise “encrypted”), then you must remove the tag (or decrypt them) before passing them to MPS_FIX1() and MPS_FIX2().

The reference passed to MPS_FIX2() must be the address of the base of the block referred to (unless the referent belongs to an object format of variant auto-header, in which case it must be a reference to the address just after the header).

However, MPS_FIX1() allows some leeway: if you pass it a reference to the interior of an allocated block, then MPS_FIX1() correctly determines whether a reference to the block is of interest to the MPS.

This means that if your tag is in the low bits of the reference, you may not have to remove it before calling MPS_FIX1(). For example, if you use three tag bits, then your reference is at most base + 7, and if your objects are at least 8 bytes long, then the reference is within the object and need not be stripped. So your code might look like this:

if (MPS_FIX1(ss, obj->ref)) {
    /* strip the tag */
    mps_addr_t p = obj->ref & ~0x7;
    mps_res_t res = MPS_FIX2(ss, &p);
    if (res != MPS_RES_OK) return res;
    /* restore the tag and update reference */
    mps_word_t tag = obj->ref & 0x7;
    obj->ref = (obj_t)((char *)p + tag);
}

This saves the cost of stripping the tag in the case that obj->ref is not of interest to the MPS.

Similarly, if you use interior pointers, you do not need to convert them to base pointers before calling MPS_FIX1() (or, indeed, before calling MPS_FIX2(), if the target of the referent belongs to an object format of variant auto-header).

7.3. Critical path

Scanning is an operation on the critical path of the MPS and so it is vital that it runs fast. The scanning protocol is designed to ensure that as much of the scanning code can be run inline in the client program as possible. In particular, the macro MPS_FIX1() does not need to call into the MPS.

The purpose of MPS_FIX1() is to provide a fast check as to whether a reference is “of interest” to the MPS. It is legitimate to call this on any word: it does not even have to be an address. So if you have a mixture of references and non-references, it might turn out to be faster to call MPS_FIX1() on each word before you even determine whether or not the word is a reference.

Whether this is in fact an optimization depends on the proportion of references to non-references, on how often genuine references turn out to be “of interest”, and what kind of code the compiler has generated. There is no substitute for measurement.

See The critical path.

Note

In one application with a high proportion of unboxed values, it turned out to be fastest to check the tag and reject non-references before calling MPS_FIX1().

Warning

If you passed a word that might not be a reference to MPS_FIX1(), and it returned true, this might be a false positive. You must be certain that the alleged reference is genuine as well as “of interest” before passing it to MPS_FIX2().

Another technique that can speed up scanning is to segregate objects into pools whose object formats contain different scan methods. In particular, if you can segregate objects that do not contain any references into leaf object pools like AMCZ (Automatic Mostly-Copying Zero-rank), these objects do not need to be scanned at all.

7.4. Ambiguous references

If the references in the object being scanned are ambiguous then MPS_FIX2() does not update the reference (because it can’t know if it’s a genuine reference). The MPS handles an ambiguous reference by pinning the block pointed to so that it cannot move.

You could use this fact to optimize the scan by avoiding the need to reassemble and store the updated reference after calling MPS_FIX2()

Note

The MPS currently has no pools that support ambiguous references, so this cannot arise for the scan method in an object format, but root scanning functions may encounter this case.

7.5. Unfixed references

The MPS does not require you to fix all your references. But if a reference is not fixed:

1. it does not keep its target alive (this might be acceptable if you know that the target is being kept alive for another reason, for example if it is in a manually managed pool, or if there is always another reference to the target that is fixed);

2. it does not get updated if the target moves (this might be acceptable if you know that the target cannot move, for example if it is in a non-moving pool, or if it is pinned by an ambiguous reference).

These optimizations can be tricky to make correct, and can make the system fragile (for example, it may break if you start using a different pool class), so it usually safest to fix all references.

7.6. Example: Scheme objects

Scanning tends to be a repetitive procedure and so you’ll find it is usually helpful to define macros to reduce the size of the source code. The MPS provides a convenience macro MPS_FIX12() for the common case of calling MPS_FIX1() and then immediately calling MPS_FIX2() if the reference is “of interest”.

Warning

Some compilers generate better code if you use MPS_FIX12(), and some if you use MPS_FIX1() and MPS_FIX2(). There’s no substitute for measurement.

Here’s the macro FIX defined by the toy Scheme interpreter:

#define FIX(ref)                                                        \
    do {                                                                \
        mps_addr_t _addr = (ref); /* copy to local to avoid type pun */ \
        mps_res_t res = MPS_FIX12(ss, &_addr);                          \
        if (res != MPS_RES_OK) return res;                              \
        (ref) = _addr;                                                  \
    } while(0)

Note

The comment refers to a temptation to write non-portable code that presents itself here. MPS_FIX2() takes a pointer to a location containing the reference (an argument of type mps_addr_t *). It is tempting to take the address of the reference and cast it to this type. The behaviour of such a cast is not defined by the C standard. See Type punning.

Here’s the Scheme scanner:

static mps_res_t obj_scan(mps_ss_t ss, mps_addr_t base, mps_addr_t limit)
{
    MPS_SCAN_BEGIN(ss) {
        while (base < limit) {
            obj_t obj = base;
            switch (obj->type.type) {
                case TYPE_PAIR:
                    FIX(obj->pair.car);
                    FIX(obj->pair.cdr);
                    base = (char *)base + ALIGN(sizeof(pair_s));
                    break;
                case TYPE_VECTOR: {
                    size_t i;
                    for (i = 0; i < obj->vector.length; ++i)
                        FIX(obj->vector.vector[i]);
                    base = (char *)base +
                        ALIGN(offsetof(vector_s, vector) +
                              obj->vector.length * sizeof(obj->vector.vector[0]));
                    break;
                }
                /* ... and so on for the other types ... */
                default:
                    assert(0);
                    fprintf(stderr, "Unexpected object on the heap\n");
                    abort();
                    return MPS_RES_FAIL;
            }
        }
    } MPS_SCAN_END(ss);
    return MPS_RES_OK;
}

Note

This scanner is a simple example intended to make the process clear to the reader. The scanning code and the object layout are not at all optimized.

7.7. Scanning interface

mps_ss_t

The type of scan states.

A scan state represents the state of the current scan. The MPS passes a scan state to the scan method of an object format when it needs to scan for references within a region of memory. The scan method must pass the scan state to MPS_SCAN_BEGIN() and MPS_SCAN_END() to delimit a sequence of fix operations, and to the functions MPS_FIX1(), MPS_FIX2() and MPS_FIX12() when fixing a reference.

MPS_SCAN_BEGIN(mps_ss_t ss)

Within a scan method, set up local information required by MPS_FIX1(), MPS_FIX2() and MPS_FIX12(). The local information persists until MPS_SCAN_END().

ss is the scan state that was passed to the scan method.

Note

Between MPS_SCAN_BEGIN() and MPS_SCAN_END(), the scan state is in a special state, and must not be passed to a function. If you really need to do so, for example because you have an embedded structure shared between two scan methods, you must wrap the call with MPS_FIX_CALL() to ensure that the scan state is passed correctly.

MPS_SCAN_END(mps_ss_t ss)

Within a scan method, terminate a block started by MPS_SCAN_BEGIN().

ss is the scan state that was passed to the scan method.

Note

MPS_SCAN_END() ensures that the scan is completed, so successful termination of a scan must invoke it. However, in case of an error it is allowed to return from the scan method without invoking MPS_SCAN_END().

Note

Between MPS_SCAN_BEGIN() and MPS_SCAN_END(), the scan state is in a special state, and must not be passed to a function. If you really need to do so, for example because you have an embedded structure shared between two scan methods, you must wrap the call with MPS_FIX_CALL() to ensure that the scan state is passed correctly.

MPS_FIX_CALL(ss, call)

Call a function from within a scan method, between MPS_SCAN_BEGIN() and MPS_SCAN_END(), passing the scan state correctly.

ss is the scan state that was passed to the scan method.

call is an expression containing a function call where ss is one of the arguments.

Returns the result of evaluating the expression call.

Between MPS_SCAN_BEGIN() and MPS_SCAN_END(), the scan state is in a special state, and must not be passed to a function. If you really need to do so, for example because you have a structure shared between two object formats, you must wrap the call with MPS_FIX_CALL() to ensure that the scan state is passed correctly.

In example below, the scan method obj_scan fixes the object’s left and right references, but delegates the scanning of references inside the object’s data member to the function data_scan. In order to ensure that the scan state is passed correctly to data_scan, the call must be wrapped in MPS_FIX_CALL().

mps_res_t obj_scan(mps_ss_t ss, mps_addr_t base, mps_addr_t limit)
{
    obj_t obj;
    mps_res_t res;
    MPS_SCAN_BEGIN(ss) {
        for (obj = base; obj < limit; obj++) {
            if (MPS_FIX12(ss, &obj->left) != MPS_RES_OK)
                return res;
            MPS_FIX_CALL(ss, res = data_scan(ss, &obj->data));
            if (res != MPS_RES_OK)
                return res;
            if (MPS_FIX12(ss, &obj->right) != MPS_RES_OK)
                return res;
        }
    } MPS_SCAN_END(ss);
    return MPS_RES_OK;
}

Warning

Use of MPS_FIX_CALL() is best avoided, as it forces values out of registers. The gains in simplicity of the code need to be measured against the loss in performance.

7.8. Fixing interface

mps_bool_t MPS_FIX1(mps_ss_t ss, mps_addr_t ref)

Determine whether a reference needs to be passed to MPS_FIX2().

ss is the scan state that was passed to the scan method.

ref is the reference.

Returns a truth value (mps_bool_t) indicating whether ref is “interesting” to the MPS. If it returns true, the scan method must invoke MPS_FIX2() to fix ref.

This macro must only be used within a scan method, between MPS_SCAN_BEGIN() and MPS_SCAN_END().

Note

If your reference is tagged or otherwise “encrypted”, you must ensure that it points to a location within the target block before calling MPS_FIX1(). (Therefore, a small tag in the low bits need not be stripped.)

Note

In the case where the scan method does not need to do anything between MPS_FIX1() and MPS_FIX2(), you can use the convenience macro MPS_FIX12().

mps_res_t MPS_FIX12(mps_ss_t ss, mps_addr_t *ref_io)

Fix a reference.

This macro is a convenience for the case where MPS_FIX1() is immediately followed by MPS_FIX2(). The interface is the same as MPS_FIX2().

mps_res_t MPS_FIX2(mps_ss_t ss, mps_addr_t *ref_io)

Fix a reference.

ss is the scan state that was passed to the scan method.

ref_io points to the reference.

Returns MPS_RES_OK if successful: in this case the reference may have been updated, and the scan method must continue to scan the block. If it returns any other result, the scan method must return that result as soon as possible, without fixing any further references.

This macro must only be used within a scan method, between MPS_SCAN_BEGIN() and MPS_SCAN_END().

Note

If your reference is tagged (or otherwise “encrypted”), you must remove the tag (or otherwise decrypt the reference) before calling MPS_FIX2(), and restore the tag to the (possibly updated) reference afterwards.

The only exception is for references to objects belonging to a format of variant auto-header (see mps_fmt_auto_header_s): the header size must not be subtracted from these references.

Note

In the case where the scan method does not need to do anything between MPS_FIX1() and MPS_FIX2(), you can use the convenience macro MPS_FIX12().

mps_res_t mps_fix(mps_ss_t ss, mps_addr_t *ref_io)

Deprecated

starting with version 1.111.

Use MPS_FIX1() and MPS_FIX2() instead.

Fix a reference.

This is a function equivalent to:

MPS_SCAN_BEGIN(ss);
MPS_FIX12(ss, ref_io);
MPS_SCAN_END(ss);

Because scanning is an operation on the critical path, we recommend that you use MPS_FIX12() (or MPS_FIX1() and MPS_FIX2()) to ensure that the “stage 1 fix” is inlined.

Note

If you call this between MPS_SCAN_BEGIN() and MPS_SCAN_END(), you must use MPS_FIX_CALL() to ensure that the scan state is passed correctly.