kmeisthax / retrogram Goto Github PK

e.g. writes to 0x4000 on an MBC GB cart should set the bank number so we can trace bank calls.

Let's say we encounter a jump table, like so:

ld a, [JMP_PARAM]
ld hl, JMP_TABLE
add a, l
ld l, a
ld a, h
adc a, 0
ld h, a
ld a, [hli]
ld h, [hl]
ld l, a
jp [hl]

Right now tracing will hit the last line of the stream and then bail as the entropy of jp [hl] is maximal. It can literally go anywhere if you don't know the value of the jump table's parameter, which goes from 8 to 9 to 16 to up to 25 bits of entropy as to where the jump lands. If we instead forked at the start on the value of a, then we'd only have 8 bits total (although with still plenty of invalid values, which we'd have to discover by some other means).

In order to do entropy minimization, we need to actually track data values through the trace, as well as have a way to analyze data dependencies (so we can discard irrelevant forks).

When a new parent label is created in between a parent and it's child, the child's parent is not updated to the new parent.

Example: In Telefang, disassemble $A2A - this creates a child block at $A3D. Then disassemble $A34 - the label .LOC_A3D will still be parented to ENTER_A2A despite not being actually local anymore.

More generally we should enforce a restriction that all local labels must be parented to the last parent block; and furthermore prevent creation of local labels on symbols that have nonlocal crossreferences. Some assemblers do not allow you to use non-local symbols as global symbols (e.g. ENTER_A34.LOC_A3D)

In Telefang, disassemble 4:4000. The resulting code will have excessively contextualized pointers, like so:

ENTER_4_4000:
    ld a, [$PR_4!c3e1]
    ld hl, 16394
    call $PR_4!e97
    jp [hl]

The PR_4 is the internal string form of a ROM context on this platform; those are WRAM and HOME pointers though, so their contexts should be stripped when analyzing.

ProjectDatabase, which is our in-memory representation of retrogram.db, should not be specific to one architecture type. It should be architecture-agnostic, so that programs from multiple architectures can be analyzed in the same database.

Database<AR> should be separated into an architecture-erased trait that includes Any and a concrete database type. ProjectDatabase will store boxed, type-erased databases and use Any to type-check AR for methods that need to interact with an architecture.

Platform constructors should have a trait Platform which allows handing the constructor a filesystem and getting a memory bus out of it. This should be roughly equivalent to the existing construct_platform methods, but with the ability to hand more than one file to the constructor.

with_architecture! can then lose the file parameter, as not all architecture-specific code needs to have a working memory bus.